Background: Malnutrition and systemic inflammation are highly prevalent among critically ill patients with bone infections and may independently predict adverse outcomes. Five composite nutrition-inflammation indices-the modified Nutrition Risk in the Critically Ill (mNUTRIC) score, Prognostic Nutritional Index (PNI), Geriatric Nutritional Risk Index (GNRI), Index of Nutritional Assessment (INA), and Hemoglobin-Albumin-Lymphocyte-Platelet (HALP) score-have been proposed as prognostic tools; however, their comparative performance in the bone infection intensive care unit (ICU) population has not been systematically defined.
Material And Method: A dual-cohort retrospective study design was employed. The internal derivation cohort comprised ICU patients with bone infections from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database (n = 917; 28-day ICU mortality 12.1%), and the external validation cohort was drawn from the Affiliated Hospital of Qingdao University (n = 256; 28-day mortality 16.0%). All five indices were calculated from laboratory and clinical data collected within 24 h of ICU admission. Multivariable logistic regression (three hierarchical adjustment models), restricted cubic spline (RCS) analysis, and subgroup analyses were performed to examine the independent associations of each index with 28-day ICU mortality. For predictive model development, a dual LASSO-Boruta feature selection strategy was applied to the derivation cohort, yielding 14 core features (events-per-variable ratio [EPV] = 7.9). Five machine learning (ML) models - logistic regression (LR), random forest (RF), XGBoost, LightGBM, and support vector machine (SVM) - were trained with synthetic minority over-sampling technique (SMOTE) and five-fold stratified cross-validation. Discriminative performance was assessed by the area under the receiver operating characteristic curve (AUROC). The optimal model was further evaluated using SHapley Additive exPlanations (SHAP) values computed directly for the final LightGBM model with external stability validation, calibration curves, and decision curve analysis (DCA). A fully nested cross-validation with feature selection repeated within each training fold confirmed the robustness of the external performance.
Result: A total of 917 ICU patients with bone infections were included in the derivation cohort. In the fully adjusted model, HALP lowest tertile (T1, ≤ 171.11; odds ratio [OR] = 3.41, 95% confidence interval [CI] 2.13-5.46, p < 0.001) and low PNI (< 45; OR = 2.48, 95% CI 1.09-5.66, p = 0.031) remained independently associated with 28-day ICU mortality, whereas the associations of mNUTRIC, GNRI, and INA were attenuated to non-significance after full adjustment. RCS analysis confirmed a significant nonlinear dose-response relationship for HALP (p-nonlinearity < 0.001). In both cohorts, PNI (derivation AUROC 0.761; external 0.742) and HALP (0.745; 0.713) demonstrated the highest discriminative performance among the five indices, while GNRI lost discrimination owing to a near-universal at-risk classification (94.9% of the derivation cohort). Subgroup analysis showed the HALP-mortality association was consistent across all clinical strata (all interaction p > 0.27). Among ML models, LightGBM achieved the highest external AUROC (0.956, 95% CI 0.921-0.985). SHAP analysis of the final LightGBM model identified APS III, acute kidney injury (AKI), and arterial oxygen tension (PO2) as the most important predictors, with the ranking highly stable in the external validation cohort. DCA demonstrated superior net clinical benefit for the ML model over any single index across clinically relevant threshold probabilities.
Conclusion: Among five nutrition-inflammation indices evaluated in critically ill patients with bone infections, HALP and PNI were the leading independent predictors of 28-day ICU mortality, with HALP showing a robust, strongly nonlinear threshold effect that was consistent across all subgroups and adjustment models. In contrast to its dominant role in cardiothoracic populations, GNRI lost discriminative value here because nearly all patients were classified as at-risk. An interpretable ML model integrating 14 routinely available clinical features achieved excellent external discrimination (AUROC = 0.956) and clinical utility, supporting its potential as a decision-support tool for early risk stratification in this high-risk population.
Background: Pathological complete response (pCR) is a key endpoint following neoadjuvant chemotherapy (NAC) in breast cancer, with potential to guide treatment adaptation and surgical de-escalation. However, limited generalizability across institutions remains a major challenge due to heterogeneous imaging protocols (scanner vendors, field strengths, and acquisition parameters) and diverse patient populations, which can cause significant domain shifts in model performance. This study develops and externally validates a multimodal framework for post-neoadjuvant, preoperative prediction of pCR using multiparametric MRI (mpMRI) and routinely available clinical data.
