Prediction of disability onset at different time intervals among Chinese middle-aged and older adults with arthritis: an interpretable machine learning approach

This study aimed to build models using machine learning approaches to predict disability onset at different time intervals among middle-aged and older adults with arthritis and free of baseline disability. This study retrospectively collected data from the China Health and Retirement Longitudinal Study (CHARLS). We included 1,925 and 1,800 community-dwelling adults aged 45 years or older with arthritis and no baseline disability in 2015 for the 2015–2018 and 2015–2020 predictions, respectively. Machine learning (ML) models were constructed and evaluated to assess the risk of basic activities of daily living (BADL) disability, instrumental activities of daily living (IADL) disability, activities of daily living (ADL) disability (BADL or IADL disability), and both disabilities (BADL and IADL disability). Shapley Additive exPlanations (SHAP) was further deployed to explore the crucial predictors in model decisions. Overall, the candidate prediction algorithms achieved moderate performance, and logistic regression (LR) exhibited better performance than more flexible machine learning methods with the maximum area under the receiver operating characteristic curve (AUROC) for BADL, IADL, ADL, and both disabilities predictions of 0.666, 0.699, 0.700, 0.717 in the 2015–2018 period, and 0.719, 0.721, 0.722, 0.766 in the 2015–2020 period. The results of SHAP revealed that depression score, age, weight, height, lung function, and cognitive function played a crucial role in disability prediction, especially depression score and age. The findings suggest that prediction models may help identify arthritis patients at high risk of disability. The important predictors identified in this analysis may facilitate future prediction studies, but their causal roles require dedicated causal analyses.