Clinical AI Radar

★ A parsimonious machine-learning model using routine laboratory data to differentiate nontuberculous mycobacterial disease from tuberculosis: Development and independent external validation.

Zhang R, Yang Z, Ding Y, Han T, Zheng G, Ren Q — Fri, 06 Nov 2026 00:00:00 +0000

99% · External validation 99%

In tuberculosis-endemic settings, suspected mycobacterial disease may reflect either nontuberculous mycobacteria (NTM) or Mycobacterium tuberculosis (MTB), yet routine clinical findings do not reliably distinguish them before pathogen-specific confirmation. We developed an eight-variable Extra Trees model in Hospital A (n = 1,090; 130 NTM) and independently evaluated the frozen model in Hospital B (n = 1,160; 128 NTM). Internal out-of-fold (OOF) AUC/AUPRC were 0.892/0.588; external AUC/AUPRC were 0.849/0.505. Calibrated Brier scores were 0.072 and 0.075. At the Hospital A-derived threshold of 0.13179, sensitivity/specificity were 80.0%/80.3% internally and 67.2%/83.7% externally; 42 of 128 external NTM cases were classified below threshold. Because feature selection preceded the original OOF evaluation, a post-hoc fully nested analysis was performed (AUC 0.875; AUPRC 0.525), indicating modest internal optimism. A post-hoc benchmark did not show an overall advantage for Extra Trees over other algorithms. Routine variables may help prioritize confirmatory testing, but the frozen threshold cannot support stand-alone diagnosis or exclusion of NTM. Prospective evaluation must verify patient-level measurement timing, reference diagnoses and clinical impact.

★ Miscarriage Risk Prediction for Women With Polycystic Ovary Syndrome Who Achieve Clinical Pregnancy After Fresh Embryo Transfer: Multicenter Development, Geographic External Validation, and Prospective Temporal Validation Study.

Tong J, Zhang Y, Liu Y, Liu L, Shang X, Wu F — Fri, 06 Nov 2026 00:00:00 +0000

98% · External validation 98%, Prospective 94%

Background: Women with polycystic ovary syndrome (PCOS) who achieve clinical pregnancy following fresh in vitro fertilization (IVF)/intracytoplasmic sperm injection (ICSI) carry an elevated miscarriage risk. However, individualized risk stratification after clinical pregnancy confirmation but before outcome determination remains challenging in clinical practice. Objective: This study aimed to develop, externally and prospectively validate over time, and interpret a machine learning-based risk stratification model for women with PCOS who had achieved a confirmed clinical pregnancy after fresh IVF/ICSI. Methods: This multicenter study included 8911 clinical pregnancy IVF/ICSI cycles across development (n=7241; miscarriages: n=973), external validation (n=1170; miscarriages: n=146), and prospective temporal validation (n=500; miscarriages: n=62) cohorts. Prespecified predictors contained no observed missing values; training-fold median imputation was kept only as a leakage-safe fallback. A total of 13 candidate machine learning and traditional statistical algorithms were compared internally using repeated stratified 5-fold cross-validation (5 repeats) within the development cohort; preprocessing and hyperparameter tuning were performed exclusively in training folds. A prelocked Extreme Gradient Boosting (XGBoost) model was carried forward unchanged for external and prospective temporal validation. Cross-fitted logistic recalibration of averaged XGBoost out-of-fold probabilities was evaluated as a secondary development-sample diagnostic analysis. Model performance encompassed discrimination, precision-recall, calibration, threshold-based metrics, decision curve analysis, risk stratification, and Shapley additive explanations-based interpretability. Results: Miscarriage rates were 13.4% (973/7241), 12.5% (146/1170), and 12.4% (62/500) for the development, external validation, and prospective temporal validation cohorts, respectively. In repeated cross-validation, XGBoost (area under the receiver operating characteristic curve [AUC] 0.885, 95% CI 0.874-0.896) and LightGBM (AUC 0.882, 95% CI 0.871-0.893) showed comparable discrimination; the paired XGBoost-LightGBM DeLong test was not significant (Holm-adjusted P=.07). Raw XGBoost out-of-fold probabilities overestimated absolute risk (mean predicted risk 30.3% vs observed risk 13.4%); cross-fitted recalibration within the development cohort improved calibration but was not propagated to validation cohorts. The locked XGBoost model achieved AUCs of 0.787 (95% CI 0.748-0.827; Brier score 0.099) in external validation and 0.763 (95% CI 0.709-0.816; Brier score 0.104) in prospective temporal validation, showing cross-cohort performance attenuation. Risk-quartile stratification using fixed development-cohort cutoffs showed consistent event enrichment across validation cohorts, but positive predictive values remained modest. Shapley additive explanations analysis identified anti-Müllerian hormone, endometrial thickness on the day of human chorionic gonadotropin administration, and basal luteinizing hormone as leading predictors. Conclusions: The locked XGBoost model using routine preoutcome variables may support miscarriage risk stratification and counseling for women with PCOS with clinical pregnancy after fresh IVF/ICSI. Given cross-cohort discrimination loss and calibration drift, it should serve only as a risk-ranking tool, not for diagnosis or treatment selection. Local calibration monitoring is mandatory before any real-world clinical deployment.

★ Development and Validation of an Interpretable Machine-Learning Model for Predicting Suboptimal Clinical Response After Sleeve Gastrectomy: A Multicenter Study.

Nie Y, Ping A, Wang T, Tang H, Huang Y, Wang Z — Fri, 06 Nov 2026 00:00:00 +0000

98% · External validation 98%

Suboptimal clinical response (SCR) after sleeve gastrectomy (SG) remains a major clinical challenge, yet reliable preoperative risk stratification is lacking. Therefore, we developed and externally validated an interpretable machine-learning model for predicting SCR after SG using multicenter data from three tertiary hospitals in China. Overall, 1312 patients were included, comprising a derivation cohort (n = 1094) and two external validation cohorts (n = 84 and n = 134). SCR, defined as percentage of total weight loss < 25% at 1 year, occurred in 374 (34.2%), 34 (40.5%), and 45 (33.6%) patients, respectively. Among six machine-learning algorithms, XGB achieved the best performance. The final model incorporated eight routinely available predictors. In internal validation, the model showed excellent discrimination, with an AUC of 0.821 (95% confidence interval [CI], 0.752-0.871), sensitivity of 0.803, and specificity of 0.714. External validation confirmed satisfactory discrimination (AUC 0.761; 95% CI, 0.688-0.823), good calibration, and clinical benefit. Notably, the XGB model significantly outperformed 10 clinicians across experience levels and remained superior in clinically challenging cases. SHAP analysis enhanced interpretability, with predictor importance ranked as BMI, age, fasting glucose, type 2 diabetes, vitamin B12, calcium, total cholesterol, and gamma-glutamyl transferase. This interpretable XGB model may enable precise preoperative risk stratification and support personalized management for patients undergoing SG.

★ Local Conformal Auditing Exposes and Restores Reliability of AI-ECG Atrial Fibrillation Detection Under External Deployment Shift.

Abbasi T, Abbas A, Mehmood UH — Fri, 06 Nov 2026 00:00:00 +0000

98% · External validation 98%

Background: AI-enabled electrocardiography (AI-ECG) can detect atrial fibrillation (AF), but safe deployment depends on calibration, subgroup reliability, and explicit error guarantees, not on discrimination alone. Objective: To test whether the conformal-prediction coverage guarantee of an AF ECG classifier survives external deployment shift, and whether a small local labeled audit restores it without retraining. Methods: A four-class 12-lead rhythm classifier trained on PTB-XL (Germany; 7,831 ECGs, 7,164 patients) was evaluated unchanged on Chinese (Chapman-Shaoxing; n=2,289) and US (Georgia; n=1,990) cohorts. We measured discrimination, expected calibration error (ECE), 90%-target split-conformal coverage, subgroup coverage, label-free shift corrections, local recalibration across audit sizes, and an AF auto-clear policy with finite-sample risk control. Results: AF AUC remained 0.998 (95% CI 0.997-0.999) in China but fell to 0.936 (0.919-0.952) in the US cohort. ECE rose from 0.051 to 0.197 and 0.237, and source-calibrated coverage fell from 0.901 to 0.758 and 0.722 (both P<0.001). Coverage was lower in Chinese men than women (0.707 vs 0.822; difference 0.115, 95% CI 0.081-0.148). Label-free corrections failed. Local recalibration restored mean coverage to 0.90 with 50 labeled ECGs but with wide variability (95% simulation interval 0.82-0.97), narrowing to 0.86-0.94 at 200 ECGs. Local AF risk control certified auto-clearing of 68% of held-out ECGs in China but abstained in the US cohort. Conclusion: AI-ECG uncertainty can fail silently despite high AUC. A local conformal audit detects when uncertainty estimates are untrustworthy and recalibrates the uncertainty layer without changing the classifier, converting trust into a measurable, subgroup-aware, task-specific guarantee.

★ A graph attention and SAINT-inspired transformer model for predicting postoperative atrial fibrillation: A bidirectional external validation study.

Tao Z, Pan S, Zhang S, Gao Z, Zhang Y, Fan G — Fri, 06 Nov 2026 00:00:00 +0000

97% · External validation 97%

Background And Objective: Postoperative atrial fibrillation (POAF) is a common surgical complication. We tested whether a two-stage within-patient architecture-graph attention followed by a SAINT-inspired feature Transformer-adds information beyond strong tabular models and transports across healthcare systems. Methods: We used the independent MIMIC-IV and INSPIRE databases. Sixty leakage-controlled scalar features were encoded as tokens for piecewise-linear encoding, graph attention, feature-wise Transformer blocks, and learned pooling. Reciprocal external validation was performed in both directions. Comparators were logistic regression, random forest, XGBoost, LightGBM, and a baseline-only stack. Component ablations, dense versus training-derived sparse graphs, paired bootstrap differences, and efficiency were evaluated. Results: In Experiment 1 (MIMIC-IV development, INSPIRE external), GAT-SAINT had an external AUC of 0.844 (95% CI, 0.829-0.858), versus 0.839 for logistic regression; the paired difference was 0.005 (95% CI, -0.001 to 0.011). In Experiment 2 (INSPIRE development, MIMIC-IV external), its AUC was 0.791 (95% CI, 0.768-0.813), versus 0.788 for random forest (paired difference, 0.003; 95% CI, -0.008 to 0.014) and 0.797 for the baseline-only stack. External Brier scores were 0.096 and 0.162. The Experiment 2 validation-locked threshold transported poorly (sensitivity, 0.025; specificity, 0.999). Component and graph-sensitivity effects were direction-dependent, and removal of some modules was compatible with little or no external AUC loss. Conclusions: Reciprocal external validation showed transportable rank discrimination but did not establish a consistent advantage over simpler tabular models or a transportable operating threshold. The architecture warrants further methodological study, but prospective validation, recalibration, and clinical-impact assessment are required before deployment.

★ External Validation of a Machine Learning Model for Delirium Risk.

Friedman JI, Parchure P, Fu W, Cheertirala S, Timsina P, Raut G — Fri, 06 Nov 2026 00:00:00 +0000

97% · External validation 97%, Prospective 89%

Importance: Although external validation of machine learning (ML)-based delirium risk models is a key stage in the model development pipeline, few published reports exist. Moreover, ML-based delirium risk models commonly degrade in performance during external validation. Objective: To examine the performance of a fusion multimodal with natural language processing (NLP) delirium risk stratification model in live clinical practice with cross-hospital external validation. Design, Setting, And Participants: This quality improvement study evaluated live clinical deployment of the model among patients 60 years or older who were admitted to non-intensive care units at Mount Sinai Hospital (MSH), with external validation performed at Mount Sinai Morningside Hospital (MSM), from December 1, 2024, to December 31, 2025. Main Outcomes And Measures: Primary outcomes were performance of the risk stratification model, assessed by area under the receiver operating characteristic curve (AUROC), and monthly delirium detection rates. A 2 × 2 confusion matrix was created, and the 4 cells were used to calculate model performance metrics and compare associated diagnostic yield between premodel deployment and active model deployment periods. Results: The overall sample included 33 095 admissions to non-intensive care units at MSM and MSH (mean [SD] age, 74.86 (9.22) years; 15 601 women [47.1%], 17 271 men [52.2%]). A total of 9020 MSM inpatient admissions were assigned a delirium risk score by the risk stratification model during live clinical operations and were compared to 21 186 MSH inpatient admissions that were assigned a risk score by the same model. Age at admission significantly differed between patients at MSM and MSH (MSM: median [IQR] age, 75.79 [68.25-83.88] years vs MSH: 72.62 [66.64-79.65] years; P < .001). Patients from MSM and MSH also significantly differed by sex, race and ethnicity, surgical vs medical primary diagnosis, length of stay, and dementia prevalence rate. Among 11 909 MSM patients, patients in the pre-ML cohort (median [IQR] age, 78.00 [71.00-85.00] years) were significantly older than those in the post-ML cohort (median [IQR] age, 75.79 [68.25-83.88] years) (P < .001). The AUROC of the model was 0.91 (95% CI, 0.90-0.92) for MSM and 0.94 (95% CI, 0.93-0.94) for MSH. The positive predictive value of the model was 25.9% for MSM and 19.4% for MSH. For MSM, the mean monthly delirium detection rate significantly increased from 9.23% (95% CI, 8.07%-10.38%) to 19.59% (95% CI, 17.32%-21.85%) after model deployment (P < .001). Conclusions And Relevance: In this quality improvement study, an ML-based delirium risk stratification model maintained acceptable discriminative performance during external deployment without retraining and was similar to its performance at the development site without retraining. Model deployment was also associated with increased diagnostic yield but a high false positive rate. These findings provide further evidence of the potential generalizability of the model.

★ Construction of a Predictive Model for Cerebrospinal Fluid Leakage After Endoscopic Transsphenoidal Pituitary Tumor Surgery Based on Machine Learning Algorithm.

Shao S, Ye C, Du Y, Liu T, Wu Y, Chen M — Fri, 06 Nov 2026 00:00:00 +0000

97% · External validation 97%

Objective: To develop ML models predicting postoperative CSF leakage after endoscopic transsphenoidal pituitary adenectomy (TSPA). Methods: Four hundred and twenty patients from five hospitals (January 2021-December 2022) comprised the training cohort; 140 from another hospital were the validation cohort. The training cohort included 60 CSF leak and 360 nonleak cases. Risk factors were analyzed. Decision tree, logistic regression, and SMOTE-based models were built and their diagnostic value compared. Results: Significant differences ( p < 0.05) existed in tumor diameter, tumor texture, operative time, intraoperative diaphragma sellae rupture, reoperation, and intraoperative CSF leakage grade. Logistic regression identified tumor texture, reoperation, intraoperative CSF leakage grade, and diaphragma sellae rupture as risk factors. All models had high AUCs in training: logistic regression (0.792), decision tree (0.835), and SMOTE (0.839). The Delong test showed a decision tree, and SMOTE outperformed logistic regression. SMOTE had the highest sensitivity/specificity. Validation cohort AUCs were: logistic regression (0.775), decision tree (0.809), and SMOTE (0.831), with SMOTE superior. Calibration curves indicated good agreement between predicted and actual values. Conclusion: Prediction models for CSF leakage postendoscopic TSPA were developed using decision trees, SMOTE, and logistic regression. Model efficacy was evaluated via AUC, specificity, and sensitivity. Comparing algorithms reduced single-algorithm bias, with SMOTE demonstrating the best performance.

★ Machine learning models integrating 2D and 3D intratumoral heterogeneity features from multi-sequence MRI: accurate prediction of perineural invasion in pancreatic ductal adenocarcinoma.

Ouyang Z, Song Z, Li Z, Cui S, Liu X, Shen W — Fri, 06 Nov 2026 00:00:00 +0000

97% · External validation 97%

Introduction: Perineural invasion (PNI) occurs in the majority of pancreatic ductal adenocarcinoma (PDAC) cases and is strongly associated with local recurrence, neuropathic pain, and poor survival, yet no reliable method exists for its preoperative identification. Here, we developed and externally validated a multi-sequence MRI-based intratumoral heterogeneity (ITH) scoring framework integrated with machine learning for the non-invasive prediction of PNI in PDAC. Methods: Using preoperative T2-weighted, diffusion-weighted, and contrast-enhanced T1-weighted MRI from 205 surgically resected patients across four medical centers (100 training, 43 internal test, and 28 and 34 external validation patients), we extracted two-dimensional and three-dimensional ITH features via K-means clustering-based subregion partitioning and incorporated independent clinical predictors tumor location and clinical T-stage to construct combined predictive models. Eleven machine learning algorithms were systematically evaluated. Results: The combined clinical-ITH model achieved an area under the curve (AUC) of 0.907 in the internal test set and AUCs of 0.860 and 0.869 in the two external validation cohorts. Corresponding AUCs were 0.881, 0.860, and 0.850 for the ITH model and 0.758, 0.342, and 0.700 for the clinical logistic regression model. Discussion: SHAP analysis ranked three T2WI-derived features as the largest contributors to the combined model output, feature effects varied in direction and were not uniformly monotonic. These findings demonstrate that multi-sequence MRI-based ITH quantification provides an interpretable approach to preoperative PNI assessment with consistent discrimination across the two external surgical cohorts, offering a potential tool to guide surgical planning and adjuvant therapy selection in pancreatic cancer.

★ CT-based bone mineral density estimation using HU, radiomics, and deep learning.

Tanioka S, Aydin OU, Hilbert A, Ito S, Miyazaki T, Yamamoto A — Fri, 06 Nov 2026 00:00:00 +0000

96% · External validation 96%

Purpose: To develop and test models that estimate bone mineral density (BMD) from thoracolumbar junction CT images using Hounsfield unit (HU) assessment, radiomics, and convolutional neural networks (CNNs). Methods: We retrospectively identified patients who underwent lumbar dual-energy x-ray absorptiometry and CT covering T12 and L1 within 6 months at four hospitals between 2014 and 2025. In total, 465 patients were eligible for T12 analysis, and 448 for L1. Data from two institutions were used for training/validation and data from the other two for external testing. We developed five models: region-of-interest-based HU regression; volume-of-interest (VOI)-based HU regression; radiomics; CNN; and an ensemble combining VOI, radiomics, and CNN. Performance was assessed using mean absolute error (MAE), root mean squared error (RMSE), coefficient of determination (R2), and Bland-Altman limits of agreement (LoA). Results: For T12, the CNN model performed best among the single models (MAE: 0.122, RMSE: 0.157, R2: 0.603), whereas the ensemble model showed a slightly higher MAE (0.128) but lower RMSE (0.154) and higher R2 (0.620). For L1, the radiomics model was the best single model (MAE: 0.124, RMSE: 0.155, R2: 0.582), while the ensemble showed numerically improved performance (MAE: 0.116, RMSE: 0.145, R2: 0.633). Pairwise comparisons showed no significant difference in absolute prediction errors between the ensemble and the best single model at either T12 or L1. Conclusion: We developed an ensemble framework integrating volumetric HU assessment, radiomics, and CNNs to estimate BMD from thoracolumbar junction CT images. External testing across two independent institutions demonstrated promising performance for BMD estimation.

★ Creation and verification of an interpretable machine learning model for forecasting in-hospital results in Cardiorenal syndrome patients in critical condition.

Xia Z, Peng Z, Xu P — Fri, 06 Nov 2026 00:00:00 +0000

96% · External validation 96%

Background: Cardiorenal syndrome (CRS) encompasses a spectrum of disorders in which acute or chronic dysfunction of the heart or kidneys induces dysfunction of the other. The heart-kidney crosstalk underlying CRS has since been extended into the broader cardiovascular-kidney-metabolic (CKM) syndrome by incorporating metabolic risk factors.CRS is associated with significantly increased mortality in intensive care unit (ICU) patients. However, dedicated tools for predicting short-term outcomes remain lacking, which limits timely risk stratification and treatment optimization. Methods: This multicenter cohort study enrolled 6,668 CRS patients from MIMIC-IV (v2.2) for model development and 3,602 from eICU-CRD (v2.0) for external validation. In patients remaining in the ICU for ≥24 hours, candidate predictors were ascertained within the first 24 hours, including early treatment exposures. Subsequent all-cause in-hospital mortality was predicted at this 24-hour landmark. Eleven machine learning algorithms were systematically compared after 1:1 random undersampling of the development cohort. SHapley Additive exPlanations (SHAP) was applied to enhance interpretability and identify key prognostic features. Results: The artificial neural network (ANN) model performed best. The final model incorporated six readily obtainable variables: AKI stage, age, mechanical ventilation, vasopressor use, β-blocker use, and ACEI/ARB use. It achieved an AUC of 0.810 (95% CI 0.757-0.860) in internal validation and 0.704 (95% CI 0.679-0.729) in external validation, indicating good internal but moderate external discrimination. Conclusions: We developed the first interpretable machine learning model specifically for predicting in-hospital mortality in critically ill CRS patients. SHAP provides transparent feature contribution analysis, and a web-based interface facilitates clinical implementation. When applied 24 hours after ICU admission, the model enables early identification of high-risk patients. It may also support risk stratification and resource allocation in the ICU.

★ A Prediction Model for Heart Failure with Improved Ejection Fraction Phenotype Based on Electrocardiogram Images and Clinical Baseline Features.

Wei X, Xiao X, Zhang S, Dong C, Yan H, Ma Y — Fri, 06 Nov 2026 00:00:00 +0000

96% · External validation 96%

Objective: To develop and externally evaluate machine learning models for predicting the heart failure with improved ejection fraction (HFimpEF) phenotype in patients with ischemic cardiomyopathy, using electrocardiogram (ECG) image features, all eligible clinical baseline features, and their combination. Methods: This two-center retrospective study included patients hospitalized between March 2018 and February 2025 with baseline left ventricular ejection fraction (LVEF) ≤40%. HFimpEF was defined as an increase in LVEF of at least 10 percentage points and a follow-up LVEF >40%, with at least three months between echocardiograms. The development cohort comprised 415 patients (194 with HFimpEF and 221 with persistent HFrEF); an independent cohort of 100 patients (55 and 45, respectively) was used for external validation. Four classifiers were evaluated using ECG image features alone, all eligible clinical baseline features alone, and their combination. LASSO was used separately to rank clinical variables. Results: In the reported internal evaluation, the combined random forest model achieved an AUC of 0.825, accuracy of 0.747, precision of 0.753, specificity of 0.791, and F1 score of 0.721. In external validation, the combined-model AUCs (95% confidence intervals) were 0.624 (0.513-0.734) for random forest, 0.607 (0.496-0.718) for logistic regression, 0.653 (0.545-0.761) for SVC-RBF, and 0.565 (0.454-0.676) for k-nearest neighbors. β-Blocker use ranked highest in the exploratory LASSO analysis. Combined models did not consistently outperform single-feature-set models externally. Conclusion: Models combining ECG image features and clinical baseline features showed promising internal discrimination for identifying subsequent HFimpEF in patients with ischemic cardiomyopathy. However, external discrimination was modest, and the combined models did not consistently outperform single-source models. Further model refinement and validation in larger, independent cohorts are needed before clinical application.

★ Development and validation of a machine learning-based model for predicting amputation risk in patients with lower extremity arterial disease and diabetes.

Tang L, Wang L, Li J, Liu X, Chen Q, Song P — Fri, 06 Nov 2026 00:00:00 +0000

96% · External validation 96%

Background: Patients with lower extremity arterial disease (LEAD) complicated by diabetes face substantial amputation risk. Pre-amputation death acts as a critical competing event that distorts risk estimation. Validated personalized prediction tools for this population remain scarce and hinder precise clinical interventions. This study aimed to develop and validate two complementary interpretable models for 12-month amputation risk in patients with LEAD and diabetes to support clinical decision-making. Methods: This retrospective multicenter study enrolled 1,919 patients with LEAD and diabetes from four Chinese hospitals (January 2018-September 2024), with follow-up ending in September 2025. The primary outcome was 12-month amputation (major or minor). Prior to model construction, candidate predictors were screened via LASSO regression, random forest classifier and Boruta algorithm. Ten machine learning algorithms were then trained and evaluated by discrimination and calibration metrics. SHapley Additive Explanations (SHAP) was applied for interpretability. A Fine-Gray competing risk regression model was built using the same screened predictor set. Results: Patients were split into training (n = 967), internal validation (n = 414), and external validation (n = 538) cohorts. The 12-month amputation rate was 14.0%. Median follow-up was 12 months; mean age was 72.7 ± 9.5 years and 65.7% were male. The optimal GraBoost model achieved internal validation AUC = 0.860 and external validation AUC = 0.828, incorporating 14 clinical predictors, including: Rutherford classification, financial burden, age, Barthel Index, diabetic foot, statin, fibrinogen (FIB), neutrophil (NEUT), residence, prealbumin (PALB), C-reactive protein (CRP), activated partial thromboplastin time (APTT), body mass index (BMI) classification and wound. SHAP revealed higher grade Rutherford classification, lower Barthel Index, younger age, wounds, diabetic foot, and elevated FIB, APTT, CRP, NEUT correlated with higher amputation risk. Urban residence, normal BMI classification, lower financial burden, statin use and higher PALB were protective. Shared predictors showed consistent directional effects across GraBoost and Fine-Gray models. A dual-model web-based calculator was constructed for risk assessment. Conclusion: We developed and validated complementary interpretable GraBoost machine learning and Fine-Gray competing risk models for 12-month amputation risk. The dual-model web calculator may help clinicians identify high-risk patients and guide clinical management.

★ Explainable hybrid vision transformer-convolutional neural network with cross-attention fusion for multi-institutional breast cancer receptor status prediction from H&E histopathology images.

Albalawi ES, Qayyum J, Qayyum J — Fri, 06 Nov 2026 00:00:00 +0000

95% · External validation 95%

The integration of artificial intelligence (AI) into digital pathology has revolutionised breast cancer diagnostics, offering transformative opportunities to automate and refine molecular profiling, which is essential to precision oncology. Accurate determination of receptor status-estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2)-remains pivotal in guiding personalised treatment planning and prognostic assessment. However, traditional immunohistochemistry (IHC) and molecular assays are limited by resource intensity, lengthy turnaround times, and inter-observer variability, underscoring the need for scalable computational alternatives. This study proposes a novel integrative AI-driven framework for predicting breast cancer receptor profiles directly from hematoxylin and eosin (H&E)-stained histopathology images. The framework employs a hybrid vision transformer-convolutional neural network (ViT-CNN) architecture with a cross-attention fusion mechanism to capture both global contextual information and fine-grained local features. Experiments show AUCs of 0.967, 0.952, and 0.941 for ER, PR, and HER2 prediction, respectively, representing relative improvements of 12.4%, 9.8%, and 10.8% over the best-performing prior method (DCS-ST), re-evaluated under our identical protocol (absolute AUC differences of 0.107, 0.085 and 0.092). Furthermore, the framework enables automated molecular subtyping into luminal A, luminal B, HER2-enriched, and triple-negative categories with a macro-averaged F1-score of 95.7%. A comprehensive multi-institutional validation was performed on 2,847 external whole-slide images from three partner institutions (not including the primary BINO Hospital test set), showing consistent performance (a mean AUC reduction of 0.83 percentage points relative to the internal test set) across the staining protocols and scanners represented. Integrated explainability modules, including attention visualisations and Grad-CAM, enhance interpretability, supporting pathologist review. In a single-institution retrospective workflow audit, AI assistance was associated with an estimated 72.7% reduction in diagnostic turnaround time (from a conventional mean of 72.4 hours to 19.8 hours; Section 6.10, where approximately 65% of the AI-assisted time is un-itemised queueing and logistics) and, in a 120-WSI reader study, with higher inter-observer agreement (kappa 0.847 to 0.967). These clinical-impact estimates are hypothesis-generating and require prospective validation. The findings position the framework as a promising, explainable approach to AI-assisted digital pathology that warrants prospective, multi-centre clinical evaluation.

★ CT-based prediction of EGFR mutation subtypes and TP53 co-mutation status in lung adenocarcinoma.

Liu H, Li Z, Zhang Y, Yu G, Miao Y, Teng X — Fri, 06 Nov 2026 00:00:00 +0000

95% · External validation 95%

Background: Epidermal growth factor receptor (EGFR) mutation status is a key determinant of targeted treatment in lung adenocarcinoma, whereas co-occurring TP53 alterations may further affect therapeutic response and prognosis. This study aimed to develop a noninvasive CT-based approach for predicting EGFR mutation subtypes and TP53 co-mutation status in lung adenocarcinoma, thereby providing imaging support for pretreatment risk stratification and clinical decision-making. Methods: Patients with pathologically confirmed lung adenocarcinoma who underwent pretreatment chest CT and molecular testing were retrospectively enrolled from a single-center cohort. A total of 1,633 patients were included and categorized into five molecular groups: EGFR wild type, EGFR 19del, EGFR L858R, EGFR 19del with TP53 co-mutation, and EGFR L858 R with TP53 co-mutation. Based on chest CT images, key slice extraction, lesion ROI generation, automatic segmentation, and image standardization were performed. Semantic CT features, radiomics features, and deep CT representations were then derived. Under a unified data partitioning and standardized evaluation framework, four models were constructed and compared for the five-class classification task: a random forest model based on semantic CT features, a support vector machine model based on radiomics features, a Qwen-based semantic classification model, and a MedicalDINOv3-based deep learning model. To eliminate confounding bias induced by multimodal feature fusion, two ablation experiments were supplemented to split the model input features, and two MedicalDINOv3 models were built with inputs restricted to semantic features only and CT images only, respectively. Model performance was assessed using accuracy, macro-averaged area under the receiver operating characteristic curve (AUC), and macro-F1 score. External validation was further conducted using public and multicenter datasets. Results: Among the 1,633 patients in the internal cohort, 731 were EGFR wild type, 295 had EGFR 19del, 399 had EGFR L858R, 118 had EGFR 19del with TP53 co-mutation, and 90 had EGFR L858 R with TP53 co-mutation. To quantify the independent predictive value of different input modalities, ablation experiments on three variants of MedicalDINOv3 were performed. Ablation results revealed distinct performance among three input strategies. The MedicalDINOv3 with only semantic input obtained a test accuracy of 31.55% and an AUC of 0.6305 with limited predictive ability. The branch solely fed with CT images reached an accuracy of 55.61% and a macro-average AUC of 0.6347. The complete MedicalDINOv3 integrating CT deep features, clinical and semantic features performed best, with a test accuracy of 0.6905 and a macro-average AUC of 0.8788. This comparison proved that CT deep representations outperform traditional handcrafted imaging and semantic features, and multi-modal fusion further improves subtype stratification performance. The support vector machine model achieved an accuracy of 0.527, a macro-AUC of 0.792 and a macro-F1 score of 0.524; the random forest model achieved an accuracy of 0.455, a macro-AUC of 0.819 and a macro-F1 score of 0.412; and the Qwen-based model achieved an accuracy of 0.339, a macro-AUC of 0.700 and a macro-F1 score of 0.316. Receiver operating characteristic analyses showed that the deep learning model provided relatively stable discrimination for EGFR wild type and TP53 co-mutated categories, whereas overlap remained between the two major EGFR-sensitive subtypes, EGFR 19del and EGFR L858 R. In external validation, the public dataset used for EGFR binary classification yielded an overall accuracy of 67.0% and an AUC of 0.6666. The raw overall accuracy of the cohort was 75.5%, and raw overall accuracy is prone to overestimation when the cohort has severe class imbalance dominated by EGFR wild-type samples, which can only serve as an auxiliary reference. Multicenter five-class external validation yielded a balanced accuracy of 58.2% and a macro F1 of 0.574. The subtype-specific AUCs for EGFR wild type, EGFR 19del, EGFR L858R, EGFR 19del with TP53 co-mutation, and EGFR L858 R with TP53 co-mutation were 0.8342, 0.6504, 0.7221, 0.6586, and 0.7563, respectively. Conclusion: A CT-based deep learning model can provide noninvasive stratification of EGFR mutation subtypes and TP53 co-mutation status in lung adenocarcinoma. Compared with semantic CT features and radiomics features, deep CT representations showed superior overall discriminative performance. This approach may serve as an adjunctive tool to molecular testing for pretreatment assessment in lung adenocarcinoma. However, the model showed inconsistent generalization performance across different molecular subtypes, with notably limited predictive efficacy for EGFR 19del and EGFR 19del+TP53 subtypes. Therefore, predictions for these categories should be interpreted with caution, and the model requires further validation in larger, multicenter, and standardized datasets before clinical application.

Ultrasound-Based Peri-nodular and Intra-nodular Habitat Radiomics for Preoperative Prediction of Lymph Node Metastasis in Papillary Thyroid Carcinoma.

Huang X, Qi Q, Cai S, Yuan X, Xu P — Fri, 06 Nov 2026 00:00:00 +0000

94% · External validation 94%

Introduction: This study aimed to develop and validate a model integrating ultrasound-based radiomic features from intra-nodular habitats and the peri-nodular region for predicting cervical lymph node (LN) metastasis in papillary thyroid carcinoma (PTC). Methods: A total of 433 PTC patients (256, 111, and 66 in the training, internal, and external validation cohorts, respectively) from two institutions from January 2020 to December 2024 were retrospectively enrolled. The intra-nodular region was segmented and clustered into three heterogeneous habitats by the K-Means method, while the peri-nodular region was defined as the 1-mm area surrounding the nodule. After feature extraction and selection, a predictive model (ITH-Peri model) combining intra-nodular habitat and peri-nodular features was developed using the optimal classifier selected from five machine learning algorithms in the training cohort and was evaluated in the validation cohorts. For comparison, single-habitat, all-habitats, intra-nodular, and peri-nodular radiomic models were constructed. Results: The ITH-Peri model yielded the area under the curve (AUC) of 0.857 and 0.850 in the internal and external validation cohorts, respectively. It demonstrated significantly higher discriminative ability than comparator models, whose AUCs ranged from 0.637 to 0.800 (internal validation) and from 0.641 to 0.787 (external validation) (p < 0.05). Calibration curves and the Hosmer-Lemeshow test (p = 0.307 and 0.583) confirmed the model's good calibration, while decision curves demonstrated its higher net benefit. Additionally, the model showed markedly enhanced accuracy and sensitivity for junior radiologists. Discussion: The ITH-Peri model significantly outperformed comparator radiomic models in both validation cohorts, with good calibration and higher net benefit, suggesting that integrating intra-nodular habitat and peri-nodular features captures complementary predictive information. It could also improve junior radiologists' accuracy and sensitivity, thereby supporting more precise preoperative risk stratification in clinical management. Conclusion: The ITH-Peri model shows promise as a non-invasive and accurate tool for preoperatively predicting cervical LN status in PTC patients.

Deep Learning for AD Risk Screening and Progression Prediction.

Lu B, Chen YR, Li RX, Zhang MK, Yan SZ, Chen GQ — Fri, 06 Nov 2026 00:00:00 +0000

93% · External validation 93%

The rise of Alzheimer's disease (AD) immunotherapies underscores the need for accessible, time-efficient biomarkers for early diagnosis. MRI-based deep learning shows promise in this regard, yet few models have undergone comprehensive clinical validation. Using the SILCODE cohort (722 Chinese participants, 1,105 scans), we test the transportability of our classifier (pretrained on Western samples) to an independent Chinese cohort without retraining. The associations between risk scores and disease progression, cognitive performance, and plasma biomarkers are examined. In addition, an interpretable framework is developed to generate individualized brain risk maps and identify subtypes. The model classifies AD with an AUC of 0.913 and 95.2% sensitivity, identifying 90.5% of high-risk individuals 2-3 years before clinical diagnosis. Risk scores correlate significantly with cognitive and plasma indices, particularly p-tau 217. Brain risk maps reveal AD subtypes, including a subgroup with rapid cognitive decline. These findings establish a population-robust, clinically relevant MRI-based model for early AD detection and risk stratification.

Development and Validation of a Mixture-of-Experts-Based Model for Postoperative In-Hospital Mortality Risk Assessment: Multicenter Retrospective Study.

Huang S, Xia J, Nie X, Zheng Z, Luo J — Fri, 06 Nov 2026 00:00:00 +0000

93% · External validation 93%

Background: Postoperative mortality remains a key indicator of surgical safety and quality. Traditional risk models based on static preoperative or intraoperative variables often fail to capture dynamic physiological changes during surgery and generalize poorly across surgical types. To address these limitations, we developed and validated a deep learning framework capable of learning heterogeneous temporal patterns from intraoperative signals while maintaining interpretability and adaptability. Objective: This study aimed to develop a mixture-of-experts-based model (Postoperative Risk Modeling Expert network, PRIME-MoE) to provide dynamic, surgery-adaptive postoperative in-hospital mortality risk prediction. The primary objective was to improve prediction accuracy, cross-database robustness, and interpretability across heterogeneous surgical populations and institutions. Methods: We retrospectively analyzed 173,990 surgical cases from the INSPIRE (Informative Surgical Patient Dataset for Innovative Research Environment) and HUAXI databases. PRIME-MoE used 4 one-layer unidirectional long short-term memory experts with dense softmax routing based on static surgical context. We evaluated discrimination, calibration, threshold-specific operating points, parameter count, graphics processing unit inference latency, routing patterns, and a sensitivity analysis that truncated the final 30 minutes of each operation without retraining. Results: PRIME-MoE achieved internal-test area under the receiver operating characteristic curves (AUROCs) of 93.8 (95% CI 92.3-96.0) in INSPIRE and 90.7 (95% CI 87.1-94.8) in HUAXI. Reciprocal evaluation retained discrimination (AUROCs 90.8 and 87.8) but showed imperfect calibration (Brier scores 0.035-0.049; calibration slopes 0.28-0.49), supporting site-specific recalibration. Graphics processing unit inference required a median of 11.4 ms (9.1-18.0; 3.3-45.3) per case. Without retraining, scoring 30 minutes before the recorded end of surgery retained AUROCs of 93.1 in INSPIRE and 90.2 in HUAXI. Conclusions: PRIME-MoE provides an interpretable framework for postoperative in-hospital mortality prediction and retained discrimination in reciprocal evaluation across 2 heterogeneous perioperative databases. The 30-minute analysis supports near-end-of-surgery scoring, but calibration was insufficient for direct deployment. Prospective validation, local recalibration, threshold selection, and workflow evaluation are required before clinical use.