Methods: In a multicenter retrospective setting, 1291 patients from five institutions were enrolled and institutionally divided into a development cohort pooled from two centers (n=683) and three independent, single-center external validation cohorts (n=608). Model inputs comprised post-treatment mpMRI (dynamic contrast-enhanced and diffusion-weighted imaging) for deep feature extraction, clinicopathological variables, radiomics features from both pre- and post-treatment scans, and delta-radiomics features capturing longitudinal within-sequence changes (Δ = post - pre). A dual-tower attention network incorporating a Feature Quality Enhancer (FQE) module-which regularizes the latent space by promoting intra-class compactness, inter-class separability, and feature consistency-was designed to learn discriminative imaging representations. Deep, radiomic, and clinical features were then integrated through a hierarchical feature selection strategy, with robustness assessed via repeated experiments.. Predictive performance was evaluated using the area under the receiver operating characteristic curve (AUC) in both development and external validation cohorts.
Results: The proposed AIFPS achieved an AUC of 0.882 (95% CI: 0.855-0.905) in the development cohort. In external validation, AIFPS achieved the highest numerical AUCs (0.853, 0.888, and 0.895) and demonstrated statistically significant improvements over all single-modality models in all cohorts. While its superiority over certain dual-modality models did not reach statistical significance after correction for multiple comparisons, the full multimodal approach consistently yielded the numerically highest discrimination.
Conclusion: AIFPS integrates quality-enhanced deep imaging features with radiomics, including delta-radiomics, and clinicopathological variables to enable accurate and generalizable preoperative prediction of pathological complete response, particularly in HER2-positive disease, while further validation is required for triple-negative and luminal subtypes. Its application in luminal breast cancer requires further dedicated subtype-specific modeling. With prospective validation, this approach may support individualized treatment stratification following neoadjuvant chemotherapy.
Background: Differentiating papillary thyroid carcinoma (PTC) from benign nodules in patients with Hashimoto's thyroiditis (HT) remains a significant clinical challenge due to the complex sonographic background and overlapping inflammatory features. This study aimed to develop and validate an interpretable machine learning (ML) model integrating ultrasound features and immune-inflammatory biomarkers for PTC identification in HT patients.
Methods: A multicenter retrospective study was conducted, enrolling 1, 200 HT patients with thyroid nodules from Wuxi People's Hospital (January 2020 to December 2025), randomly divided into a training cohort (n=840) and a testing cohort (n=360). An independent external validation cohort (n=550) was recruited from Nanjing Gaochun People's Hospital during the same period. A total of 38 candidate variables, including clinical characteristics, ultrasound features, and immune-inflammatory indices, were evaluated. Least Absolute Shrinkage and Selection Operator (LASSO) and multivariate logistic regression were utilized for feature selection. Seven ML algorithms were compared to construct the optimal predictive model. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) were employed to enhance model interpretability.
Results: Six variables were selected as the most informative features for predicting the pathological outcome of PTC: Systemic Immune-Inflammation Index (SII), Thyroid Peroxidase Antibody (TPOAb), pseudonodule formation, shear wave elastography maximum elasticity (SWE-Emax), vascularity grade, and TI-RADS category. The Random Forest (RF) model demonstrated superior performance, achieving an AUC of 0.907 in the training cohort, 0.872 in the testing cohort, and 0.823 in the external validation cohort. Calibration curves and DCA confirmed the model's high clinical net benefit. SHAP analysis revealed that SII, TPOAb, and pseudonodule formation were the top three contributing features. A web-based risk calculator was subsequently deployed for clinical application.
Conclusion: The proposed interpretable RF model effectively integrates ultrasound features and immune-inflammatory biomarkers to accurately identify PTC in HT patients. The developed web-based calculator provides a practical, non-invasive tool to assist clinicians in personalized decision-making.
Lu SJ, Wei ZK, Li GC et al. · Front Med (Lausanne)
External validation 98%
Abstract
Background: Early hemodynamic stabilization after vasopressor initiation in septic shock is heterogeneous and difficult to predict. We aimed to develop and externally validate interpretable machine learning models for predicting early hemodynamic stabilization.
Methods: This retrospective study used MIMIC-IV for model development and internal validation and eICU for external validation. Adult ICU patients with septic shock receiving vasopressors were included. Dynamic clinical features during the first 6 h after vasopressor initiation were used to predict stabilization during 6-24 h, defined by sustained mean arterial pressure (MAP) control without vasopressor escalation or increased norepinephrine-equivalent dose. Seven machine learning models were compared. Performance was assessed using discrimination, calibration, Brier score, decision curve analysis, external validation, and SHAP.
