2026-10-01

Objectives: To identify risk factors for ureteral stricture after endoscopic lithotripsy for impacted ureteral stones and to develop and externally validate a machine-learning-based prognostic model. Methods: This multicenter study enrolled patients with impacted ureteral stones in China. During model construction, controls were matched 1:1 with stricture cases based on gender, age, year of surgical procedure, and surgeon to minimize the influence of variability in surgical equipment and techniques. The primary outcome was postoperative ureteral stricture. The model development cohort consisted of patients who underwent endoscopic laser lithotripsy at the Second Hospital of Tianjin Medical University between 2012 and 2021. External validation was performed using independent datasets from the Tongji Hospital of Tongji Medical College, the Huazhong University of Science and Technology, and the Tianjin Medical University General Hospital between 2022 and 2023. Results: Among the six machine learning models, the least absolute shrinkage and selection operator regression model demonstrated the best overall performance, achieving an area under the receiver operating characteristic curve (AUC) of 0.814 in the internal cohort and AUCs of 0.794 and 0.782 in the two external validation cohorts, respectively. The final model incorporated seven readily available clinical features: residual stone, ureteral wall thickness, stone burden, stone culture, surgical method, history of ureter-related procedure, and intraoperative ureteral injury grade. To enhance clinical usability, the model was further translated into a user-friendly predictive tool. Conclusions: The developed model accurately predicts the risk of ureteral stricture using perioperative clinical variables. This tool may assist clinicians in surgical planning and postoperative management for patients with impacted ureteral stones.

Background: To develop and independently validate a deep learning framework for multiclass differentiation of osteoarthritis (OA), rheumatoid arthritis (RA), gouty arthritis (GA), and normal controls on hand digital radiographs. Methods: This retrospective multicenter study included 1,179 hand radiographs from 520 participants at four institutions. A three-center development cohort comprised 1,001 radiographs from 431 participants and was evaluated using patient-level stratified five-fold cross-validation. ResNet50, ResNet101, DenseNet121, and Vision Transformer (ViT) were trained using an identical pipeline. The internally best model, ViT, was retrained on the complete development cohort and evaluated once on an independent fourth-center cohort of 178 radiographs from 89 participants that was held out from training and model selection. Results: ViT achieved the numerically highest internal performance, with an accuracy of 0.937 ± 0.029 and a macro-F1 score of 0.937 ± 0.029. On independent external validation, ViT correctly classified 133 of 178 radiographs, with accuracy of 0.747, macro-precision of 0.756, macro-recall of 0.747, macro-specificity of 0.916, macro-F1 of 0.744, and macro-AUC of 0.917. External class-wise AUCs were 0.948 for RA, 0.873 for OA, 0.888 for GA, and 0.959 for normal controls. Conclusion: The transformer-based model showed strong internal performance and retained useful discriminative ability in an independent external cohort; however, the reduction in external accuracy and macro-F1 indicated limited cross-center transferability. These findings support the feasibility of the proposed approach while emphasizing the need for broader prospective validation and reader-performance studies before clinical deployment.

Objective: This study aimed to construct a new machine learning model based on multiple risk factors to predict the prognosis of patients with biliary tract malignancies after surgery. Materials And Methods: This study included 6,951 patients in the SEER database from 2004 to 2015, and 245 patients in our medical center from 2015 to 2022. The patients in the SEER database were divided into a training set and a test set in a 7:3 ratio. The data from this research center serves as the external validation set. The clinical outcome was overall survival (OS). Univariate and multivariate Cox regression analyses were used in the training set to screen out independent prognostic factors. Then, multiple machine learning models were constructed to predict the survival rates of patients at 1 year, 3 years and 5 years. The performance of the model was evaluated using the area under the curve (AUC), calibration curves, and the decision curve analysis (DCA). SHAP analysis was used to identify features with greater contributions to model prediction. Subgroup analysis and survival curves were used to evaluate the benefits of chemotherapy for patients with positive lymph nodes. Results: Age, sex, marital status, tumor site, tumor differentiation, AJCC stage, regional nodes examined, regional nodes positive, and LN surgery Scope were all independent risk factors affecting the OS of patients with biliary tract malignancies. Among the various models constructed, the random forest (RF) model showed good AUC values across different datasets. Moreover, the calibration curves and decision curve analysis of the RF model in different sets indicate that it has excellent performance. SHAP analysis indicates regional nodes positive, AJCC stage, tumor site, and age play a crucial role in the survival analysis of biliary tract malignancies. Chemotherapy was associated with improved OS among patients with positive lymph nodes in the training and test cohorts. Conclusions: This study developed a new machine learning model based on multiple risk factors, which can effectively predict the prognosis of patients with biliary tract malignancies after surgery. This model may assist postoperative risk stratification and individualized adjuvant treatment decision-making.

