Risk prediction for venous thromboembolism among older adult internal medicine inpatients: development and validation of a machine learning-based model

To develop and validate a machine learning-based risk prediction model for venous thromboembolism (VTE) in older adult internal medicine inpatients, thereby providing a research basis for subsequent external validation and clinical utility evaluation. This retrospective study enrolled 2,315 older adult patients admitted to the Department of Internal Medicine at the Affiliated Hospital of Guangdong Medical University between January 2024 and October 2025. The cohort was partitioned at a pre-specified temporal cutoff of May 2025: patients admitted between January 2024 and April 2025 ( n = 1,640) constituted the development cohort, whereas those admitted between May and October 2025 ( n = 675) constituted the temporal validation cohort. Sixteen candidate predictors were pre-specified prior to model fitting. Eight machine learning algorithms—logistic regression (LR), random forest (RF), gradient boosting (GB), support vector machine (SVM), K-nearest neighbors (KNN), neural network (NN), soft voting (Voting), and stacking (Stacking)—were compared using patient-ID-grouped nested cross-validation with a 5-fold outer loop and a 3-fold inner loop. The NN was pre-specified as the primary model; among the remaining algorithms, the best performer in the development-cohort cross-validation was designated as the secondary comparator. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC), average precision (AP), and the Brier score; confidence intervals were estimated with 1,000 patient-ID-clustered bootstrap resamples. The SHAP (SHapley Additive exPlanations) method was applied to the temporal validation cohort to interpret the secondary comparator selected in the development cohort. In the temporal validation cohort, the pre-specified primary model (NN) achieved an AUROC of 0.773 (95% CI: 0.721–0.825) and a Brier score of 0.106, whereas the secondary comparator (Voting) achieved an AUROC of 0.790 (95% CI: 0.741–0.837) and a Brier score of 0.102; the paired AUROC comparison showed no statistically significant difference between the two models. The 18-variable extended logistic regression model (LR18), incorporating pharmacological and mechanical prophylaxis indicators, demonstrated higher discrimination, with an AUROC of 0.841 (95% CI: 0.803–0.878). Calibration curves revealed a systematic underestimation of absolute risk across all models in the temporal validation cohort, as reflected by positive calibration-in-the-large (CITL) values (NN CITL: 0.655, 95% CI: 0.397–0.877). SHAP analysis identified prior VTE, log D-dimer, self-care ability, and mobility as the key variables influencing model predictions. Both the NN and Voting models demonstrated moderate discrimination under retrospective temporal evaluation but exhibited underestimation of absolute risk and, under high-sensitivity thresholds, a substantial number of false-positive predictions. The models still require independent external validation, recalibration, and clinical utility evaluation.