Results: The MIMIC-IV and eICU cohorts included 3,445 and 453 patients, with stabilization rates of 41.77% and 32.67%, respectively. Gradient boosting machine (GBM) showed the best overall performance, with an AUC of 0.720 and Brier score of 0.209 in internal validation and an AUC of 0.730 and Brier score of 0.192 in external validation. At the training-derived threshold, external sensitivity and specificity were 0.466 and 0.797. SHAP identified early MAP stability, lactate clearance, platelet count, renal function, inflammatory burden, and vasopressor intensity as major predictors. The stabilization endpoint was associated with ICU and hospital mortality and remained robust to alternative MAP definitions. However, predictive performance was substantially driven by proximal hemodynamic features, particularly MAP trajectories, and the full GBM did not consistently outperform simpler hemodynamic benchmarks externally.
Conclusions: Early dynamic clinical features provided moderate discrimination for predicting hemodynamic stabilization after vasopressor initiation, with similar external performance. The model should be interpreted as a structured risk-stratification tool rather than a measure of pharmacological vasopressor responsiveness. Prospective validation is required before clinical implementation.
Background: Mechanical ventilation requires repeated adjustment to changing patient physiology, but consistent individualized management remains challenging. We developed VentPilot, an artificial intelligence-based system for recommending ventilator settings, and evaluated it in multicenter retrospective validation cohorts.
Methods: VentPilot was developed using offline reinforcement learning on high-resolution physiologic and ventilator trajectories from Seoul National University Hospital (SNUH), with a reward combining ventilator-free days and intensivist-derived preference feedback. We evaluated VentPilot in an internal validation cohort from SNUH and an external validation cohort from Mayo Clinic. Fitted Q-evaluation assessed the estimated return of VentPilot relative to observed clinician behavior under the prespecified reward function. In a complementary inverse probability-weighted analysis, clinical outcomes were compared between patients with higher versus lower concordance between observed ventilator settings and VentPilot recommendations.
Results: Among 4296 mechanically ventilated adults, 3002 were included in the derivation cohort, 283 in the internal validation cohort, and 1011 in the external validation cohort. Under the prespecified reward function, fitted Q-evaluation estimated higher returns for VentPilot than for observed clinician behavior in both validation cohorts, with differences in overall reward of 1.6 (95% CI 0.8-2.5) and 1.3 (95% CI 0.6-2.0), respectively. In inverse probability-weighted analyses, higher concordance with VentPilot recommendations was associated with more ventilator-free days within 28 days, with mean differences of 5.0 days (95% CI 2.4-7.6) and 3.1 days (95% CI 1.7-4.6), respectively. Higher concordance was also associated with lower 28-day mortality and shorter ICU length of stay.
Conclusions: In multicenter retrospective validation cohorts, fitted Q-evaluation estimated higher returns for VentPilot than for observed clinician behavior under the prespecified reward function, while greater concordance between observed care and VentPilot recommendations was associated with more favorable clinical outcomes. Further clinical evaluation is warranted to establish VentPilot's safety, usability, and clinical impact.
Tseregounis IE, Shin E, Henry SG et al. · JAMIA Open
External validation 97%Prospective 54%
Abstract
Objective: Prescription drug monitoring programs (PDMPs) offer a promising data source for the development of clinical decision support (CDS) systems to guide opioid prescribing, but most PDMP agencies lack capacity for in-house model development-instead relying on third-party tools. This study aims to externally validate a PDMP-based model, developed using 2018 California PDMP records, to predict new long-term use among Kentucky opioid-naïve patients and compare its performance to Kentucky-specific machine learning (ML) models.
Materials And Methods: We identified opioid-naïve Kentucky residents who filled an index pill opioid prescription between 2021 and 2023 (development cohort: 2021-2022; validation cohort: 2023). The California model was updated via logistic calibration and evaluated against 5 newly developed Kentucky ML models (least absolute shrinkage and selection operator [LASSO], elastic net, deep neural network, random forest, XGBoost) using discrimination and calibration metrics.
Results: Among 985 317 opioid-naïve patients, 33 298 (3.4%) transitioned to new long-term opioid use. The updated California model achieved excellent discrimination (area under the receiver operating characteristic curve [AUROC]: 0.919 [0.915-0.923]) and was calibrated up to the highest risk levels (calibration intercept: 0.010, [-0.021, 0.041]; slope: 0.997, [0.979, 1.014]), with negative predictive values exceeding 0.99. Performance was comparable to Kentucky-specific ML models, including the highest performer XGBoost (AUROC: 0.930 [0.926-0.933]). All models struggled with low positive predictive values (0.15-0.19).
Conclusion: A California PDMP-derived prediction model transported successfully to Kentucky with minimal updating. Its strong comparative performance provides a feasible foundation for PDMP agencies to develop transparent, in-house CDS systems with potential clinical implications for safer and more individualized opioid prescribing.