Machine Learning-Based Risk Scoring Using Clinical and Surgical Variables for Severe Complications after Rectal Cancer Surgery.

Song SH, Kwon HK, Park JS, Park SY, Kim HJ, Lee J — Fri, 06 Nov 2026 00:00:00 +0000

93% · Prospective 93%

Background: Severe postoperative complications after rectal cancer surgery increase morbidity, delay recovery, and may compromise oncologic outcomes. Existing risk models show limited accuracy and generalizability. Machine learning can integrate complex clinical and surgical variables to improve individualized risk prediction. Objective: To develop and prospectively validate machine learning-based models for predicting severe postoperative complications after rectal cancer surgery. Design: Retrospective model development and prospective temporal internal validation study. Setting: Single tertiary referral center in South Korea. Patients: A total of 3,503 patients who underwent curative rectal cancer resection with anastomosis between 2011 and 2025 were included, comprising 3,402 retrospective and 101 prospective cases. Main Outcome Measures: Severe postoperative complications defined as Clavien-Dindo grade ≥ 3 within 30days after surgery. Results: Severe complications occurred in 9.8% of the retrospective and 13.9% of the prospective cohort. Four machine-learning algorithms - logistic regression, random forest, artificial neural network, and gradient boosting - were trained using 5-fold crossvalidation. In internal validation, the random forest model achieved the best performance (area under the receiver operating characteristic Curve 0.898; Sensitivity 0.685), while gradient boosting performed best in the temporal internal validation cohort (area under the receiver operating characteristic Curve 0.910; Sensitivity 0.707). The most influential predictors were tumor perforation, intraoperative transfusion, protective stoma, bowel obstruction, and cerebrovascular accident. Limitations: Validation was single institution (temporal internal, not external) with a small prospective cohort. Conclusions: Machine learning-based models accurately predicted severe postoperative complications after rectal cancer surgery, with strong discrimination maintained on temporal internal validation. Applied immediately after surgery, they may support individualized postoperative risk stratification and monitoring; external validation is warranted. See Video Abstract.

Artificial Intelligence-Driven Extraction of Patient-Reported Pain Outcomes in Osteoarthritis Using Large Language Models.

Doshi J, Batter S, Wu Y, Salerno W, Santilli A, Kannayiram S — Fri, 06 Nov 2026 00:00:00 +0000

92% · Language models 92%

Objective: Patient-reported outcome measures (PROMs) are central to assessing symptom burden and treatment response in Osteoarthritis (OA) but are often embedded in unstructured clinical notes. We evaluated whether prompt-engineered large language models (LLMs) could accurately extract or infer pain scores on a 0-10 scale from routine OA clinical documentation. Methods: We retrospectively analyzed 163 paired internal medicine and orthopedic clinic notes from patients in a prospective OA registry. A trained abstractor extracted explicitly documented pain scores from notes. For notes lacking explicit pain scores, physicians estimated pain severity from narrative documentation. The Llama 3.2-90B model, using Chain-of-Thought and Few-Shot prompting, inferred pain scores. Registry-collected pain scores served as a reference standard. Agreement was assessed using Kendall's Tau-b. Results: Agreement between LLM-extracted and abstracted scores was strong for IM notes (n =132, τ = 0.76, p < 0.0001) and excellent for orthopedic notes (n = 65, τ = 0.95, p < 0.0001). The LLM identified 15 EMR-generated default zero scores in IM notes; after excluding them, agreement was perfect (τ = 1.0, p < 0.0001). For notes without explicit pain scores (31 IM and 98 orthopedic notes), LLM-based pain inference showed weak correlation with registry scores (IM: τ = 0.17; orthopedic: τ = 0.11), but performed comparably to physician estimation. LLM-physician agreement ranged from 54.2%-88.0% within ±1 point and 66.7%-100% within ±2 points for IM notes, and from 54.8%-70.5% within ±1 point and 84.5%-97.7% within ±2 points for orthopedic notes. Conclusion: LLMs accurately extract explicitly documented PROMs from clinical notes and reliably identify documentation artifacts. Inference from narrative documentation was less precise when notes lacked sufficient contextual information and detailed pain descriptions.

★ An integrated CT/MRI diagnostic system for pancreatic cystic lesions that supports pathological classification and risk stratification.

Cui X, Liu H, Zheng S, Wang H, Liu J, Gao Z — Wed, 04 Nov 2026 00:00:00 +0000

97% · Prospective 97%, External validation 79%

Objective: To develop a pathology-based preoperative decision-support system (Structured Imaging Report System for Pancreatic Cystic Lesions, SIRS-PCL) for subtype classification and risk stratification of PCLs. Methods: We conducted a multicentre retrospective study involving 1,102 patients treated between 1 January 2011 and 31 December 2024. The study cohort was divided into a development dataset (n = 858) and a temporal validation dataset (n = 244) from the periods 2011-2022 and 2023-2024, respectively. Three logistic prediction models were used to classify the five PCL subtypes and to perform separate three-tier risk stratifications (benign/low, medium, and high) for intraductal papillary mucinous neoplasm (IPMN) and mucinous cystic neoplasm (MCN) cases. SIRS-PCL was prospectively implemented at Tianjin Medical University Cancer Institute and Hospital from 25 June 2025 to 25 June 2026, and the diagnostic accuracy, sensitivity, and specificity were evaluated in a cohort of 76 patients. Results: A total of 1,102 patients with PCLs were included in the study (median age 56 years, interquartile range 44-64), of whom 714 were female (64.8%). The five-category model included in SIRS-PCL achieved a mean area under the receiver operating characteristic curve (AUC) of 0.93 [95% confidence interval (CI) 0.91-0.95] in temporal validation. The IPMN model showed good discrimination for risk stratification (mean AUC 0.86, 95% CI 0.79-0.93), with the highest overall accuracy of 77.1% exceeding current guidelines (range 50.0%-58.6%). The MCN risk model also showed high performance (mean AUC 0.97, 95% CI 0.91-1.00). In comparison to current practice, SIRS-PCL use reduced unnecessary surgeries 11.5% vs. 23.1% and over-examinations 6.7% vs 33.7%. In the prospective implementation cohort, SIRS-PCL achieved accuracy of 86.2% for five-category pathological classification, 73.7% for IPMN risk stratification, and 100% for MCN risk stratification. Conclusions: SIRS-PCL is a pathology-based, quantitative decision-support system for preoperative pathological subtype classification and pathological risk stratification of surgically evaluated neoplastic PCLs.

★ Zaizao Shengxue capsule for anemia: a prospective multicenter real-world study of hemoglobin response and predictive modeling.

Xing F, Wu J, Cheng C, Chen S, Qin Z, Wang Y — Wed, 04 Nov 2026 00:00:00 +0000

96% · External validation 96%, Prospective 61%

Objective: The aim in this study was to characterize the 8-week hemoglobin response among patients with anemia receiving Zaizao Shengxue capsule (ZSC) in real-world care and develop and externally evaluate a model for estimating the individual probability of response. Methods: This prospective multicenter study included 440 patients from 14 centers who received ZSC for 8 weeks. Response was defined as an increase in hemoglobin of at least 20 g/L from baseline at week 8. The development cohort (n = 338) was divided into a training cohort (n = 238) and an internal validation cohort (n = 100). Baseline predictors were selected in the training cohort. Ten candidate models were trained using different algorithms and compared in the internal validation cohort. After model selection and locking, the final model was independently evaluated in 102 patients from four external centers. SHAP and LIME were used for model interpretation. Results: Overall, 204 of 440 patients (46.4%) achieved an 8-week hemoglobin response. Response rates varied across anemia etiologies: 24.0% in cancer-related anemia (60/250), 74.4% in iron-deficiency anemia (67/90), and 77.0% in anemia of chronic disease (77/100). Anemia etiology, age, and estimated glomerular filtration rate (eGFR) were retained as predictors in the final Extra Trees model. The model achieved an area under the receiver operating characteristic curve (AUC) of 0.821 (95% CI, 0.729-0.902) in internal validation and 0.769 (95% CI, 0.670-0.860) in external validation. SHAP analysis identified anemia etiology as the largest contributor to model predictions, followed by age and eGFR. Conclusion: Eight-week hemoglobin response varied across anemia etiologies among patients receiving ZSC. A model incorporating anemia etiology, age, and eGFR may support individualized pretreatment estimation of response probability based on routinely available clinical information.

★ DeepFusion-HSK: a multimodal deep learning for subtype classification of herpes simplex keratitis with multicenter validation.

Zhang Z, Lv T, Tang J, Xu K, Li FF, Yu X — Wed, 04 Nov 2026 00:00:00 +0000

96% · External validation 96%

Introduction: To develop and evaluate a multimodal deep learning (DL) model for subtype classification of viral keratitis. Methods: This retrospective multicenter study included 3,501 slit-lamp images (807 eyes) as the internal development set and two external validation cohorts (2,957 images total). DeepFusion-HSK integrates three complementary illumination modalities-diffuse light imaging (DLI), slit-beam imaging (SBI), and fluorescein staining imaging (FSI)-through a novel fusion architecture. The framework uses a tri-way higher-order interaction layer to capture cross-modal pathological signatures, dual-task learning with auxiliary heads to prevent modality-specific feature degradation, and a dynamic gating mechanism that uses high-confidence coarse priors to calibrate fine-grained predictions and mitigate softmax probability spillover. To simulate real-world data deficiency, we evaluated model resilience under complete FSI dropout. Results: On the internal test set, DeepFusion-HSK achieved an accuracy of 85.1% (95% CI: 78.5%-90.0%) and AUCs of 0.947-0.949 for discriminating epithelial herpes simplex keratitis (E-HSK) and non-epithelial herpes simplex keratitis (NE-HSK), significantly outperforming Swin-Transformer, EfficientNetV2, and ResNet50. External validation confirmed robust generalization, with accuracies of 81.7% on Dataset I and 78.2% on Dataset II, demonstrating clinical resilience. Ablation studies confirmed the hierarchical synergy of multimodal inputs: the best single modality (FSI) yielded 78.5% accuracy, the optimal dual-modality combination (DLI + FSI) reached 81.7%, and full tri-modal fusion improved performance to 85.1%, validating that integrated global morphology, depth contour, and superficial ulcer signatures are essential for robust HSK subtyping. Discussion: DeepFusion-HSK offers a robust, generalizable, and clinically deployable solution for therapeutic HSK subtyping. Its resilience to missing imaging modalities makes it well suited for heterogeneous real-world ophthalmic practice, providing a decision-support tool to reduce vision-threatening misdiagnoses and guide therapy.

Radiographic manifestations of pneumatosis intestinalis and construction of an AI-assisted diagnostic system.

Zhang X — Wed, 04 Nov 2026 00:00:00 +0000

93% · External validation 93%

Pneumatosis intestinalis (PI) has a wide range of etiologies, and the imaging differentiation between benign and ischemic PI is difficult, leading to a high risk of missed diagnosis. This study systematically summarizes the CT (Computed Tomography) imaging manifestations of PI and constructs an AI (Artificial Intelligence)-assisted diagnostic model that integrates multidimensional features. This study retrospectively collected CT and x-ray data from 612 patients with PI (398 benign and 214 ischemic cases) from three tertiary hospitals, and included an external validation set of 106 cases. Imaging analysis established a five-level PI-CT severity scoring system based on gas morphology, distribution, portal venous gas (PVG), intestinal wall enhancement, mesenteric edema, and ascites. The AI system adopted a two-level architecture: The first level used EfficientNet-B3 for rapid screening on x-ray films; the second level used 3D nnU-Net (No-new-Net) to segment intestinal wall gas, and then fused the data through a three-branch network (gas morphology, accompanying signs, and clinical metadata) via cross-attention to output the probability of ischemic PI. In the internal test set, the model achieved an AUC (Area Under the Curve) of 0.947 for ischemic PI (sensitivity 91.3%, specificity 93.6%), significantly outperforming senior radiologists (AUC = 0.876); the AUC on the external validation set was 0.931. SHapley Additive exPlanations analysis showed that PVG and blood lactate were the most important predictors. This two-tiered AI system can significantly improve the early identification of the causes of critical PI, providing a reliable method for intelligent diagnosis of emergency imaging.

Beyond Risk Scores: Concordance of a Multidisciplinary Decision Support Instrument with Real Life Treatment Decisions for Patients with Ruptured Abdominal Aortic Aneurysm.

Wed, 04 Nov 2026 00:00:00 +0000

92% · Prospective 92%, External validation 44%

Objectives: Treatment decisions for patients with ruptured abdominal aortic aneurysm (rAAA) are ethically complex and must be made under time pressure. This prospective observational study assessed how a previously developed decision support instrument (DSI) for rAAA aligns with real life treatment decisions and determined a reference threshold for interpreting its output. The DSI estimates the proportion of experts from the original multidisciplinary expert panel who would be expected to favour operative treatment (OT) for a given patient profile. Methods: Consecutive rAAA patients were included between March 2024 and January 2026 from two Dutch hospitals. Treating vascular surgeons completed the DSI after the actual treatment decision (OT or palliative care). DSI outputs were analysed in relation to observed treatment decisions. A threshold sweep analysis was performed to identify a contextual reference threshold reflecting the level of DSI estimated expert support at which OT was generally pursued. Post-operative outcomes were assessed in OT patients for clinical context. Results: Fifty patients were included; OT was chosen in 43 patients (86%). DSI estimated expert support was higher for patients receiving OT compared with palliative care (median 98% [interquartile range 97, 99%] vs. 83% [59, 89%]; p < .001). The contextual reference threshold was 93%, with concordance of 94% (95% confidence interval 83.8 - 97.9%). Case analysis highlighted challenges in capturing unresponsiveness due to profound hypovolaemic shock and assessing cardiac burden acutely Among patients surviving OT, relatively high health related quality of life was observed. Conclusions: The DSI shows concordance with real life treatment decisions for rAAA. A contextual reference threshold above which operative treatment was generally pursued could be empirically derived. Although refinement is needed to better identify certain high risk patients, these findings represent an important first step in the validation pathway and provide proof of concept for behavioural artificial intelligence technology methodology.

★ Automatic Ultrasound Habitat Radiomics for Differentiating Henoch-Schonlein Purpura Nephritis from IgA Nephropathy in Children: A Multicenter Study.

Gong S, Zhang W, Xu T, Zhang XF, Zhang MJ, Gao P — Tue, 03 Nov 2026 00:00:00 +0000

97% · External validation 97%

Rationale And Objectives: To assess a noninvasive ultrasound habitat radiomics approach with automatic kidney segmentation for distinguishing pediatric Henoch-Schonlein purpura nephritis (HSPN) from immunoglobulin A nephropathy (IgAN). Materials And Methods: This retrospective multicenter study included 592 biopsy-confirmed pediatric patients from three centers (HSPN, n = 338; IgAN, n = 254). The development cohort (n = 453) was split into training (n = 362) and internal validation (n = 91) cohorts, and an independent external validation cohort (n = 139) was used for final testing. Kidney regions of interest were automatically segmented using nnU-Net. K-means clustering defined three echogenic habitat subregions, and 408 radiomics features were extracted from the whole-kidney region and habitat subregions. Feature selection was performed in the training cohort, and three machine-learning classifiers were evaluated. The selected radiomics model was compared with clinical and combined models. Results: nnU-Net achieved high kidney segmentation performance, with a mean Dice similarity coefficient of 0.957 and a mean intersection over union of 0.918 in the independent healthy-child test set. Fifteen radiomics features were retained, including six whole-kidney features and nine habitat-derived features. Logistic regression showed stable performance, with areas under the receiver operating characteristic curve (AUCs) of 0.806 (95% confidence interval [CI], 0.715-0.897) internally and 0.825 (95% CI, 0.758-0.892) externally. The combined model outperformed the clinical model in AUC in both cohorts (0.846 vs 0.717; 0.805 vs 0.733). Conclusion: This deep-learning-assisted ultrasound habitat radiomics approach enabled automatic kidney delineation and showed encouraging performance as a reproducible imaging adjunct for differentiating pediatric HSPN from IgAN.

★ Predicting infection risk in rheumatoid arthritis patients receiving biological or targeted synthetic disease-modifying anti-rheumatic drugs: an application of machine learning and healthcare big data.

Peng K, Yang D, Wang J, Shen CY, Cheng MC, Chan SCW — Tue, 03 Nov 2026 00:00:00 +0000

97% · External validation 97%

Background: Patients with rheumatoid arthritis (RA) initiating biologic or targeted synthetic disease-modifying antirheumatic drugs (b/ts DMARDs) face elevated risk of serious infections, necessitating tools for individualized risk stratification. Objectives: Primary objective was to develop and validate a clinically interpretable machine learning (ML) model to predict 1-year risk of serious infection after b/ts DMARD initiation; secondary objectives were to estimate infection incidence and identify key predictors associated with risk. Methods: We performed a retrospective cohort study using territory-wide EHR from Hong Kong's Clinical Data Analysis and Reporting System (CDARS) for model development and internal validation, with external validation in the U.S. All of Us database. The outcome was first serious infection requiring hospitalization within 1 year. Candidate predictors included demographics, comorbidities, prior infections and medications, laboratory markers. Multiple ML algorithms were trained; model selection was based on AUROC, and interpretability was assessed using SHAP. Results: A total of 3,159 patients from CDARS (8.8% with serious infections) and 1,845 from All of Us (2.8% with serious infections) were included. The model demonstrated the highest AUROC in internal validation (0.840, 95% CI: 0.793-0.888) and maintained robust performance in external validation (AUROC: 0.729, 95% CI: 0.665-0.793). Key predictors included prior infections, diabetes, b/ts DMARD type, and inflammatory markers. Rituximab was linked to the highest infection risk, while tofacitinib and upadacitinib had the lowest. Conclusion: This study developed and validated an ML model using routine clinical data to predict serious infection risk in RA patients, supporting personalised treatment and proactive infection management.

★ Translating the familial hypercholesterolemia pediatric diagnostic score (FH-PeDS) into clinical practice: Development of an open-access online tool.

Sikonja J, Kafol J, Groselj U — Tue, 03 Nov 2026 00:00:00 +0000

95% · External validation 95%

Background: More than 90% of children with familial hypercholesterolemia (FH) remain undiagnosed. Although genetic testing is the diagnostic gold standard, limited availability necessitates the use of clinical diagnostic criteria, which have suboptimal performance in children. Objective: To introduce the recently developed familial hypercholesterolemia pediatric diagnostic score (FH-PeDS) and its online implementation. Methods: FH-PeDS was developed using real-world data from Slovenian children identified through universal cholesterol screening and externally validated in an independent Portuguese pediatric cohort. Results: FH-PeDS demonstrated improved diagnostic accuracy compared with traditional clinical scoring systems and was recently recognized as promising tool in the European Atherosclerosis Society consensus statement on FH in children and adolescents. It is now freely available at https://fh-peds.eu in two versions: a semi-quantitative clinical scoring system and a machine learning model. Conclusion: The online FH-PeDS provides healthcare professionals with an accessible tool to estimate the likelihood of pediatric FH, potentially facilitating earlier diagnosis and treatment. Further external validation is warranted before widespread implementation.

A model of brain function to predict tau across Alzheimer's disease and related disorders.

Corriveau-Lecavalier N, Zhang G, Botha H, Barnard LR, Wiste HJ, Dicks E — Tue, 03 Nov 2026 00:00:00 +0000

91% · External validation 91%

We developed a machine learning framework grounded in a computational model of brain function and FDG-PET imaging to predict Alzheimer's disease (AD)-type tau across AD and related disorders (ADRD). The model was trained on 1088 research participants spanning normal aging and ADRD with FDG-PET and flortaucipir-PET (FTP-PET). Validation was conducted in: (1) a longitudinal subset of the training cohort (n = 215), (2) Alzheimer's Disease Neuroimaging Initiative (n = 295), and (3) a heretogeneous clinical cohort (n = 264; n = 28 with FTP-PET). Results showed that the model accurately predicted FTP-PET when applied at the participant-level (R2 ~ 0.34-0.71). It performed well in clinical scenarios, including classification of low versus high tau stages (79-95% accuracy) and prediction of tau in a heterogeneous clinical cohort without FTP-PET. These findings advance the understanding of the relationship between neurodegeneration and AD-type tau and provide a framework for estimating AD-type tau across ADRD.

★ Development and external validation of interpretable machine learning models for 7-, 14-, and 28-day mortality in sepsis-associated liver injury.

Yao H, Li Z, Lü J, Shen W — Mon, 02 Nov 2026 00:00:00 +0000

99% · External validation 99%

Background: Mortality in sepsis-associated liver injury (SALI) remains high, and early risk stratification is difficult. Existing prognostic models commonly focus on a single endpoint, usually 28-day mortality, and rarely undergo independent external validation. Methods: We developed machine-learning models using clinical features recorded during the first 24 h to predict all-cause in-hospital mortality within 7, 14, and 28 days. In MIMIC-IV (v3.1), 1,550 adults meeting the prespecified SALI criteria were allocated to a training cohort (n = 1,119) and a held-out internal validation cohort (n = 431). An independent Chinese cohort from Xiaogan Central Hospital (n = 229; 28-day in-hospital mortality, 45.9%) was used for external validation. One hundred candidate pipelines were evaluated using stratified 10-fold cross-validation. SHAP was used to describe feature attribution across prediction horizons. Results: AUROCs in the held-out internal validation cohort were 0.840, 0.841, and 0.850 for 7-, 14-, and 28-day mortality, respectively; corresponding external AUROCs were 0.720, 0.717, and 0.763. The external 28-day model outperformed modified SOFA (0.763 vs. 0.673; P = 0.011) and MELD-Na (0.763 vs. 0.636; P = 0.002). A parsimonious logistic-regression sensitivity model achieved an external 28-day AUROC of 0.762, not significantly different from the full model (P = 0.970). The 14-day model was miscalibrated externally; recalibration improved the calibration slope from 0.563 to 0.970 but did not materially improve the Brier score. SHAP rankings varied across horizons and were interpreted descriptively rather than causally. Conclusion: The models showed moderate transportability to one independent retrospective cohort. Multicentre prospective validation is needed before clinical implementation.

★ Predicting persistent fever in patients with cancer receiving antibiotics: a machine learning approach.

Pucher G, Kanjilal S, Kopp K, Deep A, Rostalski T, Schuler M — Mon, 02 Nov 2026 00:00:00 +0000

98% · External validation 98%

Persistent fever after initiating intravenous antibiotics is common in hospitalized patients with cancer and often leads to additional imaging and antimicrobial escalation. Early identification of patients unlikely to remain febrile could provide additional information during clinical reassessment. We developed and externally validated machine learning models to predict persistent fever 48-72 h after antibiotic initiation using electronic health records from two academic hospitals. Adult cancer patients with fever within 24-48 h of treatment were included. Models incorporated demographics, comorbidities, laboratory results, vital signs, and temperature patterns. The best-performing model achieved an AUROC of 0.82 (95% CI 0.79-0.84) internally and 0.78 (95% CI 0.77-0.79) externally. Temperature-related variables were the dominant predictors. Decision curve analysis demonstrated positive net benefit across much of the evaluated threshold range. Retrospective assessment suggested that approximately one quarter of antibiotic escalations and chest CT scans occurred in patients correctly predicted not to remain febrile, highlighting the potential contextual value of fever predictions. These findings indicate that short-term fever persistence can be predicted with clinically meaningful accuracy using routine data. However, clinical use should be limited to an adjunctive role rather than standalone rule-out decision-making, and prospective studies are required to evaluate safety, clinician uptake, and real-world impact.

★ Development and validation of a prognostic prediction model for elderly patients with sepsis-associated acute kidney injury based on interpretable machine learning.

Zhang N, Wang JZ, Ma Q, Tian JH, Zhou WJ — Mon, 02 Nov 2026 00:00:00 +0000

98% · External validation 98%

Background: Elderly patients with sepsis-associated acute kidney injury (SA-AKI) have high short-term mortality; early identification of high-risk individuals is critical for personalized intervention. Traditional clinical scoring systems relying on linear assumptions and conventional machine learning models often lack interpretability, hindering translation to clinical practice. This study aimed to develop and externally validate an interpretable prognostic model for 28-day in-hospital mortality in elderly SA-AKI using routinely available clinical variables. Methods: Clinical data of 11,468 elderly SA-AKI patients were retrospectively extracted from the MIMIC-IV database and divided into training and internal test sets at a 7:3 ratio using stratified random sampling. A dual feature selection strategy combining LASSO regression and random forest was applied. Five machine learning models-decision tree, logistic regression, random forest, LightGBM, and XGBoost-were constructed in parallel. Model discrimination and calibration were evaluated by AUC, Brier score, accuracy, sensitivity, specificity, and F1 score. External independent validation was performed using a clinical cohort of 257 patients from the General Hospital of Ningxia Medical University. The SHAP method was used for global and individual-level interpretability of the optimal model. Results: Dual feature selection identified 10 independent predictors: age, weight, red blood cell distribution width, arterial partial pressure of oxygen, blood lactate, prothrombin time, activated partial thromboplastin time, blood urea nitrogen, serum creatinine, and respiratory rate. Among the five models, XGBoost demonstrated the best performance, with training AUC 0.9898 (accuracy 0.9472) and external validation AUC 0.9156 (accuracy 0.8560), indicating strong generalizability. SHAP global importance ranking showed age as the primary driver, followed by serum creatinine and arterial partial pressure of oxygen. Each core feature exhibited a nonlinear association with 28-day mortality risk, with clear model-derived risk inflection points rather than clinical intervention thresholds. Conclusion: The XGBoost-based interpretable model exhibits excellent discriminative performance and robust external generalizability for predicting 28-day mortality in elderly SA-AKI. The SHAP attribution framework renders model decisions transparent, providing quantitative evidence-based support for early risk stratification and hypothesis generation for individualized organ-protective strategies in the ICU.

★ Rapid and label-free analysis of biological characteristics of hepatocelluar carcinoma tissues based on histological images.

Zhang H, Li J, Sheng J, Li Z, Jiang W, Xu K — Mon, 02 Nov 2026 00:00:00 +0000

97% · External validation 97%

Introduction: Hepatocellular carcinoma (HCC) exhibits substantial molecular heterogeneity, complicating treatment selection and clinical management. Objectives: We developed an artificial intelligence-based histopathological framework to predict HCC biological characteristics from routine hematoxylin and eosin (H&E)-stained slides. Methods: Using The Cancer Genome Atlas Liver Hepatocellular Carcinoma (TCGA-LIHC) cohort, transcriptomic phenotypes associated with angiogenesis, stemness, and autophagy were identified through gene set analysis, differential expression analysis, and consensus clustering, and evaluated for overall survival and drug sensitivity. Whole-slide images were processed into 224 × 224-pixel tumor tiles for phenotype prediction using feature extraction (UNI, ResNet50, CTrans, GigaPath, and Virchow2) combined with classic multiple instance learning and clustering-constrained attention-based multiple instance learning.. Models were externally validated at Renmin Hospital of Wuhan University and Tongji Hospital. Interpretability was assessed using attention heatmaps, and predictions were evaluated against molecular markers and tumor differentiation grades. Results: Unsupervised clustering identified two subgroups per phenotype, with significant difference in survival (all p < 0.004) and sensitivities to sorafenib or transcatheter arterial chemoembolization (all p < 0.0001). The tumor detection model achieved an area under the receiver operating characteristic curve (AUC) of 97.53%. The phenotype prediction models achieved internal optimal AUCs of 85.51%-91.05% and accuracies of 79.63%-88.89%. External optimal AUCs were 70.94%, 88.64%, and 91.17% for angiogenesis, stemness, and autophagy, with accuracies of 62.00%, 78.00%, and 84.00%, respectively. Attention heatmaps revealed distinct morphological features among subgroups, including neovascularization, increased nuclear-to-cytoplasmic ratios, and cytoplasmic vacuolization. Model predictions showed concordance with CD31/CD105 (κ = 0.398-0.432), CD133/LGR5 (κ = 0.428-0.629), Beclin-1/p62-LC3B (κ = 0.592-0.762) (all p < 0.05), and tumor differentiation grade (κ = - 0.167 to - 0.175, p < 0.05). Conclusion: This proof-of-concept AI framework predicted transcriptome-defined HCC phenotypes from routine histology, with predictions associated with prognosis and therapeutic response. It may provide a cost-effective approach for assessing HCC heterogeneity and complement molecular profiling.

★ Non-Invasive Assessment of Perineural Invasion Density: A Novel Marker for Prognosis and Treatment Stratification in Prostate Cancer.

Liu B, Zhang L, Zhu W, Yang M, Liu X, Li X — Mon, 02 Nov 2026 00:00:00 +0000

96% · External validation 96%

We established and validated perineural invasion density (PNID) as a superior prognostic marker and developed an MRI-based AI model for its preoperative prediction. In this multicenter study of 1091 PCa patients undergoing radical prostatectomy, we evaluated PNID's association with biochemical recurrence (BCR) and radiographic progression-free survival (rPFS). We also analyzed concurrent β-blocker or antimuscarinic use with adjuvant ADT. A deep learning model was developed in a primary cohort (n = 523) and validated in two external cohorts (n = 317, 251). High-PNID was an independent prognostic factor for BCR (HR 4.06, 95% CI: 2.75-5.97, P < 0.001) and rPFS (HR 4.61, 95% CI: 2.97-7.16, P < 0.001). Notably, concurrent β-blocker use was significantly associated with favorable BCR outcomes exclusively in the High-PNID subgroup (HR 0.32, 95% CI: 0.12-0.84, P = 0.021), with no benefit in the Non-High PNID population or with antimuscarinics. The AI model showed excellent performance predicting High-PNID, with AUCs of 0.917 (training), 0.842 (internal validation), 0.889 and 0.915 (external validations). Model-identified High-PNID patients had significantly worse outcomes. PNID surpasses traditional PNI in prognostic accuracy and identifies a high-risk phenotype. High-PNID may define a subgroup benefiting from adjuvant β-blocker therapy, providing an exploratory clinical rationale for targeting the nerve-tumor microenvironment in stratified subpopulations. Hence, this model may be a reliable, non-invasive tool for preoperative identification of this high-risk subgroup, aiding clinicians in personalized risk stratification and treatment planning.

Pathology-guided multimodal transformer predicts postoperative survival benefit from adjuvant chemotherapy in glioma: from DNA damage repair to immune characterization.

Li Z, Duan X, Lian J, Li X, Liang Q, Liu X — Mon, 02 Nov 2026 00:00:00 +0000

94% · External validation 94%

Introduction: High-grade gliomas (HGGs) are aggressive primary brain tumors with high recurrence rates and poor prognosis. Survival benefit from adjuvant temozolomide (TMZ) varies substantially across patients, and reliable preoperative biomarkers for treatment stratification remain limited. This study aims to develop an interpretable multimodal deep learning model integrating whole-slide histopathological images and multiparametric MRI to preoperatively predict survival benefit from adjuvant TMZ in HGG patients. Methods: A Transformer-based multimodal fusion model was proposed. Key tumor regions were identified from whole-slide pathology images via weakly supervised clustering, and pathological features were extracted using attention-based multiple instance learning. Concurrently, 2.5D tumor subregional features were derived from multiparametric MRI sequences. These multimodal features were deeply fused using a Transformer encoder. The model was developed and validated on a multicenter cohort (504 patients), and its biological basis was explored through transcriptomic analysis, gene set enrichment analysis, protein-protein interaction networks, and immune deconvolution. Results: The fusion model achieved AUCs of 0.910, 0.861, and 0.833 in the training, internal validation, and external validation cohorts, respectively. Transcriptomic analysis implicated DNA damage repair processes in high-benefit patients, suggesting that enhanced DNA repair capacity may contribute to TMZ sensitivity. Immune deconvolution showed higher M1 macrophage infiltration and lower regulatory T cell abundance in the high-benefit group, consistent with a less immunosuppressive tumor microenvironment. Importantly, low-benefit patients who completed TMZ chemotherapy had shorter survival than those treated with concurrent chemoradiotherapy alone or surgery alone. Conclusion: This model identifies patients unlikely to benefit from standard TMZ chemotherapy, providing a tool for postoperative personalized treatment stratification and optimization of therapeutic strategies in HGGs.

Systemic immune dysregulation and local barrier failure as predictors of septic shock in acute appendicitis: a large-scale multicenter interpretable machine learning study.

Chen Y, Rui Y — Mon, 02 Nov 2026 00:00:00 +0000

94% · External validation 94%

Background: The transition from localized acute appendicitis to systemic septic shock is driven by complex immune and metabolic dysregulation combined with anatomical barrier failure. Accurately predicting early hemodynamic deterioration remains clinically challenging. This study aimed to develop an interpretable multimodal machine learning model by integrating systemic immunological profiles with local sonographic characteristics. Methods: This multicenter cohort study included 3000 patients with acute appendicitis. Following rigorous data reduction using the least absolute shrinkage and selection operator regression, eight supervised machine learning algorithms were constructed. The optimal predictive algorithm was comprehensively evaluated across training, internal testing, and independent external validation cohorts. The SHapley Additive exPlanations framework was utilized to provide transparent model interpretability, and the final algorithm was deployed as a dynamic digital calculator. Results: Six independent predictors were identified, including the Systemic Immune Inflammation Index, C-reactive protein to albumin ratio, glucose to lymphocyte ratio, Prognostic Nutritional Index, periappendiceal fluid, and the sonographic loss of the submucosal echogenic layer. The Random Forest model demonstrated the highest discriminative performance among the evaluated algorithms, achieving an area under the curve of 0.914. The model's discrimination, calibration, and decision-analytic performance were consistently evaluated in the internal testing and external validation cohorts. SHAP analysis indicated that systemic inflammatory indices and sonographic features reflecting local disease severity made important contributions to the model's predicted risk of septic shock. These findings should be interpreted as predictive associations rather than evidence of biological causality or synergy. Conclusion: This multicenter study developed and externally validated an interpretable machine learning model for early risk stratification of septic shock. The publicly accessible calculator may support individualized risk assessment; however, prospective impact evaluation is required to determine whether its clinical implementation facilitates earlier resuscitation or improves patient outcomes.

Deep learning-based prediction of regional brain Tau PET burden from MRI.

Kim D, Ho BC, van Voorst H, Liu Y, Benzinger TLS, Sha S — Mon, 02 Nov 2026 00:00:00 +0000

94% · External validation 94%

Introduction: Regional tau is strongly associated with Alzheimer's disease (AD) symptoms, but tau positron emission tomography (PET) is not routinely available, limiting early risk stratification. Methods: We trained a deep learning model to predict regional tau PET standardized uptake value ratios (SUVRs) from magnetic resonance imaging (MRI) with and without basic demographics using public cohorts (Alzheimer's Disease Neuroimaging Initiative [ADNI], Open Access Series of Imaging Studies 3 [OASIS3-tau], Anti-Amyloid Treatment in Asymptomatic Alzheimer's Disease Study [A4], Harvard Aging Brain Study [HABS]). Mean absolute error (MAE), Pearson r, area-under-the-receiver-operating-curve (AUC), and comparison with cognitive testing were assessed in external test sets (SCAN, Stanford University). Results: The model achieved an MAE of 0.14 (95% confidence interval [CI]: 0.12-0.16) and r = 0.72 for an early uptake region (Braak I-IV), and an MAE of 0.14 (95% CI: 0.12-0.17) and r = 0.71 for a late uptake region (Braak V-VI). AUCs for binary tau positivity in these regions were 0.87 (95% CI: 0.83-0.90) and 0.88 (95% CI: 0.84-0.91), respectively. Discussion: Predicting regional tau burden from magnetic resonance imaging (MRI) is feasible and may enable scalable tau-informed clinical AD care.

★ Explainable Machine Learning for Predicting Major Adverse Renal and Cardiovascular Events in Acute Myocardial Infarction With Chronic Kidney Disease: Algorithm Development and Validation.

Feng W, Jiang R, Huang W, Wang C, Wang Z, Chang J — Fri, 30 Oct 2026 00:00:00 +0000

98% · External validation 98%

Background: Major adverse renal and cardiovascular events (MARCE) have been proposed as a comprehensive end point reflecting both cardiac and renal outcomes in patients with acute myocardial infarction (AMI) and chronic kidney disease (CKD). However, existing prediction tools lack accuracy and generalizability for this high-risk population. Objective: This study aimed to develop and externally validate explainable machine learning models and a nomogram for predicting in-hospital MARCE among patients with AMI and CKD. Methods: We enrolled patients with AMI and CKD who underwent emergency treatment at the First Affiliated Hospital of Wenzhou Medical University as the discovery cohort and those at Wenzhou TCM (traditional Chinese medicine) Hospital of Zhejiang Chinese Medical University as the external validation cohort. Clinical features, including demographic details, comorbidities, laboratory results, echocardiography measurements, procedural details, and medication parameters, were collected for all participants. Important variables were selected using least absolute shrinkage and selection operator regression in the training set. We developed extreme gradient boosting (XGBoost), random forest, decision tree, and logistic regression (LR) models and evaluated their discrimination, calibration, and classification metrics in the hold-out test set and external cohort. Results: In the training set, we identified 6 important variables (Killip class, lactate, estimated glomerular filtration rate, use of diuretics, use of beta-blockers, and use of renin-angiotensin system inhibitors). In the test set, model discrimination was comparable across XGBoost (area under the receiver operating characteristic curve [AUC] 0.882, 95% CI 0.816-0.947), LR (AUC 0.875, 95% CI 0.799-0.950), and random forest (AUC 0.870, 95% CI 0.800-0.940), while decision tree showed lower performance (AUC 0.814, 95% CI 0.720-0.886). Shapley additive explanations (SHAP) values illustrated both global and individual feature contributions in the XGBoost model. A nomogram based on LR achieved an AUC of 0.873 (95% CI 0.714-0.926) in external validation. Conclusions: Our study identified 6 important predictors and provided a series of prediction models to enhance risk stratification of MARCE for patients with AMI and CKD. These models might aid in stratifying risk and choosing appropriate therapeutic strategies in the management of patients with AMI and CKD.