Background: Trauma-induced coagulopathy (TIC) is characterized by marked heterogeneity, and conventional classification methods have limited ability to capture its dynamic progression, frequently failing to account for inter-individual variability. This study aimed to develop a framework for identifying glucose trajectory phenotypes based on early blood glucose trajectories and to evaluate their association with 28-day mortality. Methods: A retrospective analysis was performed using data from 905 trauma patients in the discovery cohort obtained from the MIMIC-IV database and 260 trauma patients in a validation cohort from Hebei Medical University Third Hospital. Least absolute shrinkage and selection operator regression analysis was applied to identify prognostic variables, followed by the development and comparative evaluation of 12 machine learning models. Group-based trajectory modeling (GBTM) was used to characterize dynamic blood glucose trajectory patterns during the first 7 days following admission, and their associations with clinical outcomes were examined. Results: Six key prognostic variables were identified: red cell distribution width (RDW), white blood cell count, blood glucose, anion gap, lactate, and pH. The logistic classifier model achieved an area under the curve of 0.785 in the discovery cohort and 0.761 in the validation cohort, with acceptable calibration (ECE = 0.119). Among 184 eligible patients in the discovery cohort. GBTM identified three distinct blood glucose trajectory patterns: declining-then-stabilizing, persistently fluctuating, and stabilizing-then-rising. The stabilizing-then-rising trajectory pattern was significantly associated with the highest 28-day mortality (p = 0.034, adjusted OR 3.879, 95% CI 3.101-4.886, p < 0.001 in the pooled analysis, n = 1,165), a finding that was further confirmed in the validation cohort (p < 0.001). Conclusion: Application of GBTM to early blood glucose trajectories enables the identification of glucose trajectory phenotypes associated with differential mortality risk. This approach may provide a novel framework for dynamic risk stratification and targeted clinical management.

Objective: To evaluate the impact of incorporating community-level social determinants of health (SDoH) on the performance and algorithmic fairness of artificial intelligence (AI) models trained to predict glaucoma progression to surgery. Materials And Methods: This multicenter study included 63 268 glaucoma patients across 12 centers. Using electronic health records and community-level SDoH features (Distressed Communities Index [DCI] and rural-urban commuting area [RUCA] codes), we developed XGBoost models to predict progression to incisional glaucoma surgery. We compared SDoH-unaware and SDoH-aware models across an internal test set and 3 diverse external test sets, evaluating area under the receiver operating characteristic curve (AUROC) (performance) and generalized equalized odds (GEO) (fairness). Results: Including SDoH features resulted in negligible changes in predictive performance across the internal and all external test sets. However, the SDoH-aware model improved algorithmic fairness (achieving lower GEO scores) in the internal set (0.098 vs 0.216) and a mixed-population external site (0.172 vs 0.378). Conversely, inclusion of SDoH marginally improved at a highly urbanized, low-DCI external site (0.491 vs 0.504) and failed to improve at a rural, high-DCI external site (0.241 vs 0.228). Discussion: Incorporating SDoH features can improve algorithmic fairness regarding socioeconomic disparities without sacrificing predictive accuracy; however, these benefits do not universally transfer to external sites with demographically dissimilar cohorts. Conclusion: Incorporating community-level SDoH features into glaucoma prediction models improved algorithmic fairness with respect to socioeconomic disparities, without significantly altering predictive accuracy. However, fairness benefits did not universally generalize to demographically dissimilar external cohorts. AI models require rigorous external validation across diverse populations to ensure equity prior to clinical implementation.