Huang H, Feng L, Zhong ME et al. · PLOS Digit Health
External validation 96%
Abstract
MRI-detected extramural vascular invasion (mrEMVI) is an important prognostic biomarker in rectal cancer, reflecting tumor invasiveness and metastatic potential. To address the subjectivity and inter-observer variability inherent in manual mrEMVI assessment, this study trained and validated an nnUNet-based segmentation model for automated voxel-level localization and visualization of mrEMVI. This multi-center retrospective study included a total of 2,501 rectal cancer patients, comprising 1,830 in the training cohort (with 5-fold cross-validation) and 671 in two independent external test cohorts. The Dice similarity coefficient was used to evaluate segmentation performance; the inter-reader agreement for mrEMVI identification was assessed using Cohen's kappa (κ). The prognostic value of mrEMVI status identified by the artificial intelligence (AI) model was evaluated using Kaplan-Meier survival analysis and multivariable Cox regression. In internal five-fold cross-validation, the model yielded Dice scores of 0.850, 0.442, and 0.335 for tumor, intravascular tumor signal, and dilated vessel segmentation, respectively. The model demonstrated strong classification performance, with accuracies of 81.5% (95% CI: 76.7%-85.8%) and 84.7% (95% CI: 80.7%-88.2%) in the two external test cohorts, and achieved substantial agreement with senior radiologists (κ = 0.713-0.736). Patients identified as AI-mrEMVI+ had significantly lower 3-year disease-free survival (DFS) and 5-year overall survival (OS) rates than AI-mrEMVI- patients (DFS: 62.3% vs. 84.9%, HR = 2.67, 95% CI: 1.95-3.66; OS: 68.7% vs. 87.1%, HR = 2.64, 95% CI: 1.75-3.97; both p < 0.001). This work establishes a scalable, objective framework for mrEMVI assessment based on voxel-level segmentation, inter-observer agreement analysis, and comprehensive prognostic validation, with direct implications for risk stratification and treatment individualization in rectal cancer.
Yin K, Su H, Zhang X et al. · J Imaging Inform Med
External validation 96%
Abstract
Accurate preoperative classification of pulmonary nodules (PNs) is important for clinical decision-making and avoiding overtreatment. However, marked heterogeneity among pathological subtypes limits the performance of conventional single-model classification approaches. This study aimed to develop and validate an expert committee-driven adaptive framework based on AI-derived quantitative CT features for differentiating four clinically relevant PN subtypes: atypical adenomatous hyperplasia and adenocarcinoma in situ (AAH + AIS), minimally invasive adenocarcinoma (MIA), invasive adenocarcinoma (IAC), and inflammatory nodules (IN). This retrospective study included 491 pathologically confirmed PNs from a single center as the internal cohort. Quantitative CT features were automatically extracted using an AI-assisted platform. A heterogeneous pool of candidate classifiers was constructed, and an expert committee-driven arbitration mechanism was implemented to enable subtype-specific optimal model selection based on validation performance. Diagnostic performance was primarily assessed using the area under the receiver operating characteristic curve (AUC). Two independent multicenter cohorts were used for external validation. In internal validation, the proposed framework achieved AUCs of 0.918 for AAH + AIS, 0.860 for MIA, 0.913 for IAC, and 0.913 for IN. External validation demonstrated consistent performance across centers, with improved sensitivity for heterogeneous subtypes, particularly MIA and IN, compared with conventional single-model approaches. An expert committee-driven adaptive framework enables subtype-specific model selection for pulmonary nodule classification and demonstrates stable diagnostic performance across multicenter cohorts. This approach supports the potential clinical utility of AI-derived quantitative CT features for noninvasive preoperative assessment of pulmonary nodules.