★ CT habitat and peritumoral radiomics for predicting progression after repeat TACE in hepatocellular carcinoma: a multicenter study.

Wang F, Hu Y, Shi ZX, He XY, Sui YH, Wen FK — Fri, 30 Oct 2026 00:00:00 +0000

98% · External validation 98%

Background: This study developed and validated an machine-learning-based radiomics model integrating tumor habitat radiomics, peritumoral radiomics, and clinical features to predict progression-free survival (PFS) following repeat transarterial chemoembolization (TACE) in patients with hepatocellular carcinoma (HCC). Methods: This retrospective multicenter study included 270 patients with HCC. Patients from Center 1 were assigned to training (n = 94) and internal validation (n = 41) cohorts, while Centers 2 (n = 82) and 3 (n = 53) served as external validation cohorts. Tumor habitats and 5-mm peritumoral regions were generated from contrast-enhanced CT images. Radiomics features were selected using univariate Cox and LASSO-Cox regression. A fusion model and HPC risk score were constructed by integrating selected radiomics signatures and clinical variables. Performance was assessed using time-dependent ROC, C-index, calibration, decision curve analysis, and Kaplan-Meier analyses. Results: The optimal habitat cluster number was K = 2. Eleven habitat features, three peritumoral features, and alpha-fetoprotein were retained in the fusion model. The model achieved AUCs of 0.923, 0.921, and 0.897 for 3-, 6-, and 9-month progression-free survival in the training cohort, and 0.932, 0.863, and 0.916 in the internal validation cohort, respectively; all external AUCs exceeded 0.82. The fusion model showed higher C-index values than the clinical model across cohorts. High-risk patients had significantly shorter PFS than low-risk patients (all P < 0.001). Calibration and decision curve analyses demonstrated good agreement and clinical utility. Conclusion: The CT-based fusion model and HPC risk score effectively predicted progression risk in HCC.

★ CT-based deep learning integrating pancreatic imaging and body composition for severity prediction in acute pancreatitis.

Zeng D, He M, Lin B, Liu Y, Wang X, Lu J — Fri, 30 Oct 2026 00:00:00 +0000

97% · External validation 97%

Objectives: To develop and validate a deep learning (DL)-based CT model integrating pancreatic imaging and CT-derived body composition for severity prediction in acute pancreatitis (AP). Materials And Methods: In this retrospective dual-center study, patients with AP who underwent CT within 72 h of symptom onset were enrolled from January 2021 to September 2025 at Center 1 and January 2023 to September 2024 at Center 2. The primary dataset (n = 426) was randomly divided into a training set (n = 298) and an internal test set (n = 128), while 110 patients from Center 2 served as the external test set. A three-dimensional DL model was developed for pancreatic imaging, and two-dimensional DL models were constructed for L3-level visceral adipose tissue (VAT) and skeletal muscle (SM). Six predictive models were constructed: clinical, pancreas, VAT, SM, pancreas-clinical (PC), and combined. Performance was evaluated using receiver operating characteristic analysis, DeLong testing, calibration, decision curve analysis, net reclassification improvement (NRI), and integrated discrimination improvement (IDI). Results: The combined model achieved area under the curve (AUCs) of 0.897 (95% CI: 0.862-0.932), 0.909 (95% CI: 0.861-0.958), and 0.910 (95% CI: 0.856-0.964) in the training, internal test, and external test sets, respectively. The combined model outperformed the PC model in the training set (p = 0.002) and external test set (p = 0.020). NRI and IDI analyses confirmed the incremental predictive value of body composition features. Conclusions: A CT-based DL model integrating pancreatic imaging, body composition, and clinical variables demonstrated good performance for AP severity prediction. Key Points: Question Early severity stratification in AP remains challenging because conventional clinical and imaging assessments may not fully capture pancreatic injury and systemic disease status. Findings A CT-based DL model integrating pancreatic imaging, body composition, and clinical variables demonstrated good performance for severity prediction in AP. Critical Relevance Statement This study evaluates the value of integrating pancreatic imaging and CT-derived body composition information and supports the use of routinely acquired CT data for early risk stratification, potentially facilitating more individualized monitoring and management of AP.

★ Machine learning-based personalized prediction of developmental dysplasia of the hip using risk factors and clinical examination.

Zhang SS, Caggiari G, Chen XY, Huang BX, Deng HS, Xu N — Fri, 30 Oct 2026 00:00:00 +0000

97% · External validation 97%

Background: Machine learning-based disease prediction models can assist with disease diagnosis and early detection. This study aimed to identify the risk factors for developmental dysplasia of the hip (DDH) and develop multiple machine learning-based models to support screening and risk stratification for DDH. Methods: A retrospective analysis of 12,105 infants screened for DDH between January 2020 and December 2022 at Shenzhen Children's Hospital was performed to collect information on physical examination and risk factors. After univariate and multivariate screening, age-stratified sampling was performed on 11,898 DDH-negative patients to match the age distribution of 207 DDH-positive patients, thereby reducing potential bias. These balanced cohorts were randomly split into training and test sets (at an 8:2 ratio) under both 1:1 and 1:2 positive-to-negative ratios. Predictive modeling was based on random forest, support vector machine, logistic regression, k-nearest neighbor, light gradient boosting machine, and extreme gradient boosting (XGBoost). Results: Twelve variables were included in the model. The XGBoost model trained on 1:1 data was chosen as the best model [area under the curve (AUC): 0.883, accuracy: 0.807, precision: 0.744, F1 score: 0.800, recall (sensitivity): 0.865, and specificity: 0.761]. External validation using data from Jinhua Municipal Central Hospital provided preliminary evidence of transferability [AUC: 0.909, accuracy: 0.838, precision: 0.829, F1 score: 0.840, recall (sensitivity): 0.850, and specificity: 0.825]. Skin pattern asymmetry, restriction of hip abduction, swaddling, breech presentation, and female sex were the top five most important predictive variables. Conclusions: The XGBoost model can be used to support early clinical screening and risk stratification for DDH, helping clinicians identify infants who should undergo hip ultrasonography.

★ Development and multi-database validation of machine learning models for 28-day mortality prediction in ICU patients with COPD and pulmonary hypertension.

Xia Y, Yang Y, Yang J — Fri, 30 Oct 2026 00:00:00 +0000

97% · External validation 97%

Background: Early mortality prediction in critically ill patients with chronic obstructive pulmonary disease (COPD) coexisting pulmonary hypertension (PH) remains challenging, particularly because prediction models developed in intensive care unit (ICU) database may not perform consistently across other healthcare systems. Methods: This retrospective multicenter study employed two large-scale databases to establish and validate ML models for predicting early mortality in ICU patients with COPD and PH. The MIMIC-IV database served as the model-development cohort (2555 patients), with external validation conducted on eICU CRD (292 patients) database. Six ML algorithms were adopted: Random Forest (RF), Extreme Gradient Boosting (XGBoost), Logistic Regression (LR), Light Gradient Boosting Machine (LightGBM), Support Vector Machines (SVM), and Artificial Neural Network (ANN).The baseline variables were screened to identify potential predictors via least absolute shrinkage and selection operator (Lasso) regression. Model interpretability was evaluated via SHapley Additive exPlanations (SHAP) analysis. Results: The overall cohort consisted of 2,847 ICU patients with COPD and PH. Twelve consistently crucial clinical features were identified, including acute kidney injury, pneumonia, age, weight, heart rate, red blood cell distribution width (RDW), white blood cell (WBC), anion gap, chloride, creatinine, base excess, potential of hydrogen (pH). The Random Forest algorithm achieved the optimal performance, with a training Area Under Curve (AUC) of 0.938, internal validation AUC of 0.748, and eICU CRD external validation AUC of 0.800. SHAP analysis indicated that acute kidney injury was the most influential predictor, followed by base excess and RDW. Additionally, a web tool was developed in this study to facilitate the application of early death risk prediction in clinical practice. Conclusions: This study is the multi-database validation of ML models for early mortality prediction in ICU patients with COPD and PH. The Random Forest model integrated with SHAP interpretability offers a reliable tool for clinical decision-making support, with consistent validity verified across two databases, including our internal validation cohort. This work provides a decision support tool to assist in identifying high-risk individuals within this complex comorbid population, and establishes a foundation for future studies with a more dynamic and personalized prognostic approach.

★ BioWaitIndex: a novel machine learning-derived biomarker for improved prediction of liver transplant waitlist outcomes.

Patel K, Connor AA, Sanghvi M, Elaileh A, Rizk E, Semaan S — Fri, 30 Oct 2026 00:00:00 +0000

97% · External validation 97%, Prospective 35%

Background: Liver transplantation remains limited by organ shortage, necessitating precise waitlist risk stratification. The Model for End-Stage Liver Disease (MELD) remains the cornerstone of liver allocation and effectively predicts near-term mortality when assessed dynamically during waitlist follow-up. However, MELD measured at initial transplant evaluation may not fully capture baseline multisystem vulnerability, particularly among candidates with lower MELD scores who may remain on the waitlist for prolonged periods. Methods: The BioWaitIndex (BWI) was developed using a random forest model in an NHANES development cohort comprising 13,830 participants. The trained model was first temporally validated in a separate, non-overlapping sample from subsequent NHANES cycles comprising 8,151 participants. It was then independently clinically validated in a separate, non-overlapping, IRB-approved cohort of 1,502 adult liver transplant candidates from Houston Methodist Hospital. In the clinical validation cohort, associations between BWI measured at initial transplant evaluation and subsequent waitlist outcomes were assessed using competing-risk regression. Results: In multivariable competing-risks analyses with BWI and MELD both measured at initial evaluation, each unit increase in BWI was associated with a greater likelihood of transplantation (sHR 1.02, 95% CI 1.01-1.03; p < 0.001) and a lower risk of waitlist death or clinical deterioration (sHR 0.97, 95% CI 0.97-0.98; p < 0.001). Adding BWI to MELD at evaluation improved net reclassification for waitlist mortality by 31.2% (p < 0.001) and for transplantation by 11.1% (p = 0.026), with consistent associations across sexes. Conclusion: BWI may complement dynamic MELD assessment by providing earlier information about baseline multisystem vulnerability and identifying candidates whose risk may not be fully reflected by their MELD score at evaluation.

★ Machine learning-based risk prediction for invasive pulmonary aspergillosis in non-neutropenic intensive care unit patients, with an exploratory analysis of pathogen-related risk factors.

Bai J, Feng T, Liu B, Wang J, Wang Z, Liu F — Fri, 30 Oct 2026 00:00:00 +0000

96% · External validation 96%

Background: Invasive pulmonary aspergillosis (IPA) is increasingly recognized in non-neutropenic patients, but validated bedside risk stratification tools remain scarce. We developed a machine learning model to predict IPA and tested whether corticosteroid effect depends on bacterial co-pathogen. Methods: This multicenter retrospective cohort study enrolled 837 non-neutropenic adults hospitalized with pulmonary infections across three tertiary hospitals in China between December 2020 and December 2025. The discovery cohort was split into training and internal validation sets. Four classifiers were compared, with Random Forest selected for optimal performance. The final model was externally validated in an independent cohort. Feature importance was assessed using SHAP (SHapley Additive exPlanations) analysis. Corticosteroid-bacteria interactions were examined using subgroup analyses and propensity score matching, with the synergy index S calculated to quantify biological interaction. A simplified bedside score was derived from the top-ranking predictors. Results: The Random Forest model achieved external validation AUC of 0.865. Six predictors consistently dominated across all assessments: Pseudomonas aeruginosa isolation, hypoxemia (PaO2/FiO2 <200 mmHg), diabetes mellitus, organ failure (Sequential Organ Failure Assessment, SOFA >5), lymphopenia (lymphocyte count <0.8×109/L), and ICU (intensive care unit) admission. Corticosteroid exposure did not rank among the top six predictors. However, its association with IPA diverged sharply by bacterial context: OR below 1 with P. aeruginosa (OR 0.52, 95% CI 0.214-1.283) but OR above 1 with other Gram-negative bacilli (OR 2.17, 95% CI 0.722-6.535); P for interaction = 0.131; a directional pattern that persisted after propensity matching. The six-variable score derived from these predictors showed AUCs of 0.869 (training), 0.774 (internal validation), and 0.719 (external validation). Conclusions: IPA risk in non-neutropenic ICU patients is shaped more by bacterial co-pathogen context than by corticosteroid exposure alone. The divergent corticosteroid effects across bacterial species challenge the conventional assumption of a uniform steroid-associated risk. The externally validated six-variable score offers a practical, software-free tool to support pathogen-informed, individualized risk stratification and clinical decision-making in the ICU.

★ Admission immune-inflammatory biomarker profiles for trauma severity stratification: an externally validated machine learning study.

Dai X, Zhang L, Wu Z, Xu H, Chen B, Zhou Y — Fri, 30 Oct 2026 00:00:00 +0000

95% · External validation 95%

Background: Traumatic injury induces rapid systemic immune-inflammatory alterations, but whether routinely measurable admission biomarkers can provide an early biological estimate of anatomical injury burden remains uncertain. This study aimed to characterize trauma-associated immune-inflammatory profiles across trauma severity levels defined by injury severity score (ISS) and evaluate their combined value for early severity stratification. Methods: This two-center retrospective study included healthy participants as a physiological reference group, non-trauma acute-care patients as a clinically relevant comparator, and trauma patients classified as mild (ISS <16), moderate (ISS 16-25), or severe (ISS >25). Nineteen admission laboratory biomarkers and demographic variables were evaluated in a development cohort of 775 participants and an independent external validation cohort of 349 participants. Multinomial least absolute shrinkage and selection operator regression was used to identify severity-associated biomarkers. Ten classifiers were compared on the same predefined trauma-only development split, and the selected model was evaluated in an independent external trauma cohort. Results: A 10-feature panel was retained following feature selection and collectively defined an immune-inflammatory profile associated with increasing trauma severity. Increasing injury burden was associated with heterogeneous and partly nonlinear inflammatory and hematologic alterations, including changes in neutrophil-related indices, lymphocyte-related measures, and albumin. The multinomial logistic regression model achieved the best overall performance, with an accuracy of 0.941 and 0.973 in the internal holdout and external validation cohorts, respectively, with an external macro-AUC of 0.993. Model interpretation consistently identified neutrophil count and RDW-CV as major contributors to severity stratification. Conclusion: Admission-stage blood biomarkers capture reproducible immune-inflammatory alterations associated with increasing anatomical injury severity. Their combined analysis may provide a complementary approach for early biological assessment of injury burden and supports further investigation of clinically accessible biomarkers of trauma-associated immune-inflammatory dysregulation.

★ Development and External Validation of An Interpretable Machine Learning Model for Severity Stratification of Current Periodontitis Using Routine Blood Biomarkers, With Exploratory Analyses of Age- and Sex-Specific Model Behaviour.

Shi H, Yao J, Jiang Z, Chen M, Zhuge Q, Jiang C — Fri, 30 Oct 2026 00:00:00 +0000

95% · External validation 95%

Aim: To develop and externally validate an interpretable machine learning model using routine laboratory data to distinguish current Stage III-IV from Stage I-II periodontitis among adults with established periodontitis. Materials And Methods: This two-centre retrospective study included 1321 patients for development and 584 for external validation at two hospitals of the Zhejiang University School of Medicine. The stacking model used age, sex, smoking status and 13 blood-derived predictors and was frozen before external validation. Performance was assessed by discrimination, calibration, classification measures, decision-curve analysis and comparison with simpler models. SHAP and restricted cubic spline analyses were exploratory. Results: The stacking model had an out-of-fold AUC of 0.827 (95% CI: 0.802-0.851) in development and an external AUC of 0.774 (95% CI: 0.730-0.818). External calibration was less favourable (intercept 0.785, slope 0.822). External AUC was 0.745 for age, sex and smoking status and 0.763 after adding the 13 blood-derived predictors. Conclusions: Among adults with established periodontitis and sufficiently complete predictor data, the model showed moderate external discrimination but less favourable calibration. Blood-derived predictors added modest discrimination beyond age, sex and smoking status. The model does not replace periodontal examination. Further multi-centre validation and calibration are needed before clinical use.

Predicting the risk of chemotherapy-induced myelosuppression in clinical practice: an interpretable machine learning approach based on routine laboratory data.

Fan H, Wu X, Cheng Z, Tang X, Huang H, Chen K — Fri, 30 Oct 2026 00:00:00 +0000

94% · External validation 94%, Prospective 46%

Objective: Chemotherapy-induced myelosuppression (CIM) is the most common adverse reaction of tumor chemotherapy and may lead to treatment delay, dose reduction, life-threatening infection, and poor prognosis. Current predictive methods remain suboptimal, with insufficient interpretability and limited clinical application. This study aimed to develop and validate an interpretable machine learning model for early CIM risk prediction using routine laboratory data. Methods: A total of 1,000 cancer patients from Northeast Yunnan Central Hospital between January 2024 and January 2025 were enrolled and divided into a case group (n=500, with myelosuppression) and a control group (n=500, without myelosuppression).All observed samples were randomly split into a training set (70%) and a validation set (30%). The Least Absolute Shrinkage and Selection Operator (LASSO) was used to select optimal variables for model construction. The predictive performance of 15 different ML models was evaluated using sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV), F1-score, Youden's index, area under the receiver operating characteristic curve (AUC), decision curve analysis (DCA), and residual boxplots. The best-performing model was selected for SHAP-based interpretability analysis and internally validated using a confusion matrix, clinical impact curve, and calibration curve.Additionally, 488 patients treated at Northeast Yunnan Provincial Central Hospital between June 2025 and November 2025 were selected as an independent external validation cohort for external validation. Results: After screening, 15 clinical indicators were used for ML model development. All models demonstrated excellent discrimination. The LightGBM model showed the best overall performance (AUC: 0.999 [0.998-1.000]; sensitivity: 0.986; specificity: 0.989; accuracy: 0.987; PPV: 0.989; NPV: 0.986; F1-score: 0.987; Youden's index: 0.974). According to the LightGBM model, CIM risk increased when the following feature thresholds were met: white blood cell count <5×109/L, mean hemoglobin concentration <325 g/L, monocyte percentage (MONO%) >10%, red blood cell count <4×1012/L, mean platelet volume <8 fL, and plateletcrit <0.1%.In the independent external validation (n=488), the LightGBM model remained stable and reliable, with an AUC of 0.940 (95% CI: 0.920-0.959), sensitivity of 0.818, specificity of 0.888, accuracy of 85.2%, and favorable calibration and clinical net benefit. Conclusion: The LightGBM prediction model developed in this study can enable early identification of high-risk patients, providing a reference for healthcare providers to implement early interventions.

A personalized hybrid federated learning framework for glucose prediction in type 1 diabetes.

Gallardo-García C, Fuertes C, Hernando ME, Subías D, Rigla M, García-Sáez G — Fri, 30 Oct 2026 00:00:00 +0000

93% · External validation 93%

Deep learning methods for glucose prediction in type 1 diabetes (T1D) are often constrained by data availability, restricted access to data, and substantial heterogeneity across treatment settings and institutions. Federated Learning offers a privacy-preserving alternative to centralized training, but its performance under clinically realistic, non-independent and non-identically distributed (non-IID) data remains challenging. This work proposes a hybrid federated-personalized framework that combines cross-silo training with patient-level personalization for multi-horizon glucose prediction. Clinically defined silos based on treatment modality and insulin type are used to explicitly model real-world heterogeneity. A Temporal Fusion Transformer is trained using federated optimization (FedProx), incorporating clinician-prescribed basal insulin profiles as known future inputs. The resulting global model is subsequently adapted at the patient level, enabling personalization. The proposed approach is evaluated on the T1-DEXI dataset and externally validated on the OhioT1DM cohort. Centralized and federated personalization are compared to quantify performance differences between in-distribution and external evaluation and to assess patient-specific adaptation. Results show that the federated global models reduce prediction errors at longer horizons compared to centralized training on unseen data distributions, whereas centralized training achieves lower in-distribution errors. After patient-level personalization, centralized models show slightly lower absolute errors than federated models. These findings are consistent with a balance between population-level behavior and individual adaptation under the evaluated configurations. They support the feasibility of combining cross-silo training with patient-level personalization for privacy-preserving glucose prediction in heterogeneous clinical environments. Furthermore, incorporating scheduled basal insulin profiles enhances in-distribution prediction accuracy, underscoring the importance of clinically meaningful inputs in model design.

Large Language Model-Assisted Radiology Report Quality Assurance: Real-World Multicenter Evaluation.

Wang L, Zhang Y, Peng L, Liu P, Wang L, Cui Y — Fri, 30 Oct 2026 00:00:00 +0000

92% · External validation 92%, Language models 89%

Background: Radiology reports serve as the definitive clinical record of imaging diagnoses, with their accuracy being directly critical to patient safety. However, amid escalating workloads, manual review remains inadequate for capturing subtle errors in reports on a large scale. Although large language models (LLMs) have demonstrated considerable promise in medical text processing, previous studies have predominantly been confined to radiology report text from a single imaging modality or datasets with synthetic errors. Objective: This study aimed to evaluate the feasibility, accuracy, and generalizability of general-purpose LLMs for quality assurance of radiology report text from multiple imaging modalities (x-ray, computed tomography [CT], and magnetic resonance imaging [MRI]), multi-anatomical regions, and multicenter clinical settings, with head-to-head comparisons conducted against radiologists of varying seniority. Methods: Radiology report texts from 3 aforementioned imaging modalities and 5 anatomical regions (head, chest, abdomen, spine, and knee) were retrospectively collected from center 1 (March 2025-May 2025) and reviewed using a clinician-annotated 5-category error taxonomy (type 1: interpretive; type 2: factual). Five LLMs (Qwen3, GPT-5.4 Pro, Gemini 3.1 Pro, Llama 4, and DeepSeek-V3.1) were benchmarked on 10,000 randomly selected reports, and the best-performing model was evaluated on the full cohort. A blinded head-to-head comparison with 3 radiologists (3-10 years of experience) used a modality-stratified, error-enriched challenge set. The radiologists subsequently rated the model-generated explanations using a 5-point Likert scale. Cross-center performance was evaluated in an independent external cohort. The same 2 senior radiologists labeled reports from both centers using the same protocol, resolving disagreements by consensus. Results: In total, 4.6% (4431/95,582) of the reports in center 1 contained at least one confirmed text internal error. Qwen3, the selected model, achieved a report-level sensitivity, specificity, positive predictive value, and F1-score of 0.99, 0.99, 0.90, and 0.94, respectively, with high performance across reports from the 3 imaging modalities and multiple anatomical regions and better detection of factual than interpretive errors. In the error-enriched challenge set, its error instance exact category recall exceeded that of the radiologists (0.96 vs 0.45-0.67; P<.001), with a shorter mean task completion time (5.80 vs 55.80-67.80 seconds per report). Explanations had a mean Likert score of 4.57, and all inter- and intraobserver intraclass correlation coefficients exceeded 0.89. In center 2, balanced accuracy was 0.94 (95% CI 0.92-0.95). Before consensus, Cohen κ for report-level error presence was 0.783 (center 1) and 0.850 (center 2). Conclusions: This study suggests that LLMs can serve as efficient adjunctive tools for text-level quality control of radiology reports, particularly for detecting explicit factual errors. Performance remained stable across centers. In the text-only review task, the model achieved higher exact category recall and shorter task completion times than the participating radiologists. These findings support the potential value of LLM-assisted radiology report text review in real-world clinical environments.

A Transformer-Based Risk Prediction Model for Contrast Media Extravasation During High-Pressure Injection: A Dual-Center Study with External Validation.

Zhu X, Bao Z, Hu Z, Zhao Z, Wang C, Zou J — Fri, 30 Oct 2026 00:00:00 +0000

92% · External validation 92%

Purpose: To assess whether a Transformer-based deep learning model can predict contrast media extravasation (CME) risk before high-pressure intravenous iodinated contrast injection, and to provide an automated screening tool for high-risk patients before contrast-enhanced CT. Patients And Methods: We retrospectively analysed 35,537 patients who underwent contrast-enhanced CT between November 2023 and June 2025 at Northern Jiangsu People's Hospital (internal cohort, n = 26,126; 0.88% CME-positive) and its South Branch (external validation, n = 9,411; 0.99% CME-positive), an overall CME-positive rate of 0.91%. Sixteen demographic, clinical, injection and nursing variables were used. A Transformer encoder model (three stacked encoder blocks) was trained with a batch-adaptive hard negative sample-corrected Focal Loss for extreme class imbalance, with class weights recalculated per mini-batch and an extra penalty on hard negative samples; perturbation-based feature attribution was also performed. Internal five-fold cross-validation was followed by independent external validation. Metrics included accuracy, recall, specificity, positive and negative predictive value, F1-score, Cohen's Kappa, Matthews correlation coefficient (MCC) and area under the ROC curve (AUC). Results: Internal five-fold cross-validation gave an accuracy of 93.07% ± 0.76%, recall of 66.80% ± 2.16%, specificity of 93.31% ± 0.79% and AUC 0.907; external validation gave 92.71%, 63.44%, 93.00% and 0.862. Cohen's Kappa was 0.12-0.15 (internal mean 0.13) and MCC 0.21-0.23 (internal mean 0.22). Puncture site, contrast agent viscosity, total injection volume and venous thrombosis were high-contribution features (>10%); injection rate, diabetes, sex, age, BMI, VIA scale, malignant tumor history and patient source contributed 1-10%. Conclusion: The model showed acceptable discriminative ability and generalizability across two centers, but remains a retrospective research prototype: because a rare-event model with high specificity still generates many false-positive alerts, its outputs require local recalibration and the alert threshold has not been clinically optimized. Prospective multicenter validation is therefore needed before clinical deployment.

Machine Learning-Based Prediction of Major Postoperative Complications After Noncardiac Surgery Under General Anesthesia Using Routinely Available Electronic Health Record Data: Retrospective Multicenter Development and Validation Study.

Nishino T, Mase H, Kim C, Sugita S, Kondo Y, Ishikawa M — Fri, 30 Oct 2026 00:00:00 +0000

91% · External validation 91%

Background: Postoperative complications after noncardiac surgery under general anesthesia remain a major cause of morbidity, mortality, and resource-intensive care. Conventional risk assessment tools rely mainly on preoperative information and may not capture perioperative physiological and procedural factors associated with postoperative risk. Objective: This retrospective multicenter observational study aimed to develop and validate electronic health record (EHR)-based machine learning (ML) models for predicting major postoperative complications and to evaluate their discrimination, overall probabilistic performance, calibration, and potential clinical utility. Methods: We analyzed 35,702 adults who underwent noncardiac surgery under general anesthesia at 3 tertiary hospitals between January 2021 and March 2025. Preoperative, intraoperative, and early postoperative variables were extracted from diagnosis procedure combination (DPC) claims, EHRs, and anesthesia information systems. The primary outcome was a 30-day composite of in-hospital death, unplanned reoperation, prolonged mechanical ventilation, new initiation of renal replacement therapy, and postoperative infection. Logistic regression (LR), extreme gradient boosting (XGB), and light gradient boosting machine (LGBM) models were developed using preoperative variables alone or with 7 additional intraoperative and early postoperative variables. Leave-one-facility-out internal-external cross-validation (IECV) was the primary validation framework, with a stratified random 80:20 hold-out analysis as secondary validation. Performance was assessed using the area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (PR-AUC), Brier score, scaled Brier score, calibration measures, and decision curve analysis. Results: The primary composite outcome occurred in 3884 of 35,702 patients (10.9%). In pooled IECV, perioperative models consistently outperformed the corresponding preoperative models. AUROC increased from 0.784-0.791 to 0.810-0.817, PR-AUC increased from 0.356-0.361 to 0.405-0.413, Brier scores decreased from 0.082-0.083 to 0.078-0.080, and scaled Brier scores increased from 0.144-0.151 to 0.175-0.191. Calibration varied across facilities; pooled calibration slopes for the perioperative models ranged from 0.875 to 0.902, and calibration intercepts ranged from -0.367 to -0.222, indicating overly extreme predictions and overall risk overestimation. Exploratory decision curve analysis generally showed greater net benefit for perioperative models across threshold probabilities of 0.01-0.30, although the curves overlapped substantially. Secondary and sensitivity analyses were consistent with the primary findings. Conclusions: Adding routinely available intraoperative and early postoperative variables produced modest but consistent improvements in postoperative risk prediction compared with preoperative variables alone. However, facility-specific variation in calibration limits direct transportability of absolute risk estimates. The models may support prioritization for postoperative clinical reassessment, but exploratory decision curve analysis did not establish a specific intervention or triage threshold. Local recalibration and prospective workflow evaluation are required before clinical implementation.

An AI-derived inflammation-metabolism ecosystem signature reveals ATP2A2 as a microenvironment-remodeling driver of immune evasion in lung adenocarcinoma.

Hu X, Zhang G, Zhang Y, Zhao Y, Zhou F, Li G — Fri, 30 Oct 2026 00:00:00 +0000

91% · External validation 91%

Background: Chronic inflammation and metabolic dysregulation are increasingly recognized as interconnected drivers of tumorigenesis and immune evasion, yet their combined prognostic and mechanistic relevance in lung adenocarcinoma (LUAD) remains poorly defined. Methods: A curated 117-gene set representing systemic inflammation-metabolism dysregulation was obtained from MSigDB. Core prognostic genes were identified through differential expression analysis and univariate Cox regression. Using TCGA-LUAD as the training cohort and six independent GEO cohorts for external validation, 101 machine learning algorithm combinations were systematically evaluated to construct a robust prognostic model. Model performance was assessed by C-index, time-dependent ROC, and PCA. Immune characteristics were analyzed using TCIA, ESTIMATE, and TIMER-based deconvolution. ATP2A2, identified as the gene most strongly correlated with risk score, underwent multi-cohort transcriptomic and immunohistochemical validation, in vitro functional characterization, and mechanistic investigation via IOBR-integrated immune profiling. Single-cell RNA sequencing and multiplex immunofluorescence provided cellular-resolution verification of microenvironmental associations. Results: The RSF + SuperPC model achieved optimal performance and effectively stratified patients into high- and low-risk groups with significantly divergent survival across all seven cohorts. High-risk tumors showed reduced immune/stromal infiltration and elevated tumor purity. ATP2A2 was consistently overexpressed in tumor tissue, correlated with poor differentiation, advanced stage, and shorter survival, and predicted immunotherapy resistance. Mechanistically, high ATP2A2 expression was associated with suppressed immune cell infiltration, attenuated chemokine signaling, impaired cancer-immunity cycle activity, and an "immune-desert" phenotype confirmed by single-cell and spatial analyses. Conclusion: This study establishes a validated inflammation-metabolism-based prognostic signature in LUAD and identifies ATP2A2 as a key driver linking tumor aggressiveness to immunosuppressive microenvironment remodeling, positioning it as a promising prognostic biomarker and potential therapeutic target for overcoming immunotherapy resistance.

★ Interpretable machine learning for preoperative prediction of WHO/ISUP grade in clear cell renal cell carcinoma using contrast-enhanced CT-derived extracellular volume fraction.

Wang Y, Wang Y, Bai G, Huang W, Huang M, Chen W — Thu, 29 Oct 2026 00:00:00 +0000

98% · External validation 98%

Objective: To develop and validate an interpretable machine learning (ML) model that integrates extracellular volume fraction (ECV) derived from contrast-enhanced CT with clinical and imaging features for preoperative prediction of World Health Organization/International Society of Urological Pathology (WHO/ISUP) grade in clear cell renal cell carcinoma (ccRCC). Materials And Methods: A retrospective cohort of 263 patients with pathologically confirmed ccRCC from two institutions was analyzed. Patients from Hospital A (n=184) were randomly divided into training (n=128) and internal validation (n=56) cohorts, while patients from Hospital B (n=79) served as an independent external validation cohort. CT-derived ECV, arterial enhancement fraction (AEF), and conventional imaging features were extracted. Six machine learning models (Logistic Regression [LR], eXtreme Gradient Boosting [XGBoost], Gaussian Naive Bayes [GNB], Support Vector Machine [SVM], K-Nearest Neighbors [KNN], Light Gradient Boosting Machine [LightGBM]) were constructed, and the best-performing model was identified based on bootstrap resampling. Model interpretability was assessed using SHapley Additive exPlanations (SHAP). Results: Tumor size (low-grade: 4.99 ± 1.98 cm; high-grade: 6.20 ± 2.72 cm, P = 0.002), ECV (low-grade: 0.46 ± 0.09; high-grade: 0.57 ± 0.12, P<0.001), and AEF (low-grade: 1.37 ± 0.24; high-grade: 1.56 ± 0.25, P<0.001) were significantly higher in high-grade ccRCC, while hematocrit (HCT, low-grade: 0.41 ± 0.05; high-grade: 0.38 ± 0.06, P<0.001) was lower. The optimal LR model achieved robust performance: training cohort AUC = 0.995 (95% CI: 0.990-0.999), internal validation AUC = 0.903 (95% CI: 0.818-0.988), external test AUC = 0.857 (95% CI: 0.756-0.958). SHAP analysis showed ECV, tumor diameter, and AEF were the top 3 predictors. Conclusion: An interpretable ML model incorporating CT-derived ECV and imaging features effectively predicts preoperative WHO/ISUP grade of ccRCC, providing a noninvasive tool for individualized risk stratification and clinical decision-making.

★ Interpretable machine learning for identifying mild cognitive impairment among Chinese patients with obstructive sleep apnea: a multicenter cross-sectional study.

Li E, Ai F, Cai P, Wen K, Guo B, Wen H — Thu, 29 Oct 2026 00:00:00 +0000

98% · External validation 98%

Background: Mild cognitive impairment (MCI) is clinically relevant in obstructive sleep apnea (OSA). We developed and externally validated an interpretable machine-learning model for concurrent clinically defined MCI to support prioritization for formal cognitive assessment rather than diagnosis. Methods: This multicenter cross-sectional diagnostic-prediction study included 400 eligible patients with OSA from Changchun, divided into a training set (n = 280) and an internal algorithm-selection hold-out (n = 120), plus an independent Shenzhen external-validation sample (n = 192). MCI classification integrated reported cognitive decline, objective cognitive evidence, largely preserved everyday independence, dementia exclusion, and assessment-team consensus with physician participation. Of 27 candidate predictors, redundancy review left 24 and LASSO-Boruta consensus retained nine. Feature selection was performed once on the full training set before repeated cross-validation and was not nested within resampling folds. Eight algorithms were tuned using repeated 10-fold cross-validation with training-fold-only SMOTE and PR-AUC as the primary metric. The RF threshold (0.3228) was derived from training out-of-fold probabilities and locked before validation. Results: MCI was classified in 93/280 (33.2%), 38/120 (31.7%), and 57/192 (29.7%) participants. The nine RF predictors were coffee consumption, physical activity, AHI, MSaO2, PHQ-9, GAD-7, ESS, CPSS-14, and PSQI total scores. In the internal hold-out used for algorithm selection, RF achieved ROC-AUC 0.953, PR-AUC 0.936, and Brier score 0.060; these selection-conditioned estimates are not unbiased generalization estimates. In independent external validation, ROC-AUC was 0.957, PR-AUC 0.809, and Brier score 0.051. Calibration intercept - 0.616 and slope 0.701 indicated systematic probability overestimation despite strong discrimination. Conclusions: The RF model discriminated strongly in Shenzhen, but external miscalibration and systematic probability overestimation mean that high AUC does not make raw probabilities transportable; local calibration assessment and, if needed, recalibration are required before clinical use. The MCI outcome cannot be reproduced from questionnaire/scale scores alone, and Supplementary Table S10 is a retrospective operational specification rather than the original CRF. Performance evidence comes almost entirely from male OSA patients; validity in women is unestablished, and the model should not be used for clinical assessment in females. It is a research screening-prioritization tool, not a diagnostic, causal, prognostic, or clinical-impact instrument.

★ Development and Multicenter Validation of a Hybrid Radiomics-Deep Learning Model for Risk Stratification of O-RADS 4 Adnexal Masses.

Sun Y, Shen H, Gu Y, Sun R, Dai Y, Zheng Q — Thu, 29 Oct 2026 00:00:00 +0000

98% · External validation 98%, Prospective 96%

Rationale And Objectives: To develop and validate an artificial intelligence-driven hybrid model integrating ultrasound radiomics, deep learning (DL) signatures, and clinical predictors for differentiating benign from borderline/malignant Ovarian-Adnexal Reporting and Data System (O-RADS) 4 adnexal masses, and to evaluate its utility as a radiologist decision-support tool. Materials And Methods: This multicenter study retrospectively enrolled 653 patients from four centers (521 from Center A [training, n = 405; internal validation, n = 116]; 132 from Centers B-D [external validation]), and prospectively enrolled 126 patients from Center A. Pathological results served as the gold standard. Radiomics features were extracted to generate a Rad-score using machine learning classifiers. A transfer learning-based DL model using convolutional neural networks produced a DL-score. A hybrid model integrating Rad-score, DL-score, and clinical predictors was constructed. Diagnostic performance of models was assessed using the area under the receiver operating characteristic curve (AUC). In the prospective validation set, six radiologists interpreted cases with and without hybrid model assistance. Diagnostic differences were assessed using McNemar's χ² test, and inter-reader agreement by Cohen's kappa. Results: The hybrid model demonstrated good performance with AUCs of 0.88 (95% confidence interval [CI]: 0.82-0.93) and 0.86 (95% CI: 0.78-0.92) in the prospective and external validation sets. Model assistance significantly improved accuracy and specificity across all readers (p < 0.05), with inter-reader agreement improving from kappa 0.49 to 0.73. The model stratified lesions into low-risk (4a) and high-risk (4b) subcategories with consistently distinct malignancy rates across all datasets (4a: 4.5%-16.1%; 4b: 75.4%-97.1%). Conclusion: The hybrid model demonstrated good performance for risk stratification of O-RADS 4 adnexal masses and improved radiologist performance across experience levels.

