2026-10-16

Background: Assessment of prognosis is crucial for managing non-traumatic subarachnoid hemorrhage (NSAH) patients. The study aimed to explore the association between platelet-to-white blood cell ratio (PWR) and all-cause mortality in this high-risk population and to develop machine learning models to improve prognostic accuracy. Methods: Using data from the Medical Information Mart for Intensive Care (MIMIC-IV). The observational endpoints were 30-day and 365-day mortality. Patients were categorized into four groups based on the quartiles of PWR. Kaplan-Meier curves, multivariate Cox regression analyses, restricted cubic splines (RCS), and subgroup analyses were applied. We also developed machine learning models that include PWR and clinical features selected with the Boruta and LASSO algorithms to predict 30-day mortality. Model performance was evaluated by the receiver operating characteristic (ROC) curve and the area under the curve (AUC). Shapley Additive exPlanations (SHAP) method was utilized to provide model interpretability. External validations in the eICU-Collaborative Research Database (eICU-CRD) and the Hospital Information System (HIS) of the First Affiliated Hospital of Harbin Medical University were used to assess generalizability. Results: A total of 1,176 patients with critically ill NSAH were enrolled from MIMIC-IV, predominantly male (58.8%), with a median age of 62 years. In multivariable Cox models, higher PWR quartiles were independently associated with lower risk of 30-day (HR = 0.55, 95% CI 0.35, 0.86, p = 0.009) and 365-day (HR = 0.53, 95% CI 0.37, 0.77, p = 0.001) mortality after fully adjusting. RCS suggested a linear association. No significant interaction was observed by subgroup analysis across all subgroups. The top-performing machine learning model attained an AUC of 0.868 during internal validation (MIMIC-IV). External validations yielded AUC values of 0.874 (eICU-CRD) and 0.763 (HIS). SHAP analysis revealed a correlation between decreased PWR and an elevated predicted risk of mortality. Conclusion: PWR is an independent predictor of short- and long-term mortality in ICU patients with NSAH. Machine learning models incorporating PWR significantly improve risk stratification, thereby providing precious support for the early identification and personalized management of patients in the ICU.

Background: Sepsis patients with type 2 diabetes mellitus (T2DM) face a substantially elevated risk of venous thromboembolism (VTE), yet prediction tools tailored for this population are lacking. This study aimed to develop and validate interpretable machine learning models for predicting in-hospital VTE in septic patients with T2DM. Methods: A total of 648 sepsis patients with T2DM admitted to Shaanxi Provincial People's Hospital between January 2020 and October 2024 were included in this retrospective cohort study, an internal validation cohort of 303 patients was retrospectively collected from Shaanxi Provincial People's Hospital between November 2024 and December 2025, and an independent external validation cohort of 260 patients was retrospectively collected from Xi'an No. 3 Hospital (a separate tertiary care center) between November 2023 and December 2025. Six machine learning algorithms-Logistic Regression, Support Vector Machine (SVM), Multilayer Perceptron (MLP), LightGBM, XGBoost, and Random Forest-were developed and compared. Feature selection was performed using the Boruta algorithm combined with forward sequential feature selection. Model performance was evaluated by discrimination (area under the receiver operating characteristic curve, AUC), calibration (calibratin plots and Brier scores), and clinical utility (decision curve analysis). Model interpretability was enhanced using SHAP analysis. Results: Among the 648 internal patients, 115 (17.7%) developed VTE. The Boruta algorithm and forward sequential feature selection identified five key predictors: D-dimer, C-reactive protein (CRP), activated partial thromboplastin time (APTT), international normalized ratio (INR), and fibrin degradation products (FDP). Among the six algorithms, LightGBM achieved the best overall performance, with an AUC of 0.9025 in the test set, 0.8048 in the internal validation cohort, and 0.8902 in the external validation cohort. Calibration curves demonstrated good agreement between predicted and observed outcomes, and decision curve analysis confirmed positive net clinical benefit across a wide range of threshold probabilities. SHAP analysis revealed that elevated D-dimer, CRP, INR, and FDP, together with shortened APTT, were the strongest drivers of VTE risk. Sensitivity analyses excluding D-dimer-driven cases, removing the D-dimer feature, and restricting to symptomatic VTE yielded consistent results (AUC range: 0.84-0.89). Conclusions: The LightGBM model incorporating five readily available coagulation and inflammatory markers accurately predicts in-hospital VTE in sepsis patients with T2DM, with robust discrimination, calibration, and clinical utility. SHAP-based interpretability enhances transparency and may facilitate personalized risk stratification and informed thromboprophylaxis decisions in this high-risk population.