Complications occurring early after robot-assisted radical prostatectomy (RARP) can hinder postoperative recovery. Existing risk instruments, however, generally incorporate few predictors and assume linear associations. We aimed to derive and externally validate an interpretable machine-learning model that combines clinical, laboratory, and intraoperative information and to translate the model into an online risk calculator. Records were reviewed for 1,254 patients with prostate cancer (PCa) who received RARP at the First, Third, or Fifth Medical Center of the Chinese PLA General Hospital. Among them, 1,001 patients from the First and Fifth centers were randomly allocated 7:3 to training (n = 701) and internal testing (n = 300), while 253 patients from the Third center provided an independent external cohort. Sixty-four candidate variables underwent univariate screening and least absolute shrinkage and selection operator regression. The tested algorithms were gradient boosting decision trees (GBDT), random forest, logistic regression, support vector machine, AdaBoost, Gaussian naive Bayes, and multilayer perceptron. We examined discrimination with the area under the receiver operating characteristic curve (AUC), agreement between predicted and observed risk with calibration plots, and clinical usefulness with decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) were used to interpret predictions. The leading model was also benchmarked against the Comorbidity Score for Robotic Surgery (CRS). Feature selection yielded eight predictors for model training. For the multilayer perceptron (MLP), the AUC was 0.880 on internal testing and 0.843 in the independent external cohort. Across both evaluations, this algorithm offered the most balanced combination of performance measures, good calibration, and favorable clinical net benefit. SHAP ranked crystalloid infusion as the leading predictor; operative time and thrombin time (TT) ranked next. In an exploratory comparison in the external cohort, MLP showed greater potential than CRS in discrimination, calibration, and net benefit. We visualized the final model and deployed it as a web-based calculator. The interpretable MLP model accurately estimated the likelihood of early postoperative complications among patients with PCa undergoing RARP. Its accompanying online calculator may assist clinicians in recognizing high-risk patients as soon as surgery is completed and may inform individualized risk classification, postoperative monitoring, and clinical care.
Yang G, Huang K, Luo H et al. · Front Cell Infect Microbiol
External validation 95%
Abstract
Background: The lactate dehydrogenase-to-albumin ratio (LDAR) reflects tissue damage, inflammation, and nutritional status. Its prognostic value in urosepsis remains unexplored. We evaluated LDAR's association with short-term mortality and developed an LDAR-based prediction model.
Methods: This retrospective dual-cohort study used MIMIC-IV (n=3,060) and an external cohort (n=370). Exposure was log2-transformed LDAR (logLDAR); outcomes were 28-day intensive care unit (ICU) and in-hospital mortality. Cox regression with three adjustment models, restricted cubic splines (RCS), and Kaplan-Meier curves assessed associations; Fine-Gray models addressed competing risks. DeLong tests compared area under the curve (AUC) of six conventional scores with and without logLDAR. Net reclassification improvement (NRI) and integrated discrimination improvement (IDI) quantified incremental value of adding logLDAR; Machine learning selected predictors for a parsimonious model validated internally and externally. decision curve analysis (DCA) evaluated the 7-variable model's utility.
Results: Among 3,060 patients, higher logLDAR was associated with significantly elevated 28-day ICU (23.4% vs. 9.02%) and in-hospital (22.1% vs. 8.24%) mortality (both P<0.001). In the primary multivariable-adjusted model, per-unit increases in logLDAR yielded hazard ratios (HRs) of 1.26 (95% CI: 1.17-1.36) for ICU mortality and 1.35 (95% CI: 1.25-1.45) for in-hospital mortality. Compared with the lowest tertile, the highest tertile showed HRs of 1.79 (1.40-2.30) and 2.29 (1.77-2.96), respectively. RCS analysis revealed linear dose-response relationships (P<0.001), and the Fine-Gray model confirmed robust associations (subdistribution hazard ratio (sHR)=1.36, 1.27-1.46). Adding logLDAR significantly improved the AUCs of all six conventional scores (DeLong P<0.01), with NRI/IDI showing substantial reclassification improvement (P<0.001). A parsimonious 7-variable model achieved AUCs of 0.674, 0.677, and 0.783 in the internal training, internal validation, and external cohorts, respectively, with adequate calibration (Hosmer-Lemeshow P > 0.05) and DCA demonstrating net clinical benefit. These findings were fully validated in an external real-world cohort.
Conclusions: LDAR independently predicts short-term mortality in critically ill patients with urosepsis, supplementing conventional severity scores. The parsimonious risk model showed moderate performance with acceptable cross-cohort generalizability.
Background: Aortic valve calcification (AVC) quantification is a guideline-recommended imaging biomarker for aortic stenosis (AS) severity and progression yet remains underreported on routine chest computed tomography (CT). With nearly 20 million nongated chest CTs performed annually in the United States, automated AVC detection offers potential for opportunistic AS screening without added radiation or cost.
Objectives: This study evaluated whether deep learning can accurately quantify AVC from nongated, noncontrast chest CT, with performance comparable to expert assessment.
Methods: We evaluated a convolutional neural network across a multicenter consortium of 33 sites spanning 8 health systems in the United States and Brazil. The algorithm was developed using training (n = 1,446) and validation (n = 361) sets and evaluated on a separate holdout test set (n = 239) acquired from 2021 to 2024. Reference standards were manual segmentations independently verified by at least 2 board-certified radiologists. Primary outcomes were Pearson correlation and Bland-Altman analysis of algorithm-estimated vs manual Agatston scores. Secondary outcomes were sensitivity and specificity for a high calcium burden associated with moderate-to-severe AS using validated sex-specific thresholds.
