A Machine Learning-Based Fall Risk Prediction Model for Middle-Aged and Older Adults with Comorbid Cardiovascular Disease and Arthritis: Development and Geographic Split-Sample Validation.

Falls impose a major public health burden on middle-aged and older adults. Coexisting cardiovascular disease (CVD) and arthritis further increases fall risk, yet targeted prediction models remain scarce, especially in Chinese populations. We analyzed 633 China Health and Retirement Longitudinal Study participants aged ≥45 years with comorbid CVD and arthritis at the 2015 baseline. Seven predictors were selected through univariate analysis, LASSO logistic regression (LR), and clinical review. Eight machine-learning algorithms were compared. Among models with an area under the receiver operating characteristic curve (AUC) ≥0.72, the model with the highest specificity was preferred to reduce false-positive referrals. A geographically defined holdout group (n = 110) from different provincial-code regions, but within the same CHARLS source population, was treated as a geographic split-sample rather than independent external validation. Performance was assessed using AUC, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1-score, calibration, decision-curve analysis, precision-recall curves, and SHapley Additive exPlanations (SHAP). During cumulative follow-up, 296 participants (46.8%) reported at least one fall. Random forest (RF) and LR achieved similar internal-validation AUCs (0.730 vs. 0.729); RF was selected because of higher specificity. Internally, RF showed 87.5% specificity, 37.7% sensitivity, 70.3% PPV, 64.2% NPV, and an F1-score of 0.491. Its AUC in the geographic holdout sample was 0.735. Despite similar AUCs, sensitivity and specificity differed substantially between samples, suggesting limited regional stability of the fixed threshold. SHAP identified fall history, ankle pain, activities of daily living impairment, and sleep duration as the most influential predictors. The RF model showed moderate discrimination and a conservative, high-specificity profile. However, its low internal sensitivity means it should not be used alone to rule out falls or as a diagnostic tool. Pending independent validation and regional threshold calibration, it may serve only as adjunctive stratification within a stepped assessment pathway and should not replace routine clinical fall-risk screening.