Integrating histopathology, genomic, and clinical phenotypes holds promise for improving prognostic patient stratification, including within treatment-defined subgroups, in colorectal cancer (CRC). However, ensuring the availability of all modalities in routine clinical practice remains challenging. Comprehensive molecular profiling is often constrained by cost and turnaround time, leaving histopathology as the most consistently available modality across institutions. In addition, variability in staining protocols, scanning devices, patient demographics, and outcome distributions across centers undermines the generalizability of models trained under controlled conditions with complete inputs. To address these challenges, we developed the Domain-Adaptive Incomplete Multimodal Stratification framework (DAIMS). DAIMS leverages paired histopathology, genomic, and clinical data during training to learn generalizable disease representations and uses histology-derived surrogate latent representations when auxiliary modalities are unavailable, enabling flexible histology-only inference. Trained on TCGA samples and validated on two independent European and East Asian cohorts comprising 842 patients, DAIMS consistently outperformed state-of-the-art methods. Adapted DAIMS improved the C-index by 5.6%-13.9% on F1CRC and 4.1%-9.0% on SURGEN compared with baseline methods. DAIMS improved within-cohort prognostic discrimination and localized prognostically relevant morphological patterns that remained stable across institutions. Our evaluation showed that DAIMS achieved generalizable survival stratification across disease stages, molecular subgroups, and treatment-defined patient groups.

Objective: To develop and externally validate a machine-learning model for classifying documented pressure injury status during hospitalization among intensive care unit (ICU) patients receiving invasive mechanical ventilation. Methods: The development cohort included 1601 patients from MIMIC-IV version 3.1, including 390 with and 1211 without documented pressure injury. Mean values of continuous clinical variables recorded during the 24 h before ICU admission were used. Variables and patients with >30% missing data were excluded, and remaining missing values were imputed using random forest imputation. Features were selected using Boruta and least absolute shrinkage and selection operator (LASSO) regression. Six machine-learning models were compared using 10-fold cross-validation, and XGBoost was subsequently optimized using GridSearchCV with ROC-AUC as the optimization metric and a random seed of 42. SHapley Additive exPlanations (SHAP) were used for model interpretation. External validation was performed in 1576 eICU patients, including 221 with and 1355 without documented pressure injury; the first value recorded after ICU admission was used because continuous pre-ICU data were unavailable. Results: XGBoost showed the best overall performance among the six candidate models. After optimization, the model achieved a 10-fold cross-validated area under the receiver operating characteristic curve (AUC) of 0.853 (95% CI, 0.765-0.941) in MIMIC-IV and an AUC of 0.703 (95% CI, 0.669-0.737) in the independent eICU cohort. The final model included sex, white blood cell count, blood glucose, hematocrit, blood urea nitrogen, and anion gap. SHAP analysis identified blood urea nitrogen and anion gap as the strongest contributors to model output. Conclusion: The optimized XGBoost model demonstrated moderate external discrimination for classifying documented pressure injury status among ICU patients receiving invasive mechanical ventilation and may provide supplementary information for clinical assessment.

Objectives: Tertiary lymphoid structures (TLS) reflect the tumor immune microenvironment and may provide prognostic information in hepatocellular carcinoma (HCC), but assessment requires pathology. This study aimed to develop and externally validate a CT radiomics-based machine-learning model for preoperative prediction of TLS status and explore disease-free survival across pathological TLS/microvascular invasion (MVI) phenotypes. Materials And Methods: This multicenter retrospective study included 959 patients with HCC (training, n = 792; external validation, n = 167). Radiomic features and an intratumoral heterogeneity (ITH) score were extracted from contrast-enhanced CT images. A stacking ensemble model (ITH-Ensemble) incorporating clinicoradiological variables was developed to predict TLS status and compared with clinical and radiomics models. SHapley Additive exPlanations (SHAP) analysis was used for interpretation. Disease-free survival (DFS) was compared among four TLS/MVI phenotypes. Transcriptomic support was assessed using TCGA-LIHC data (n = 366). Results: The ITH-Ensemble achieved an external-validation AUC of 0.842, versus 0.828 for the radiomics model and 0.732 for the clinical model. SHAP identified the Radscore and inflammatory markers, including monocyte count and alanine aminotransferase, as contributors. The four TLS/MVI phenotypes showed distinct DFS distributions (log-rank p < 0.0001); TLS+/MVI- had the most favorable and TLS-/MVI+ the least favorable outcome. In TCGA-LIHC, CD2 expression correlated with natural killer cell infiltration (Spearman ρ = 0.766, p < 0.001) and longer overall survival (p = 0.0247). Conclusions: The CT-based ITH-Ensemble provides a noninvasive preoperative estimate of TLS status in resectable HCC. Pathological TLS/MVI phenotypes were associated with DFS, although a fully preoperative dual-factor framework requires independent MVI prediction. Level Of Evidence: 3 (retrospective multicenter cohort study).