Results: The algorithm showed high correlation with expert reference standards (r = 0.99; 95% CI: 0.98-0.99; P < 0.001). Bland-Altman analysis showed minimal bias (mean difference: 5.2 AU; 95% CI: -7.6 to 17.9). For moderate-to-severe AS, sensitivity was 0.92 (95% CI: 0.80-0.97) and specificity 0.98 (95% CI: 0.95-0.99). Performance remained consistent across demographic subgroups and CT technical parameters.
Conclusions: This automated deep learning algorithm quantifies AVC from routine chest CT with expert-comparable performance, potentially enabling opportunistic AS screening. Prospective studies are needed to determine whether it improves clinical outcomes.
Duan X, Wang L, Fan Y et al. · Front Cell Infect Microbiol
External validation 94%
Abstract
Background: Infectious spondylitis, encompassing spinal tuberculosis (STB), pyogenic spondylitis (PS), and brucellar spondylitis (BS), is difficult to diagnose preoperatively because the three entities present with overlapping clinical features. Differentiation is often delayed by prolonged culture turnaround times and the unavailability of serological testing.
Methods: In this multicenter retrospective diagnostic study at five tertiary hospitals in China (2015-2023), patients with STB (n = 1,000), PS (n = 579), and BS (n = 255) were enrolled and assigned to derivation (n = 1,439) and external validation (n = 395) cohorts. We developed and validated a multimodal machine learning model that integrates routine preoperative clinical and laboratory variables with radiomic features extracted from non-contrast MRI. Unimodal, early fusion, and late fusion architectures were compared. Pathogen-specific serological tests were excluded from the primary model to avoid data leakage. We evaluated discrimination, calibration, clinical utility, and incremental value; SHAP analysis provided case-level interpretability.
Results: The late fusion LightGBM model achieved the highest external validation performance (macro-averaged AUC 0.852, 95% CI 0.822-0.882). Radiomic integration provided significant incremental value over the clinical-only model (ΔAUC 0.080; NRI 0.398; IDI 0.072). Calibration was good (Brier score 0.158), and decision curve analysis demonstrated positive net benefit across clinically relevant thresholds. SHAP analysis revealed class-specific, biologically plausible predictors: low serum albumin and elevated T1W GLCM contrast for STB; high CRP and NLR for PS; and residence in an endemic region for BS. Secondary analysis showed that adding serological tests yielded only modest incremental improvement (ΔAUC 0.042).
Conclusions: This preoperative multimodal explainable AI model achieves accurate and interpretable differentiation of STB, PS, and BS using only routine clinical data and non-contrast MRI, although its sensitivity for BS remains limited and prospective validation is warranted. It may serve as a clinically actionable decision-support tool for preoperative triage, particularly when serological confirmation is pending or unavailable.
Background: The long-term survival outcomes of older patients with esophagogastric junction adenocarcinoma (EGJA) remain unclear. This study aimed to develop and validate a machine learning model to predict postoperative survival in patients aged ≥ 75 years with EGJA.
Methods: This retrospective study analyzed patients aged ≥ 75 years who underwent surgical resection for EGJA between 2009 and 2018 across nine tertiary hospitals in China. A random survival forest (RSF) model was developed to predict overall survival (OS) and disease-free survival (DFS) using key clinicopathological variables. Model performance was compared with traditional Cox proportional hazards (CPH) regression and TNM staging.
Results: The study included 1,202 patients. In the training set, the RSF model demonstrated superior discriminatory ability for 5-year OS (time-dependent AUC: 0.887 vs. 0.78) and DFS (0.836 vs. 0.767) compared to the CPH model. The RSF model also showed better calibration with higher C-index (CiD) for OS (0.723 vs. 0.704) and DFS (0.746 vs. 0.701). In the test set, the RSF model maintained robust performance, with CiD values of 0.719 for OS and 0.738 for DFS, outperforming the CPH model (0.704 and 0.697, respectively). In the independent external validation cohort, the RSF model maintained favorable performance and outperformed the CPH model for OS and DFS prediction, with CiD values of 0.720 for OS and 0.711 for DFS, compared with 0.687 and 0.679 for the CPH model, respectively. The RSF model exhibited particularly strong predictive power for both OS and DFS between 12 and 36 months post-surgery. Key predictors included pN stage, pT stage, age, and age-adjusted Charlson Comorbidity Index.
