Machine Learning-Based Multidimensional Health Decline Prediction Framework: Data-Driven Modeling for the Middle-Aged and Elderly Population

Background: With the worsening of global population aging, individuals aged 45 and above face numerous challenges, including disability, pain, cognitive impairment, hearing, and depression. This study aims to utilize machine learning and deep learning methods to develop predictive models for forecasting the health outcomes and disease risk among middle-aged and elderly individuals, while identifying the key factors influencing disease prevalence and quality of life in this demographic. Methods: This research is grounded in the China Health and Retirement Longitudinal Study (CHARLS) database (2015–2018), encompassing 20,967 participants. Machine learning and deep learning techniques were employed to create predictive models for disability, pain, cognitive impairment, hearing, and depression, utilizing 69 variables. A minimum redundancy maximum relevance (MRMR) feature selection strategy and incremental feature selection (IFS) were applied to identify key variables. The optimal predictive models were evaluated for accuracy using the area under the receiver operating characteristic curve (ROC-AUC). Results: The analysis resulted in 15 predictive models, with TabPFN showing the highest performance in predicting disability, pain, cognitive impairment, hearing, and depression, achieving AUCs of 0.802, 0.860, 0.814, 0.769, and 0.856, respectively. Shapley interpretability analysis identified the top five critical features affecting these health outcomes. Conclusions: The proposed models can assess the disease risk of various conditions in middle-aged and elderly individuals, suggesting that early prevention and effective intervention may reduce incidence rates. Additionally, the study provides insights into key factors influencing health, offering a scientific foundation for tailored health management and precision intervention strategies.