Accurate preoperative differentiation of pediatric acute appendicitis subtypes is relevant to treatment planning. We developed and externally validated a machine-learning framework combining habitat-based computed tomography (CT) radiomics with clinical variables for three-class classification (simple, complex, and appendiceal abscess). This retrospective dual-center study included 317 children undergoing appendectomy at Institution A (training n = 222, internal validation n = 95) and 174 children at Institution B (external validation). K-means + + habitat segmentation (k = 3) of preoperative noncontrast CT volumes of interest yielded 1218 radiomic features; ten features selected on the training cohort were combined with 20 preoperative clinical variables. Eight classifiers were compared using macro-averaged one-versus-rest area under the receiver operating characteristic curve (AUC), accuracy, and macro-F1, with bootstrap confidence intervals (CIs), DeLong tests, calibration statistics, and decision curve analysis. In external validation, naive Bayes achieved the highest macro-AUC (0.946, 95% CI 0.910-0.974), accuracy 0.776 (0.713-0.833), and macro-F1 0.766 (0.674-0.837), significantly exceeding all seven comparators among the configurations evaluated (p < 0.001); the internal validation macro-AUC was 0.940 (0.893-0.977). Per-class external AUCs were 0.920 (simple), 0.948 (complex), and 0.971 (abscess). Calibration slopes (0.23-0.30) indicated that probability outputs require recalibration before probabilistic use. In a five-arm ablation of the input feature blocks, the fusion model combining clinical variables with habitat radiomics performed best (external macro-AUC 0.971), ahead of habitat radiomics alone (0.952), clinical variables plus conventional whole-VOI radiomics (0.776), clinical variables alone (0.763) and conventional whole-VOI radiomics alone (0.582). Habitat radiomics combined with machine learning provided accurate, externally validated three-class classification of pediatric acute appendicitis in children undergoing appendectomy. Naive Bayes offered the most balanced subtype discrimination. The model is a subtype-stratification aid rather than a stand-alone diagnostic or treatment-selection tool.

Background: Valvular heart disease, particularly aortic stenosis (AS), often remains asymptomatic until advanced stages, making early detection challenging. Although cardiac auscultation is a traditional screening tool, its diagnostic accuracy is limited. We aimed to externally validate a deep learning-based model (eAS model) using phonocardiogram (PCG) and single-lead electrocardiogram (ECG) signals to detect moderate-or-greater AS. Methods: PCG and ECG recordings were acquired using a PCG device (Cardio-EGG). Transthoracic echocardiography served as the reference standard. The external validation cohort consisted of 951 patients from 3 medical institutions, including 70 with moderate-or-greater AS. The primary outcome was sensitivity for detecting moderate-or-greater AS; secondary outcomes included the area under the receiver operating characteristic curve (AUROC), specificity, predictive values, and likelihood ratios (LR). Results: The external validation cohort (n = 951, median age 73 years, male 65%) exhibited more advanced valvular disease than the training/development cohort. The eAS model achieved an AUROC of 0.9632 (95% CI: 0.9471-0.9774). Using the pre-determined threshold, sensitivity was 92.9% and specificity 87.9%, yielding a positive LR of 7.65 and negative LR of 0.08. Model-derived probabilities showed strong associations with echocardiographic measures of AS severity (adjusted R2 = 0.5051 for peak transvalvular velocity). Exploratory subgroup analyses showed generally consistent performance, although estimates were imprecise in small subgroups. Conclusion: The eAS model demonstrated high diagnostic performance for detecting moderate-or-greater AS in a multicenter external validation cohort. AI-based PCG/ECG analysis may support the decision to refer patients from auscultation to echocardiography for AS evaluation.