Conclusion: The machine learning-based RSF model provides individualized and accurate survival predictions for older patients with EGJA after surgical resection. This model may assist clinicians in postoperative risk stratification and follow-up planning.
Liang J, Liu Y, Wang W et al. · Cyborg Bionic Syst
External validation 93%Language models 84%
Abstract
Multimorbidity is strongly associated with mortality, but existing comorbidity indices often generalize poorly across populations because they depend on fixed disease lists and context-specific weights. We developed the large-language-model-based multimorbidity score (LLM-MMS), a prompt-based large-language-model-derived multimorbidity phenotype score, as a bio-inspired artificial intelligence score for human-machine clinical risk stratification and evaluated it for all-cause mortality prediction in 770,439 adults from 9 international longitudinal cohorts. Harmonized participant-level data were converted into standardized health narratives and processed without model fine-tuning. In the UK Biobank, LLM-MMS achieved strong discrimination, calibration, and prediction accuracy. Across 8 external cohorts, it maintained consistent performance, with C-indices of 0.67 to 0.86 and observed-to-expected ratios close to 1.00. This outperformed recalibrated conventional comorbidity indices in every cohort with absolute C-index gains of 0.12 to 0.21. Performance remained stable across sex and age subgroups, including adults younger than 60 years, and decision-curve analysis showed greater or comparable clinical net benefit in most cohorts. Proteomic and interpretability analyses supported the biological plausibility of LLM-MMS, linking high multimorbidity burden to inflammatory, metabolic, and tissue-damage pathways and identifying clinically coherent drivers of risk. These findings suggest that LLM-MMS offers a transportable and interpretable bio-inspired framework that integrates clinician-like multimodal reasoning with population-level validation for multimorbidity risk stratification across diverse populations.
While most biomarkers currently rely on invasive laboratory testing, which limits large-scale or repeated screening, scalable non-invasive methods could transform population screening and personalized health management. Facial photographs, as a ubiquitous and non-invasive data source, offer such potential but remain underexplored for clinical biomarker prediction. Existing supervised approaches are constrained by the availability of clinically labeled data, whereas general-purpose vision models are optimized for non-clinical tasks and may inadequately capture subtle, multiscale, and spatially distributed clinical facial features. To address these limitations, we developed MedicalFaceFound, a facial foundation model pretrained on over 10 million images and evaluated across 62 biomarkers spanning eight physiological systems. It performed best among the Swin-Large and ResNet18 for 45 (73%) biomarkers and generalized across four independent external cohorts (median Pearson's r = 0.172), with strong performance for RBC, eGFR, and HDL-C (median r = 0.478, 0.442, and 0.430, respectively). MedicalFaceFound also outperformed polygenic risk score models for 14 of 26 biomarkers, and face-estimated cardiovascular biomarkers were strongly associated with coronary stenosis (AUC = 0.66). The model retained predictive performance with 400 labeled samples and was feasible to deploy as a smartphone application, supporting scalable, non-invasive biomarker assessment and personalized risk screening.
Objective: To evaluate the predictive information contained in the restricted surveillance feature set (demographic, temporal and specimen variables) for respiratory pathogen identification, and to examine factors associated with model performance.
Methods: We retrospectively analysed 24,689 acute respiratory infection cases from seven sentinel hospitals in Sichuan, China; after quality control, 21,395 samples with complete 21-pathogen testing were included. Per-pathogen multi-label binary classifiers were trained using 15 features available at presentation; logistic regression, random forest, back-propagation neural network and XGBoost were compared under an identical split and threshold-selection protocol, with thresholds derived from out-of-fold probabilities. Label-definition sensitivity, leave-one-hospital-out cross-validation and prevalence-performance association were analysed.
Results: Random forest and XGBoost performed comparably (XGBoost the primary model), with Macro-F1 0.1553 (95% CI 0.104-0.2091), 0.2202 for common pathogens and Macro-AUC 0.7702; the four-algorithm spread was 0.0611. Log-prevalence correlated strongly with per-pathogen F1 (r = 0.851); the seven rare pathogens reached Macro-F1 0.0255. Leave-one-hospital-out validation gave Macro-F1 0.1073 ± 0.0289 (-30.9%) and Macro-AUC 0.6841 (-11.2%). Feature ablation showed temporal and non-temporal features were largely redundant (Macro-F1 0.1153 vs. 0.1147; AUC 0.6892/0.6865 vs. 0.7702). Treating untested as negative lowered Macro-AUC to 0.7653, whereas simulated false-negative label contamination had minimal effect.
