Development and international validation of machine learning models for predicting incident disability in older adults: a multi-cohort study using CHARLS, HRS, and ELSA

Several machine learning models have been developed using China Health and Retirement Longitudinal Study (CHARLS) data to predict incident disability. However, previous studies have generally used a single development period, relied on random-split internal validation, or focused on restricted clinical subgroups. The temporal stability and cross-national transportability of these models, and the incremental value of objective functional measures over readily obtainable non-invasive predictors, remain uncertain. We aimed to develop and evaluate two explainable models for predicting incident disability in older adults. We analyzed CHARLS participants aged ≥ 60 years who were free of ADL/IADL disability at baseline and had an observed next-survey outcome. Cohort 1 (2011–2013) was used for model development, and Cohort 3 (2018–2020) provided participant-ID-disjoint temporal evaluation after exclusion of IDs allocated to Cohort 1. An Objective-enriched Model and a Non-invasive Core Model were developed using XGBoost with 5 × 5 nested cross-validation and training-only calibration. Performance was evaluated in internal, temporal, HRS and ELSA samples. The primary CHARLS analysis included 8,267 unique participants and 2,070 disability outcomes. The Core Model achieved AUCs of 0.671 (95% CI 0.636–0.704) in internal validation and 0.711 (0.693–0.729) in temporal evaluation, with a temporal Brier score of 0.166 and an observed-to-expected ratio of 1.016. The Objective-enriched Model achieved an internal-validation AUC of 0.684 (0.649–0.716). Its paired AUC difference from the Core Model was 0.0131 (− 0.0046 to 0.0316). Across HRS and ELSA, the models retained moderate discrimination (AUCs, 0.692–0.729) despite differences in cohort measurement systems, which meant that several predictors were unavailable or only approximately mapped. Predicted risks exceeded observed event rates (O/E ratios, 0.441–0.495). The seven-predictor Non-invasive Core Model showed moderate discrimination for next-survey ADL/IADL disability in temporal and cross-population evaluations. Selected objective measurements provided modest additional predictive information. Cross-population implementation should be preceded by cohort-specific predictor harmonization, recalibration, and prospective evaluation. Not applicable.