Researchers have developed and internationally validated machine learning models that predict whether older adults will develop disability in activities of daily living by the time of the next survey wave, using data from large aging studies in China, the United States and England. The open-access study, published in BMC Geriatrics, reports that a simple seven-predictor model built from readily obtainable, non-invasive information achieved moderate discrimination both over time within China and across the two other national cohorts.
The stated goal was to test whether objective measures add meaningful predictive value over non-invasive variables, a question the authors say previous single-cohort studies had left unresolved.
The primary CHARLS analysis included 8,267 unique participants and 2,070 incident disability outcomes. In internal validation, the Core Model achieved an area under the receiver operating characteristic curve (AUC) of 0.671 (95% CI 0.636 to 0.704). In the temporal evaluation on the later, participant-disjoint cohort, the AUC rose to 0.711 (0.693 to 0.729), with a Brier score of 0.166 and an observed-to-expected ratio of 1.016, indicating that predicted risks closely matched observed event rates over time.
The Objective-enriched Model reached an internal-validation AUC of 0.684 (0.649 to 0.716). Its paired AUC difference from the Core Model was 0.0131, with a confidence interval of -0.0046 to 0.0316 that crossed zero. The authors conclude that the selected objective measurements provided only modest additional predictive information, suggesting that a lean set of non-invasive predictors captures most of the available signal for this outcome.
Despite differences in cohort measurement systems, which meant several predictors were unavailable or could only be approximately mapped, the models retained moderate discrimination, with AUCs between 0.692 and 0.729 across the two external samples.
Calibration, however, did not transfer as cleanly. In the HRS and ELSA evaluations, predicted risks exceeded observed event rates, with observed-to-expected ratios ranging from 0.441 to 0.495. This means the models systematically overestimated disability risk in these populations, likely reflecting differences in how disability and its predictors were measured across the cohorts.
The authors emphasize that cross-population implementation should be preceded by cohort-specific predictor harmonization, recalibration, and prospective evaluation. They also note that earlier CHARLS-based models generally used a single development period, relied on random-split internal validation, or focused on restricted clinical subgroups, leaving temporal stability and cross-national transportability uncertain until now.
The findings are limited by the observational cohort design, the survey-based definition of disability, and the imperfect mapping of predictors across measurement systems, so the reported performance should be interpreted as a benchmark for further validation rather than a ready-to-deploy clinical tool.
Subject of Research: Machine learning prediction of incident disability in older adults using CHARLS, HRS and ELSA cohort data
Article Title: Development and international validation of machine learning models for predicting incident disability in older adults: a multi-cohort study using CHARLS, HRS, and ELSA
Article References: Liu, H., Chen, Y., Yang, X., & Wang, S. (2026). Development and international validation of machine learning models for predicting incident disability in older adults: a multi-cohort study using CHARLS, HRS, and ELSA. BMC Geriatrics. https://doi.org/10.1186/s12877-026-08595-0
Image Credits: AI Generated
DOI: 10.1186/s12877-026-08595-0
Keywords: machine learning, disability prediction, older adults, CHARLS, HRS, ELSA, XGBoost, ADL, IADL, model validation, BMC Geriatrics, aging research
News Source: Teresa Odom. (October 28, 2026). Machine learning models predict disability risk in older adults across three countries. Scienmag.