Conclusion: Models built on the restricted surveillance feature set showed limited diagnostic performance and were not demonstrated suitable for clinical deployment. Performance was strongly associated with pathogen prevalence and constrained by limited feature informativeness, while algorithm choice contributed little and cross-centre generalisability was limited. Priorities include multimodal feature integration, rare-pathogen accrual and prospective multicentre external validation.
Guo Y, Wang S, Lai J et al. · Front Endocrinol (Lausanne)
External validation 92%
Abstract
Objective: To develop an explainable hybrid deep learning features framework using low-dose chest CT (LDCT)-derived pectoralis features for opportunistic type 2 diabetes mellitus (T2DM) opportunistic screening, evaluating its incremental value over conventional biomarkers.
Methods: This multi-center study analyzed LDCT images from 1,209 individuals. Following automated MedSAM segmentation, conventional metrics (e.g., Mean_HU, SMI) and deep learning (DL) features (ResNet152) were extracted. To prevent data leakage, PCA, LASSO selection, and ExtraTrees model optimization were strictly confined to the training set. A hybrid (COM) model integrating both feature types was compared against a baseline Muscle model. SHapley Additive exPlanations (SHAP) provided clinical interpretability.
Results: The COM model achieved an area under the curve (AUC) of 0.830 (training), 0.759 (internal validation), and 0.746 (external validation). While performing comparably to the DL model without significant difference across validation cohorts (DeLong p > 0.05), it significantly outperformed the baseline Muscle model (external AUC: 0.515; DeLong test, p < 0.001). SHAP analysis confirmed that lower Mean_HU correlated with higher T2DM risk, while DL signatures provided robust, independent diagnostic value.
Conclusion: This explainable hybrid model and DL model capture sub-visual pectoralis alterations, offering significant incremental diagnostic value over simple clinical and imaging metrics. They serves as an efficient, non-invasive tool for opportunistic T2DM screening during routine LDCT.
Timely detection of fetal distress during labor is a central pursuit of modern obstetric care. The primary measure used in this regard is fetal heart rate and contraction monitoring (echocardiotocography or CTG), which provides a continuous signal during labor. The vast majority of machine learning applications to CTG assume access to a 30 to 60 minute window just prior to delivery, which makes an untenable assumption that the timing of delivery is known. In this work, we reframe the problem as a time-to-event prediction task. We develop deep time-to-event models to predict the joint probability of imminent delivery and fetal acidosis, a biomarker of fetal hypoxia collected from umbilical cord gas after delivery. We develop a framework for this problem, drawing on deep discrete-time survival methods and marked point processes, adapting each to this particular setting (i.e., continuous physiological signals, fully observed events, partially observed labels). One approach dubbed Marked DeepHit accurately estimates delivery within one hour at multiple elapsed-time landmarks, with test-set AUROC of 0.837 [95% CI 0.692-0.970] and 0.935 [0.858-0.964] at 6 and 12 hours into labor, respectively. We find that these models also achieve competitive performance near delivery time compared to prior models trained specifically on the end of tracing, and outperform other clinical feature-based proposals, both on internal and external validation at a second site. These results give a proof-of-principle to support future prospective deployment of AI-EFM during labor.
Background & Aims: We previously developed multiple machine-learning and regression-based risk prediction models using retrospective electronic health records (EHR) to estimate near-term risk of pancreatic cancer (PC). This pilot study aimed to assess the feasibility of an early detection strategy combining initial EHR-based risk prediction with downstream imaging and blood-based testing.
Methods: This prospective study included adults aged 50-84 years with an estimated ≥1% 18-month risk of PC, identified between January 2021 and March 2023 using four previously validated models. Patients with a history or active suspicion of PC were excluded. Participants completed cross-sectional imaging and CA19-9 labs at baseline and again at 18 months. Outcomes included enrollment rates, test completion rate, and burden of incidental findings.
Results: Of 1,251 eligible patients, 102 (8.2%) enrolled (median age 69 years [IQR 62-75]; 58% women; 48% Hispanic, 34% White, 11% Black, and 7% Asian). At baseline, 89 (87%) completed imaging and 98 (96%) completed CA19-9. At 18 months, 79% completed both imaging and CA19-9. Ten (10%) participants underwent additional clinical evaluations triggered by imaging findings. Evaluations included an endoscopic ultrasound, renal imaging, and FibroScan. CA19-9 results were normal in most participants, though 18% had elevated levels at baseline.
Conclusions: Real-time EHR-based ML algorithms can be combined with blood- and imaging-based surveillance for early detection of PC with acceptable burden of incidental findings. However, improved patient engagement is a key aspect that will need to be addressed in order to successfully scale an AI-guided strategy for early detection.