BACKGROUND: Arthritis is a common chronic disease among middle-aged and older adults and is strongly related to functional decline. METHODS: The research sample and data were derived from the China Health and Retirement Longitudinal Study (CHARLS) 2015. We employed the least absolute shrinkage and selection operator (LASSO) and multifactor logistic regression analysis to identify features for model construction. We proposed six machine learning (ML) predictive models. The optimal model was selected using various learning metrics and was further interpreted using the SHapley Additive exPlanations (SHAP) method. RESULTS: A total of 5111 subjects were included in the analysis, of which 1955 developed functional disability. Among the six models, XGBoost showed the best performance, achieving a test set area under the curve (AUC) of 0.74. SHAP analysis ranked the features by their contribution as follows: waist circumference, handgrip strength, self-reported health status, age, body pains, depression, history of falls, sleeping duration, and availability of care resources. SHAP dependence plots indicated that individuals over 60 with increased waist circumference (>
85 cm), short sleeping duration (<
5 h), and lower handgrip strength (<
25 kg) had a higher probability of functional disability. CONCLUSION: This study presents an interpretable machine learning-based model for the early detection of functional disability in patients with arthritis and informs the development of care strategies aimed at delaying functional disability in this population.