Qi-Shuai Ma, Li-Qun Jiang, Buong-O Chun
Background: Functional impairment is common among older adults with diabetes, but multidimensional machine learning classification in this population remains underexplored. This study compared machine learning algorithms for identifying concurrent functional impairment among older adults with self-reported physician-diagnosed diabetes. Methods: This cross-sectional study included 1213 adults aged ≥ 65 years from the nationally sampled 2015 China Health and Retirement Longitudinal Study cohort. Functional impairment was defined as limitations in ≥2 activities of daily living or instrumental activities of daily living. Training set feature selection yielded 18 input features. Eight algorithms were evaluated in a held-out internal test set. Repeated stratified nested 10-fold cross-validation with five repeats assessed stability conditional on the locked feature set. SHAP assessed feature contributions and ranking stability. Results: Functional impairment was present in 34.21% of participants. Random forest achieved the numerically highest test-set AUC (0.774; 95% CI, 0.714-0.833), with sensitivity of 0.512 and a Brier score of 0.181. Its advantage over other models was modest. Under repeated nested cross-validation, random forest achieved a mean AUC of 0.808 ± 0.043 and a mean Brier score of 0.166 ± 0.016. SHAP rankings were stable (Kendall's W = 0.901); leading contributors included depressive symptoms, self-rated health, 2.5-m walking test completion time, distance vision, history of falls, executive function, and bilateral grip strength. Excluding direct physical performance inputs retained most discrimination but modestly reduced threshold-dependent performance. Conclusions: Random forest provided the most favorable overall internal performance for classifying concurrent functional impairment, although sensitivity remained modest. External and prospective validation is required before clinical implementation.