★ Harnessing Clinical Characteristics to Identify High-Risk of Drug-induced QTc Prolongation in Hospitalized Patients.

Zhao Y, Weng BB, Lan G, Jin NS, Liao PL, Lai XD — Thu, 29 Oct 2026 00:00:00 +0000

97% · External validation 97%

Background: Drug-induced QTc-prolongation is a major concern for developing ventricular arrhythmias, which may eventually lead to sudden cardiac death. Identifying high-risk patients is essential to support healthcare providers in handling automatically generated alerts in clinical practice. The aim of study was to develop clinical data-driven machine learning methods to predict drug-induced QTc-prolongation. Methods: Inpatients with paired electrocardiograms were enrolled from Southwest Hospital (2015-2022). Six machine learning (ML) algorithms were developed to predict drug-induced QTc prolongation (ΔQTc > 60 ms). Model performance was evaluated using an internal test set and an independent external validation cohort. Model interpretations were recognized by feature importance and SHapley Additive exPlanations (SHAP) methods. Finally, we benchmarked the ML models against the traditional Tisdale risk score. Results: A total of 2451 participants were recruited, of whom 292 developed QTc-prolongation, while 150 patients from a separate center served as the external validation cohort. The eXtreme Gradient Boosting (XGBoost) model achieved the best performance with recall of 0.884 and AUC of 0.814 in internal test set. During external validation, the AUC remained robust at 0.864, although recall declined to 0.650. SHAP analysis identified baseline QTc and QT intervals, inflammation, high-risk drugs, and hypertension are the primary drivers of model predictions. Conclusions: The XGBoost algorithm provides a highly sensitive framework for predicting drug-induced QTc prolongation. Integrating this ML tool with identified clinical drivers into hospital decision support systems would facilitate the optimization of treatment strategies and early ECG monitoring strategies.

★ Development and external validation of an interpretable machine learning model for predicting 4-year incident frailty in middle-aged and older adults with arthritis: A retrospective longitudinal study based on the CHARLS and KLoSA cohorts.

Xu X, Cesar N, Ntawuyamara E, Su SH — Thu, 29 Oct 2026 00:00:00 +0000

97% · External validation 97%

ObjectiveBased on retrospective longitudinal data from the China Health and Retirement Longitudinal Study (CHARLS) and Korean Longitudinal Study of Ageing (KLoSA) cohorts, we developed and externally validated an interpretable machine learning model to predict 4-year incident frailty in middle-aged and older adults with arthritis.MethodsModel development was performed using the CHARLS cohort (n = 2126) and included participants aged ≥45 years who had been diagnosed with arthritis and were nonfrail at their initial survey. New-onset frailty was tracked over a 4-year interval. Thirty-one candidate predictors were identified through a literature review, none of which overlapped with the deficits constituting the frailty index. Ten core features, including age, marital status, left- and right-hand grip strength, number of painful sites, knee pain, leg pain, back pain, number of hospitalizations, and antirheumatic medication use, were selected via combined least absolute shrinkage and selection operator-Boruta feature selection. Six algorithms were compared based on their area under the curves in the internal validation dataset, specificity, sensitivity, and other clinical applicability metrics. Tenfold cross-validation was employed. The selected model was frozen and transported to KLoSA without refitting. A prespecified recalibration-in-the-large was applied to account for the lower baseline frailty incidence in the Korean cohort. Model interpretability was achieved with SHapley Additive exPlanations global and local analyses, and an online calculator was developed.ResultsLogistic regression was selected as the final model because it achieved competitive discrimination with the least shrinkage between training and validation sets. Its discriminative ability, measured as the area under the receiver operating characteristic curve, was 0.764 (95% confidence interval: 0.707-0.816) internally and remained at 0.747 (95% confidence interval: 0.701-0.793) when the model was transported to KLoSA. Performance was stable across 10-fold cross-validation (mean, 0.772 ± 0.046). Discrimination was well preserved, although a prespecified recalibration-in-the-large was required to correct for the lower baseline frailty incidence in the Korean cohort (Brier score 0.205→0.065). In the SHapley Additive exPlanations ranking, the number of painful sites, antirheumatic medication use, leg pain, age, and grip strength contributed the most. Once a user enters the 10 easily collected variables, the calculator reports that person's probability of becoming frail within 4 years. In post hoc analyses, pain burden and low grip strength emerged as two independent risk dimensions. High pain burden remained associated with falls (adjusted odds ratio = 1.60) and hospitalization (1.66) even at a given level of frailty, whereas low grip strength and medication use were associated with an increased frailty risk without a corresponding excess of events.ConclusionsThis explainable machine learning model, developed using data from Chinese middle-aged and older patients with arthritis and externally validated in a KLoSA cohort, combines acceptable predictive accuracy with cross-population robustness. The tool estimates an individualized 4-year frailty risk probability using 10 routinely available indicators. Given the observational nature of both cohorts, this model should be viewed as a candidate risk stratification aid rather than a validated clinical tool. Formal impact evaluation through prospective validation is necessary before the model can be recommended for use in community settings.

★ Multi-Backbone U-Net segmentation and interpretable hybrid deep learning classification for lung cancer detection in chest CT images.

Golkarieh A, Kiashemshaki K, Boroujeni SR, Ansari S, Isakan NA, Najafzadeh H — Thu, 29 Oct 2026 00:00:00 +0000

96% · External validation 96%

Background: Lung cancer remains the leading cause of cancer-related mortality worldwide, and its early and accurate diagnosis from chest CT imaging is complicated by overlapping radiological features between benign and malignant lesions, as well as by the limited cross-institution validation and interpretability reported in much of the existing computer-aided diagnosis literature. This study evaluates the effectiveness of U-Net architectures integrated with different pre-trained convolutional neural network (CNN) backbones for automated lung segmentation and lung cancer classification in chest CT images, with an emphasis on rigorous, leakage-free validation and cross-institution generalizability relevant to clinical practice. Methods: A balanced internal dataset of 832 chest CT images (416 cancerous and 416 non-cancerous cases, NIDCH, Bangladesh) was preprocessed using Contrast Limited Adaptive Histogram Equalization (CLAHE) and resized to 224 × 224 pixels. U-Net models with three CNN backbones (VGG16, ResNet50, and Xception) were implemented for lung segmentation, with pseudo-labels verified against expert-annotated masks. A standalone CNN and three hybrid classifiers combining CNN feature extraction with machine learning algorithms (Support Vector Machine, Random Forest, and Gradient Boosting) were then evaluated, alongside a Vision Transformer baseline. All models were assessed using subject-level fivefold cross-validation, followed by fully independent external testing on a second, structurally distinct dataset (IQ-OTH/NCCD; 110 subjects) without retraining. Performance was measured using Dice coefficient, IoU, accuracy, precision, recall, F1-score, and AUC, with 95% confidence intervals and statistical significance testing. Results: For segmentation, U-Net with the Xception backbone achieved the best overall performance (Dice: 0.9581 ± 0.0104; IoU: 0.9163 ± 0.0167). For classification, the CNN + SVM hybrid model achieved the highest internal cross-validation accuracy (98.21% ± 0.68%) and AUC (0.9881 ± 0.0041), significantly outperforming the standalone CNN (p = 0.041), and retained the strongest performance under external validation (92.73% accuracy), with the smallest accuracy drop among all evaluated models. Ablation analyses confirmed substantial contributions from CLAHE preprocessing and U-Net segmentation, and Grad-CAM analysis showed that segmentation constrains classifier attention to anatomically plausible lung regions. Conclusion: Integrating U-Net segmentation with hybrid CNN-based classification provides an accurate and reasonably generalizable framework for lung cancer detection, offering a methodological foundation for future clinically oriented computer-aided diagnosis research, pending prospective, multi-institution validation.

★ Noninvasive Detection and Subtyping of Esophageal Cancer by Circulating Cell-Free mtDNA Fragmentomics: A Multi-Center Validation Study.

Liu Y, Yang W, Li Y, Wang S, Li X, Peng F — Thu, 29 Oct 2026 00:00:00 +0000

95% · External validation 95%

Esophageal cancer (EC) is highly lethal due to late diagnosis, and accurate non-invasive tools for early detection and subtyping remain lacking. Circulating cell-free mitochondrial DNA (ccf-mtDNA) fragmentomics offers a promising biomarker source. In this large multi-center study enrolling 1,260 participants spanning healthy control (HC), benign esophageal diseases (BED), precancerous lesions and EC patients, we developed two machine learning models based on plasma ccf-mtDNA fragmentomic profiles: an EC detection (ED) model and a pathological subtyping prediction (PSP) model. The ED model achieved an AUC of 0.9706 (EC vs. non-EC) in the training cohort, with high performance in internal (0.9762), external (0.9748), and Chinese minority (0.9668, Uyghur and Kazakh) validation cohorts. At the optimal cut-off, it yielded 92.16% sensitivity for stage I EC in the combined validation cohort and correctly identified 93.10% of early EC cases missed by preoperative biopsy. It also discriminated high-grade intraepithelial neoplasia from non-EC with an AUC of 0.9597. The PSP model distinguished ESCC from EAC with 89.71% accuracy, improving to 90.74% after integrating key clinical parameters. In summary, ccf-mtDNA fragmentomics provides a robust, non-invasive platform for early EC detection, precancerous risk stratification, and histological subtyping, demonstrating strong potential to transform clinical management.

★ Predicting cardiovascular disease in peritoneal dialysis patients using glycated proteomics and machine learning: A prospective multicentre study.

Ji J, Nie Z, Jiang N, Xie T, Zhang L, Liu X — Thu, 29 Oct 2026 00:00:00 +0000

95% · External validation 95%, Prospective 74%

Background: Cardiovascular disease (CVD) is the leading cause of death among peritoneal dialysis (PD) patients, yet PD-specific biomarkers and prediction models remain limited. This study aimed to identify glycated peptide biomarkers associated with CVD in PD patients and develop a prediction model integrating glycated proteomics and clinical indicators using machine learning. Methods: In Phase I, a matched discovery cohort (48 PD patients with CVD and 48 without CVD) was used to identify glycated peptides by liquid chromatography-tandem mass spectrometry (LC-MS/MS), with glmnet used to prioritise candidate biomarkers. In Phase II, candidate biomarkers were quantified in a prospective internal derivation cohort (n = 240) and a multicentre external validation cohort (n = 102). The combined model was developed using F-test feature selection and elastic-net logistic regression. Its predictive performance was compared with three published cardiovascular risk models. Patients were stratified using a training-derived Youden cutoff. Results: In the discovery cohort, 257 glycated peptides were differentially abundant between groups. Four candidate biomarkers were prioritised, and two peptides, GEAFTLK(g)ATVLNYLPK and K(g)QHLFVK, were retained in the final model together with six clinical indicators: phosphate, diabetes, total Kt/V, albumin, cardiac troponin T, and glycated albumin. The combined model achieved an area under the receiver operating characteristic curve (ROC-AUC) of 0.784 in the training set, 0.781 in the internal testing set, and 0.822 in the external validation cohort, respectively. Using the training-derived cutoff of 0.373, higher 1-year CVD incidence was observed in the high-risk group in both validation cohorts. The combined model also achieved higher ROC-AUCs than the published cardiovascular risk models. Conclusion: We developed and externally validated a PD-specific CVD prediction model by integrating glycated peptide biomarkers with clinical indicators, which may facilitate earlier risk stratification and targeted prevention.

★ Construction and validation of a unified 3D convolutional neural network framework for automated identification of cervical spinal cord injury based on CT imaging.

Ou D, Mo Y, Huang Y, Ye Y, Lin S, Pan L — Thu, 29 Oct 2026 00:00:00 +0000

95% · External validation 95%

Objective: To construct and validate a standardized three-dimensional convolutional neural network (3D CNN) framework based on cervical spine CT imaging for automated identification of cervical spinal cord injury (CSCI), and to evaluate its stability across multicenter data and potential clinical applicability. Methods: Cervical spine CT images were retrospectively collected from two independent clinical centers, including cases with and without cervical spinal cord injury. All images were converted to NIfTI format through a unified preprocessing pipeline, and three-dimensional spinal cord regions of interest (ROIs) were delineated using ITK-SNAP to assess interobserver consistency. A standardized 3D CNN architecture was constructed, and the dataset was divided into training, internal validation, and external validation sets using a stratified random strategy. Model training was performed using the PyTorch framework. Classification performance across different datasets was systematically evaluated, and ROC curves, precision-recall (PR) curves, prediction probability distributions, and threshold sensitivity analyses were used to explore the model's clinical application characteristics. Results: The model demonstrated stable discriminative performance across the training, internal validation, and external validation sets, with AUC-ROC values of 0.943, 0.953, and 0.948, respectively, and AUC-PR values of 0.936, 0.939, and 0.943, respectively. Performance metrics showed minimal variation among datasets, with no evident overfitting or performance degradation. Prediction probability distributions indicated good separation between injured and non-injured samples. Threshold sensitivity analysis showed that the model could achieve performance trade-offs under different threshold settings, suggesting a degree of flexibility for clinical decision-making. Conclusion: Based on standardized data processing and network design, the 3D CNN effectively extracted latent spatial information associated with cervical spinal cord injury from CT images. The proposed framework demonstrated good stability and reproducibility across multicenter data and may serve as a preliminary research framework for CT-based CSCI risk evaluation, particularly in scenarios where MRI is unavailable or limited. Prospective validation is required before clinical deployment. It also provides a methodological reference for the standardized construction of deep learning models in medical imaging research.

Real-time Voice-Guided Artificial Intelligence Improves Biliopancreatic Endosonography Quality: A Randomized Trial.

Wang R, Tian L, Chen X, Yang Y, Qu W, Zheng Q — Thu, 29 Oct 2026 00:00:00 +0000

92% · Prospective 92%, External validation 92%

Background And Aims: Endoscopic ultrasonography (EUS) quality is operator-dependent, and adherence to guideline-recommended multi-station biliopancreatic scanning remains suboptimal. We developed a device (named EUS-GUIDEQ), an artificial intelligence (AI)-based real-time EUS navigation and quality-control system providing station recognition, voice-guided prompts, and quantitative completeness feedback. We evaluated its performance and clinical utility. Methods: EUS-GUIDEQ was developed from 27,072 annotated biliopancreatic EUS images obtained from 189 patients and validated on independent datasets consisting of 1,500 internal images and 2,062 multicenter external images. We then conducted a prospective, single-center, randomized controlled trial (1:1) comparing AI-assisted versus standard EUS. The primary endpoint was completeness of standardized scanning stations. Secondary endpoints included station-level and landmarks completeness, procedure duration (observation time excluding additional interventions), and adverse events. Outcomes were adjudicated by blinded expert review of recorded videos. Results: Station-classification model achieved accuracies of 94.2% internally and 78.8% externally. Between December 12, 2023, and May 16, 2025, 266 patients were analyzed (AI-assisted n=132; control n=134). Overall station completeness was higher with EUS-GUIDEQ assistance (median [IQR] 1.00 [0.94-1.00] vs 0.81 [0.75-0.88]; P<0.001), with benefit observed across participating endoscopists. Completeness was higher for most anatomical landmarks in nominal comparisons. Procedure time was longer with AI assistance (12.70±3.23 vs 10.55±3.62 min; P<0.001). No serious procedure-related adverse events occurred. Conclusions: In a randomized trial, voice-guided real-time AI navigation improved guideline-level biliopancreatic EUS process quality, as measured by station completeness, but increased procedure time. Multicenter, multi-vendor prospective validation is warranted.

FLAT-WHITE: from target trial emulation to federated high-throughput causal machine learning using decentralized real-world data.

Tong J, Li L, Zhang D, O'Brien KS, Lu Y, Wu Q — Thu, 29 Oct 2026 00:00:00 +0000

92% · External validation 92%

Randomized clinical trials remain the gold standard for evaluating treatment effects but are often infeasible for rapidly assessing drug repurposing opportunities, particularly in complex conditions such as Alzheimer's disease. Target trial emulation (TTE) provides a principled framework for causal inference with electronic health records and, when coupled with modern machine learning, enables scalable analyses across distributed health systems. We present FLAT-WHITE (Federated Lossless Algorithm for Target Trial Emulation with Overlap Weighting for Healthcare Data Integration via Trusted and Efficient Communication), a novel federated causal machine learning pipeline for high-throughput treatment effect estimation without sharing patient-level data. FLAT-WHITE is a one-shot, lossless algorithm, requiring a single communication round while achieving results equivalent to centralized pooled analyses. With overlap weighting on the propensity score, it robustly addresses heterogeneity in treatment assignment and ensures valid causal inference in decentralized settings. We demonstrate FLAT-WHITE in a Alzheimer's drug repurposing study, scanning 181 candidate drugs using data from four health learning systems encompassing approximately 150,000 patients with mild cognitive impairment. Findings were validated externally in 339,180 patients from the TriNetX global research network. FLAT-WHITE enables rapid, large-scale, and secure evaluation of treatment effects and establishes a foundation for collaborative clinical research across diverse healthcare systems.

Real-time artificial intelligence-assisted assessment in a home-based pelvic floor muscle training device for pelvic organ prolapse and stress urinary incontinence: A single-center prospective study.

Horng HC, Chen JF, Chang CP, Chao WT — Thu, 29 Oct 2026 00:00:00 +0000

90% · Prospective 90%

Objective: Pelvic floor muscle training (PFMT) is the first-line conservative treatment of pelvic organ prolapse (POP) and stress urinary incontinence (SUI), but its effectiveness is often limited by poor adherence and performance. We developed a third-generation home-based PFMT device incorporating real-time artificial intelligence (AI)-assisted assessment. This study aimed to evaluate the efficacy of this device in women with POP or SUI. Materials And Methods: This prospective single-arm study was conducted at Taipei Veterans General Hospital. Women with stage I-II POP or SUI performed home-based PFMT using a noninvasive wireless pressure-sensing device with real-time AI-assisted assessment. Outcomes were assessed at baseline (M0), 1 (M1), and 2 (M2) months. Primary outcomes included the Pelvic Organ Prolapse Distress Inventory-6 (POPDI-6), Urogenital Distress Inventory-6 (UDI-6), Incontinence Impact Questionnaire-7 (IIQ-7), and the 1-h pad test. Contraction patterns were analyzed and categorized as optimal or non-optimal. Results: A total of 27 women (mean age 59.0 ± 14.3 years) were enrolled and completed the study. POPDI-6 decreased from 8.6 ± 2.6 to 5.2 ± 3.0 (p < 0.01), and UDI-6 from 6.4 ± 4.9 to 3.4 ± 3.4 (p < 0.05). The 1-h pad test decreased from 30.9 ± 41.9 g to 14.3 ± 23.5 g (p < 0.01); IIQ-7 reduction was also significant. Subjective and objective improvement rates were 89% and 54%, respectively. Physician-directed pattern optimization significantly improved contraction patterns, with 52.9% of patients converting from non-optimal to optimal patterns and increased the optimal contraction rate from 37.0% to 70.4% (p < 0.01). Conclusion: Real-time AI-assisted PFMT significantly improved symptom scores, objective urinary leakage, and contraction quality in women with POP or SUI. While limited by a small sample size, a short follow-up period, and a single-arm design, this approach effectively identifies and optimizes dysfunctional contraction patterns to enhance pelvic floor rehabilitation.

★ Contrast-enhanced CT-based multimodal modeling for overall survival prediction after resection of retroperitoneal liposarcoma: a bicentric retrospective study.

Fan P, Zhang J, Luo W, Pan H, Wang J, Ma L — Wed, 28 Oct 2026 00:00:00 +0000

98% · External validation 98%

Objective: Postoperative risk estimation in retroperitoneal liposarcoma (RPLS) remains challenging because conventional clinicopathologic models capture only part of its biological heterogeneity. We aimed to develop and externally evaluate a postoperative prognostic model integrating preoperative CT features with postoperative clinicopathologic information to estimate overall survival (OS) after RPLS resection. Materials And Methods: This bicentric retrospective study included 367 patients who underwent complete RPLS resection between 2009 and 2021. Institution 1 provided the training cohort (n = 128) and internal validation cohort (n = 55); Institution 2 provided the external test cohort (n = 184). A radiomics signature was constructed from three-dimensional tumor features. A deep-learning signature was derived from the largest axial tumor section using an ImageNet-pretrained ResNet-101 trained with Cox partial-likelihood loss. Clinical variables underwent univariable and multivariable Cox regression analysis. The retained clinicopathologic variables and imaging signatures were integrated into a combined model termed RDCS. Results: During follow-up, 71, 30, and 107 patients died in the training, internal validation, and external test cohorts, respectively. The corresponding RDCS C-indices were 0.775, 0.704, and 0.725, each higher than those of the individual radiomics, deep-learning, and clinical models. Calibration curves showed agreement between predicted and observed survival. Decision-curve analysis suggested potential clinical utility within the evaluated thresholds. In a separate exploratory comparison in the external primary-tumor subset (n = 82), C-indices were 0.746 for RDCS and 0.590 for the official Sarculator application. Conclusion: RDCS estimates OS after complete RPLS resection using preoperative CT-derived imaging signatures and postoperative clinicopathologic information. RDCS may complement established postoperative prognostic tools, but its application in surveillance and clinical management requires prospective validation and clinical impact assessment.

★ Externally validated MRI radiomics machine learning models for differentiating between myxoid liposarcoma and myxoma.

Chen YH, Bazanjani S, Wang Z, Bauer F, Sharp A, Hung YP — Wed, 28 Oct 2026 00:00:00 +0000

98% · External validation 98%

Objective: To develop and externally validate a radiomics-based machine learning model for differentiating myxoid liposarcoma (MLS) from myxoma on MRI and compare its diagnostic performance with that of experienced musculoskeletal radiologists. Methods: In this retrospective study, T1-weighted and fluid-sensitive MRI sequences from 109 patients with histologically confirmed MLS or myxoma were used to train support vector machine (SVM), random forest (RF), and XGBoost (XGB) classifiers. Models were evaluated in an independent external cohort of 32 patients from a second institution using different scanner hardware and protocols. Diagnostic performance was compared with that of two fellowship-trained senior musculoskeletal radiologists. Secondary analysis assessed MLS grading. Results: The T1-weighted SVM achieved the strongest external binary performance (accuracy, 0.84; AUC, 0.91), outperforming RF and XGB (p = 0.005 and p < 0.001, respectively). Its performance did not differ significantly from that of radiologists (AUC, 0.88-0.94; p > 0.05), although the study was not powered to establish equivalence. Fluid-sensitive SVM performance was limited (accuracy, 0.56; AUC, 0.60), indicating sequence-dependent generalizability. Models and radiologists demonstrated limited MLS grading accuracy, reflecting tumor heterogeneity. Conclusion: Externally validated radiomics-based machine learning supported noninvasive differentiation of MLS from myxoma primarily on T1-weighted MRI. T1-weighted SVM performance was not significantly different from that of experienced radiologists, although equivalence was not established, whereas fluid-sensitive performance was substantially weaker. T1-weighted quantitative imaging biomarkers may complement soft tissue tumor characterization, but extrapolation to fluid-sensitive sequences requires caution.

★ Incremental Value and Cross-Center Validation of Pretreatment Tri-Modal Ultrasound Deep Learning for Predicting Breast Pathologic Complete Response to Neoadjuvant Chemotherapy: A Multicenter Study.

Wei C, Pan T, Li T, Zheng C, Duan Y, Nie F — Wed, 28 Oct 2026 00:00:00 +0000

98% · External validation 98%

Objective: To evaluate the incremental predictive value of pretreatment tri-modal ultrasound deep learning integrating two-dimensional ultrasound (2DUS), color Doppler flow imaging (CDFI), and contrast-enhanced ultrasound (CEUS) for predicting breast pathologic complete response (bpCR) after neoadjuvant chemotherapy (NAC), and to assess performance stability across independent centers. Methods: This retrospective multicenter study included 1089 patients with primary invasive breast cancer. Center 1 provided the training cohort (n = 572) and internal validation cohort (n = 246), whereas centers 2 (n = 160) and 3 (n = 111) served as independent external validation cohorts. Using pretreatment 2DUS, CDFI, and CEUS within a standardized intratumoral and 10-mm peritumoral framework, single-modality, all three dual-modality combinations, and tri-modal models were developed. The best-performing dual-modality model was identified and locked during training as the primary imaging comparator. A Tri-modal + Clinical model further integrated pretreatment clinical and biomarker variables. Performance was assessed using AUC, Brier score, calibration, decision curve analysis, and ablation experiments. Results: Overall, 383 patients (35.2%) achieved bpCR. CEUS was the strongest single modality, and 2DUS + CEUS was the best-performing dual-modality model. Compared with 2DUS + CEUS, the tri-modal model increased the AUC from 0.895 to 0.919 in the combined external cohort (ΔAUC, +0.023; 95% CI, +0.010 to +0.038; p = 0.001), with corresponding improvements in Brier score (0.124 to 0.105) and F1 score (0.747 to 0.786). The AUC point estimate increased in both external centers, although the 95% CI for the incremental AUC included zero in external validation cohort 2. Despite modest standalone discrimination, CDFI provided measurable incremental information when added to 2DUS + CEUS. Further integration of pretreatment clinical and biomarker variables increased the combined external AUC to 0.935. Reduced-input retraining showed performance decreases after omission of any modality, with the largest reduction after CEUS omission. Conclusions: Pretreatment tri-modal ultrasound deep learning provided incremental predictive value for bpCR beyond the best dual-modality model, with a consistent direction of benefit across two independent external centers. Clinical and biomarker integration provided additional predictive information, supporting further evaluation of this multimodal framework for individualized pretreatment bpCR prediction.

★ A clinical-deep learning model for prolonged length of stay following endovascular aortic repair.

Hu C, Bai T, Qin X, Tang K, Peng T, Yang Q — Wed, 28 Oct 2026 00:00:00 +0000

98% · External validation 98%

Background: Prolonged length of stay (PLOS) after endovascular aortic repair (EVAR) for abdominal aortic aneurysm (AAA) is associated with increased complications, resource burden, and patient distress. Early identification of high-risk patients remains challenging. This study aimed to develop and externally validate a predictive model integrating clinical variables and three-dimensional deep learning (3D-DL) features from preoperative computed tomography angiography (CTA) to predict PLOS after EVAR. Methods: A total of 788 patients with AAA who underwent elective EVAR were enrolled from two hospitals. The internal training cohort consisted of 556 patients, and the external validation cohort included 232 patients. PLOS was defined as a LOS above the 75th percentile of the distribution. A clinical nomogram was developed using LASSO regression followed by multivariate logistic regression. Deep learning (DL) features were extracted from manually segmented computed tomography angiography (CTA) images using a 3D ResNet200 network, and LASSO regression was applied to select the most informative features. An ensemble model integrating the clinical predictors and the selected DL features was constructed. Model performance was assessed using receiver operating characteristic curves, calibration plots, and decision curve analysis. Results: LASSO combined with multivariate logistic regression identified five independent clinical predictors: age, hemoglobin, albumin, maximum aneurysm diameter, and diabetes. The clinical model achieved an area under the curve (AUC) of 0.917 in the internal cohort and 0.944 in the external cohort. From the 3D ResNet200 network, 32 DL features were extracted. LASSO regression further selected two of these features. The ensemble predictive model, which integrated the clinical predictors and DL features, demonstrated improved discrimination, with an AUC of 0.971 in the internal cohort and 0.958 in the external cohort. Internal validation with 1,000 bootstrap resamples yielded optimism-corrected AUCs of 0.911 for the clinical model and 0.966 for the ensemble model. Subgroup analyses demonstrated consistent model performance across age and diabetes strata, while performance in the female subgroup warrants further evaluation in larger cohorts. Conclusion: This study developed an ensemble predictive model to predict PLOS following EVAR. The model demonstrated excellent discrimination and provided significant prognostic value beyond clinical variables alone. This nomogram may serve as a preoperative tool for risk stratification and resource planning in this population. Future studies should focus on prospective validation and clinical integration.

★ UWMT-net: An uncertainty-weighted multi-task network for ultrasound-based three-class classification of thyroid nodules.

Li Q, Mao Z, He Y, Wang J, Yang Z — Wed, 28 Oct 2026 00:00:00 +0000

97% · External validation 97%

Background: Risk stratification of thyroid nodules on ultrasound remains inconsistent in routine practice, particularly for borderline presentations. Many deep-learning studies report discrimination alone, while probability reliability and threshold-based clinical utility receive less attention. Methods: We retrospectively included 364 development patients, each contributing one thyroid nodule labeled using surgical histopathology or fine-needle aspiration cytology. The primary analysis used a patient-level fixed training/validation/test framework, and robustness was additionally assessed with stratified patient-level five-fold cross-validation. A locked model was independently evaluated in 121 patients from the Affiliated Hospital of Shandong University of Traditional Chinese Medicine (Shandong Provincial Hospital of Traditional Chinese Medicine), Jinan, China. The task was formulated as three-class classification: benign, indeterminate or suspicious, and malignant. Results: The multimodal model outperformed clinical-only and image-only baselines, with the largest gains observed on class-balanced metrics. On the held-out test set, UWMT-Net achieved an accuracy of 0.838 and a macro-AUC of 0.943. Post-hoc temperature scaling reduced ECE from 0.040 to 0.015 without changing hard-label metrics. Five-fold cross-validation yielded a macro-AUC of 0.936 ± 0.011. In the independent external cohort, the locked model achieved an accuracy of 0.810, macro-F1 of 0.804, macro-AUC of 0.921, and ECE of 0.031. The added fusion baselines and paired bootstrap analysis supported a modest, but not statistically significant, numerical advantage of uncertainty-weighted multi-task learning. Conclusions: Integrating ultrasound with structured clinical information improved three-class classification and probability reliability in the development cohort, with retained performance in the independent external cohort. Larger prospective multicenter validation remains required before clinical use.

★ A Fast and Generalizable Deep Neural Network for the Detection of Atrial Fibrillation.

Mishra H, Adam F, Makwana N, Tiwari P, Kandasamy S, Hari KVS — Wed, 28 Oct 2026 00:00:00 +0000

97% · External validation 97%

Atrial fibrillation (AFib) is the most prevalent sustained cardiac arrhythmia, yet most deep learning detectors are validated on a single cohort and few report uncertainty. We developed a compact 14M-parameter convolutional neural network for binary AFib classification from 10 s, 12-lead ECG at 500 Hz, trained exclusively on the HDXML dataset (67,432 ECGs; 9628 AFib) with on-the-fly augmentation for noise, drift and missing channels. Using one unchanged checkpoint, we evaluated six independently sourced public cohorts totaling over 1.25 million recordings, including MIMIC-IV critical care and two ambulatory Holter cohorts. All metrics are reported with 95% confidence intervals from a cluster bootstrap at the patient, recording or subject levels. ROC-AUC exceeded 0.95 in every cohort, from 0.960 (95% CI 0.956-0.963) on CODE-15 to 0.999 (0.998-0.999) on SPH, including 0.966 (0.965-0.967) on MIMIC-IV with 75.3% sensitivity and 97.7% specificity, 0.963 (0.929-0.987) on CPSC 2021 and 0.995 (0.988-0.999) on MIT-BIH. Performance was stable under progressive lead masking without retraining, and median inference took 110 ms per segment on a server-class CPU without a GPU. A compact network trained on one curated source can therefore generalize across hospitals, countries and recording hardware. This study is retrospective; prospective validation is required before clinical use.

★ Selection Instability and External Performance of Transcriptomic Models Beyond Age, Sex and Stage in Resected Lung Adenocarcinoma.

Lei Z, Zou C, Zhao W — Wed, 28 Oct 2026 00:00:00 +0000

97% · External validation 97%

Background/Objectives: We evaluated molecular selection stability and external prediction beyond age, sex and broad pathological stage in resected stage I-II lung adenocarcinoma. Methods: Models were developed in 228 patients from two sources and frozen before evaluation in four historical cohorts containing 778 patients. The primary comparison added pathway scores using elastic-net Cox regression. Secondary analyses examined a gene model, 200 development bootstrap resamples, alternative modeling specifications and locally updated baselines. Prediction error was measured by the integrated Brier score over 1-3 years (IBS); negative molecular-minus-clinical differences (ΔIBS) favored molecular models. Evaluation intervals used 2000 paired bootstrap resamples. Results: The pathway model retained no molecular terms. The final gene model retained 16 genes, whereas both source-held-out fits and 159/200 development resamples retained none. Penalties of 0.5, 0.15 and 0.05 were selected in 160, 38 and 2 resamples. Alternative specifications changed gene retention from 0 to 61 and pathway retention from 0 to 5. Frozen gene-model ΔIBS ranged from -0.0034 to 0.0055 across cohorts. Its equal cohort mean was 0.0016 (95% CI, -0.0012 to 0.0042), compared with -0.0008 (-0.0024 to 0.0007) after cross-fitted local baseline updating. The clinical model predicted a 3-year mortality of 19.6% in Japan and 22.6% in Moffitt, versus observed values of 9.1% and 28.6%. Conclusions: Molecular selection was unstable, and incremental performance depended on specification and population. The results characterize these learning procedures and their limited development sample; broader transcriptomic prognostic value and contemporary clinical utility remain unestablished.

★ A Pathological Image-Based Pathomics Signature Combined with Clinicopathological Features for Predicting Lymph Node Metastasis in Esophageal Squamous Cell Carcinoma: A Development and External Validation Study.

Peng K, Lu J, Zhang P, Zhang W, Fang J, Lin J — Wed, 28 Oct 2026 00:00:00 +0000

97% · External validation 97%

Background: Esophageal squamous cell carcinoma (ESCC) accounts for over 80% of esophageal cancer cases globally. Lymph node status is the strongest determinant of survival after resection, yet occult disease is found in about 17% of patients classified as node-negative. We developed a postoperative model estimating the probability of lymph node metastasis (LNM) from the resection specimen, combining nuclear features with clinicopathological variables. Methods: Of 477 consecutive ESCC patients undergoing radical esophagectomy without neoadjuvant therapy, 416 were analyzed (LNM 217, 52.2%). Nuclear features were extracted from H&E whole-slide images (20×) by QuPath optical-density detection; LASSO logistic regression selected 17. A nomogram combining the pathomics score with age, sex, tumor location, pT category and grade was validated by 10-fold cross-validation (out-of-fold, OOF) and applied without refitting to 101 patients from another institution. Results: OOF AUC was 0.821 (95% CI 0.779-0.858) for the combined model, versus 0.791 (clinical) and 0.762 (pathomics alone). The pathomics score remained an independent predictor (adjusted OR 2.47, 95% CI 1.80-3.39; p < 0.001). LNM prevalence rose across tertiles (15.8%, 55.1%, 85.6%; p < 0.001). Among 199 node-negative patients, high pathomics risk predicted worse 3-year overall survival (77.2% vs. 94.3%; p = 0.005) and progression-free survival (75.5% vs. 88.9%; p = 0.007). Externally (n = 101; LNM 24.8%), AUCs were 0.812, 0.801 and 0.697, and adding pathomics did not significantly improve discrimination. Conclusions: Within the development cohort, pathomics added modest but significant incremental value and identified node-negative patients at elevated risk. Because the model requires the resection specimen, it cannot guide preoperative treatment selection. External validation did not confirm incremental value: the clinical component; the pathomics signature did not.

★ Development and validation of an interpretable machine learning model for early prediction of mortality in ICU patients with chronic obstructive pulmonary disease and heart failure.

Shen J, Liu Y, Chen Y — Wed, 28 Oct 2026 00:00:00 +0000

97% · External validation 97%

Chronic obstructive pulmonary disease and heart failure frequently coexist, and the mortality risk for patients with these comorbidities is significantly elevated in the intensive care unit (ICU) setting. Therefore, developing an interpretable model for early mortality prediction is crucial to improving the prognosis and management of these patients. This study developed models based on the Medical Information Mart for Intensive Care IV database (3167 patients) and performed external validation using the eICU Collaborative Research Database (1781 patients). Least absolute shrinkage and selection operator regression combined with the Boruta algorithm was used to screen clinical features within the first 24 hours of ICU admission. Based on the selected features, 5 machine learning models (eXtreme Gradient Boosting, Categorical Boosting, Light Gradient Boosting Machine, Random Forest, and Support Vector Machine) were constructed and compared for their performance in predicting 28-day all-cause mortality. The optimal model was interpreted using the Shapley Additive Explanations framework. In internal validation, the Categorical Boosting model demonstrated the best overall performance, achieving an area under the receiver operating characteristic curve of 0.788, with a sensitivity of 0.482 and a specificity of 0.900. In independent external validation, the model's area under the receiver operating characteristic curve was 0.739. Feature selection identified 14 key predictive variables. Shapley Additive Explanations analysis revealed that the Sequential Organ Failure Assessment score, Simplified Acute Physiology Score II score, anion gap, vasopressor use, and respiratory rate were the top 5 features contributing most to the model's predictions. This study developed and validated an interpretable machine learning model based on multicenter data for the early prediction of 28-day mortality risk in ICU patients with chronic obstructive pulmonary disease and heart failure. The model demonstrates good discriminative ability and reasonable generalizability, with its interpretability analysis clarifying core predictive factors to aid in clinical risk stratification and intervention.

★ Multi-sequence magnetic resonance imaging radiomics integrated with clinical features predicts pelvic lymph node metastasis in cervical cancer: a retrospective-prospective cohort study.

