Machine learning-based early prediction of multiple chronic disease risk in aging Chinese population: A longitudinal analysis using CHARLS data

Population aging has intensified the burden of multimorbidity among older adults, necessitating effective tools for the early identification of high-risk individuals. This retrospective cohort study utilized data from 8552 participants in the China Health and Retirement Longitudinal Study (2011-2018) to develop and validate machine learning models for predicting incident multimorbidity, defined as the new-onset co-occurrence of 2 or more chronic conditions among participants free of multimorbidity at baseline, assessed across 14 physician-diagnosed diseases. Five algorithms, including logistic regression, random forest, extreme gradient boosting, support vector machine, and k-nearest neighbors, were compared, with temporal validation conducted using an independent cohort of 3218 participants from the 2013 China Health and Retirement Longitudinal Study wave. Extreme gradient boosting achieved the best discrimination (area under the receiver operating characteristic curve = 0.803 in testing, 0.779 in temporal validation) with acceptable calibration (Hosmer-Lemeshow P = .189). Baseline chronic condition status, age, self-rated health, and depressive symptoms were the most influential predictors, with health status indicators collectively contributing the largest proportion of predictive importance. Machine learning algorithms can effectively stratify multimorbidity risk in aging Chinese populations, and the identified predictive factors offer potential directions for risk-focused surveillance and preventive strategies in primary care settings.