Background: Anemia represents a frequent and clinically burdensome complication in patients with primary lung cancer receiving platinum-based regimens, yet robust risk stratification tools that integrate multidimensional clinical data and undergo external validation remain scarce. This multicenter study sought to develop and externally validate a machine learning-based predictive model for CIA in this population. Methods: A total of 3,434 patients with primary lung cancer were retrospectively enrolled from two geographically distinct institutions, with 3,097 assigned to the training/internal validation set (7:3 split) and 337 retained as an independent external validation set. Variables with missing rates exceeding 30% were excluded, and remaining missing data were imputed using multiple imputation by chained equations (MICE). Thirteen candidate predictors were preselected via univariate screening and least absolute shrinkage and selection operator (LASSO) regression. Nine machine learning algorithms were developed and comparatively evaluated using area under the receiver operating characteristic curve (AUC), calibration plots, and decision curve analysis. Model interpretability was achieved through SHapley Additive exPlanations (SHAP). The primary outcome was defined as hemoglobin levels < 100 g/L (CTCAE Grade ≥ 2) recorded during the first two cycles of chemotherapy. Results: Among the nine candidate models, LightGBM achieved the highest discriminative performance, with AUCs of 0.871 (95% CI: 0.852-0.891) in the training set, 0.797 (95% CI: 0.763-0.831) in internal validation, and 0.684 (95% CI: 0.626-0.742) in the external validation set derived from Linfen Central Hospital. Calibration and decision curve analyses consistently demonstrated superior agreement and net clinical benefit for LightGBM across all three datasets. SHAP-based interpretation identified lymphocyte count, D-dimer, alanine aminotransferase, total bilirubin, sex, and albumin as the six most influential predictors, with D-dimer, female sex, and outpatient admission positively associated with anemia risk, whereas lymphocyte count, liver function indices, and albumin showed inverse associations. Conclusions: The LightGBM-based model developed herein demonstrates acceptable and moderate predictive performance, satisfactory calibration, and potential clinical utility across geographically diverse populations, offering a preliminary risk-stratification tool for early identification of lung cancer patients at elevated anemia risk prior to chemotherapy initiation. However, given the modest external validation performance, cautious interpretation and further refinement are warranted before widespread clinical deployment.

Purpose: The objective of this study was to develop and externally validate complementary magnetic resonance imaging (MRI)-based models and their stacked combinations for preoperative World Health Organization (WHO) grading of intracranial meningioma. Methods: This retrospective multicenter study included 455 histopathologically confirmed meningiomas: a four-center development cohort (n = 346) with leave-one-center-out cross-validation and an external cohort (n = 109). Four base models used contrast-enhanced T1-weighted MRI: tumor + peritumoral radiomics, an nnU-Net segmentation-based classifier, frozen DINOv3 transfer-learning features, and clinical variables. All 15 combinations were prespecified and stacked using a linear meta-learner trained on development out-of-fold predictions. Frozen models were externally evaluated without modification. Results: WHO 2-3 tumors comprised 24.6% overall (112/455; development 20.5%; external 37.6%). External AUCs were 0.732 for nnU-Net, 0.703 for DINOv3, 0.645 for clinical, 0.628 for radiomics. The multimodal ensemble (nnU-Net+DINOv3+Clinical) achieved the highest external AUC (0.742), exceeding radiomics (ΔAUC +0.114, p = 0.008) but not nnU-Net (ΔAUC +0.011, p = 0.737). Compared with the clinical model, discrimination was higher but nonsignificant at the prespecified threshold (ΔAUC +0.098, p = 0.050), while reclassification improved (IDI +0.109, p < 0.001). Radiomics-containing combinations performed strongly in development but lost most discrimination externally. The multimodal ensemble had an external calibration slope of 1.32 (Brier 0.190) and net benefit above default strategies at thresholds ≥ 0.23. Conclusion: The multimodal ensemble achieved the highest external AUC but did not significantly exceed nnU-Net, which was the most transportable single predictor of meningioma grade. Clinical variables did not improve discrimination but enabled interpretable feature-level attribution. Handcrafted radiomics lacked external transportability.