Gong W, Liu X, Wang M, Hu X, Yan G, Liu K — Wed, 28 Oct 2026 00:00:00 +0000

97% · Prospective 97%

Background: Pelvic lymph node metastasis is a major determinant of staging, recurrence risk, and treatment selection in cervical cancer, yet conventional magnetic resonance imaging criteria based on nodal size and morphology remain insufficient for occult nodal disease. We developed and temporally validated a multi-sequence magnetic resonance imaging radiomics-clinical fusion framework for pretreatment prediction of pelvic lymph node metastasis. Methods: This single-center retrospective-prospective cohort study included 200 women with histologically confirmed cervical cancer who underwent pretreatment pelvic magnetic resonance imaging. A retrospective training cohort (n = 100) was accrued from January 2018 to December 2022, and a prospective temporal validation cohort (n = 100) was accrued from January 2023 to December 2024. Reference standards comprised surgical histopathology in 154 patients and composite clinical-imaging adjudication in 46 patients. Primary tumors were manually segmented on axial T2-weighted imaging and registered to T1-weighted imaging, diffusion-weighted imaging/apparent diffusion coefficient maps, and dynamic contrast-enhanced imaging. An Image Biomarker Standardisation Initiative-compliant radiomics pipeline was used, including resampling, intensity discretization, reproducibility filtering, correlation pruning, and nested feature selection. Clinical, radiomics, clinical-radiomics, and deep fusion models were evaluated in the prospective cohort using discrimination, calibration, decision-curve, and reclassification analyses. Results: The deep fusion model achieved the best validation discrimination, with an area under the receiver operating characteristic curve of 0.910 (95% confidence interval, 0.845-0.959) and an area under the precision-recall curve of 0.849. At the training-derived threshold, accuracy was 0.84, sensitivity was 0.79, specificity was 0.86, positive predictive value was 0.75, negative predictive value was 0.89, and Brier score was 0.125. Deep fusion outperformed the clinical model (area under the curve, 0.730; mean delta area under the curve, 0.180; P < 0.001) and improved reclassification relative to the clinical model. Conclusion: Multi-sequence magnetic resonance imaging radiomics-clinical fusion improved pretreatment risk stratification for pelvic lymph node metastasis in cervical cancer. External multicenter validation is warranted before clinical implementation. These preliminary single-center findings should be interpreted cautiously and require multicenter external validation before broad clinical implementation.

★ Deep learning based detection of active pulmonary tuberculosis from chest radiographs in a retrospective prospective multicenter study.

Tuersun A, Abudoubari S, Xu J, Ma J, Nijiati P, Warisi X — Wed, 28 Oct 2026 00:00:00 +0000

96% · External validation 96%, Prospective 95%

Active tuberculosis remains a major global health problem, and distinguishing active from non-active disease on chest radiographs could accelerate triage where radiology resources are limited. We developed ShuffleNetV2 using 4,613 radiographs from two centers and evaluated it in retrospective external (1,318 patients) and prospective (560 patients) cohorts. The model achieved AUCs of 0.917 and 0.896, respectively, and performance was stable across prespecified demographic and scanner subgroups. In the prospective cohort, artificial-intelligence assistance increased radiologist AUCs to 0.897-0.937 across experience levels. These findings support ShuffleNetV2 as a rapid decision-support tool for prioritizing confirmatory clinical and microbiological assessment; positive predictions should not be used alone to initiate treatment.

An Interpretable Multimodal Machine-Learning Model for Predicting Cognitive Impairment from Routine MRI: A Study with External Validation.

Chen X, Zhang B, Li M, Wang S, Yang M, Zhang J — Wed, 28 Oct 2026 00:00:00 +0000

92% · External validation 92%

Existing MRI-based prediction models for cognitive impairment often rely on a single biomarker or simply compare algorithms, act as black boxes, and lack external validation, which limits their clinical translation. We developed and externally validated an interpretable multimodal machine-learning model that classifies a MoCA-defined screening outcome as a binary task: cognitive impairment (CI, MoCA < 26) versus no cognitive impairment (NCI, MoCA ≥ 26). Sixteen candidate features were considered, including AI-derived medial temporal atrophy (MTA), regional white matter hyperintensity (WMH) volumes and proportions, and Fazekas ratings from routine T1WI and T2-FLAIR, together with age and sex. In a single-center cohort of 433 subjects (302 CI, 131 NCI) with a 7:3 training/hold-out split, five algorithms were compared and Random Forest performed best; SHAP-guided recursive feature elimination within five-fold cross-validation then reduced the set from 16 to 11 features. The optimized 11-feature Random Forest reached a cross-validated AUC of 0.883 ± 0.024 on the training folds and an internal hold-out AUC of 0.966, and 0.860 in an independent ADNI cohort of 139 subjects (100 CI, 39 NCI). SHAP identified the left MTA score as the most influential feature, followed by near-ventricular (juxtaventricular and periventricular) WMH burden and the right MTA score. The left MTA score was most strongly correlated with MoCA (r = -0.481). The model may offer an objective, routine-MRI-based screening aid, pending prospective multicenter validation.

Real-Time AI-Based ACR TI-RADS Classification of Thyroid Nodules on the Ultrasound Console: A Prospective Agreement Study in a Region with Historical Iodine Deficiency.

Wehmann M, Freesmeyer M, Kühnel C, Gühne F, Seifert P — Wed, 28 Oct 2026 00:00:00 +0000

92% · Prospective 92%

Background/Objectives: To assess the agreement between an artificial intelligence (AI)-based decision support system (DS System), applied in real time directly on the ultrasound console, and an expert panel for ACR TI-RADS classification of thyroid nodules (TNs) in a region with a history of iodine deficiency and a consequently high prevalence of TNs. Methods: A single-center prospective study including 118 consecutive patients (with 276 TNs) referred for thyroid ultrasound between March and May 2024. DS System ultrasound features, ACR TI-RADS classifications (TR1-TR5) and management recommendations were recorded per TN, with and without the 'AI-Adapter' (AI-based modification of the TI-RADS score), and compared to the consensus assessment of a TI-RADS expert panel (paired tests of marginal homogeneity, Cohen's unweighted and weighted κ; all analyses adjusted for clustering of TNs within patients). Results: No malignancy was identified during 24-month follow-up (composite reference standard; cytology or histology in six TNs). The DS System could be applied to all TNs without technical failure. Marginal distributions differed between the expert panel and the DS System for all ultrasound feature categories except echogenicity (p = 0.23). Interrater agreement ranged from poor ("lobulated or irregular", κ = -0.08) to substantial ("cystic or almost completely cystic", κ = 0.62), most frequently slight to fair. TI-RADS classifications showed slight agreement with (κ = 0.04; linear-weighted κ = 0.14) and without AI-Adapter (κ = 0.13; linear-weighted κ = 0.21). The AI-Adapter changed the ACR TI-RADS score by -1.01 ± 1.04 points. Management recommendations showed fair agreement with (κ = 0.23) and without the AI-Adapter (κ = 0.39). Although the number of TR5 ratings was similar (expert panel: 27; DS System: 26 without and 30 with the AI-Adapter), only seven and eight of the 27 TR5 TNs of the expert panel, respectively, were also rated TR5 by the DS System. Conclusions: Real-time application of the DS System on the ultrasound console was technically feasible, but agreement with an expert panel was low. As all TNs were benign, diagnostic accuracy and clinical safety could not be assessed. Validation in cohorts including malignant TNs is required before the DS System can be considered a substitute for expert assessment in regions with a high prevalence of TNs.

Cross-Dataset Generalization in Phonocardiogram Murmur Classification: A Benchmark Study of Domain-Conditional Performance and the Role of Multi-Representation Fusion.

Kutlu H, Çolak C — Wed, 28 Oct 2026 00:00:00 +0000

91% · External validation 91%

Background/Objectives: Automated phonocardiogram (PCG) analysis is a candidate tool for scaling up cardiac auscultation, but the internal accuracy reported in many studies may not reflect performance when a model is applied to recordings from a different source. This study evaluated how a broad range of classifiers generalize across PCG datasets and examined whether a multi-representation architecture improves cross-dataset transfer. Methods: Eleven approaches were trained on the pediatric CirCor DigiScope dataset and evaluated, without target-domain fine-tuning, on two external datasets: a heterogeneous adult/mixed dataset (the PhysioNet/Computing in Cardiology Challenge 2016 database) and a homogeneous five-class open-access heart sound dataset. The approaches spanned classical machine learning (spport vector machine, random forest), standard deep models (1D-CNN, 2D-CNN, ResNet1D, CNN-BiLSTM, and two hybrid networks), image-based backbones (EfficientNet-B0, MaxViT), and a proposed four-branch fusion model (PCG-MambaConformer) combining raw-signal, mel-spectrogram, and continuous wavelet transform (CWT) representations. Threshold-transfer and precision-recall analyses, adversarial domain adaptation, an uncertainty-quantification benchmark, a disease-specific transfer analysis, multi-architecture ensembling, a branch ablation, a label-grouping sensitivity analysis, a reciprocal (reverse-direction) transfer experiment and explainability analyses (feature-space t-SNE with class-separability metrics and Grad-CAM saliency) were also performed. Performance was summarized using the area under the receiver operating characteristic curve (AUROC) with sensitivity and specificity. Results: Internal AUROC was similar across deep models (approximately 0.82-0.90) and largely independent of parameter count, whereas classical models were lower (0.67-0.69). On the heterogeneous external dataset, the highest AUROC was 0.738 and none of the nine deep models reached a sensitivity and a specificity both of at least 0.80 at any of the three thresholds examined. On the homogeneous external dataset, PCG-MambaConformer reached an AUROC of 0.985 (sensitivity 0.718, specificity 0.995 at the 0.5 threshold), whereas other models tended to collapse toward either always-normal or always-abnormal predictions. Adversarial adaptation improved performance on the homogeneous target but reduced it on the heterogeneous target. On the heterogeneous target, even an oracle threshold raised balanced accuracy by only 0.022 on average, whereas on the homogeneous target a source-derived threshold produced a balanced operating point only for PCG-MambaConformer (sensitivity 0.863, specificity 0.980). Models trained in the reverse direction (PhysioNet/CinC Challenge 2016 to CirCor) also transferred poorly (AUROC 0.612-0.677). Branch ablation showed that three of the four single-branch configurations collapsed to a one-class operating point and that the full fusion model achieved the highest AUROC, and a feature-space analysis showed that successful transfer was associated with the class-separability of the target domain (silhouette score highest for the homogeneous dataset) rather than with the proximity of dataset centroids. Conclusions: In the two external datasets evaluated, successful cross-dataset PCG transfer was associated with the homogeneity of the target domain; because these datasets also differed in population, recording protocol, disease spectrum, class distribution and signal quality, homogeneity cannot be isolated as the sole determinant of transferability. Internal accuracy was a weak indicator of external behavior, model complexity alone did not close the heterogeneous transfer gap, and balanced cross-dataset performance on a homogeneous target was associated with multi-representation fusion. The near-perfect result on the homogeneous target should be read as technical cross-dataset transfer under favorable, highly curated conditions rather than as clinical external validation, whereas the heterogeneous result is the more realistic indication of current limitations. These findings support multi-metric external evaluation on independent data, followed by prospective clinical validation, before any clinical use is considered.

Explainable vision transformer framework for breast lesion classification with Grad-CAM support.

Kumar A, Khan M, Mittal G, Husain I, Shukla S, Ali SA — Wed, 28 Oct 2026 00:00:00 +0000

90% · External validation 90%

Background: Given that breast cancer remains the most lethal malignancy among women globally, there is an urgent need for reliable and comprehensible diagnostic technology. Deep learning models have demonstrated encouraging outcomes in breast imaging; nevertheless, existing methodologies struggle to encompass global context and deliver clinically relevant justifications for their conclusions. Method: The ExViT-Breast framework was trained and internally validated on CBIS-DDSM (n = 1,566 subjects; masses and calcifications), externally validated on INbreast, and evaluated across modalities on the BUSI ultrasound dataset. A Vision Transformer backbone with a cross-view attention fusion module integrated paired craniocaudal (CC) and mediolateral oblique (MLO) mammographic views. Four explainability methods were compared: Grad-CAM, Grad-CAM++, LayerCAM, and Transformer-LRP. Classification performance was assessed by AUC and sensitivity at one false positive per examination. Explanation quality was evaluated using faithfulness (deletion/insertion curves), clinical plausibility (pointing game against expert-defined lesion regions), and ensemble-based uncertainty estimation. Model calibration was quantified using Expected Calibration Error (ECE). Results: ExViT-Breast achieved superior diagnostic performance with an AUC of 0.94 ± 0.02 on CBIS-DDSM, significantly outperforming conventional CNN architectures (ResNet-50: AUC = 0.89, p < 0.001) and single-view Vision Transformers (AUC = 0.91, p = 0.003). External validation on INbreast demonstrated robust generalization (AUC = 0.91 ± 0.03), while cross-modality evaluation on BUSI yielded AUC = 0.88 ± 0.04. Transformer-LRP consistently provided the highest explanation quality across all datasets, achieving faithfulness scores of 0.71 ± 0.05 and plausibility rates of 79.5% on CBIS-DDSM, significantly superior to Grad-CAM variants (p < 0.001). Conclusion: The ExViT-Breast framework demonstrates the potential of explainable Vision Transformers in breast imaging, combining high classification performance with clinically interpretable explanations. The integration of multi-view analysis and quantitative explainability assessment positions this approach as a promising tool for computer-aided diagnosis in breast cancer screening and detection. Clinical Trial: Clinical trial number: not applicable.

★ Development and validation of prediction models for heart failure with improved ejection fraction.

Wang Q, Wang G, Lv G, Shen G, Li J, Shao Z — Tue, 27 Oct 2026 00:00:00 +0000

98% · External validation 98%

Background: Approximately 10-60% of heart failure with reduced ejection fraction (HFrEF) patients may transition to heart failure with improved ejection fraction (HFimpEF), a clinically meaningful phenotype associated with better prognosis. However, existing prediction tools demonstrate suboptimal discrimination and poor clinical utility. We aimed to develop and validate interpretable models for HFimpEF prediction by comparing logistic regression with ensemble machine learning. Methods: This retrospective cohort study included 1,257 HFrEF patients from the First Affiliated Hospital of USTC (March 2015-October 2023), split into training (n = 1,006) and internal validation (n = 251) sets. External validation used data from 310 patients across eight hospitals in Anhui Province (January 2024-January 2025). HFimpEF was defined as LVEF increase ≥10% to >40%. Five models were developed: data-driven logistic regression (Model 1); literature-based logistic regression (Model 2) with 5 common predictors validated in both BIOSTAT-CHF and ASIAN-HF cohorts; random forest (Model 3) and XGBoost (Model 4), both tuned via 5-fold cross-validation; and a simplified ensemble (Model 5) that averaged the predicted probabilities of Models 1, 3, and 4, then calibrated the average via logistic regression. Performance was assessed using AUC for discrimination, Brier score with calibration plots for calibration, decision curve analysis for clinical utility, and DeLong's test for model comparisons; SHAP analysis was applied for interpretability of Models 3 and 4. Results: Median age was 65 years (IQR 53-73); 69.9% were male. Multivariable analysis identified seven predictors: age (OR 0.97, 95% CI 0.96-0.99), BMI (OR 1.04, 95% CI 1.00-1.09), systolic blood pressure (per 10 mmHg increase, OR 1.13, 95% CI 1.01-1.25), prior PCI (OR 0.45, 95% CI 0.26-0.79), HDL-C (OR 2.48, 95% CI 1.36-4.50), FT3 (OR 1.21, 95% CI 1.02-1.45), and LVESD (OR 0.92, 95% CI 0.87-0.99). In external validation, Model 5 showed superior discrimination (AUC 0.67 vs. 0.60 for Model 2, P = 0.004) but poorer calibration (Brier score 0.34). Model 1 demonstrated balanced performance (AUC 0.66, Brier score 0.23). Conclusions: Our findings indicate that ensemble learning methods provide a moderate enhancement in discrimination for HFimpEF prediction, while logistic regression models exhibit superior robustness and interpretability. We have developed an online tool (https://alslyy-wq.shinyapps.io/USTC-HF-HFimpEF-Prediction/) that integrates both methodologies to facilitate practical HFimpEF risk assessment.

★ Machine learning-based radiomic analysis of T2-FLAIR white matter hyperintensity predicts hemorrhagic transformation after endovascular thrombectomy for acute ischemic stroke.

Wu Q, Shi Y, Zhang Z, Zhang J, Wu J, Yang Z — Tue, 27 Oct 2026 00:00:00 +0000

98% · External validation 98%

Objective: Hemorrhagic transformation (HT) remains a major complication after endovascular therapy (EVT) in patients with acute ischemic stroke (AIS). This study aimed to develop and validate a machine-learning model based on T2-FLAIR white matter hyperintensity (WMH) radiomics combined with clinical variables to predict HT after EVT. Methods: In this dual-center retrospective study, 446 patients (mean age 61.88 ± 10.98 years; 332 [74.4%] male) with AIS who underwent EVT were enrolled from two centers. Patients from the primary center were randomly assigned to a training cohort and an internal test cohort, and patients from the second center served as an independent external validation cohort. Clinical risk factors were identified by univariate and multivariate logistic regression. Preoperative T2-FLAIR images were used to segment WMH lesions and extract radiomics features. After feature selection, three predictive models were then developed: a clinical model based on the independent clinical predictors, a radiomics model based on the radiomics signature, and a combined model integrating both. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration analysis, DeLong testing, and decision curve analysis. Results: Among the 446 patients, 153 (34.3%) developed HT within 24-48 h after EVT. Multivariate analysis identified NIHSS score, occlusion site, and atrial fibrillation as independent clinical predictors of HT. In the internal test cohort, the combined model achieved the best discriminative performance (AUC = 0.818, 95% CI: 0.651-0.986), outperforming the clinical model (AUC = 0.740, 95% CI: 0.562-0.917) and the radiomics model (AUC = 0.806, 95% CI: 0.621-0.990). In the external validation cohort, the combined model maintained robust performance (AUC = 0.817, 95% CI: 0.746-0.889), which was comparable to the clinical model (AUC = 0.818, 95% CI: 0.742-0.894) and higher than the radiomics model alone (AUC = 0.764, 95% CI: 0.682-0.845). Conclusions: A combined model incorporating T2-FLAIR WMH radiomics and key clinical variables provides a feasible tool for individualized prediction of HT after EVT in patients with AIS, although its incremental discriminatory value over the clinical model alone was limited in the external validation cohort. This approach may help improve perioperative risk stratification and support personalized clinical decision-making.

★ Early Prediction and Prognosis in Acute Kidney Injury Using a Biomarker-Aware Transformer: Retrospective Multicenter Study.

Xu Z, Li D, Xu Q, Yu AS, Chrischilles EA, Cowell LG — Tue, 27 Oct 2026 00:00:00 +0000

97% · External validation 97%, Language models 49%

Background: Acute kidney injury (AKI) is a common, life-threatening complication among hospitalized patients. Timely prediction is crucial for prevention, while accurate prognosis following onset is essential for guiding care. Although electronic health records (EHRs) provide rich longitudinal data, their high dimensionality, heterogeneity, and temporal complexity pose challenges for traditional machine learning. Objective: The aim of this study was to develop and evaluate a transformer-based model for generalizable AKI prediction and prognosis by modeling long-range temporal dependencies and complex relationships in heterogeneous EHR data and to examine its interpretability. Methods: We developed the Biomarker-Aware Transformer for AKI (BAT-AKI), which integrates multidimensional embeddings capturing temporal dynamics, textual descriptions, hierarchical ontology structures, and semantic relationships in EHRs. To enhance sensitivity to AKI-related biomarker fluctuations, BAT-AKI combines masked language modeling with a novel biomarker abnormality-aware pretraining task. The model was pretrained and fine-tuned for 4 tasks: AKI onset, early reversal, recovery at discharge, and 90-day post-AKI mortality. Pretraining and internal validation used data from the National Patient-Centered Clinical Research Network (PCORnet) site, with external validation across 3 additional sites. Results: BAT-AKI achieved the strongest overall discrimination, outperforming baseline models in area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC) on most tasks. For AKI onset prediction, it achieved an AUROC of 0.8549 (95% CI 0.8433-0.8658) and an AUPRC of 0.6041 (95% CI 0.5786-0.6292), exceeding the second-best baseline by 2.8% in AUROC and 8.1% in AUPRC. Across prognostic tasks, BAT-AKI achieved AUROCs of 0.6006 (95% CI 0.5703-0.6318), 0.6080 (95% CI 0.5745-0.6397), and 0.8017 (95% CI 0.7708-0.8311) and AUPRCs of 0.7036 (95% CI 0.6690-0.7388), 0.7417 (95% CI 0.7082-0.7765), and 0.5063 (95% CI 0.4422-0.5684) for early reversal, recovery at discharge, and 90-day post-AKI mortality, respectively, corresponding to relative gains of 2.4% to 6% in AUROC and 3.7% to 4.1% in AUPRC over the second-best baseline. Performance remained stable across external cohorts, with AKI onset AUROCs ranging from 0.7754 (95% CI 0.7677-0.7827) to 0.7989 (95% CI 0.7885-0.8092) and AUPRCs ranging from 0.4777 (95% CI 0.4631-0.4924) to 0.5687 (95% CI 0.5495-0.5871), corresponding to relative improvements of 3.7% to 5.3% in AUROC and 16.8% to 25.3% in AUPRC over the transformer baseline. In a low-resource setting using only 2% of the fine-tuning data, BAT-AKI achieved an AUROC of 0.7898 (95% CI 0.7759-0.8036) and an AUPRC of 0.4903 (95% CI 0.4640-0.5178) for AKI onset prediction, exceeding the second-best baseline by 4.2% in AUROC and 10.9% in AUPRC. Visualization of embeddings, attention patterns, and integrated-gradient feature attributions highlighted clinically meaningful relationships and patient-specific trajectories. Conclusions: BAT-AKI provides generalizable AKI prediction and prognosis with consistent discrimination across institutions, while offering interpretability insights into its predictions, supporting its potential for early warning and AKI care optimization.

★ Artificial Intelligence-Enhanced Auscultation to Identify Left Ventricular Dysfunction - Development and External Validation in Prospective Multicenter Studies.

Sato E, Izumi Y, Saito Y, Saijo Y, Kitai T, Hayashida A — Tue, 27 Oct 2026 00:00:00 +0000

97% · External validation 97%, Prospective 87%

Background: Left ventricular dysfunction (LVD) is common and associated with adverse prognosis, even in asymptomatic individuals. Early treatment improves outcomes, underscoring the need for efficient strategies to identify patients who may benefit from echocardiographic evaluation. This study aimed to develop and externally validate an AI auscultation method to detect LVD among patients referred for clinically indicated echocardiography. Methods: In this prospective, multicenter study, adult patients referred for echocardiography were enrolled to undergo digital auscultation at four standard auscultation sites with simultaneous single-lead ECGs. LVD was defined as a left ventricular ejection fraction (LVEF) less than or equal to 40% on concurrent echocardiography. An ensemble model combining a transformer-based network and a one-dimensional convolutional neural network was trained on multicenter data and externally validated. The model's performance was compared with a seven-cardiologist reader substudy and clinical logistic regression models incorporating clinical variables, single-lead ECG-derived features, and B-type natriuretic peptide levels. Results: The multicenter training cohort included 2359 patients (mean age, 69.7±13.5 years; 40.4% female; 8.9% with reduced LVEF). The external validation cohort of 964 patients (mean age, 68.6±15.4 years; 38.9% female; 7.6% with reduced LVEF) was obtained from another independent multicenter registry. The four-site AI auscultation outperformed ECG-only models and the clinical logistic models, achieving an area under the receiver operator curve (AUROC) of 0.926 (95% confidence interval [CI], 0.911 to 0.939) in the training cohort. In the external validation cohort, AUROC was 0.904 (95% CI, 0.876 to 0.930) and sensitivity, specificity, and accuracy were 86.3%, 74.2%, and 75.1%, respectively. In a 100-patient reader substudy, AI auscultation outperformed cardiologist consensus in accuracy (82.0% vs. 62.0%; P=0.004). Model performance was consistent across prespecified subgroups. The simplified two-site auscultation model retained high accuracy in the external validation cohort (AUROC, 0.903; 95% CI, 0.878 to 0.927). Conclusions: AI auscultation accurately detected LVD among patients referred for clinically indicated echocardiography, supporting its potential as a scalable and accessible pre-echocardiographic risk-enrichment tool. Further validation is warranted before broader screening applications. (Funded by AMI Inc. and the Japan Society for the Promotion of Science KAKENHI program [grant number 25K19371].).

★ A deep learning model for detection and staging of osteonecrosis of the femoral head on x-ray images: a retrospective, large-scale, multicentre study in China.

Wei Z, Hu S, Sun H, Tan J, Nommi Kemajou N, Zhou R — Tue, 27 Oct 2026 00:00:00 +0000

95% · External validation 95%

Background: Accurate detection and staging of osteonecrosis of the femoral head (ONFH) are essential for effective treatment yet frequently require specialist expertise and advanced equipment. To address this requirement, we aimed to develop a deep learning-based diagnostic model for the automatic detection and staging of ONFH using x-ray images, and compared its performance with clinicians to allow them to improve their diagnostic performance. Methods: In this retrospective, large-scale, multicentre study, we developed our deep learning model, Osteonecrosis Artificial Intelligence Diagnostician (OsteoAID), which was based on the You Only Look Once version 8 (YOLOv8) detection network and trained on a multicentre x-ray dataset. The dataset, dated from Jan 1, 2015, to Dec 31, 2023, and obtained from 16 399 patients across six Chinese tertiary hospitals, included 20 722 anteroposterior hip x-ray images of 28 785 hips, including hips with ONFH, healthy hips, and hips with other conditions radiographically similar to ONFH. To establish a rigorous ground truth for model training and evaluation, expert orthopaedists annotated each image to differentiate ONFH from confounding diseases and additionally classified the confirmed ONFH lesions on the basis of the Ficat and Arlet staging system. To verify the diagnostic performance of OsteoAID, we compared the model's accuracy, sensitivity, specificity, F1 score, and area under the curve (AUC) with those of the original YOLOv8 model on the internal testing set for both the staging and detection tasks. Furthermore, the performance of OsteoAID was compared with the performance of radiologists and orthopaedists on two external testing sets, and its diagnostic performance was also benchmarked against MRI-the gold standard for ONFH diagnosis-with a third external testing set in which x-ray images were staged by corresponding MRI scans. Diagnostic metrics were compared with those of clinicians using t tests and McNemar's tests, with significance set at p<0·05. Findings: On the internal testing set, OsteoAID improved the staging accuracy from 87·42% (95% CI 85·88-89·20) to 96·44% (94·73-97·17) and the AUC from 88·54% (86·74-90·37) to 96·94% (95·75-97·98) compared with the original YOLOv8 model. For detection the model had a mean average precision of 92·97% (91·73-94·43) at an intersection-over-union threshold of 50%. In comparisons with clinicians on the external testing set I, the model showed staging accuracy of 96·60% (94·27-97·81), which was significantly higher than radiologists (74·96% [73·38-76·85]; p<0·0001) and junior orthopaedists (83·98% [81·27-84·06]; p<0·0001) and similar to that of expert orthopaedists (96·54% [94·18-97·79]; p=0·20). Furthermore, model-assisted diagnostics improved the accuracy of junior orthopaedists by an absolute difference of 12·70% and of radiologists by 21·47%. Although the accuracy improvement for expert orthopaedists was non-significant (an absolute difference of 0·74%; p=0·080), the diagnostic time was substantially reduced across all groups. On external testing set II (which included challenging cases: low-quality x-ray images affected by improper exposure, poor contrast, incomplete field of view, or foreign-object shadows, as well as smartphone photographs of x-ray films), the model showed robust performance even under extreme conditions, achieving an overall staging accuracy of 93·76% (93·09-95·82). Additionally, combining x-ray imaging with the deep learning model produced diagnostic results similar to MRI on external testing set III with an overall accuracy of 96·63% (95·68-97·86) and an AUC of 97·33% (96·65-98·75). Interpretation: The developed deep learning model shows expert-level diagnostic performance in the dual tasks of detecting and staging ONFH by use of anteroposterior x-rays, which can substantially enhance both the diagnostic proficiency and efficiency of clinicians. Consequently, the model could serve as a highly accessible decision-support tool to optimise clinical triage and guide appropriate treatment strategies. Funding: National Natural Science Foundation of China, National High Level Hospital Clinical Research Funding, and Beijing Natural Science Foundation.

★ Closing the Automation Gap in HPV-Based Cervical Cancer Screening: Independent External Validation of an AI Model for Dual-Stain Triage.

Lahrmann B, Keil A, Miranda Ruiz F, Clarke MA, Egemen D, Grewal KK — Tue, 27 Oct 2026 00:00:00 +0000

95% · External validation 95%, Prospective 54%

Background: Primary human papillomavirus (HPV) screening has enabled scalable molecular detection in cervical cancer prevention, but downstream triage of HPV-positive individuals remains dependent on manual p16/Ki-67 dual-stain (DS) cytology interpretation. Although DS is guideline-endorsed and biologically well-defined, scalable implementation is limited by reader-dependent variability and lack of standardized automation. Methods: We developed Cytoreader-Fusion, a transparent, ensemble-based AI system for automated slide-level DS classification within HPV-based screening workflows that produces well-calibrated probabilistic outputs. The model was trained on 1919 cases and evaluated under blinded honest-broker oversight on 9006 cases across four independent cohorts, including a fully independent post-lock external validation cohort (EXVC; n=4996). All preprocessing steps, aggregation rules, and decision parameters were prespecified and irreversibly locked prior to validation. Performance was evaluated against independently ascertained histologically confirmed cervical intraepithelial neoplasia grade 3 or worse (CIN 3+) (or anal intraepithelial neoplasia grade 3 or worse [AIN 3+]) end points rather than cytologic concordance. A single unified classifier was applied across both ThinPrep and SurePath preparations without platform-specific retraining. Results: Cytoreader-Fusion demonstrated stable and reproducible slide-level classification across imaging variability, preparation platforms, and clinical settings. Across blinded screening and referral cohorts, sensitivity and specificity were higher than atypical squamous cells of undetermined significance or worse (ASC-US+) Papanicolaou (Pap) cytology in HPV-based screening populations and comparable to or exceeded manual DS interpretation. In the fully independent post-lock cohort (n=4996), the locked system achieved a sensitivity of 88.3% (90% UI, 85.3%-91.3%) and a specificity of 54.4% (90% UI, 52.9%-55.9%) for detection of CIN 3+, without any model retraining, threshold tuning, or configuration adjustment after lock. Operating characteristics remained consistent across cohorts and preparation platforms. Conclusions: This study establishes a high-stringency validation framework for clinically deployable artificial intelligence in population-based cancer screening. By combining prespecified algorithm locking, independent blinded evaluation under honest-broker supervision, end point-based validation against histologically confirmed disease, and cross-platform generalization, Cytoreader-Fusion demonstrates robust and scalable automation of DS triage within HPV-based screening programs. (Funded by the U.S. National Cancer Institute and others.).

★ Machine Learning-Based Prediction of Long-Term Mortality in Patients with Implantable Cardioverter Defibrillators.

Pan L, Xiao Z, Yu Z, Li S, Zhang W, Li R — Tue, 27 Oct 2026 00:00:00 +0000

95% · External validation 95%

Aims: Accurate prediction of the clinical course in implantable cardioverter-defibrillator (ICD) and cardiac resynchronization therapy-defibrillator (CRT-D) patients is crucial, as their disease progression varies significantly due to comorbidities and individual characteristics. Methods: This retrospective study included 3321 patients from four hospitals in China, with 3207 for model development and internal validation and 114 for external testing. Five machine learning (ML) models were evaluated using the C-index, time-dependent AUROC, and Brier score. Decision curve analysis (DCA) assessed clinical utility, while Kaplan-Meier curves stratified patients into high-, intermediate-, and low-risk groups. SHAP analysis identified key predictors and explained model outputs. Results: The Random Survival Forest (RSF) model demonstrated favorable numerical performance across multiple evaluation metrics, achieving the highest time-dependent AUROC (0.74, 0.74) and C-index (0.71, 0.70) in internal and external test sets. It also had the lowest time-dependent Brier score (0.17) in the internal test and ranked second (0.13) in the external test. DCA demonstrated numerically higher net benefit for the RSF model across evaluated threshold probabilities. Kaplan-Meier curves revealed distinct survival trajectories, with high-risk patients showing significantly higher mortality than low-risk patients (HR: 4.50, 95% CI: 3.10-6.55, p < 0.0001), while intermediate-risk patients also had increased mortality (HR: 2.49, 95% CI: 1.69-3.68, p < 0.0001). LVEDD was the most influential predictor of mortality. Conclusions: The RSF survival model enables individualized, time-resolved prediction of all-cause mortality risk in ICD and CRT-D patients, supporting risk stratification and future evaluation of clinical utility, while requiring cautious interpretation of longer-term predictions due to limited numbers at risk.

TARGET-AI: A Foundational Approach for the Targeted Deployment of Artificial Intelligence Electrocardiography in the Electronic Health Record.

Oikonomou EK, Batinica B, Dhingra LS, Aminorroaya A, Coppi A, Khera R — Tue, 27 Oct 2026 00:00:00 +0000

94% · External validation 94%, Prospective 35%

Background: Artificial intelligence (AI) applied to electrocardiograms (ECGs) offers promise for structural heart disease (SHD) screening, yet clinical integration remains challenged by high false-positive rates and a lack of tailored deployment strategies. Methods: We developed TARGET-AI, a multimodal approach that integrates longitudinal electronic health records (EHRs) with ECGs to define patient phenotypes and support targeted SHD screening opportunities. TARGET-AI uses an EHR foundation model applied to 118 million events from 159,322 individuals, to generate temporal patient-level embeddings and identify screening candidates, and a contrastive vision-language model trained on 754,533 ECG image-echocardiogram report pairs, to detect distinct SHD subtypes with tunable characteristics. We evaluated this in 5198 individuals referred for a transthoracic echocardiogram (TTE) within 90 days of an ECG (temporal validation) and in geographically distinct cohorts, including 33,518 UK Biobank participants undergoing protocolized ECG and cardiac magnetic resonance imaging, and an inpatient cohort of 3628 patients from the Medical Information Mart for Intensive Care (MIMIC)-IV database with ECG-TTE pairs. We compared discrimination metrics between targeted and untargeted strategies, with bootstrap-derived 95% confidence intervals excluding zero considered significant. Results: The TARGET-AI ECG model discriminated 26 SHD subtypes, including left ventricular systolic dysfunction (area under the receiver operating curve [AUROC] of 0.90), severe aortic stenosis (AUROC of 0.85), and elevated right ventricular systolic pressure (AUROC of 0.82). Compared with untargeted screening, targeted screening in the temporal validation set (n=5198) was associated with a significant increase in F1 scores (median: 0.25; range: 0.09 to 0.75) and a decrease in the number of false positives (median: -303; range: -715 to -77) across 26 SHD labels. Similar results were seen in the UK Biobank (n=33,518; median change in false positives: -819 [range: -3521 to -459] across seven SHD labels) and MIMIC-IV (n=3628; median false-positive change: -255 [range: -716 to -86] across five SHD labels) participants. Conclusions: TARGET-AI may guide the targeted deployment of AI-ECG for SHD screening across health systems. (Funded by the National Heart, Lung, and Blood Institute of the National Institutes of Health and others.).

International Retrospective Observational Study of Continual Learning for AI on Endotracheal Tube Placement from Chest Radiographs.

Chen E, Saenz A, Banerjee O, Marklund H, Zhang X, Johri S — Tue, 27 Oct 2026 00:00:00 +0000

91% · External validation 91%

Background: Medical artificial intelligence (AI) models often underperform when deployed at new hospitals despite strong performance during development, creating a need for effective adaptation strategies that maintain institutional privacy. Continual learning (i.e., repeatedly retraining a model at each new hospital where it is deployed) can improve performance during and after deployment. Methods: To determine whether continual learning could improve model performance across diverse clinical settings without requiring data exchange between hospitals or complex computational setups, we performed a retrospective study of a convolutional neural network model for endotracheal tube (ETT) assessment, comparing three deployment approaches (original model deployment, hospital-specific fine-tuning, and continual learning across hospitals). The study included 2313 intensive care unit chest radiographs across 23 hospitals spanning 12 countries and 5 continents, with 1 hospital held out for external validation. Participants were adults (≥18 years of age) with ETTs placed in 2021 (or earlier if a hospital found too few cases from 2021), identified through standardized key word searches in hospital systems and verified by radiology reports where available. Each hospital contributed approximately 100 anteroposterior radiographs, with 50 used for training, as well as associated deidentified reports and metadata. Each image was preprocessed by the contributing hospital to remove identifiable features and was independently reviewed and annotated by two radiologists. Results: Continual learning models achieved a significantly lower average ETT-carina error of 10.58 mm (standard deviation [±SD], 12.98 mm), compared with 12.49 mm (SD±, 15.37 mm) for models fine-tuned only at their deployment hospital and 16.39 mm (±SD, 21.56 mm) for the original model. In addition, continual learning models outperformed the original model at all hospitals, and they outperformed models fine-tuned only at their deployment hospital at 21 of 22 hospitals. Conclusions: Despite using only 50 training data points from each hospital, continual learning surpassed traditional fine-tuning as a method for improving medical AI model generalization.

Silent Validation of a Longitudinal Model for Predicting Delayed Cerebral Ischemia in Real Time after Subarachnoid Hemorrhage.