We developed and externally validated an interpretable machine-learning model for observed continuous renal replacement therapy (CRRT) initiation in sepsis-associated acute kidney injury (SA-AKI) using a strict 24-hour landmark framework. This multicenter retrospective study used MIMIC-IV and eICU-CRD data. Predictors were restricted to the period from sepsis diagnosis to 24 h, and the outcome was first observed CRRT initiation after the landmark and within 7 days. Patients who died, initiated CRRT, or were no longer under observation by 24 h were excluded. Eight algorithms were evaluated, and a seven-predictor Gradient Boosting model was selected. Discrimination, calibration, incremental value, SHAP-based interpretation, and exploratory prediction-subgroup outcomes were assessed. The final MIMIC-IV cohort included 5,238 patients, with 239 CRRT events; 3,666 were assigned to training and 1,572 to internal validation. The eICU-CRD cohort included 4,683 patients and 157 events. The model retained serum creatinine, AKI stage, SOFA score, urine output, lactate, red blood cell distribution width, and peripheral oxygen saturation. AUCs were 0.905 internally and 0.816 externally, with an external calibration slope of 0.567. In an exploratory age- and sex-matched eICU analysis, the high predicted-risk/no observed CRRT subgroup had higher ICU, 7-day, and 28-day mortality than the true-negative subgroup. The model provided interpretable risk estimates for observed CRRT initiation, but external calibration and incremental value were limited. Findings should not be interpreted as evidence of CRRT indication, undertreatment, or treatment benefit. Prospective validation and local recalibration are required.

Purpose To develop and validate a deep learning (DL) model based on dual-modality US videos to differentiate benign from malignant superficial lymphadenopathy (LA) and stratify malignant subtypes. Materials and Methods This multicenter study included patients with pathologically confirmed LA from five centers (June 2019-December 2025). The Dual-Modality US Video Lymphadenopathy Diagnostic Network (DMUVL-DiagNet) integrated B-mode and color Doppler flow imaging videos with basic clinical information. A retrospective video dataset (n = 1616) was used for training and internal testing, while external testing used a prospective dual-modality US video set (n = 454) and a static US image set (n = 1956). Model performance was compared against clinical baseline and US-based DL models and an artificial intelligence (AI)-assisted reader study was conducted. Areas under the receiver operating characteristic curves (AUCs) were calculated to evaluate diagnostic performance. The study was prospectively registered at the Chinese Clinical Trial Registry (ChiCTR2400090592). Results Overall, 4026 patients (median age, 52.6 [IQR 41-62] years; 2175 males) were included. DMUVL-DiagNet consistently outperformed clinical and US-based models across all test sets, achieving AUCs of 0.91-0.95 for benign versus malignant classification and 0.87-0.91 for lymphoma versus metastasis classification. The model showed good generalizability (subgroup AUCs, 0.85-0.93) and surpassed the average performance of all radiologist experience groups in both tasks. Additionally, DMUVL-DiagNet assistance improved junior radiologist AUCs from 0.72 to 0.88 and from 0.58 to 0.77 for the two tasks, respectively (all P < .001). Conclusion DMUVL-DiagNet enabled accurate characterization of superficial LA and improved junior radiologist diagnostic performance. © The Authors 2026. Published by the Radiological Society of North America under a CC BY 4.0 license.

Objective: Electronic health record (EHR) time series are irregularly sampled and frequently incomplete, with missingness potentially carrying clinically informative signals. We developed Masked-Attention-based timestamp-wise Data Embedding (MADE), a missingness-aware model for continuous prediction of deterioration in general wards, and evaluated it across heterogeneous hospitals. Materials And Methods: Observed variables were tokenized into feature-specific embeddings and processed by a masked-attention Transformer, followed by bi-LSTM temporal aggregation for risk prediction. MADE was evaluated for major adverse event (MAE) and sepsis in 1 internal (SVH) and 2 external cohorts (AMC, HUMC) against conventional and deep learning baselines. Ablation studies compared no imputation with last observation carried forward (LOCF) and multiple imputations by chained equations (MICE). Discriminative performance was assessed using the area under the receiver operating characteristic curve (AUROC) and precision-recall curve (AUPRC). Clinical utility and reliability were assessed using decision curve analysis, Brier score, and expected calibration error (ECE). Results: MADE achieved strong discrimination for MAE (AUROC 0.966/0.902/0.825 for SVH/AMC/HUMC) and sepsis (AUROC 0.934/0.874/0.776). The no-imputation configuration generally outperformed LOCF and MICE, particularly for MAE and for sepsis AUPRC, supporting preservation of informative missingness. Decision curves showed favorable net benefit for MAE and competitive benefit for sepsis. Calibration indicated modest underestimation at higher risks, suggesting site-specific recalibration. Discussion: MADE mitigates limitations of fixed grid and imputation-dependent EHR modeling by jointly leveraging observed values and missingness patterns under heterogeneous sampling. Conclusion: MADE supports scalable observation-aligned risk monitoring, particularly for MAE, while sepsis prediction may require site-specific recalibration or feature-set adaptation under cross-institution domain shifts.