Background: Intraoperative hypotension (IOH) is a common and potentially preventable complication strongly associated with postoperative acute kidney injury, myocardial injury, and mortality. Current approaches to IOH management remain largely reactive, and existing prediction tools have variable clinical utility. This study aimed to develop and externally validate a dynamic machine-learning model using routinely collected perioperative data to perform one-step-ahead prediction of intraoperative hypotension on a regularized 5-minute intraoperative grid. Methods: This retrospective multicenter study included 702 adults undergoing non-cardiac surgery under general anesthesia for model development and internal validation and an independent external cohort of 248 patients. A dynamic prediction framework was constructed using routinely collected preoperative variables and intraoperative 5-minute time-series data. Individualized baseline features and temporal predictors were derived from rolling observation windows. Four machine-learning algorithms were compared using a fully nested grouped cross-validation framework with external validation. Model performance was assessed by discrimination, calibration, decision-curve analysis, and SHapley Additive exPlanations (SHAP)-based interpretability. Results: All four machine-learning models demonstrated good discrimination in the primary fully nested case-grouped internal validation, with AUROC values ranging from 0.842 to 0.857. Random forest achieved the highest AUROC (0.857, 95% CI 0.841-0.868), the highest AUPRC (0.413, 95% CI 0.367-0.452), and extra trees achieved the lowest Brier score (0.093, 95% CI 0.088-0.098). In the independent external validation cohort, random forest remained the best-performing model, yielding an AUROC of 0.894 (95% CI 0.875-0.907), an AUPRC of 0.540 (95% CI 0.457-0.596), and a Brier score of 0.092 (95% CI 0.083-0.102). SHAP analysis consistently identified MAP-derived features, including current MAP, recent hypotensive burden, recent MAP measurements, relative MAP reduction, and MAP variability, as the most influential predictors across all four models. Conclusions: A dynamic machine-learning model using routinely collected perioperative time-series data accurately predicted MAP <65 mmHg at the next scheduled 5-minute intraoperative time point in an independent external cohort. These findings support further evaluation of this short-horizon risk stratification model in prospective clinical workflow studies.

Objectives: To assess the feasibility of estimating a radiographic Decayed-Missing-Filled-Teeth score (rDMFT), a proxy for the clinical DMFT index, from a single panoramic radiograph with a position-aware deep-learning pipeline. Materials And Methods: A tooth-status detector and a tooth-numbering detector (32 FDI positions), trained on DENTEX 2023, were combined in a position-aware read-out (one tooth per FDI position, third molars excluded, missing teeth = 28 - occupied positions) and validated on 100 panoramic radiographs from an independent hospital against a two-dentist, radiograph-based reference (intraclass correlation coefficient, ICC; mean absolute error, MAE; Bland-Altman analysis). Results: Against the reference, the pipeline reached an ICC of 0.84 (95% CI 0.74 to 0.90) and an MAE of 2.38, below the between-dentist agreement in point estimate (ICC 0.88, MAE 2.13), with larger errors against each dentist alone (MAE 3.00 and 2.73) and a bias of + 1.14 units from caries over-detection; agreement was highest for filled (r 0.90) and lowest for decayed teeth (r 0.34; dentists 0.16). In the 26 radiographs with missing teeth the missing count was exact in 42% (62% within ± 1 tooth), and missing teeth were over-called in 15 of 74 complete dentitions; tooth loss was no longer overlooked (26 of 26 detected), but the overall ICC was 0.84 versus 0.90 without numbering. Conclusions: Automated rDMFT estimation is feasible, but agreement was lower in point estimate than between two dentists, with larger errors against each dentist. The missing component now responds to actual tooth loss but remains imprecise and is over-called in some complete dentitions; the decayed component is unreliable. The pipeline is a research building block, not a decision-support or triage tool. Clinical Relevance: rDMFT is a radiographic proxy, not a clinical measure, and the pipeline is not proposed for clinical use; any application would first require validation in unselected cohorts against clinically determined reference standards.