Willms J, Schmid J, Inauen C, Bögli SY, Bienefeld-Seall N, Beck T — Tue, 27 Oct 2026 00:00:00 +0000

90% · Prospective 90%

Background: Longitudinal prediction of delayed cerebral ischemia (DCI) following subarachnoid hemorrhage (SAH) is crucial for timely interventions. Machine learning models for DCI risk prediction and stratification may support clinical decision-making, but safety concerns and regulatory constraints have precluded real-time bedside validation. We utilized the ICU (intensive care unit) Cockpit information technology infrastructure to establish a live data stream enabling real-time model predictions that were evaluated by independent research physicians in a silent validation framework, benchmarking algorithmic recommendations without influencing treatment decisions. Methods: We deployed a DCI algorithm validated on historical data in a live environment on running data streams of actual patients. Three independent physicians not involved in patient care reviewed real-time model outputs from day 2 through day 14 after SAH onset. With algorithm support, these physicians estimated DCI probability, decided on further diagnostic orders (computed tomography angiography with perfusion) and/or recommended the intervals for repetitive neurologic assessments (increased, 1-2 hourly; standard, ≥3 hourly). Researchers' decisions were compared with actual clinical decisions and the occurrence of DCI. Results: Out of 63 aneurysmal SAH patients enrolled, 17 developed DCI (27%). A total of 794 assessments were made. The number of increased neurologic assessment orders was 466 with algorithm support and 629 without (a reduction of 163; 25.91%). Although the number of true positives and false negatives for DCI prediction was similar with and without the algorithm, false positives were reduced by 164 (P<0.01). Additional diagnostic orders were 128 with algorithm support and 153 without (reduction of 25; 16.34%). Conclusions: This study demonstrates that continuous, real-time risk estimates can be generated without disrupting standard care, and that clinicians - when supported by an algorithm - may safely reduce unnecessary investigations without compromising DCI event detection. In addition, silent validation showed potential to optimize monitoring strategies and reduce clinician workload without reducing diagnostic accuracy. (Funded by the Swiss National Science Foundation [project nos. 407540-167195 and 325230_200568].).

★ Machine Learning Models Using Electronic Health Record Data to Predict Obstructive Sleep Apnea.

Hwang NH, Mazzotti DR, Mishra S, Westover MB, Hwang D, Thomas RJ — Mon, 26 Oct 2026 00:00:00 +0000

98% · External validation 98%

Rationale: Obstructive sleep apnea (OSA) is highly prevalent yet largely under-diagnosed. Current screening strategies rely on effortful questionnaires with modest accuracy, and prior machine learning models often require resource-intensive inputs and lack external validation. Objectives: To develop and validate machine learning models to predict OSA (apnea-hypopnea index [AHI4%]≥5) and moderate-severe OSA (AHI4%≥15) using only routinely collected electronic health record (EHR) data, test cross-system transferability and compare performance with clinical screeners and a Bayesian risk-updating framework. Methods: Three tree-ensemble families (XGBoost, random forests, and histogram-based gradient boosting) were trained on an adult cohort with sleep-study results at Kaiser Permanente (KPSC), using full (106-feature) and minimal (age, sex, body-mass index, race/ethnicity) predictor sets and tuned by nested cross-validation. XGBoost models were selected, internally and temporally validated, and externally validated at University of Kansas Medical Center (KUMC). Measurements And Main Results: On our internal hold-out set (n=53,048), the full XGBoost model reached a ROC-AUC of 0.791 for any OSA and 0.768 for moderate-severe OSA. The minimal model retained most performance (0.764 and 0.736). Temporal performance was stable. Externally, frozen models showed lower AUCs (0.718 and 0.690 for full models; 0.725 and 0.706 for minimal models) but similar to KUMC-native models and improved further with transfer learning. The models significantly outperformed STOP-BANG (AUC difference 0.06-0.09 for predicting any OSA; p<0.001). Bayesian updating with select patient-reported items further improved discrimination. Conclusions: Machine learning models based on EHR data can predict OSA probability that outperform existing screeners, transfers across systems, and offers potential for EHR implementation. At A Glance Commentary: Scientific Knowledge on the Subject: Obstructive sleep apnea is common and heavily underdiagnosed, yet upstream patient identification still depends on symptom and anthropometry questionnaires such as STOP-BANG, which are effort-dependent and achieve modest performance. Prior machine learning models for OSA are heterogeneous, often developed on small or demographically narrow cohorts, frequently rely on physiologic signals that reintroduce a testing step, and are rarely validated across independent health systems.What This Study Adds to the Field: We developed machine learning models from routinely collected electronic health record data, comprising a full (106-feature) and a minimal (4-feature) model, to predict any OSA and moderate-severe OSA. The study is distinguished by multiple validation efforts, including temporal validation and external validation at a second health system. We further explored adaptation strategies for external adoption using transfer learning techniques. Finally, the models outperformed existing screeners such as STOP-BANG, and we describe a Bayesian framework that optionally incorporates select patient-reported symptoms to refine individual risk estimates. Overall, these models are feasible to implement within EHR systems as a low-effort, passive screener that can support encounter-based and population health outreach screening strategies.

★ Incorporating visceral adipose tissue into dual-region deep learning radiomics on CT improves preoperative lymph node metastasis prediction in gastric cancer: A multicenter study.

Miao S, Xiong Z, Yang H, Zhang L, Qi H, Liu Z — Mon, 26 Oct 2026 00:00:00 +0000

97% · External validation 97%

Background: Visceral adipose tissue (VAT) is associated with tumor progression in gastric cancer (GC), but its added value for predicting lymph node metastasis (LNM) remains unclear. Purpose: This study aimed to construct a DR-DLR model that incorporates CT-based features from both VAT and tumor, with the goal of predicting LNM in GC patients, while investigating its biological relevance. Methods: 855 GC patients were retrospectively collected across three clinical centers, divided into a primary cohort, one internal test cohort, and two independent external validation cohorts. An Attention Residual Transformation Network (ARTNet) extracted imaging representations from CT images. Feature representations and clinical variables were integrated using a multilayer perceptron (MLP) to build the DR-DLR model. Model performance, generalizability, prognostic value, and biological relevance were evaluated using internal/external testing, radiologist comparison, survival analysis, and radiogenomics analysis. Results: Incorporation of VAT-derived features improved predictive performance compared with tumor-only models. The DR-DLR model achieved areas under the receiver operating characteristic curve (AUC) of 0.898 (95 % CI: 0.817-0.946), 0.834 (95 % CI: 0.743-0.897), and 0.848 (95 % CI: 0.775-0.901) in the internal and two external test cohorts, respectively. It outperformed junior and senior radiologists in LNM assessment. Overall survival was shorter in high-risk patients than in low-risk. Radiogenomics analysis revealed that the high-risk phenotype was associated with activation of immune-related pathways and increased infiltration of CD4 + T cells, plasma cells, dendritic cells, and mast cells. Conclusion: The DR-DLR model improved LNM prediction, with its high-risk phenotype linked to immune-related molecular alterations, supporting its value for preoperative risk stratification.

★ Head-to-head comparison of glucose-to-albumin ratio against five albumin-derived indices for predicting 28-day ICU mortality in critically ill patients with spontaneous intracerebral hemorrhage: a multicenter retrospective study.

Hao F, Huangfu L, Li W, Li Y, Hui R, Sun C — Mon, 26 Oct 2026 00:00:00 +0000

97% · External validation 97%

Background: Spontaneous intracerebral hemorrhage (ICH) carries a 30-day mortality exceeding 40%. The glucose-to-albumin ratio (GAR) reflects metabolic stress and nutritional status. Although previous studies have demonstrated its prognostic value in ICH, its utility specifically in ICU-admitted patients and its comparative performance against other albumin-derived indices remain incompletely characterized. Methods: We included 1,625 ICU ICH patients from MIMIC-IV (derivation) and 861 from two Chinese centers (external validation). Log-transformed GAR (LogGAR) was the exposure; 28-day ICU mortality was the primary outcome. Cox regression, restricted cubic splines, Kaplan-Meier, and Fine-Gray competing-risk models assessed LogGAR's independent prognostic value, compared head-to-head with five albumin-derived indices. Lasso-selected features informed a multilayer perceptron (MLP), benchmarked against severity scores and a penalized logistic regression on identical features. Results: LogGAR was independently associated with 28-day ICU mortality (derivation: HR = 1.81, 95% CI 1.33-2.47, p < 0.001; external: HR = 3.48, 95% CI 2.54-4.78, p < 0.001), with linear dose-response (RCS P-nonlinear = 0.315). Fine-Gray competing-risk analysis confirmed robustness (subdistribution HR = 2.01, 95% CI 1.58-2.57, p < 0.001). Adding LogGAR to six severity scores improved discrimination (external ΔAUC = 0.085-0.107, all p < 0.001). Among six indices, LogGAR was comparable to logLDAR in the derivation cohort (0.679 vs. 0.681, p = 0.78) and highest in the external cohort (0.691). Subgroup analyses identified significant effect modification only for AKI (FDR = 0.029). The Lasso-MLP (8 features) and penalized logistic regression achieved comparable performance (validation AUC 0.718 for both; external AUC 0.759 vs. 0.783), indicating no added value from nonlinearity. Calibration slopes were 0.66 (validation) and 0.91 (external). SHAP identified LogGAR as the top predictor. Conclusion: LogGAR is independently associated with 28-day ICU mortality in critically ill patients with ICH and shows discriminative performance comparable to logLDAR and superior to the remaining four indices. A penalized logistic regression on Lasso-selected features matched the MLP, suggesting the predictive signal lies in features rather than model complexity-a pragmatic consideration for clinical implementation. As a readily available biomarker, LogGAR may support early ICU risk stratification.

★ Radiomics-based interpretable machine learning model for identifying high-risk patients with small bowel obstruction: a multicenter study.

Zhang G, Liu X, Huang W, Zhu X, Xia C, Xia H — Mon, 26 Oct 2026 00:00:00 +0000

97% · External validation 97%

Background: Early identification of high-risk small bowel obstruction remains clinically challenging, especially when prompt intervention is needed to prevent bowel ischemia and related adverse outcomes. Methods: In this multicenter retrospective study, 274 patients with small bowel obstruction from three hospitals were included. Patients from Hospital A were randomly divided into training and test cohorts, and patients from the other two hospitals formed an external validation cohort. Radiomics features were extracted from venous-phase contrast-enhanced CT images, and logistic regression, naive Bayes, Extra Trees, Light Gradient Boosting Machine (LightGBM), and multilayer perceptron models were developed. Independent clinical predictors were identified and integrated with radiomics features to build a combined model. Results: LightGBM showed the most stable internal performance, with AUC values of 0.934 and 0.870 in the training and test cohorts, respectively. Lower diastolic blood pressure and bowel wall thickening were independent predictors of high-risk status. The combined model achieved AUC values of 0.942, 0.867, and 0.803 in the training, test, and external validation cohorts, respectively, with an external validation accuracy of 0.833 and sensitivity of 0.946. Conclusion: CT-based radiomics can identify patients with small bowel obstruction at high risk, and a combined model integrating radiomics and clinical predictors may support earlier risk stratification and intervention.

★ Integrating Magnetic Resonance Imaging-Based Deep Learning and Radiomics to Predict Treatment Response in Hepatocellular Carcinoma Receiving Combined Systemic and Locoregional Therapies: A Multicenter Study.

Wang LN, Zhou SQ, Li F, Wu LF, Xu L, Wang K — Mon, 26 Oct 2026 00:00:00 +0000

96% · External validation 96%

Background: Although combining systemic and locoregional therapies has improved outcomes for hepatocellular carcinoma (HCC), identifying appropriate candidates remains challenging because no reliable tool predicts response. Methods: This retrospective multicenter study enrolled 334 HCC patients who received first-line locoregional-systemic therapy from nine centers in China. Following automated tumor segmentation using nnU-Net on pretreatment contrast-enhanced magnetic resonance imaging, predefined radiomics features were computationally extracted, and deep learning features were obtained via a ResNet convolutional neural network. Three types of predictive models for tumor response (radiomics, deep learning, and combined) were developed and validated on the training cohort (n = 213), internal validation cohort (IVC, n = 53), and external validation cohort (EVC, n = 68). Performance was assessed using area under the receiver operating characteristic curve (AUC) and net reclassification improvement (NRI). Associations of model with progression-free survival (PFS) and overall survival (OS) were evaluated. Results: Among the three models (multi-sequence radiomics, multi-sequence deep learning, and combined), the combined model showed the best predictive performance in both the training (AUC: 0.997, 95% CI: 0.994-1.000) and validation cohorts (AUC of IVC 0.900, 95% CI: 0.806-0.994; AUC of EVC: 0.871, 95% CI: 0.790-0.951). The combined model demonstrated significantly improved predictive performance compared with the multi-sequence radiomic model (training NRI = 43.9%, p < 0.001; IVC NRI = 38.7%, p = 0.038; EVC NRI = 3.2%, p = 0.806) and multi-sequence deep learning model (training NRI = 80.8%, IVC NRI = 72.7%, EVC NRI = 50.4%; all p < 0.05). Specifically, the combined model was significantly associated with both PFS and OS. Conclusions: The combined radiomics-deep learning model outperformed both single-modality models in predicting objective response and stratifying survival outcomes in unresectable HCC patients receiving combined systemic and locoregional therapy.

★ Deep learning multimodal fusion model based on ultrasound images and clinical-ultrasound texts for differentiating benign and malignant renal tumors.

Cao H, Shi Y, Yu S, Cao X, Li C, Mou L — Mon, 26 Oct 2026 00:00:00 +0000

96% · Prospective 96%

Objectives: To develop a deep learning multimodal fusion model based on ultrasound (US) images and clinical-US texts to differentiate benign and malignant renal tumors. Methods: The retrospective dataset from January 2012 to February 2023 in a tertiary hospital was used for model development and randomly allocated to training and validation sets. An independent prospective test set was obtained from March 2023 to October 2024. We developed Model-G using grayscale US (GUS) images, Model-GC using GUS and color Doppler flow imaging images, Model-T using clinical-US texts, and Model-UT using US images and clinical-US texts. ResNet50 was used for image feature extraction, BERT for text encoding, and ViT for multimodal classification. Results: In total, 1008 patients (740 in training set, 186 validation set, and 82 prospective test set) with a single renal tumor were included. In the validation set, Model-UT achieved the highest diagnostic performance in comparison to Model-G, Model-GC, and Model-T with an area under the receiver operating characteristic curve (AUC) of 0.89 (95% CI, 0.86-0.93) (all p-values < 0.05). In the prospective test set, Model-UT yielded an AUC of 0.92 (95% CI, 0.86-0.97), outperforming high-level, mid-level, and low-level radiologists (all p-values < 0.05). Conclusions: The multimodal fusion model presents promising potential in differentiating benign and malignant renal tumors. Advances In Knowledge: Deep learning model integrating US images with clinical-US texts improved benign-versus-malignant renal tumor differentiation. Decision-curve, controlled ablation, and SHAP analyses support its potential use as a clinical decision-support tool.

★ Estimating Individualized Survival Benefit of Consolidative Metastasis-Directed Radiotherapy Following Chemoimmunotherapy in Nasopharyngeal Carcinoma With Exclusive Distant Metastasis.

Wang X, Chen S, Zhang X, Mao L, Chen J, Yan S — Mon, 26 Oct 2026 00:00:00 +0000

96% · External validation 96%

Optimal patient selection for consolidative metastasis-directed radiotherapy (MDRT) after chemoimmunotherapy (CIT) remains uncertain in nasopharyngeal carcinoma (NPC) with exclusive distant metastasis, defined as distant metastasis without synchronous locoregional recurrence. This study was aimed at estimating the survival benefit of MDRT and identifying patients most likely to benefit. A total of 1083 patients from two independent centers were analyzed using a target trial emulation with cloning, artificial censoring, and inverse probability of censoring weighting. The emulation estimated a 4.2-month gain in 3-year progression-free survival (PFS) restricted mean survival time with the MDRT strategy (95% CI: 2.4-5.9; bootstrap p < 0.001). Eight survival machine-learning algorithms were developed and compared, among which the gradient boosting machine achieved 3-year PFS AUROCs of 0.875 in internal validation and 0.877 in geographical external validation. Heterogeneous treatment effects were further estimated using a T-learner framework across risk groups defined by recursive partitioning analysis. In the external cohort, model-estimated 3-year absolute risk reductions with MDRT were 21.6%, 14.7%, and 5.7% in intermediate-, low-, and high-risk groups, respectively. The framework was subsequently implemented as a web-based clinical decision-support system. These findings support individualized selection of consolidative MDRT after CIT and provide a practical framework for multidisciplinary decision-making.

★ Development and Validation of a Machine Learning-Based Model for Elevated Depressive Symptom Risk in COPD Patients: A Multi-Cohort Cross-Sectional Study.

Cheng Z, Wu J, Yan X — Mon, 26 Oct 2026 00:00:00 +0000

96% · External validation 96%

Background: Depression is a common yet underrecognized comorbidity in chronic obstructive pulmonary disease (COPD). Existing prediction models often lack external validation, limiting cross-population generalizability. We aimed to develop and externally test machine learning models for predicting concurrent elevated depressive symptoms in COPD using three international aging cohorts. Methods: We included COPD patients from the Health and Retirement Study (HRS, USA, n=1,899) and the China Health and Retirement Longitudinal Study (CHARLS, China, n=2,206) for model development, and the English Longitudinal Study of Ageing (ELSA, UK, n=538) for external testing. Elevated depressive symptoms were defined using harmonized cutoffs (CES-D‑8≥3 for HRS/ELSA; CES-D‑10≥10 for CHARLS). Following recursive feature elimination, seven machine learning algorithms were trained and assessed for discrimination, calibration, and clinical utility. SHapley Additive exPlanations (SHAP) analysis was employed for model interpretability. Results: Eight variables were used to build the seven models. The XGBoost model achieved the best internal performance (training AUC 0.824, 95% CI: 0.809-0.838; validation AUC 0.757, 95% CI: 0.731-0.783) and maintained robust external validity in ELSA (test AUC 0.785, 95% CI: 0.748-0.824). Calibration and decision curve analyses supported clinical utility. SHAP analysis identified functional dependency, self‑rated health, and life satisfaction as predominant predictors. Higher dependency and poorer self-rated health had positive SHAP contributions, indicating higher model-predicted risk, whereas higher life satisfaction had negative contributions, indicating lower model-predicted risk. BMI exhibited a U‑shaped association. Conclusion: The XGBoost model generalized to an independent cohort and was deployed as a research prototype web application for risk stratification. SHAP analysis identified important model contributors for risk stratification in COPD, but these explanations describe model predictions and do not establish causal intervention targets. Prospective studies are needed to evaluate clinical utility.

★ Development and external validation of an explainable machine learning model for identifying depressive symptoms in middle-aged and older adults with low haemoglobin: a prediction-model study.

Wang Q, He J, Ma S, Zhao J, Zong J, Ma D — Mon, 26 Oct 2026 00:00:00 +0000

96% · External validation 96%

Background: Early screening for depressive symptoms in middle-aged and older adults is important for timely risk identification. Low haemoglobin may identify a clinically accessible vulnerable subgroup in which chronic disease burden, fatigue, functional impairment, and reduced physiological reserve often coexist. However, it should be understood as a high-risk phenotype rather than a direct measure of nutritional deficiency or frailty. Evidence in this population remains limited. Methods: This study developed a model using data from the 2015 China Health and Retirement Longitudinal Study (CHARLS), with CHARLS 2011 and a cross-sectional cohort from Gansu Provincial People's Hospital used for temporal and external validation, respectively. Depressive symptoms were defined as a 10-item Centre for Epidemiologic Studies Depression Scale score ≥10. Candidate variables included sociodemographic, behavioural, health status, functional, cognitive, and comorbidity-related information. After least absolute shrinkage and selection operator feature selection, 16 variables were retained. Seven machine learning models were compared, and Random Forest was retained for interpretation and online tool development based on its balanced performance and interpretability. Results: A total of 2,022 participants were included in the development cohort, 1,095 in the temporal validation cohort, and 400 in the external validation cohort. The AUROC values of the Random Forest model were 0.791, 0.776, and 0.809, respectively, with Brier scores of 0.182, 0.193, and 0.182, indicating stable discrimination and acceptable calibration. Shapley additive explanations analysis showed that sleep time, instrumental activities of daily living score, and activities of daily living score were the most important contributors. Threshold sensitivity analysis showed that lower probability thresholds increased sensitivity/recall but reduced specificity. A prototype online assessment tool was developed for individualised estimation of depressive symptom probability. Conclusion: This study developed and validated an interpretable machine learning model for identifying depressive symptoms in middle-aged and older adults with low haemoglobin. The model may support early screening and risk stratification in this clinically vulnerable population. However, the classification threshold and risk-stratification cut-offs are exploratory rather than clinically validated. Further multicentre validation is warranted before broader clinical implementation.

★ Cross-cohort generalizability of machine learning models for early-stage node-negative breast cancer relapse prediction.

Ozdemir C, Gupta S — Mon, 26 Oct 2026 00:00:00 +0000

95% · External validation 95%

Background: Adjuvant therapy decision in early-stage node-negative breast cancer remains one of the most challenging problems in oncology. Although most patients in this population achieve favorable outcomes with endocrine therapy alone, a subset remains at risk of relapse and may benefit from chemotherapy. Accurate identification of high-risk and low-risk patients is therefore essential to prevent both undertreatment and overtreatment. Methods: In this study, machine learning models were assessed using MSK-IMPACT (n=645) and METABRIC (n=503) cohorts comprising node-negative stage I-II breast cancer patients. Three experimental settings were evaluated: training on MSK-IMPACT and testing on METABRIC, the reverse direction, and five-fold cross-validation on the combined cohort. Sensitivity analysis and data harmonization were applied to identify and mitigate cohort-specific bias. Survival analysis, SHapley Additive exPlanations (SHAP) interpretability, and ablation analyses were performed to validate clinical relevance and characterize feature contributions. Results: The combined cohort model achieved an area under the receiver operating characteristic curve (AUC) of 0.788 ± 0.028 across five-fold cross-validation, while bidirectional cross-cohort validation yielded AUC of 0.741 when testing on MSK-IMPACT and AUC of 0.714 when testing on METABRIC. Survival validation on METABRIC out-of-fold predictions from the combined cohort model showed hazard ratios of 3.01 (95% CI: 2.28-3.98) for disease-free survival and 2.35 (95% CI: 1.79-3.09) for overall survival, consistent across Luminal A, Luminal B, and Triple-Negative subtypes. Conclusion: Validated across two geographically independent cohorts, the proposed model enables cost-effective, gene expression-free identification of patients at high-risk of distant relapse and may facilitate adjuvant chemotherapy decision-making in early-stage node-negative breast cancer.

A data-driven regional amyloid PET score predicts cognitive decline beyond Centiloid.

Hirose T, Akamatsu W, Kato T, Alzheimer’s Disease Neuroimaging Initiative — Mon, 26 Oct 2026 00:00:00 +0000

93% · External validation 93%

Background: The Centiloid (CL) scale standardizes global amyloid PET quantification and is widely used to define amyloid positivity. As a global summary measure, however, CL may not fully reflect the regional distribution of amyloid deposition, which can carry additional prognostic information about the rate of cognitive decline. Objective: To develop and externally validate a fixed, regional amyloid PET composite score that complements CL for predicting cognitive decline in Alzheimer's disease. Methods: The Regional Amyloid PET Score (RAPS) was derived from 82 FreeSurfer regions using machine learning with bootstrap stability selection to predict the rate of change in CDR-Sum of Boxes (CDR-SB) in 433 amyloid-positive ADNI [ 18 F]florbetapir participants. The fixed nine-region weights were applied without retraining in a cross-tracer ADNI [ 18 F]florbetaben subset ( N = 71; largely overlapping the discovery participants) and two external validation cohorts, NACC SCAN ( N = 1531; four tracers) and OASIS-3 ( N = 428). Results: RAPS comprised nine regions. In ADNI, RAPS correlated more strongly with CDR-SB slope than CL and showed higher discrimination of rapid decliners (AUC 0.813 vs 0.713). Performance was directionally consistent across validation cohorts; in NACC SCAN, RAPS and CL independently predicted clinical progression. Cross-cohort meta-analysis of the three independent cohorts supported incremental discrimination beyond CL (pooled ΔAUC +0.066; I 2 = 0%). Conclusions: RAPS, a fixed regional amyloid PET-derived score, may complement CL for prognostic stratification in Alzheimer's disease research.

Development of a large language model clinical variable extraction pipeline for the study of lymphedema.

Fisher AT, Lakshmi A, Liu BC, Woo J, Gaus R, Chaudhari AS — Mon, 26 Oct 2026 00:00:00 +0000

93% · Language models 93%

For many disease processes, clinical information is often captured in unstructured clinical notes that require manual data extraction for research. Large language models (LLMs) have the potential to automate the data extraction process, enabling study for diseases with fewer tabular metrics available. Here, we report the development and validation of an LLM-based clinical note data extraction approach for a retrospective cohort study of lymphedema etiologies and treatments. After using International Classification of Diseases, 10th Revision codes to identify potential lymphedema patients and text-matching to find potentially relevant clinical notes, an LLM was used to further filter notes, summarize for relevant content, and extract per-patient variables concerning etiology, treatments and efficacies, clinical stage, and wound status. Prompts were optimized for accuracy with an iterative review strategy, starting with zero-shot approaches and modified to few-shot approaches where necessary, validated by subset human review. A Protected Health Information-compliant Application Programming Interface version of OpenAI's GPT-5 was used for all LLM queries. Human review of LLM-extracted final variables revealed high accuracy when compared against both an intermediate summarization step (99% accurate) and against original clinical notes (98% accurate) across 2858 patients/32,288 notes. LLMs are a valuable tool for conducting clinical research at scale using unstructured text data.

Zero-Shot Extraction of Structured Clinical Data from Free-Text CT Reports Using Foundational Large Language Models.

Gupta A, Hongslo A, Nguyen Q, Caldwell J, Choi B, Harvey CJ — Mon, 26 Oct 2026 00:00:00 +0000

91% · Language models 91%

Introduction: Large language models (LLMs) are being explored for multiple applications in medical research, including medical text classification. Authors of the study evaluated the performance of three off-the-shelf LLMs in classifying unstructured free-text CT pulmonary angiography reports for pulmonary embolism (PE)-related diagnostic labels. Methods: Authors assessed human-labeled CT reports using three LLMs (ChatGPT-4o, Llama 3.3 70B, and Llama 3.1 8B) with zero-shot simple and chain-of-thought (CoT) prompts. Classification tasks included acute PE (N = 1,025) and, among reports with PE (n = 660), bilateral PE and large PE. Results: LLMs demonstrated high classification performance for acute PE and bilateral PE but lower performance for large PE. With simple prompts, the highest F1 scores were 97.0% for acute PE (Llama 3.1 8B; sensitivity, 95.6%; specificity, 97.3%), 92.9% for bilateral PE (ChatGPT-4o; sensitivity, 98.0%; specificity, 88.6%), and 72.8% for large PE (ChatGPT-4o; sensitivity, 82.9%; specificity, 82.2%). CoT prompting did not meaningfully improve acute PE classification but provided slight improvements in subclassification performance for bilateral and large PE. Conclusions: Off-the-shelf LLMs can classify free-text CT reports using simple prompt engineering and may facilitate automated data extraction for medical research following appropriate evaluation of classification performance.

CardioCLIP: sequential multimodal prediction of echocardiographic abnormalities and natriuretic peptide elevation from electrocardiogram and chest radiographs.

Shibata S, Nishimori M, Nishihara Y, Maeda E, Iiyama M, Shinohara M — Mon, 26 Oct 2026 00:00:00 +0000

90% · External validation 90%

Aims: Diagnostic reasoning in clinical practice is inherently sequential and multimodal, with physicians integrating heterogeneous examinations under constraints such as missing data and variable test order. However, most multimodal machine learning models assume simultaneous availability of all modalities and fail to reflect this workflow. Methods And Results: We propose CardioCLIP, a sequential multimodal framework that embeds electrocardiograms (ECGs) and chest X-rays (CXRs) into a shared pathological representation space. The model consists of (i) a self-supervised embedding stage aligning ECG-CXR pairs from the same patient and day, and (2) a gated recurrent unit-based prediction stage that estimates clinical indicators as data are sequentially observed. The model was trained on 9458 patients from a single tertiary centre and evaluated for B-type natriuretic peptide (BNP), left ventricular ejection fraction (LVEF), and E/e'. External validation was conducted using MIMIC-IV.CardioCLIP achieved strong cross-modal alignment and improved discrimination for selected echocardiographic abnormalities and natriuretic peptide elevation (AUC: 0.90 for BNP > 100 pg/mL, 0.87 for LVEF < 50%, and 0.80 for E/e' > 15). It also showed discrimination for subsequent severe BNP elevation (BNP > 400 pg/mL within 3 years) with an AUC of 0.79. External validation in MIMIC-IV showed an AUC of 0.89 for N-terminal pro-BNP elevation, supporting potential transportability for this endpoint. Conclusion: CardioCLIP demonstrated the feasibility of order-flexible sequential multimodal prediction. By leveraging routinely acquired ECG and CXR data, it may provide a non-invasive adjunct for identifying echocardiographic abnormalities and natriuretic peptide elevation in cardiovascular care, subject to prospective clinical validation.

Habitat informed 3D deep learning predicts immunochemotherapy response in esophageal squamous cell carcinoma.

Li T, Zou T, Li Z, Liu Z, Chen J, Xiong Y — Mon, 26 Oct 2026 00:00:00 +0000

90% · External validation 90%

Immunochemotherapy (ICT) is a critical treatment for locally advanced esophageal squamous cell carcinoma (LA-ESCC), yet predicting individual response remains challenging. We developed a three-channel 3D deep learning framework integrating tumor habitat analysis for predicting ICT response. In this multicenter retrospective study of 277 patients, voxel-level clustering of pretreatment CT images segmented tumors into habitat subregions reflecting microenvironmental heterogeneity. A ResNet-50 model incorporating habitat maps, cropped tumor images, and ROI masks achieved an AUC of 0.927 (95% CI, 0.850-0.951) in the independent external test cohort, compared with 0.869 for habitat radiomics (DeLong p = 0.343). SHapley Additive exPlanations analysis identified texture features, particularly gray-level co-occurrence matrix correlation in habitat 3, as the primary contributors. This framework provides a noninvasive approach for predicting ICT response in LA-ESCC.

★ High-volume cervical lymph node metastasis in papillary thyroid carcinoma: development and external validation of an interpretable prediction model.

Wang D, Wang J, Wu Z, Cui Z — Fri, 23 Oct 2026 00:00:00 +0000

99% · External validation 99%

Background: The transportability of models for postoperative high-volume cervical nodal disease in papillary thyroid carcinoma (PTC) and the incremental predictive value of molecular variables remain uncertain. Methods: We developed models in the public Deyang cohort (521 patients; 90 events) and performed geographic external validation in the independent Liaocheng cohort (193 patients; 73 events). The outcome was >5 pathologically confirmed metastatic cervical lymph nodes. Five algorithms were compared using repeated stratified nested cross-validation with a prespecified 10-variable postoperative feature set. Ridge logistic regression (C = 0.1) was selected and applied without refitting. Paired comparisons of 8-variable clinicopathological and 10-variable clinicopathological-molecular models assessed the incremental value of high-risk co-mutation status and mutation count in both cohorts. Results: Internal validation yielded an area under the receiver operating characteristic curve (AUROC) of 0.795 (95% CI, 0.745-0.842), average precision (AP) of 0.497 (0.406-0.598), and Brier score of 0.115 (0.098-0.134). External AUROC was 0.752 (0.678-0.821), AP 0.601 (0.498-0.728), and Brier score 0.217 (0.180-0.254). Observed and mean predicted risks were 37.8% and 23.9%, respectively; the observed-to-expected ratio was 1.582 (1.319-1.860) and the weak-calibration slope was 0.831 (0.542-1.191). At the prespecified threshold of 0.158, sensitivity was 0.795 and specificity 0.600. Net benefit was 0.254 for the model and 0.261 for classifying all patients as high risk (difference, -0.008; 95% CI, -0.051 to 0.033). External paired differences after adding molecular variables were +0.002 for AUROC (-0.005 to 0.010), -0.002 for AP (-0.016 to 0.010), -0.001 for Brier score (-0.002 to 0.000), and +0.009 for net benefit (-0.003 to 0.026). Conclusion: The model retained moderate discrimination across centers but underestimated absolute risk and did not demonstrate a net benefit advantage over classifying all patients as high risk at the prespecified threshold. High-risk co-mutation status and mutation count showed no clear incremental predictive value. Further validation with standardized lymph-node assessment and patient outcomes is required before clinical application.

★ Machine Learning Model Using Clinical and Contrast-Enhanced MRI Features Differentiates Cerebral Tuberculoma from Lung Cancer Brain Metastasis.

Wang A, Dai W, Chen Q, Qiu X, Wang J, Hu Y — Fri, 23 Oct 2026 00:00:00 +0000

98% · External validation 98%

Rationale And Objectives: Cerebral tuberculomas (CTB) and lung cancer brain metastases (BM) frequently show overlapping appearances on conventional magnetic resonance imaging (MRI) but require entirely different treatment strategies. This study aimed to develop and validate a machine learning model integrating clinical and contrast-enhanced MRI features to noninvasively differentiate these two entities. Materials And Methods: The internal dataset (n = 188, 122 CTB, 66 BM) was split into training/validation (n = 131) and internal test (n = 57) sets. Feature selection (univariate and stepwise logistic regression) was performed only on the training and validation set, yielding four predictors. Machine learning classifiers were trained using these predictors with repeated 5-fold cross-validation. An external test set (n = 51, 26 CTB, 25 BM) from another hospital was used for independent validation. Results: Multivariate logistic regression retained four independent predictors: age (OR = 1.046, 95%CI: 1.003-1.091, p = 0.036), carcinoembryonic antigen (CEA) (odds ratios [OR] = 1.711, 95%confidence intervals [CI]: 1.235-2.370, p = 0.001), multiple lesions (OR = 0.080, 95%CI: 0.015-0.438, p = 0.004), and ring with mural nodule (OR = 11.289, 95%CI: 2.551-49.952, p = 0.001). The Gradient Boosting model built on these four features achieved the best performance, with an area under the receiver operating characteristic curve of 0.970 (95%CI: 0.925-0.998) on the internal test set and 0.957 (95%CI: 0.900-0.995) on the external test set. On the external test set, the model showed a sensitivity of 88.0% and a specificity of 92.3%. Conclusion: A Gradient Boosting model integrating age, CEA, multiple lesions, and ring with mural nodule effectively distinguishes CTB from BM with robust external test performance, providing a noninvasive tool to assist clinical decision-making.

★ Interpretable Deep Learning for Opportunistic Screening of Left Ventricular Dysfunction Screening Using Chest Radiographs.

Jang S, Lee DH, Cho HS, Byeon SJ, Hur BY, Park CM — Fri, 23 Oct 2026 00:00:00 +0000

98% · External validation 98%

Purpose To develop and evaluate a deep learning model (DLLV) for identifying reduced left ventricular ejection fraction (LVEF) on chest radiographs (CXRs) and assess its potential as a prescreening tool for selective transthoracic echocardiography (TTE) in health checkup populations. Materials and Methods Same-day CXR-TTE pairs obtained between January 2016 and September 2021 were retrospectively collected for model development. Counterfactual learning was applied to enhance model performance and interpretability. External testing included one tertiary referral center cohort and two health checkup cohorts. Discrimination was evaluated using area under the receiver operating characteristic curve (AUC) and area under the precision-recall curve (AUPRC), and DLLV-based prescreening was evaluated by estimated TTE reduction and missed-case burden. Results The development dataset included 24,674 patients (32,135 CXR-TTE pairs); external cohorts included 5440, 28,034, and 6085 patients (53-62% male). DLLV achieved an AUC of 0.86 (95% CI: 0.84, 0.88) for LVEF ≤ 40% in the tertiary referral cohort. In health checkup cohorts, AUCs were 0.92-0.97 for LVEF ≤ 40%, with AUPRCs of 0.04 in both cohorts, corresponding to 87-fold (95% CI: 16, 404) and 37-fold (95% CI: 3, 251) enrichment over baseline prevalence. Counterfactual reinforcement learning produced heatmaps with greater heart localization precision (P < .001 for all) and higher AUC for LVEF < 50% (P = .04). DLLV-based prescreening estimated TTE volume reductions up to 85% while identifying all 12 and 7 individuals with LVEF ≤ 40%. Conclusion DLLV identified reduced LVEF on CXRs across external test cohorts and showed potential as a prescreening tool for selective TTE in health checkup populations. ©RSNA, 2026.

★ Generalisability of an adult-trained AI model to paediatric coeliac disease diagnosis: a multicentre external validation study.

Jaeckle F, Gillett PM, Kirkwood KJ, Natu S, Chan JYH, Bateman AC — Fri, 23 Oct 2026 00:00:00 +0000

98% · External validation 98%, Prospective 32%

Background: Coeliac disease diagnosis on duodenal biopsies is limited by interobserver variability. We have previously demonstrated pathologist-level performance with our artificial intelligence (AI) model for the histopathological diagnosis of adult coeliac disease, but not in paediatric practice. As paediatric coeliac disease screening programmes expand internationally, accurate and scalable diagnostic tools are needed. We investigated whether an AI model trained exclusively on adult whole-slide images (WSIs) can generalise to paediatric coeliac disease diagnosis across independent centres. Methods: A training and validation dataset of 9958 WSIs from 8421 adult patients (961 coeliac disease; collected Jan 1, 2014, to March 31, 2025) from five centres was used to develop an ensemble of multiple-instance learning models using features from a foundation model. Testing was performed on 708 consecutive paediatric patients (86 with coeliac disease) from two centres (Edinburgh and Southampton) not included in training. Model calibration was assessed, and probability outputs were grouped into clinically interpretable categories. Findings: In adult cross-validation, the AI model had an area under the receiver operating characteristic curve (AUC) of 99.1% (95% CI 98.9-99.4), sensitivity of 89.3% (87.2-91.1), specificity of 98.8% (98.5-99.0), and negative predictive value (NPV) of 98.6% (98.3-98.9). On testing (paediatric) datasets, performance remained high (AUC 98.8%, 95% CI 98.2-99.4; sensitivity 80.2%, 70.6-87.3; specificity 98.4%, 97.1-99.1; NPV 97.3%, 95.7-98.3). Restricting analysis to predictions outside the intermediate-probability range (predicted coeliac disease probability <10% or ≥65%; 604 [85.3%] of 708 cases) improved sensitivity to 100% (95% CI 93.7-100.0) and specificity to 98.7% (97.4-99.4). No misclassifications were observed among high-confidence predictions (<2% or ≥85%; 467 [66.0%] of 708 cases). The expected calibration error was 0.03. Performance improved significantly when biopsies from both duodenal sites (bulb [D1] and descending [D2/3]) were considered. Interpretation: Our AI model, trained on adult biopsies, generalises to paediatric coeliac disease diagnosis across centres and scanner platforms. Well calibrated probability outputs provide clinically interpretable measures of diagnostic confidence and could support safe identification of coeliac disease-negative biopsies within defined thresholds. These findings demonstrate the feasibility of applying adult-derived AI models in paediatric populations and reinforce the importance of multi-site (D1 & D2) biopsy sampling. Funding: The National Institute for Health and Care Research.