Introduction: Postpartum depression (PPD) is associated with severe adverse outcomes, yet timely risk stratification remains challenging. Existing machine learning models are often limited by insufficient external validation, limited interpretability, and lack of accessible clinical tools. This study aimed to develop and externally validate an explainable machine learning model with a web-based calculator for predicting postpartum depressive symptoms. Methods: This retrospective cohort study included 553 women from two hospitals, comprising 403 in the development cohort and 150 in an independent external validation cohort. The development cohort was divided into training and internal test sets. Screening-positive postpartum depressive symptoms at 6 weeks postpartum were defined as an Edinburgh Postnatal Depression Scale (EPDS) score ≥9. Five machine learning models were developed using demographic, psychosocial, and obstetric predictors. Model discrimination was assessed using the area under the receiver operating characteristic curve (AUC), with calibration and decision curve analyses performed additionally. SHapley Additive exPlanations (SHAP) were used for model interpretation. Results: Screening-positive rates were 38.0% in the development cohort and 26.0% in the external validation cohort. Tree-based ensemble models showed broadly comparable performance. eXtreme Gradient Boosting (XGBoost) was selected based on its numerically highest mean cross-validated AUC and balanced overall performance, with AUCs of 0.895 ± 0.033 in cross-validation, 0.845 (95% confidence interval [CI]: 0.751-0.940) in the internal test set, and 0.820 (95% CI: 0.743-0.889) in the external validation cohort. Calibration was acceptable in the internal test set, whereas absolute risk was overestimated in the external validation cohort. Decision curve analysis indicated potential net clinical benefit. SHAP identified age, marital relationship quality, prenatal depressive symptoms, monthly personal income, and sleep disturbance as important model-based contributors. A four-predictor web calculator was developed. Discussion: The explainable XGBoost model showed promising discrimination for predicting EPDS-defined postpartum depressive symptoms. However, given the modest external validation sample and evidence of risk overestimation, further prospective multicenter validation and recalibration are needed before routine clinical implementation.

Aims/Background: Sepsis is a leading cause of morbidity and mortality in intensive care units (ICUs). Early identification of patients at high risk of death is essential to guide timely interventions and optimize resource allocation. This study aimed to develop and validate an interpretable machine learning model for predicting 28-day mortality in adult ICU patients. Methods: We retrospectively analyzed data from three independent critical care databases, including 98,233 adult ICU patients. Sixteen routinely collected variables available by the 24-hour prediction landmark-including vital signs and laboratory parameters obtained during the first 24 hours after ICU admission, baseline comorbidities documented at or before ICU admission, and treatments administered during the same 24-hour period-were selected for model development. Nine machine learning algorithms were compared, and model performance was evaluated using both internal and external validation cohorts. Explainable machine learning techniques were applied to identify the most influential predictors. Results: Among the nine algorithms, extreme gradient boosting (XGBoost) demonstrated the highest predictive performance. The model achieved an area under the receiver operating characteristic curve (AUROC) of 0.884 (0.871-0.897) in the internal validation cohort and maintained comparable discrimination in the two external validation cohorts, eICU-Sepsis, derived from the eICU Collaborative Research Database (eICU-CRD), and DRECCv1.0 critical care database (DRECC)-Sepsis, with AUROCs of 0.877 (0.872-0.882) and 0.875 (0.856-0.894), respectively. Mechanical ventilation, comorbidity burden, lactate dehydrogenase, age, respiratory rate, and blood pressure made important contributions to the model's prediction of 28-day mortality. Conclusion: This interpretable machine learning model may facilitate early risk stratification of ICU patients with sepsis. Although the model shows potential for identifying high-risk individuals at the 24-hour prediction landmark, further prospective validation and clinical impact assessment are required before routine clinical implementation.