★ A dual-prompt framework combining spatial-mask priors and clinical-text conditioning for CT-only GTV delineation in esophageal cancer: A multicenter study.

Sun H, Chen C, Chen J, Wang Q, Li J, An Z — Fri, 23 Oct 2026 00:00:00 +0000

97% · External validation 97%

Background: Accurate gross tumor volume (GTV) delineation is essential for esophageal cancer radiotherapy, but CT-only automatic delineation remains difficult because of indistinct tumor boundaries. Purpose: This study aimed to develop and validate a two-stage dual-prompt framework combining a spatial-mask prior with LLaMA3-encoded clinical-text conditioning for primary esophageal GTV delineation. Methods: This multicenter study included 857 patients: 500 for development, 125 for internal validation, 50 for ablation analysis, and 182 for external validation. The proposed framework consists of two stages: (1) a CT-based automatic segmentation network generating an initial GTV as a spatial mask prompt, and (2) a refinement network that jointly incorporates spatial mask prompt and LLaMA3-encoded clinical text prompt derived from structured clinical information. Performance was evaluated using overlap, surface distance, and volumetric metrics, and compared with multiple state-of-the-art deep learning models. Ablation and radiomics analyses assessed component contributions and downstream consistency. Results: In internal validation, the framework achieved a DSC of 80.32 ± 6.65%, an HD95 of 11.42 ± 8.90 mm, and an absolute volume difference of 9.99 ± 10.49 mL; all Holm-adjusted p values were <0.001 versus the comparison models. Mean DSCs were 78.60 ± 13.39% and 78.36 ± 15.54% in the external cohorts. Mean DSCs were 78.60 ± 13.39% and 78.36 ± 15.54% in the two external cohorts, respectively. In external cohort A, ContextUNETR showed a lower mean HD95 than the proposed model (10.93 ± 8.03 vs. 13.44 ± 18.90 mm), although the paired difference was not statistically significant after Holm correction (adjusted p = 0.906). Compared with the direct Stage-1 output, the complete framework consistently improved overlap, boundary-distance, and volumetric-agreement metrics across the internal and both external validation cohorts (all adjusted p < 0.001). LLaMA3 encoding improved DSC versus an MLP using the same variables (adjusted p = 0.009), whereas ASSD did not differ significantly. Automatic and manual GTVs showed high radiomic concordance and comparable prognostic performance. Conclusion: The framework improved overlap and volumetric agreement for CT-only primary esophageal GTV delineation across multicenter validation cohorts, although boundary-distance performance varied across external cohorts.

★ Integrating 72-h BUN/Cr trajectories with interpretable machine learning for 28-day mortality after acute stroke: a dual-center landmark study.

Cheng Z, Lu Z, Wang S, He M, Lv Y, Duoji Z — Fri, 23 Oct 2026 00:00:00 +0000

97% · External validation 97%

Background: The blood urea nitrogen-to-creatinine ratio (BUN/Cr) may reflect renal perfusion and systemic stress, but its early course after acute stroke is unclear. We identified 72-h BUN/Cr trajectories and assessed their role in interpretable machine-learning evaluation of 28-day mortality. Methods: This dual-center landmark study included 2,805 ICU patients with acute ischemic stroke, nontraumatic intracerebral hemorrhage, or nontraumatic subarachnoid hemorrhage from the Medical Information Mart for Intensive Care IV (MIMIC-IV) version 3.1. Before modeling, patients were split into a training set (n = 1,964) and an internal test set (n = 841). The trajectory and prediction models developed in the training set were applied without modification to the internal test set and the Qinghai external validation cohort (n = 387). Eight algorithms and prespecified nested models were evaluated using Cox regression, discrimination, calibration, paired comparisons, decision curves, and TreeSHAP. Results: A quadratic three-class model identified low/stable, high U-shaped, and rising-to-plateau patterns, with mortality rates of 21.6, 35.9, and 39.7%. Adjusted hazard ratios versus Class 1 were 1.87 (95% CI 1.25-2.79) for Class 2 and 1.73 (1.31-2.29) for Class 3. The training-selected Light Gradient Boosting Machine (LightGBM) achieved an AUROC of 0.734 (95% CI 0.695-0.773), a calibration slope of 0.894, and an O/E ratio of 1.015 in the internal test set. In the Qinghai external validation cohort, the AUROC was 0.624 (95% CI 0.559-0.689), the calibration slope was 0.459, and the O/E ratio was 1.920. Adding trajectory class to a model already containing routine clinical and static renal variables changed the AUROC by 0.003 in the internal test set (p = 0.515) and -0.014 in the Qinghai external validation cohort (p = 0.130), without a clear gain in discrimination. Conclusion: BUN/Cr trajectories were associated with post-landmark mortality and summarized the three-day renal course. Training-selected LightGBM showed moderate internal discrimination and underwent frozen external evaluation. Trajectory class added modest information beyond routine clinical and static renal variables, while the combined approach provides an interpretable framework for 72-h reassessment.

★ Early and accurate detection of ovarian cancer by profiling cell-free mitochondrial DNA fragmentomics in multi-ethnic population.

Guo X, Gong Z, Li Z, Sun T, Fu Y, Jiao H — Fri, 23 Oct 2026 00:00:00 +0000

95% · External validation 95%

Background: Ovarian cancer (OC) remains highly lethal due to frequent late-stage diagnosis and the challenge of distinguishing malignant from benign adnexal masses (BAM) preoperatively. Reliable non-invasive diagnostic biomarkers remain an unmet clinical need. This multi-center study aimed to evaluate the diagnostic potential of plasma circulating cell-free mitochondrial DNA (ccf-mtDNA) fragmentomics for improving OC management. Methods: We developed two machine learning models based on ccf-mtDNA features: an OC Detection (OD) model to distinguishing OC from healthy controls (HC), and a Benign vs. Malignant Differential diagnosis (BMD) model to discriminate OC from BAM. Results: The OD model achieved an area under the curve (AUC) of 0.987 in the training cohort and maintained high performance (AUC above 0.979) in internal and external validation cohorts, with 91.67% sensitivity for Stage I OC at 95% specificity. The BMD model demonstrated strong discriminatory power (AUC = 0.982) and generalizability (AUC ranging from 0.961 to 0.980 in four validation cohorts), significantly outperforming serum biomarkers. Remarkably, both models performed robustly in the Chinese minority cohort. Conclusions: These findings establish ccf-mtDNA fragmentomics as a powerful liquid biopsy approach for the early and accurate detection of OC, with significant potential for clinical translation in high-risk populations and differential diagnosis of benign vs. malignant adnexal masses.

Mobile AI-Assisted Oral Lesion Triage for Community Oral Cancer Screening: Prospective Field Evaluation Study.

Liu HY, Huang YC, Chen SW, Wei LY, Chou CY, Lee JJ — Fri, 23 Oct 2026 00:00:00 +0000

94% · Prospective 94%, External validation 50%

Background: AI-based tools for oral cancer screening have shown promising performance in curated image datasets, but their performance in real-world mobile community screening workflows remains less clear. Objective: This study evaluated the prospective field performance of a parameter-locked, mobile AI-assisted oral lesion triage system embedded in a human-supervised community oral cancer screening workflow. Methods: We conducted a prospective community-based field evaluation in eastern Taiwan from June 17 to December 31, 2025. Trained nonspecialist personnel acquired standard white-light intraoral images using handheld mobile devices during routine screening. The AI system incorporated on-site image quality assessment and lesion-level risk stratification into green, yellow, and red triage categories. Technically inadequate images after repeated acquisition attempts were classified as ungradable and excluded from performance analysis. Three board-certified oral and maxillofacial specialists established an operational clinical reference standard through structured image review and consensus adjudication. The primary outcome was lesion-level identification of high-risk lesions requiring specialist referral, defined as red versus nonred triage. Diagnostic performance was estimated with Wilson 95% CIs; precision-recall performance and preconsensus specialist agreement were also assessed. Results: Among 602 screened participants, 4283 interpretable intraoral images were included in expert adjudication and lesion-level analysis. The AI system flagged 68 red lesions; 12 (17.6%) were confirmed as high risk, and 4 additional high-risk lesions were identified during expert review, yielding 16 high-risk events for the primary red-triage analysis. For high-risk lesion identification, sensitivity was 75% (12/16; 95% CI 50.5%-89.8%) and positive predictive value was 17.6% (12/68; 95% CI 10.4%-28.4%); specificity, negative predictive value, and accuracy were 98.6% (3928/3984), 99.9% (3928/3932), and 98.5% (3940/4000), respectively. Area under the receiver operating characteristic curve values were 0.957 for red, 0.931 for yellow, and 0.964 for green triage; precision-recall analysis showed area under the precision-recall curve values of 0.8235, 0.9052, and 0.9476, respectively. Preconsensus specialist agreement was high, with a Fleiss κ of 0.921 and a Gwet AC1 of 0.981. Conclusions: This mobile AI-assisted oral lesion triage system was feasibly integrated into a real-world community screening workflow operated by trained nonspecialist personnel. In this low-prevalence field setting, red triage showed moderate sensitivity and modest positive predictive value, supporting its use as human-supervised referral-prioritization support rather than autonomous diagnosis. Further multicenter studies with blinded assessment, participant-level outcomes, workflow-efficiency measures, and longitudinal follow-up are needed before clinical impact and scalability can be determined.

Development and Validation of an Interpretable Chest X-Ray Radiomic Model for Identifying PRISm in Non-COPD Individuals.

Zhou Q, Zhai W, Zhou T, Wang Y, Zhou X, Jin Q — Fri, 23 Oct 2026 00:00:00 +0000

93% · External validation 93%

Rationale And Objectives: Chest X-ray (CXR) is low-cost and widely accessible but lacks effective tools for identifying preserved ratio impaired spirometry (PRISm). This study aimed to develop and validate a CXR-based radiomics-clinical combined model for PRISm screening. Materials And Methods: Paired CXR images and clinical variables from 9 hospitals were retrospectively collected between January 2017 and December 2023. Bilateral lung fields on CXR were segmented with a fully automated segmentation algorithm. PyRadiomics was adopted to extract radiomic features. Clinical, radiomic, and combined models were established. Model performance was assessed using receiver operating characteristic analysis with DeLong's test for intermodel area under the curve (AUC) comparison. Global and local SHapley Additive exPlanations (SHAP) analyses were performed to interpret the model. Results: A total of 1563 participants (mean age, 63.1 ± 12.0 years; 857 males; 422 PRISm cases) were included. The combined model achieved optimal performance, with AUCs of 0.761, 0.753, and 0.759 across the training, internal, and external validation cohorts. In the external validation cohort, the combined model (AUC = 0.759, 95% confidence interval [CI]: 0.658-0.854) significantly outperformed individual radiomic (AUC = 0.618, 95% CI: 0.505-0.732) and clinical models (AUC = 0.655, 95% CI: 0.549-0.760; all p < 0.001). SHAP analysis revealed sex as the top predictor in the combined model, followed by radiomic features, body mass index, age, and smoking status. Local SHAP visualizations illustrated variable feature contributions for individual PRISm and normal spirometry cases. Conclusions: The CXR-based combined model exhibits favorable stability and diagnostic performance for PRISm identification, supporting its utility as a primary-screening tool. The SHAP method enhanced the model's interpretability and clinical applicability.

AI-Based Delirium Risk Prediction Compared With the Validated 4AT-Score Screening Tool in Patients With Trauma: Prospective Cohort Study.

Franceschini K, Valentini M, Leithner A, Kramer D, Jauk S, Lorenzer LD — Fri, 23 Oct 2026 00:00:00 +0000

93% · Prospective 93%, External validation 40%

Background: Delirium is a complex neuropsychiatric syndrome whose severe and lasting consequences in hospitalized patients are largely considered preventable. Prevention primarily relies on the early identification of high-risk patients prior to syndrome onset. However, this remains challenging in clinical practice due to the absence of a diagnostic gold standard and the considerable administrative burden associated with established screening tools such as the 4AT score. Objective: The aim of this study is to compare the 4AT score with an established AI-supported prognostic tool in applied clinical practice. Our goal is to demonstrate that shifting from a purely diagnostic approach toward AI-supported delirium risk prediction enables more efficient clinical workflows, reduces misclassification rates, facilitates earlier and more targeted interventions, and ultimately improves long-term patient outcomes. Methods: In this prospective study, a total of 2002 patients from a trauma ward were screened and assessed for delirium risk using the 4AT score and the AI-based tool at admission, on the evening of admission, and on the following day, with each assessment predicting the risk of delirium occurrence at any point during the remainder of the hospital stay. The AI tool based its predictions on clinical data automatically derived from electronic health records. The AI-based tool applied a previously developed and validated random forest model (not developed as part of this study). Results: Of the 2002 processed patients, the AI prediction tool correctly predicted 87.3% (165/189) of patients with delirium onset as at-risk, compared to an early detection rate of 28.0% (5/189) achieved by the 4AT score. Moreover, the AI-based tool correctly classified 73.9% (1340/1813) patients without delirium as not-at-risk, compared to 96.3% (1746/1813) for the 4AT score. Therefore, the AI-based tool achieved very good delirium predictions, with an area under the receiver operating characteristic curve of 0.884 (95% CI 0.8634-0.9039). Conclusions: The AI-based delirium prognostic tool demonstrated superior early predictive capacity for delirium risk at admission, compared with the 4AT score. The AI-based tool enables proactive risk management and targeted risk reduction.

Assisting Screening of Adolescent Depression With Large Language Models as Symptom Detectors: Algorithm Development and Validation.

Ignashina M, Bondaronek P, Shams P, Osuchukwu J, Santel D, Pestian J — Fri, 23 Oct 2026 00:00:00 +0000

92% · Language models 92%

Background: Depression prevalence is rising in adolescents and young adults (ages 10-24 years). Traditional screening methods for depression, such as the Patient Health Questionnaire (PHQ-9), are particularly challenging for adolescents. AI has potential to help with this, but the scarcity of annotated datasets highlights the need for off-the-shelf approaches. In this study, we investigate the feasibility of using state-of-the-art large language models (LLMs) to detect depressive symptoms in adolescent settings (ages 15-17 years). Objective: This study aimed to evaluate the feasibility of using in-context learning with LLMs to identify depressive symptoms in free-text clinical notes for adolescent populations, to evaluate the feasibility of off-the-shelf LLM configurations in detecting PHQ-9 symptom-related information under zero- and few-shot settings, and to demonstrate the potential clinical utility of LLM-driven evidence extraction through an interpretable AI-based depression screening tool. Methods: We examined free text of the electronic health records (EHRs) of adolescent patients with the diagnosis of depression or related mood disorders (age groups 15-17 years, 924 patients) from the Cincinnati Children's Hospital Medical Center. We leveraged the combination of the Beck Depression Inventory (BDI) and the PHQ-9 to develop tailored categories suited for adolescent depression symptoms. We applied 3 state-of-the-art LLMs (Llama, smaller FLAN-T5, and Phi) to identify these symptom categories in zero- and few-shot configurations, allowing us to assess their performance without additional task-specific training. A word-search model was used as the baseline comparator. Results: We manually annotated clinical notes for 22 patients across 16 depression-related symptom categories. Our analysis shows that all LLMs substantially outperformed the word-matching baseline. Within the evaluated prompting configurations, FLAN-T5 achieved the highest precision (average F1-score=0.65, precision=0.78) and strong performance in detecting rare symptoms such as sleep problems (F1-score=0.92) and self-loathing (F1-score=0.80). Phi offered a balanced performance (precision=0.44, recall=0.60), while Llama 3 achieved the highest recall (0.90) but tended to overgeneralize, making it less suitable for precise screening. Within this pilot study, the best-performing LLM exceeded the baseline F1-score by 0.43. Finally, we demonstrate the clinical utility of FLAN-T5 annotations as features in a machine learning model, which distinguished depression cases from controls with high precision (0.78). Conclusions: We used a range of popular LLMs to identify indicators of adolescent (ages 15-17 years) depression symptoms in mental health notes. Within the evaluated configurations, FLAN-T5 demonstrated the most favorable balance between precision and recall. These results support the feasibility of using off-the-shelf LLMs for symptom detection in adolescent clinical notes, while future work using standardized prompting strategies and task formulations will help establish more definitive comparisons between model architectures. The computational efficiency of FLAN-T5 further supports its potential for deployment in resource-limited clinical settings. Our findings require further validation for broader populations and other mental health conditions.

Replicability and Validity of a New Artificial-Intelligence Assessment of Posttraumatic Stress Disorder From Patient Language: A Sequential Evaluation With Model Preregistration.

Kjell O, Ganesan AV, Boyd RL, Oltmanns J, Rivero A, Feltman S — Fri, 23 Oct 2026 00:00:00 +0000

92% · Prospective 92%, External validation 59%, Language models 45%

Artificial intelligence (AI) shows promise in identifying psychopathology through language, but replicability in AI models remains challenging. We develop an AI-based language assessment of posttraumatic-stress-disorder (PTSD) severity and introduce the sequential evaluation with model preregistration to rigorously evaluate its validity and replicability. This design includes two phases: development with preregistration and evaluation. Data included development (N = 1,437) and prospective (N = 346) samples, in which participants described their lives during automated interviews. In the prospective sample, preregistered models correlated with PTSD CheckList scores (r = .38, p < .001) and converged with PTSD diagnosis (area under the curve [AUC] = .76; outperforming demographics and trauma exposures: AUC = .61, p < .01). We found that for each standard-deviation increase, mental-health-care expenditure rose by $696.50 (p < .001). Our preregistered PTSD model assessments are replicable in prospectively collected clinical data and showed external validity against expense criteria. With further development, such models can be used to screen for PTSD or monitor treatment response, especially in telehealth or automated interviews, in which deployment can be seamless.

★ Preoperative prediction of positive margins in breast-conserving surgery using an ensemble model based on DCE-MRI: a multi-center retrospective and prospective study.

Zhao X, Bai JW, Dong LA, Li ZH, He JZ, Chen YX — Thu, 22 Oct 2026 00:00:00 +0000

98% · External validation 98%, Prospective 95%

Background: Breast-conserving surgery (BCS) is the standard care for early breast cancer. However, positive surgical margins lead to high recurrence and require reoperation, remaining a clinical challenge. This study aims to develop an ensemble model to preoperatively predict positive margins using dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). Methods: A total of 887 patients from three Chinese medical centers were included. Patients from the Yunnan (581 patients) and Fujian (51 patients) provinces were used for model development, with external testing on 188 patients from Guangdong and prospective testing on 67 patients from Yunnan. Three base models were developed: the clinical model based on 32 features, the deep learning (DL) model using ResNet-34, and the radiomics model utilizing radiomics features. The ensemble model combined predictions from these three base models using weighted voting. Performance metrics (accuracy, AUC, sensitivity, specificity) were computed, and prognostic analyses were conducted with a public dataset. Results: The ensemble model outperformed each of the three base models, with the AUC of 0.931, accuracy of 0.872, sensitivity of 0.871, and specificity of 0.880 in the validation set. Robustness was further confirmed in independent test cohorts, yielding AUCs of 0.762 and 0.861. Furthermore, patients identified at high-risk by the model exhibited poor recurrence-free survival (RFS, HR = 8.117, p < 0.0001) and overall survival (OS, HR = 5.748, p < 0.0001), underscoring its prognostic value. Conclusions: The ensemble model based on DCE-MRI effectively predicts positive margins and prognosis for BCS patients, facilitating the clinical decision-making on the surgery. Trial registration This study was approved by the Ethics Committee of the Yunnan Cancer Hospital (KYLX2023-134), Xiang'an Hospital of Xiamen University (XDYX202305K26), and Cancer Hospital of Shantou University Medical College (2024059). The study was also registered with the Chinese Clinical Trial Registry (ChiCTR2400083298, Registered 19 April 2024, https://www.chictr.org.cn/showproj.html?%20proj=207948 ).

★ Non-contrast CT radiomics for identifying nonfunctioning kidneys: a multicenter development and external validation study.

Jiang Y, Zheng Y, Feng Z, Bao L, Wei Q — Thu, 22 Oct 2026 00:00:00 +0000

97% · External validation 97%

Purpose: In urological practice, the decision to resect or preserve a kidney with obstruction, tumor or atrophy is based on split renal function, although SPECT renal dynamic scintigraphy as the reference standard is resource-intensive. We developed and externally validated a non-contrast CT radiomics model to identify nonfunctioning kidneys and support urological decision-making in this setting. Methods: In this three-center retrospective study (TRIPOD+AI), adults underwent non-contrast CT and SPECT renal dynamic scintigraphy within 72 h. IBSI-compliant radiomics features were extracted with PyRadiomics, and between-kidney asymmetry features were constructed. After three-stage selection (Spearman, recursive feature elimination, and LASSO), 12 machine learning models were developed and externally validated in 2 independent cohorts. A Rad-score and clinical nomogram were also derived. Results: Among 1029 patients, the internal cohort included 418 patients. Twenty-three of features were retained. The random forest model yielded an internal test AUC of 0.9426, sensitivity of 85.5%, specificity of 88.3%, and negative predictive value of 95.6%. Its external AUCs were 0.875 and 0.913. The nomogram attained external AUCs of 0.920 and 0.881. At the thresholds fixed in the training data, the random forest achieved external sensitivity of 66.0 and 76.4% with negative predictive values of 94.8 and 93.3%, and the nomogram 92.5 and 93.7% with 98.3 and 97.2%. Both radiomics models outperformed logistic-regression baselines built from clinical variables and simple CT size and asymmetry measurements, although the margin over the combined baseline was not statistically significant in one external cohort. The Rad-score was correlated with single-kidney GFR (ρ = -0.738, P < 0.001) and facilitated risk stratification. Conclusion: Non-contrast CT radiomics identified nonfunctioning kidneys without contrast agents or radiotracers. Because non-contrast CT is already part of routine urological workup, this externally validated rule-out tool may reduce unnecessary SPECT examinations and support decisions between nephrectomy and nephron-sparing management.

Multi-parametric MRI radiomics to predict sensitivity to neoadjuvant chemoradiotherapy based on multi-regions of interest in locally advanced rectal cancer: a multicenter study.

Yao L, Zhu Y, Tang Y, Liao L, Pang X, Chen J — Wed, 21 Oct 2026 00:00:00 +0000

94% · External validation 94%

Background: To evaluate whether baseline multi-parametric MRI radiomics models, integrating multiple regions of interest (ROIs), can predict pathological complete response (pCR) to neoadjuvant chemoradiotherapy (NCRT) in locally advanced rectal cancer (LARC), and to validate performance externally. Methods: In this multicenter, retrospective cohort study, 510 patients with LARC treated with NCRT followed by total mesorectal excision were enrolled (330 for training, 83 for internal validation, and 97 for external validation). Radiomic features were extracted from three ROIs (primary tumor, mesorectal region, and lateral lymph node regions). After redundancy reduction and clustering, the top 30 features were selected using gradient boosting decision trees via cross-validation. Four classifiers, logistic regression, random forest, support vector machine, and extra trees, were evaluated, with SMOTE applied to address class imbalance. Model performance was compared across ROI combinations, MRI sequences, and algorithms using the area under the receiver operating characteristic curve (AUC). Results: Overall, 98 of 510 patients (19.2%) achieved pCR. The multi-ROI, multi-parametric model based on extra trees achieved an AUC of 0.717 in the external validation set. Among single-ROI models, tumor ROI yielded the highest AUC (0.630). Among MRI sequences, the ADC-based model performed best (AUC 0.629). Conclusion: Baseline multi-parametric MRI radiomics models integrating multiple ROIs showed modest predictive performance for pCR in LARC and demonstrated similar predictive utility in an independent external cohort. This approach may provide pre-treatment risk stratification and support future research on individualized treatment strategies and organ preservation.

Development and validation of a model using routine clinical features to improve diabetes diagnosis.

Hutchison AL, Narula S, Faubert B, Thomas CC, Rinella ME, Mirmira RG — Wed, 21 Oct 2026 00:00:00 +0000

94% · External validation 94%

Context: Hemoglobin A1c (A1c) and fasting plasma glucose (FPG) miss diabetes diagnosis in many individuals, but the alternative, oral glucose tolerance testing (OGTT), is difficult to obtain. Objective: This work aimed to develop and validate a model to identify diabetes missed by A1c and FPG but detected by OGTT and associate its outcomes with mortality. Methods: We trained machine-learning models using National Health and Nutrition Examination Survey (NHANES) data, applied them to corresponding National Center for Health Statistics mortality data, and externally validated the models with Diabetes Prevention Program Outcomes Study (DPPOS) data. We tested models including the A1c, vital signs, complete blood count, and comprehensive metabolic panel with and without the FPG. This retrospective analysis included NHANES study participants from 1999 to 2016 and DPPOS study participants from 1996 to 2020. Main outcome measures included the area under the receiver operating curve, sensitivity, and specificity of models to detect diabetes. Results: The A1c/FPG+ model had better sensitivity for DPPOS individuals with A1c less than 6.5% (48 mmol/mol) and FPG less than 126 mg/dL compared to a logistic regression of A1c and FPG and recalibrated FPG for individuals both on (0.69 vs 0.53 vs 0.57, respectively) and off (0.75 vs 0.65 vs 0.68, respectively) medication. NHANES participants without diabetes but with A1c/FPG+-predicted diabetes had elevated risk of mortality compared to those without A1c/FPG+-predicted diabetes (HR = 1.4; P < 2 × 10-16). Conclusion: Machine-learning models can improve the detection of individuals who would need OGTT assessment to diagnose diabetes using standard nonfasting laboratory values and vital signs with and without FPG.

Development and External Validation of a Machine Learning-Derived Risk Stratification Score for Metastatic Castration-Resistant Prostate Cancer.

Wed, 21 Oct 2026 00:00:00 +0000

93% · External validation 93%

Purpose: Metastatic castration-resistant prostate cancer (mCRPC) shows marked clinical heterogeneity and poor long-term survival, highlighting the need for simple and reliable risk stratification tools beyond the limited usability of existing complex prognostic models. In this study, we developed and validated PROGRESS (Prostate Cancer Global Risk Evaluation and Stratification Score), a simplified prognostic score derived through machine learning‑guided feature selection distilled into a simple laboratory-based tool. Methods: Baseline data from 2,035 patients with mCRPC (training set) were embedded into synthetic representations, reduced with UMAP, and clustered with DBSCAN to discover outcome-related subpopulations. Feature contribution analysis informed a multivariable model that yielded three independent predictors (prostate specific antigen [PSA], alkaline phosphatase [ALP], and AST) composing PROGRESS. External validation used two independent data sets (n = 1,239) across different therapeutic settings. Overall survival (OS) was analyzed with Kaplan-Meier and Cox regression; discrimination was assessed by ROC-AUC, AUC(t), and Harrel C-index. Results: Unsupervised modeling identified two patient risk subpopulations with significantly different OS rates (median 27.4 v 17.7 months; hazard ratio [HR], 2.20 [95% CI, 1.91 to 2.54]; P < .00001). The three retained predictors (PSA, ALP, and AST) were combined into a composite score (PROGRESS) that showed strong discrimination in the training cohort (ROC-AUC 0.89; AUC(t) 0.60; Harrel C-index 0.69). Using a prespecified cutoff, patients classified by PROGRESS as increased risk had significantly shorter survival than low-risk patients (median 18.3 v 25.6 months; HR, 1.72 [95% CI, 1.50 to 1.97]; P < .0001). PROGRESS prognostic performance was consistent across all validation cohorts (all P < .0001). Conclusion: By combining ML-based pattern discovery with classical modeling, PROGRESS provides a simple, objective, and readily implementable prognostic tool using three inexpensive routine laboratory measures. Its consistent performance across independent data sets supports clinical utility for individual risk stratification, trial enrichment, and early decision making. Prospective real-world validation is warranted.

P-WIN deployment: real-time personalized machine learning to predict critical deterioration in a pediatric intensive care unit.

Tweedy E, Mehta S, Grover S, Stinson HR, Ruiz VM, Pleskunas E — Wed, 21 Oct 2026 00:00:00 +0000

90% · Prospective 90%

Objectives: Early recognition of critical deterioration events (CDEs) in pediatric intensive care units (PICUs) is crucial to improving outcomes, but the real-time performance of machine-learning (ML)-based alert systems is rarely assessed. We prospectively evaluated the performance, operational utility, and deployment feasibility of the PICU Warning Index (P-WIN), a personalized ML alert system predicting CDE in the PICU. Materials And Methods: In this prospective cohort study at a quaternary-care children's hospital, P-WIN estimated deterioration risk for PICU patients aged 0-24 years at 30-minute intervals between January 7 and April 7, 2026 using electronic health record data collected via Fast Healthcare Interoperability Resources and proprietary EPIC API endpoints. CDEs included cardiopulmonary resuscitation, initiation of extracorporeal membrane oxygenation, dilute (10 μg/mL) epinephrine bolus, or unplanned endotracheal intubation. P-WIN was compared to baseline workflows: an automated PICU Warning Tool and a clinician-huddle-based assessment. Primary measures were CDE-level sensitivity, admission-level positive predictive value (PPV), and alerts per patient-day. Results: Among 1005 PICU patients, 99 CDEs were experienced by 62 patients. P-WIN detected 73% of these CDE 1-24 hours in advance, outperforming the PICU Warning Tool (39%, P < .05) and huddle workflow (23%, P < .05). P-WIN generated 0.21 (95% CI, 0.18-0.23) alerts per patient-day. The PICU Warning Tool and huddle workflow had similar PPV to P-WIN. Conclusion: P-WIN detected significantly more CDEs than existing PICU Warning tool and huddle workflows while maintaining a similar PPV. Our findings underscore the feasibility of ML-based personalized warning systems to enhance shared situational awareness of PICU deterioration.

Multicentric development and validation of a fully automated artificial intelligence system for planning medial open wedge high tibial osteotomy on weight-bearing anterior-posterior long leg radiographs: The ISAKOS John Joyce Award 2025.

Rupp MC, von Schacky CE, Lindner FJ, Feucht MJ, Ehmann Y, Fritsch L — Wed, 21 Oct 2026 00:00:00 +0000

90% · External validation 90%

Purpose: Preoperative planning, currently performed by orthopaedic surgeons (OS), is essential for successful osteotomy, but is labour-intensive and error-prone. As such, there is significant potential in the use of automated artificial intelligence (AI) algorithms to augment the abilities of OS in managing lower extremity pathologies in a high-volume task that demands high precision and reliability. The aim was to develop a deep learning (DL) system capable of fully autonomous planning a medial open wedge high tibial osteotomy (mow-HTO) on weight bearing anterior posterior (AP) long leg radiographs (LLR) and to validate its agreement with an expert-derived reference standard, its reliability and processing time of the DL model compared to expert OS on multicentric test data sets (two institutions). Methods: A total of 594 patients (age 41.1 ± 13.2 years, 182 female) who underwent osteotomy between 01/2014 and 01/2021 were retrospectively enrolled. Manual annotations for lower extremity alignment and high tibial landmarks were performed on AP LLR radiographs. The data set was divided into training (60%, n = 399), validation (10%, n = 59) and hold-out testing (30%, n = 136), with an additional external dataset (n = 100) for further testing. A DL system with twelve expert networks was trained to identify relevant landmarks for HTO planning. The DL model was then tested on fully automated planning of a medial open wedge HTO based on (1) the desired postoperative intersection of the mechanical axis with the tibial plateau (%-based) and (2) the desired femorotibial axis (°-based). Agreement with the expert-derived reference, reliability and processing time were compared with three expert OS using FDA-approved preoperative planning software on the test data sets. Results: For preoperative alignment parameter analysis, compared to the ground truth, human preoperative accuracy ranged from 0.13° ± 0.10° to 1.06° ± 1.4°, interrater reliability ranged from intraclass correlation (ICC) 0.84-1.0 and clinically acceptable accuracy ranged from 72% to 100%. Preoperative accuracy of the DL model ranged from 0.22° ± 0.18° to 1.04° ± 1.28°, interrater reliability ranged from ICC 0.93-1.0 and clinically acceptable accuracy ranged from 91% to 100%. For the planning of the osteotomy gap, compared to the ground truth, human accuracy ranged from 0.46 ± 0.65 mm to 0.5 ± 0.88 mm (%-based) and 0.43 ± 0.55 to 0.52 ± 0.75 mm (°-based), reliability ranged from ICC 0.9-0.96 and 0.95-0.97 and clinically acceptable accuracy was 93.6%-96.0% (%-based) and 89.8%-93.9% (°-based). For the DL model, the accuracy was 0.53 ± 0.63 mm (%-based) and 0.44 ± 0.53 mm (°-based), reliability was ICC 0.99 (both methods) and clinically acceptable accuracy was 98.1% (%-based) and 96.2% (°-based). The DL-model significantly and substantially outperformed the OS in intrarater reliability (0.88 vs. 1.0) and processing time for a fully automated alignment analysis (25.3 ± 0.8 s vs. 148.7 ± 10.1 s, p < 0.001). Conclusion: The developed DL model enabled fully automated mow-HTO planning on AP long-leg radiographs with reliability comparable to that of expert OS while significantly reducing processing time and improving measurement consistency. These suggest that DL-based systems may assist orthopaedic providers by accelerating and standardising osteotomy planning. Level Of Evidence: Level III, retrospective comparative study.

★ External validation and application of machine learning models for obesity in Chinese children and adolescents per Chinese and WHO growth standards.

Zheng F, Wang X, Xue M, Wang Q, Zhang W, Zhang Z — Tue, 20 Oct 2026 00:00:00 +0000

97% · External validation 97%

Background: Childhood and adolescent obesity is a growing global public health challenge, underscoring the need for reliable tools to identify individuals at high risk. We aimed to develop, validate, and deploy machine learning (ML) models for predicting obesity in children and adolescents aged 3-18 years using multi-domain risk factors per Chinese and World Health Organization (WHO) growth standards, with an online tool for clinical and public health translation. Methods: We conducted a population-based cross-sectional study. We used data from 22,555 participants in Beijing (2022, 2024) for training/internal validation, and data from 5,343 participants in Tangshan (2020) for external validation. We defined obesity using both Chinese and WHO growth standards. We analysed 38 features across 5 domains; we selected key predictors using least absolute shrinkage and selection operator regression. We evaluated 4 ML models using 28 performance metrics, with SHapley Additive exPlanations enhancing interpretability. Results: We retained 5 features per the Chinese growth standards and 12 per the WHO growth standards. Age, maternal body mass index (BMI), paternal BMI, screen time, and birthweight were consistent predictors. The neural network achieved optimal performance per Chinese growth standards (area under the receiver operating characteristic curve (AUROC) = 0.735), and eXtreme gradient boosting performed best per WHO growth standards (AUROC = 0.736), with the corresponding external validation AUROCs of 0.699 and 0.707. We subsequently deployed user-friendly online prediction tools for translation. Conclusions: We identified two interpretable ML models for obesity in children and adolescents per both Chinese and WHO growth standards, each demonstrating modest but reliable predictive performance.

★ MTL-LymphNet: A Lightweight Multi-Task Learning Framework for Simultaneous Lymph Node Segmentation and Metastasis Classification in Thyroid Cancer Ultrasound.

Thahiruddin M, Wahyudi I, Mustofa AN — Tue, 20 Oct 2026 00:00:00 +0000

95% · External validation 95%

Ultrasound assessment of cervical lymph node metastasis in papillary thyroid carcinoma is crucial but suffers from inter-observer variability. Existing deep learning solutions typically use separate single-task pipelines, resulting in parameter redundancy. We propose MTL-LymphNet, a lightweight, backbone-agnostic multi-task learning framework that simultaneously segments lymph nodes and classifies metastasis within a single forward pass, evaluated across 3 backbones and 4 loss weighting strategies. On 338 images from 2 independent clinical centers, MTL-LymphNet achieved an internal cross-validation Dice coefficient of 0.891 ± 0.022 and AUC of 0.948 ± 0.027. Bidirectional external validation revealed substantial domain shift, with cross-center Dice of 0.865 and 0.779, and AUC of 0.733 and 0.830. The multi-task configuration offered a favorable efficiency-performance trade-off (∆Dice = Δ, -0.014, P < .0001) for 40% fewer parameters than separate models. No backbone or weighting strategy consistently dominated, and calibration degraded substantially under domain shift (ECE: 0.049 internally vs 0.178-0.304 externally), requiring center-specific recalibration before clinical use. To our knowledge, this is the first MTL framework for cervical lymph node analysis with bidirectional multi-center validation and systematic ablation. With a compact size of 22.6 to 29.1 MB and CPU inference of 81 to 131 ms, MTL-LymphNet demonstrates technical feasibility as a CAD tool warranting prospective validation.

Super Resolution Reconstruction: A Strategy to Improve the Diagnostic Performance of Radiomics for Predicting Microvascular Invasion in Hepatocellular Carcinoma.

Liu X, Tang X, Zhao H, Gao Y, Zhou X, Li Z — Tue, 20 Oct 2026 00:00:00 +0000

94% · External validation 94%

Rationale And Objectives: Microvascular invasion (MVI) is a key factor associated with high postoperative recurrence rate in patients with hepatocellular carcinoma (HCC). This study aims to evaluate the value of super-resolution (SR) reconstruction based on hepatobiliary phase (HBP) MR images for improving the accuracy of MVI prediction. Materials And Methods: A total of 544 patients with HCC from two medical centers between 2018 and 2025 were retrospectively analyzed, including 450 patients in the training set, 51 in the internal validation set, and 43 in the external test set. Based on a lesion-aware 3D Restormer transformer architecture, SR reconstruction was performed using HBP MR images to generate super-resolution MR images. Tumor segmentations were automatically performed using the nnU-Net architecture and manually corrected by an experienced abdominal radiologist before radiomics feature extraction. Radiomics features were then extracted from original and SR images. Radiomics models based on original and SR images were developed to predict MVI status. Three machine learning algorithms, including Logistic Regression, Random Forest, and LightGBM, were used for model development. Results: A total of 851 radiomics features were extracted from the original and SR images, respectively. Redundant features were filtered, and 50 radiomics features were selected using LASSO regression with cross-validated. For MVI prediction, the pooled areas under the receiver operating curves (AUCs) of radiomics model based on the original images were 0.710, 0.764, and 0.614 in the training, validation and test sets, respectively. The corresponding AUCs of the radiomics model based on the SR images were 0.855, 0.798, and 0.827, respectively. In the test set, the AUC of the radiomics model based on the SR images was higher than that based on the original images (0.827 vs. 0.614, p <0.001). Conclusion: SR reconstruction can improve the diagnostic performance of radiomics models for predicting MVI in patients with HCC.

Composition-aware and leakage-controlled sparse log-ratio learning for mortality classification from respiratory microbiome profiles.

Hou L, Zhang X, Huang K, Wu W, Xing L, Shen W — Tue, 20 Oct 2026 00:00:00 +0000

92% · External validation 92%

Background: Respiratory microbiome profiles contain compositional information that may complement clinical variables for mortality classification in hospitalized patients with COVID-19. However, microbiome-based classifiers remain vulnerable to compositional artifacts, patient-level leakage, and cohort-specific feature construction. Objective: To develop and externally evaluate a leakage-controlled sparse log-ratio learning framework for mortality classification using respiratory microbiome profiles. Methods: This retrospective secondary analysis used three public COVID-19 respiratory microbiome cohorts. The Sparse Log-Ratio Risk Network (SLRRN) was developed using patient-level nested validation, with genus filtering, zero replacement, balance construction, model tuning, and threshold selection restricted to training partitions. The model learned 16 signed genus balances and integrated four clinical covariates available at the first eligible respiratory sampling landmark. The final specification and threshold were frozen before external evaluation. Results: The development cohort included 82 patients (20 deaths) from PRJNA678105 and PRJNA683617, and frozen external evaluation included 76 independent patients (38 fatal and 38 severe nonfatal cases) from PRJNA726205. Internal AUROC was 0.838 (95% CI, 0.765-0.911), and AUPRC was 0.674 (0.577-0.771). The paired AUROC difference versus XGBoost was 0.041 (0.009-0.073). Without refitting or threshold revision, external AUROC was 0.821 (0.738-0.904), with sensitivity of 0.816 and specificity of 0.842 at the frozen threshold of 0.37. In the outcome-enriched external cohort, the Brier score was 0.166 and the ECE was 0.065; these estimates describe the selected study sample. Conclusion: SLRRN provides a composition-aware and leakage-controlled framework for respiratory microbiome-based mortality classification. The findings support further prospective evaluation with standardized sampling and direct comparison with established clinical models before clinical translation.

Pilot study of a ChatGPT-based educational and counselling tool in a specialist HIV and sexual health clinic.

Koh MCY, Ngiam JN, Yong HYJ, Puspita T, Quah J, Mak L — Tue, 20 Oct 2026 00:00:00 +0000

91% · Prospective 91%

Objectives: Real-world implementation of large language models such as ChatGPT to enhance patient education remains limited. We evaluated the feasibility, acceptability and attitudes towards a ChatGPT-based counselling tool among patients and healthcare workers (HCWs) in a specialist infectious diseases clinic. Methods: We conducted a prospective clinical implementation study in a specialist HIV and sexual health clinic in Singapore from Jan 2025 to December 2025. Patients interacted with a custom ChatGPT-based chatbot prior to routine counselling by physicians or pharmacists. Responses were reviewed with patients subsequently during consultation. Post-interaction, patients and HCWs completed structured surveys incorporating domains from a validated survey tool: perceived risk, anxiety, technology/social influence and intention for future use. Results: The chatbot was most frequently used for counselling on antiretroviral therapy initiation or switching, and sexual health advice. A total of 35 patients and 35 HCW assessments were analysed. Patients had greater perceived risk with the use of the chatbot compared to HCWs, as indicated by lower median scores (13, IQR 10-16; vs. 18, IQR 15-21), while anxiety scores were similar. Technology/social influence scores were high across both groups (overall median 20, IQR 19-22). Most participants indicated willingness for future use. Qualitative feedback highlighted improved patient understanding and more focused consultations, but concerns included lack of contextual nuance and occasional inaccuracies. Conclusions: The pilot implementation of a ChatGPT-based counselling tool in a specialist HIV and sexual health clinic was feasible and generally acceptable to patients and HCWs. Further studies are required to evaluate accuracy and clinical effectiveness before broader implementation.

Automated screening of autism via a pretrained vision-language model in naturalistic caregiver-child interactions.

Deng R, Huang H, Lv S, Xing Y, Li Y, You C — Tue, 20 Oct 2026 00:00:00 +0000

91% · External validation 91%

Early autism screening is essential for timely intervention but remains constrained by specialist dependence, costly protocols, and limited accessibility. We present Autism-CLIP, a vision-language model that automates screening from 3-min naturalistic caregiver-child interaction videos recorded in toddlers aged 15-24 months. By integrating contrastive learning, large language model-generated supervision, domain-specific architecture, and hybrid training, Autism-CLIP aligns visual behavioral cues with autism-related assessment outcomes, enabling annotation-free inference. In multicenter evaluation, Autism-CLIP achieved an AUC of 0.903 in the internal cohort (n = 103) and 0.873 in the external cohort (n = 102), outperforming conventional machine learning methods. Attention mapping highlighted clinically relevant markers, including gaze aversion and limited social reciprocity. Requiring only smartphone-captured videos without manual coding or controlled settings, Autism-CLIP offers a scalable, low-cost approach to broaden access to early autism screening and support prioritization of clinical resources.

Foundation artificial intelligence models for health recognition by use of face photographs (FAHR-Face): a model development and validation study.

Haugg F, Lee G, He J, Nürnberg L, Bontempi D, Bitterman DS — Tue, 20 Oct 2026 00:00:00 +0000

90% · External validation 90%

Background: Facial appearance offers a non-invasive window into health. We built Foundation Artificial Intelligence for Health Recognition-Face (FAHR-Face), a foundation model trained on more than 40 million facial photographs and fine-tuned it for two distinct tasks: biological age estimation (FAHR-FaceAge) and survival prediction (FAHR-FaceSurvival). Methods: FAHR-FaceAge underwent two-stage, age-balanced fine-tuning on ten public image datasets; FAHR-FaceSurvival was fine-tuned on 34 363 photographs of patients with cancer. Model robustness (cosmetic surgery, make-up, pose, and lighting) and independence (attention mapping) were tested. Both models were clinically tested in two independent cancer patient datasets with survival analysed by multivariate Cox models and adjusted for clinical prognostic factors. Findings: For age estimation, FAHR-FaceAge had the lowest mean absolute error of 5·1 years on public datasets, outperforming benchmark models and maintaining accuracy across the full human lifespan. In patients with cancer, FAHR-FaceAge outperformed a previous age estimation model in survival prognostication. FAHR-FaceSurvival showed robust prediction of mortality, and the highest-risk quartile had more than triple the mortality of the lowest (adjusted hazard ratio 3·23; p<0·0001). These findings were validated in the independent cohort and both models showed generalisability across age, sex, race, and cancer subgroups. The two algorithms provided distinct, complementary prognostic information; attention mapping revealed that each model relied on distinct facial regions. The combination of FAHR-FaceAge and FAHR-FaceSurvival improved prognostic accuracy. Interpretation: A single foundation model can generate inexpensive, scalable facial biomarkers that capture both biological ageing and disease-related mortality risk. The foundation model enabled effective training by use of relatively small clinical datasets. Routine quantitative facial analysis could enhance risk stratification when traditional biomarkers are unavailable. Funding: EU European Research Council.

★ Use of an AI metabolomics score to predict cardiovascular risk among patients with Psoriasis in the MGB and UK Biobanks.

Kotanidis CP, Schuermans A, Choksi P, Lasky-Su J, Natarajan P, Li L — Mon, 19 Oct 2026 00:00:00 +0000

98% · External validation 98%

Background: Systemic inflammation in psoriasis accelerates atherosclerosis and confers excess cardiovascular risk that is not fully captured by traditional clinical predictors. Metabolomic profiling may identify disease-specific metabolic signatures associated with cardiovascular events and enhance risk prediction beyond conventional factors. Objectives: To determine whether plasma metabolomic profiling improves cardiovascular risk prediction in patients with psoriasis and to develop and externally validate an artificial intelligence-driven metabolomic risk score (MetaPsoAI) for incident cardiovascular events. Methods: We analyzed 1086 individuals with psoriasis and no prior coronary artery disease from the Mass General Brigham (MGB) Biobank who underwent nuclear magnetic resonance-based quantification of 168 circulating metabolites. Associations between metabolites and incident all-cause death or myocardial infarction were assessed. An extreme gradient boosting algorithm was used to develop MetaPsoAI for prediction of myocardial infarction or stroke. External validation was performed in the UK Biobank (n = 274,000), stratified by psoriasis status. Results: Multiple metabolites, particularly HDL subfractions and amino acid metabolites, were associated with cardiovascular outcomes. MetaPsoAI stratified risk across quartiles, with individuals in the highest quartile demonstrating a 3.6-fold higher risk of myocardial infarction or stroke in the MGB cohort. In the UK Biobank, the highest quartile was associated with a 1.9-fold increased risk among psoriasis patients, with stronger effect sizes compared with non-psoriasis controls. Conclusions: Our findings highlight the potential of AI-driven metabolomics to refine cardiovascular risk prediction in immune-mediated diseases and enable precision prevention strategies for patients with psoriasis.

★ New insights into tumor size in colorectal cancer: smaller size, higher risk?

Ding R, Li T, Li C, Song W, Gao X, Xu M — Mon, 19 Oct 2026 00:00:00 +0000

98% · External validation 98%

Background: In colorectal cancer, the AJCC T category is defined by the depth of bowel wall invasion rather than by the absolute size of the primary tumor. The clinical value of tumor size in assessing tumor aggressiveness, metastatic risk, and prognosis remains insufficiently investigated. Aim: To evaluate the clinical value of tumor size in assessing potential metastatic risk and prognostic stratification in colorectal cancer. Methods: We retrospectively included 10,513 patients with colorectal cancer who underwent surgical treatment at FUSCC between 2008 and 2020. We integrated tumor size with clinicopathological characteristics to assess metastatic risk using machine-learning models and SHAP-based interpretation. By modeling tumor size as both a categorical and continuous variable, we further investigated its prognostic relevance using Kaplan-Meier survival analysis and Cox proportional hazards regression incorporating restricted cubic splines. Additionally, an independent cohort of 6,739 patients from Changhai Hospital was used for external validation. Results: Machine-learning identified lymph node metastasis as the primary predictive endpoint, with the CatBoost model achieving the highest AUROC of 0.800. SHAP analysis showed that smaller tumor diameter contributed positively to predicted metastatic risk. Multivariable restricted cubic spline analysis further demonstrated a significant nonlinear association between tumor diameter and OS, with mortality risk increasing as tumor diameter decreased within the approximately 0-3 cm range. Among younger male patients with locally advanced disease, tumors ≤3 cm were associated with significantly worse OS and DFS than tumors >5 cm (both P < 0.001). External validation yielded similar findings. Conclusion: Smaller tumor size may paradoxically indicate higher lymph node metastatic risk and poorer prognosis in a subset of patients with colorectal cancer, particularly younger male patients with locally advanced disease. Tumor size may therefore serve as a complementary factor for postoperative risk stratification beyond conventional TNM staging.

★ Multiparametric MRI-based habitat and peritumoral radiomics integrating semi-automated segmentation for preoperative prediction of high-grade prostate cancer: a dual-center study.

Hua D, Ma L, Ji X, Huang L, Zhao Y, Shen W — Mon, 19 Oct 2026 00:00:00 +0000

98% · External validation 98%

Background: This study aimed to develop and validate an interpretable machine learning framework that combines semi-automated segmentation, intratumoral habitats derived from multiparametric MRI (mpMRI), and multiscale peritumoral radiomics to predict high-grade prostate cancer (HGPCa) before surgery. Methods: This retrospective dual-center study involved 274 patients, comprising 208 from Center 1 for development and 66 from Center 2 for external test. The clinical and imaging variables included patient age, serum PSA, fPSA, PSAD, maximum tumor diameter, PI-RADS scores, and anatomical zone distribution (PZ vs. TZ). The development cohort was randomly divided into a training cohort (n = 145) and an internal validation cohort (n = 63). Intratumoral habitats were delineated using semi-automated segmentation and unsupervised K-means clustering, while peritumoral radiomic features were extracted from concentric 1-, 3-, and 5-mm shells. Predictive classifiers were optimized through machine learning algorithms, particularly a multilayer perceptron (MLP) for the habitat signature. A combined clinical-radiomics nomogram was constructed, with SHapley Additive exPlanations (SHAP) employed to enhance model interpretability. Subgroup analyses were performed to assess the discriminative performance of the combined nomogram across the peripheral zone (PZ) and transition zone (TZ). Results: The combined nomogram exhibited superior diagnostic performance, achieving AUCs of 0.914, 0.903, and 0.797 in the training cohort, internal validation cohort, and external test cohort, respectively. The MLP-based habitat model showed good discriminative performance, with AUCs of 0.896, 0.897, and 0.782 across the training cohort, internal validation cohort, and external test cohort, respectively. Among the peritumoral regions, the 5-mm shell demonstrated the best discriminative performance. SHAP analysis identified Habitat 1 Shape Sphericity as the most influential feature. Subgroup analysis demonstrated the discriminatory ability of the combined nomogram in both the PZ (AUC: 0.796) and TZ (AUC: 0.753) subgroups. Conclusion: The integration of semi-automated segmentation-based habitat analysis and peritumoral radiomics provides a promising approach for preoperative risk stratification of prostate cancer.

★ A multimodal deep learning radiomics nomogram integrating clinicopathological features for predicting pathological complete response after neoadjuvant therapy in triple-negative breast cancer: a multicenter retrospective study.

Qi C — Mon, 19 Oct 2026 00:00:00 +0000

98% · External validation 98%

Background: Accurate noninvasive estimation of pathological complete response (pCR) after neoadjuvant therapy (NAT) in triple-negative breast cancer (TNBC) may support multidisciplinary discussions, but multimodal MRI-ultrasound evidence remains limited. We developed and externally validated a clinicopathological deep learning radiomics nomogram and explored its potential auxiliary value in a strictly hypothetical post-hoc surgical-planning scenario. Methods: With institutional review-board approval and waived consent, this retrospective multicenter diagnostic accuracy study included 436 biopsy-proven TNBC patients who underwent NAT and surgery: Center 1 training cohort (n = 268) and internal validation cohort (n = 86), plus historical cases from cooperating Centers 2-3 for external validation (n = 82). pCR was defined as ypT0/is ypN0. Pretreatment dynamic contrast-enhanced MRI (DCE-MRI), diffusion-weighted imaging (DWI)/apparent diffusion coefficient (ADC) maps, T2-weighted imaging (T2WI), and B-mode ultrasound were analyzed. Intratumoral and 5-mm peritumoral ring radiomics features were extracted and selected by LASSO. DenseNet-121 extracted modality-specific deep features, fused using multi-head self-attention. Age, tumor size, Ki-67, tumor-infiltrating lymphocytes (TILs), cT/N stage, and histological grade were integrated with imaging scores in a logistic-regression nomogram. Parameters and thresholds were locked after Center 1 development and applied once to external data. No model output was used clinically. Feature standardization and ComBat harmonization parameters were strictly locked to the training cohort before application to validation data. Results: pCR occurred in 182/436 patients (41.7%). The nomogram achieved areas under the receiver operating characteristic curve (AUCs) of 0.917 (95% CI, 0.881-0.947) in training, 0.883 (0.814-0.937) in internal validation, and 0.856 (0.773-0.921) in external validation, exceeding clinical (external AUC, 0.721), radiomics (0.786), and deep learning (0.819) models. External sensitivity, specificity, and accuracy at the locked threshold were 0.765, 0.813, and 0.793, respectively. Calibration was acceptable (slope, 0.93; intercept, -0.06; Brier score, 0.16). Decision curve analysis showed net benefit for thresholds of 0.20-0.70. Multi-threshold analysis revealed that a high-specificity threshold (0.65) yielded a positive predictive value of 0.85, potentially identifying a subset of patients with very high confidence of achieving pCR. Compared with the clinical model, external reclassification improved (continuous net reclassification improvement [NRI], 0.34; integrated discrimination improvement [IDI], 0.11; P = 0.003 and P = 0.008, respectively). Conclusion: Among the few multicenter TNBC studies combining MRI, ultrasound, intratumoral/peritumoral radiomics, and attention-based deep learning, this nomogram showed good retrospective accuracy for pCR prediction after NAT. The model is presented as an aid for pCR prediction and preoperative stratification rather than as direct support for surgical decision-making. Prospective validation and clinical-impact studies are required before implementation.

★ LDAR shows more consistent prognostic performance than other albumin-derived indices for 28-day ICU mortality in critically ill patients with urosepsis: evidence from a dual-cohort retrospective study.

Mu Y, Tang H, Zeng Y, Wang Q, Huang T, Li W — Mon, 19 Oct 2026 00:00:00 +0000

98% · External validation 98%

Background: Urosepsis remains a leading cause of ICU mortality. Albumin-derived composite indices, which integrate nutritional, inflammatory, and metabolic parameters, hold prognostic promise; however, their comparative performance has not been directly assessed. Methods: We conducted a dual-cohort retrospective study using MIMIC-IV (n = 3,060) for derivation and the Affiliated Hospital of Guizhou Medical University (n = 353) for validation. The primary outcome was 28-day ICU mortality in a competing-risks framework with discharge alive as competing event. Six indices-lactate dehydrogenase-to-albumin ratio (LDAR), anion gap-to-albumin ratio (AGAR), glucose-to-albumin ratio (GAR), red cell distribution width-to-albumin ratio (RAR), total bilirubin-to-albumin ratio (TAR), and urea nitrogen-to-albumin ratio (UAR)-were compared via cause-specific Cox and Fine-Gray subdistribution hazard models, cumulative incidence functions, restricted cubic splines (RCS), Kaplan-Meier curves, and ROC/Youden/cNRI analyses. LASSO-selected features trained a multilayer perceptron (MLP), benchmarked against six severity scores; a parsimonious external model (events per variable ≥ 10) was also examined. Results: In the derivation cohort, logLDAR, RAR, TAR, AGAR, and UAR were independently associated with 28-day ICU mortality in fully adjusted models (HRs = 1.292, 1.147, 1.136, 1.074, and 1.013, respectively; all P < 0.01), whereas logGAR was not. Fine-Gray models corroborated these findings (logLDAR sHR = 1.351, 95% CI: 1.238-1.473). RCS revealed a non-linear logLDAR-mortality relationship (P non-linear = 0.040), with risk accelerating beyond logLDAR > 7. logLDAR yielded the highest AUC (0.638, 95% CI: 0.609-0.667) and cNRI (0.119). External validation confirmed the independent prognostic value of logLDAR (HR = 1.524, 95% CI: 1.225-1.895; Fine-Gray sHR = 1.545, 95% CI: 1.295-1.844), with an external AUC of 0.735 (95% CI: 0.648-0.808), although between-cohort differences in case definitions may partly contribute. The MLP with eight LASSO-selected features achieved an external AUC of 0.762 (95% CI: 0.687-0.840), comparable to SOFA plus logLDAR (AUC = 0.782; DeLong P = 0.512). SHAP analysis ranked logLDAR among the top predictors. Conclusion: Among the six albumin-derived indices, logLDAR demonstrated the most consistent and generalizable prognostic performance for 28-day ICU mortality in urosepsis, reflecting both tissue hypoxia and nutritional-inflammatory depletion, and remained robust under competing-risk adjustment. logLDAR may serve as a simple bedside biomarker for early risk stratification; however, prospective studies are warranted to determine whether interventions targeting its underlying pathways can improve outcomes.

★ Development and validation of an explainable model for predicting lymph node metastasis in early gastric cancer: a multicenter retrospective study with external and prospective validation.

Xiong R, Sun X, Jiang Q, Xu Z, Sun Y, Zhou Y — Mon, 19 Oct 2026 00:00:00 +0000

97% · External validation 97%, Prospective 94%

Background: Accurate preoperative prediction of lymph node metastasis (LNM) is essential for treatment selection in early gastric cancer (EGC), particularly when considering endoscopic submucosal dissection. Existing approaches show limited sensitivity for metastatic disease and frequently lack external or prospective validation. We aimed to develop and validate an interpretable machine-learning model for individualized preoperative prediction of LNM in EGC. Methods: In this multicenter study, four cohorts from three tertiary centers were included. Patients with pathologically confirmed EGC who underwent gastrectomy with lymph node dissection were retrospectively enrolled for training and external validation. A temporally independent cohort (September 2023-December 2024) was prospectively enrolled for validation. A preoperative pathology validation cohort (January 2025-January 2026) was also analyzed. Clinicopathological variables were selected using LASSO regression. SMOTE was applied to address class imbalance. Six machine-learning algorithms were compared. Model performance was assessed using AUC and PR-AUC, and interpretability was evaluated using SHapley Additive exPlanations. A web-based calculator was developed. Results: Among 3412 patients in the training cohort, 467 (13.7%) had LNM. XGBoost achieved the best discrimination (training AUC = 0.9289) and maintained robust performance in the external (AUC = 0.8583; PR-AUC = 0.6079) and prospective validation cohorts (AUC = 0.8766; PR-AUC = 0.5137). In the preoperative pathology validation cohort, the model achieved an AUC of 0.8653 and a PR-AUC of 0.6050. Key predictors included lymphovascular/neural invasion, invasion depth, tumor location, differentiation, and tumor size. Conclusions: This externally and prospectively validated, interpretable machine-learning model provides transparent individualized risk estimates and may support personalized decision-making between endoscopic and surgical management in EGC.

★ Development, Validation, and Deployment of an Interpretable Machine Learning Model for In-Hospital Mortality Prediction in Critically Ill Patients with COPD and Sepsis: A Multicenter Cohort Study.

Ju J, Chen J, Li G, Wang J, Yang L — Mon, 19 Oct 2026 00:00:00 +0000

97% · External validation 97%

Background: Critically ill patients with chronic obstructive pulmonary disease (COPD) and sepsis are at high risk of in-hospital mortality because of respiratory and organ dysfunction. Early risk assessment may be useful for clinical decision-making. This study aimed to develop an interpretable machine learning model for this population. Methods: Data were extracted from MIMIC-IV and the eICU Collaborative Research Database (eICU-CRD). The MIMIC-IV cohort was split in a 7:3 ratio into training and internal test sets. Features were selected using least absolute shrinkage and selection operator regression and Boruta. Eight models were trained using 10-fold cross-validation and grid search. The best-performing model was externally validated using eICU-CRD, interpreted using Shapley Additive Explanations (SHAP), and compared with the Sequential Organ Failure Assessment (SOFA) score. Results: Among 4,900 MIMIC-IV patients, in-hospital mortality was 19.2%. In the internal test set, the naive Bayes (NB) model achieved an area under the receiver operating characteristic curve (AUC) of 0.743 (95% CI, 0.711-0.775), outperforming the SOFA score (AUC, 0.696, 95% CI, 0.663-0.730, P < 0.001). In external validation, AUC decreased to 0.681 (95% CI, 0.607-0.755), an absolute reduction of 0.062. SHAP identified Simplified Acute Physiology Score II and SOFA as predictors, followed by the anion gap and urine output. The model was deployed across web, desktop, and mobile platforms. Conclusion: An NB model was successfully developed to predict in-hospital mortality in critically ill patients with COPD and sepsis. Further prospective multicenter validation is warranted before routine clinical use.

★ Multicenter external validation of a multimodal MRI deep-learning model to predict meningioma bone invasion.

Wei Z, Hu W, Tang F, Ren J, Zhou J, Liu M — Mon, 19 Oct 2026 00:00:00 +0000

96% · External validation 96%

Background: Preoperative detection of meningioma bone invasion lacks objective biomarkers and relies on subjective radiological assessment, posing a risk of excessive bone resection or residual tumor. We therefore developed and externally validated a parsimonious six-feature DenseNet121 signature derived from multiparametric MRI across two independent institutions. Methods: In this retrospective diagnostic accuracy study, 438 consecutive patients (mean age, 53 ± 11 years; 77% women) with histologically confirmed meningioma were allocated to training (n = 182), internal validation (n = 77), and external test (n = 179) sets. Following N4 bias-field correction and isotropic resampling, contrast-enhanced T1-weighted, T2-weighted, and diffusion-weighted volumes were fed into a DenseNet121 architecture to extract a six-feature signature via nested cross-validation. Discrimination (area under the ROC curve, AUC), calibration (Hosmer-Lemeshow test), and clinical utility (decision-curve analysis) were evaluated. Results: In the external cohort (prevalence 11%), the signature achieved an AUC of 0.82 (95% CI 0.73-0.91), sensitivity 75%, specificity 77% and negative predictive value 96%, with an AUC improvement of 0.155 over conventional radiomics (95% CI, 0.01-0.30; p = 0.035). At a 20% risk threshold, decision-curve analysis confirmed a positive net clinical benefit, primarily driven by the high negative predictive value. SHAP analysis demonstrated the dominant and consistent contribution of low-ADC features. Conclusion: The DenseNet121-derived six-feature signature serves as a reliable rule-out aid for preoperative meningioma bone invasion, with a high negative predictive value that may safely support decisions to avoid extensive bone resection in predicted-negative cases. Positive predictions require confirmatory CT or intraoperative biopsy to mitigate false positives. This tool could assist nurses and surgeons in preoperative planning, facilitating personalized patient care and potentially reducing operative time and complications.

★ Pediatric chronological age estimation from pelvic radiographs using deep learning: a multicenter retrospective study.

Hwang J, Kim Y, Hwang JY, Jeon K, Choi YH — Mon, 19 Oct 2026 00:00:00 +0000

96% · External validation 96%

Objectives: Pelvic bones exhibit multiple secondary ossification centers during maturation, enabling automated age estimation. This study investigated whether a deep learning model could estimate chronological age from pelvic radiographs across the full pediatric age range. Materials And Methods: This retrospective study included pelvic radiographs from two hospitals for model development, from patients aged between 0 and 18 years. A DenseNet161 model was trained to classify age into 19 1-year interval categories using chronological age as a reference standard and conditional ordinal regression (CORN) loss. External validation was conducted using two independent cohorts, including 541 radiographs from a different department of the same hospital and 203 from an external hospital. External validation was performed on 744 radiographs from two independent cohorts. Model performance was evaluated using the accuracy of predictions within 1 year of the chronological age (ACC@dist = 1) and the mean absolute error (MAE). Results: A total of 3972 radiographs (median age, 2.5 years; interquartile range, 0-8 years; 2463 males) were included in model development. On the internal test set, the model achieved an MAE of 0.394 years (girls: 0.360; boys: 0.419) and an ACC@dist = 1 of 0.937 (girls: 0.944; boys: 0.931). On the external test sets, the corresponding values were 0.513 years (girls, 0.628; boys, 0.455) and 0.927 (girls, 0.896; boys, 0.943), respectively. Conclusion: Automated age estimation in children is feasible from pelvic radiography using a deep neural network. Key Points: Question Pelvic radiographs are commonly obtained in children, yet their potential for automated chronological age estimation across the full pediatric age range remains unexplored. Findings A deep learning model estimated chronological age from pelvic radiographs with a mean absolute error of 0.394 and 0.513 years on internal and external test sets, respectively. Clinical relevance Deep learning-based chronological age estimation from pelvic radiographs is feasible across the full pediatric age range. This approach allows clinically relevant skeletal maturation data to be extracted from imaging already acquired for other indications.

★ An explainable and leakage-aware machine learning model for early risk stratification of radioiodine-refractory differentiated thyroid cancer: development and single-center geographic external validation.

Tan G, Zhang Y, Mao Z, Zhang C, Fu J, Wang F — Mon, 19 Oct 2026 00:00:00 +0000

95% · External validation 95%

Background: Radioiodine-refractory differentiated thyroid cancer (RAIR-DTC) poses a major therapeutic challenge, with loss of radioiodine avidity, limited treatment options, and worse outcomes. Current risk stratification systems were not designed to predict RAIR transformation before clinical refractoriness. Methods: This retrospective two-center study included a development cohort of 302 consecutive patients and an external validation cohort of 122. Candidate predictors were restricted to baseline demographic, clinicopathological, molecular, and biochemical variables available at or before first postoperative RAI administration; all post-treatment, response-related, and time-dependent variables were excluded a priori. Six machine-learning algorithms were developed with repeated cross-validation and compared by discrimination, probability accuracy, and clinical utility. Model interpretation used SHapley Additive exPlanations (SHAP). External transportability was assessed by discrimination, calibration, decision-curve analysis, and validation of prespecified risk groups. Results: Forty-six patients (15.2%) developed RAIR-DTC in the development cohort and 18 (14.8%) in the external cohort. Random forest was selected for the most favorable balance of cross-validated discrimination and probability accuracy, despite comparable external discrimination from XGBoost and the stacked ensemble. It achieved an area under the receiver operating characteristic curve (AUC) of 0.896 and Brier score of 0.078 in the internal test cohort, with robust external discrimination (AUC, 0.892; PR-AUC, 0.597; sensitivity, 0.778; specificity, 0.798). Raw external calibration was poor (calibration slope 0.304; observed-to-expected ratio 0.613), indicating that ranking was preserved whereas absolute risk estimates required local recalibration. SHAP analysis identified age ≥55 years, TERT promoter (TERTp) mutation, vascular invasion, initial stimulated thyroglobulin (sTg), and extrathyroidal extension (ETE) as the dominant contributors. Training-derived quantile cut points generated three-tier risk groups, identifying a low-risk subgroup with zero events in the external cohort and a low event rate (3.3%) in the internal test set, while enriching the high-risk subgroup to event rates of 75.0% and 37.2%, respectively, with 95% confidence intervals of 47.6%-92.7% and 23.0%-53.3%, respectively. Conclusions: An explainable, leakage-aware random forest based exclusively on routinely available baseline variables provided robust external discrimination and clinically meaningful early risk stratification for RAIR-DTC. Preserved discrimination did not guarantee transportable probability estimates, underscoring that external calibration and site-specific recalibration are essential prerequisites for responsible clinical deployment.

Deep learning with interactive segmentation for risk stratification of cystic renal lesions on tri-phase CT: a multicenter study.

Yuan W, He P, He Q, Zhang X, Jiang Q, Yin S — Mon, 19 Oct 2026 00:00:00 +0000

94% · External validation 94%

Background: Accurate preoperative risk stratification of cystic renal lesions (CRLs) remains a clinical challenge, primarily due to morphological overlap which often leads to the overtreatment of indolent cysts. To address this, we developed and validated an integrated deep learning framework that combines interactive segmentation with tri-phase CT feature integration to refine malignancy prediction. Methods: This retrospective multicenter study enrolled 434 patients with CRLs from four institutions, partitioned into training (n=195), validation (n=100), and external test (n=139) cohorts. The reference standard was histopathological diagnosis or ≥4 years of radiological stability. For volumetric segmentation, we implemented an interactive 'human-in-the-loop' Swin UNETR model. To integrate complementary information across the three CT phases, we then deployed a Tri-Phase-Channel (TPC) Perceiver architecture to integrate features extracted by parallel SEResNet50 backbones from unenhanced, corticomedullary, and nephrographic CT phases. Diagnostic performance was evaluated against the single-phase SEResNet50 models and the Bosniak classification version 2019 using the Area Under the Receiver Operating Characteristic Curve (AUC) and Decision Curve Analysis (DCA). Results: In the external test cohort, the interactive segmentation model yielded a Dice score of 0.890. For malignancy prediction, the TPC-Perceiver attained an AUC of 0.969 (95% CI: 0.911-1.000), significantly outperforming the Bosniak classification version 2019 (AUC 0.884; p =0.016). At the fixed probability threshold of 0.5, the model achieved a sensitivity of 94.4% and a specificity of 96.7%, compared with 100.0% and 76.9%, respectively, for the clinical standard. Furthermore, decision curve analysis indicated that the deep learning framework conferred greater net benefit across a broad range of threshold probabilities. Conclusion: The integrated framework showed favorable performance for cystic renal lesion risk stratification. Exploratory pairwise AUC comparisons favored the TPC Perceiver over the single-phase models and the clinical Bosniak v2019 assessment, while its AUC was comparable to that of the previous SETD model. The sensitivity-specificity profile observed at the fixed threshold suggests potential value in reducing false-positive risk stratification. Larger prospective multicenter studies are required before clinical implementation.

Evaluating Explicit Anatomical Priors Beyond a Mature CT Detector for Six Acute Abdominal Emergencies.

Erdoğan HM, Koç U — Mon, 19 Oct 2026 00:00:00 +0000

93% · External validation 93%

To test whether anatomical priors from a general-purpose organ segmentation model (TotalSegmentator, TS), applied at inference or during training, improve an externally validated YOLOv11 detector for six acute abdominal emergencies. The frozen detector was applied to 827 external Stanford Merlin CT studies with language-model report labels; a nested 280-study radiologist-adjudicated subset was the reference standard. TS-derived slice selection (Arm A), organ-state integration (Arm B), and their combination (Arm C) were compared with the detector alone under frozen, recalibrated, and fine-tuned conditions. The detector was also retrained from scratch (identical 10-epoch budget) with image channels only, with the patient's TS organ maps as added input channels, or with another patient's maps. AUROC was primary; comparisons used a paired study-level bootstrap. The detector generalized externally (macro AUROC 0.816 cohort; 0.879 adjudicated). Slice selection was non-inferior but not beneficial: nearly all confident detections already lay within the anchored organ's axial range; organ-state fusion improved AUROC only for cholecystitis (hypothesis-generating). In single runs, training with organ maps raised internal test mAP@0.5 (0.517 versus 0.482), and the model depended on the maps, but external macro AUROC did not improve (0.822 versus 0.829; Δ -0.007, 95% CI -0.019 to +0.007) and kidney-stone AUROC fell (0.678 versus 0.768; exploratory), unlike with another patient's maps. Per-site recalibration was the most effective operating-point adjustment (macro F1 +0.021); annotation-free fine-tuning reduced discrimination. Explicit anatomy, applied after detection or during training, did not improve this detector's overall external discrimination. Per-site recalibration gave the largest operating-point gain.

TyG-ABSI and TyG-WWI as insulin-resistance body shape markers of mortality in middle-aged and older adults within the cardiovascular-kidney-metabolic framework: A multi-cohort observational study with external survival-model validation.

Xiao F, Guo H, Wang L, Qu Y, Dai C — Mon, 19 Oct 2026 00:00:00 +0000

92% · External validation 92%

Insulin resistance, adiposity, kidney dysfunction, and cardiovascular disease are closely interconnected in middle-aged and older adults and are central to the cardiovascular-kidney-metabolic (CKM) framework. Triglyceride-glucose (TyG)-related indices combine routine lipid, glucose, and anthropometric information, but their cross-population consistency and incremental prognostic value for mortality-risk assessment remain uncertain. We analyzed harmonized participant-level data from National Health and Nutrition Examination Survey (NHANES), English Longitudinal Study of Ageing (ELSA), China Health and Retirement Longitudinal Study (CHARLS), and Health and Retirement Study (HRS), with HRS used as a sensitivity cohort. Eligible participants were aged ≥ 50 years, had follow-up beyond baseline, available mortality status, and complete data for 7 TyG-related indices. Cohort-specific Cox models estimated associations of standardized TyG, TyG-body mass index, TyG-waist circumference, TyG-waist-to-height ratio, TyG-weight-adjusted waist index (WWI), TyG-a body shape index (ABSI), and TyG-body roundness index with all-cause mortality. Nonlinearity, proportional-hazards assumptions, sensitivity analyses, NHANES heart disease mortality, incremental prediction, and survival machine learning with external validation were further assessed. The primary analytic sample included 21,044 NHANES participants, 5454 ELSA participants, and 11,157 CHARLS participants, among whom 5566, 116, and 1393 deaths occurred during follow-up, respectively. The HRS sensitivity cohort included 4103 participants and 484 deaths. In fully adjusted Cox models, all 7 TyG-related indices were associated with higher all-cause mortality in NHANES, with the strongest associations observed for TyG-ABSI (hazard ratio [HR] 1.159, 95% confidence interval [CI] 1.115-1.205) and TyG-WWI (HR 1.155, 95% CI 1.113-1.199). In CHARLS, TyG-ABSI was positively associated with mortality (HR 1.088, 95% CI 1.029-1.151), whereas estimates in ELSA were imprecise. In HRS sensitivity analyses, TyG-WWI and TyG-ABSI remained positively associated with mortality. Adding TyG-related indices to a clinical Cox prediction framework produced only minimal absolute gains in discrimination. The final TyG-ABSI extreme gradient boosting survival model yielded 5-year time-dependent area under the curves of 0.808 in the NHANES internal test set and 0.760 in CHARLS. A research-use web calculator was developed for 5-year and 8-year mortality-risk estimation. TyG-ABSI and TyG-WWI were the most consistent TyG-related body shape markers of all-cause mortality across international aging cohorts. Although TyG-ABSI supported externally validated survival prediction, its incremental predictive value beyond established clinical and CKM variables was limited. These findings suggest that TyG-related body shape indices may serve primarily as interpretable, low-cost CKM risk markers rather than standalone prediction tools.