A Two-Stage Deep Learning Model for Mortality Risk Prediction in the General Elderly Population

The global aging population is rapidly expanding, with the proportion of individuals aged 65 and over projected to reach 16% by 2050. This demographic shift presents significant health challenges and socio-economic burdens, making accurate mortality risk prediction crucial for effective healthcare planning and resource allocation. Traditional predictive models, often trained on specific cohorts, may struggle to generalize across the diverse and heterogeneous general elderly population. In this study, we propose a novel two-stage deep learning framework aimed at enhancing mortality risk prediction for the general elderly population. In the first stage, data-driven clustering techniques are applied to multidimensional variables derived from real-world cohorts—including sociodemographic characteristics, physical examinations, physiological assessments, and biochemical analyses—to segment the elderly population into distinct subgroups. The second stage introduces a multi-task learning model to predict mortality risk across these subgroups, integrating shared and task-specific layers to capture both common patterns and subgroup-specific features. Using data from 502,688 elderly individuals from the Shenzhen Healthy Aging Research (SHARE), our model outperforms traditional single-task models and other state-of-the-art AI models, offering superior predictive accuracy and reliability. This framework provides a powerful tool for large-scale mortality risk screening and management in elderly populations.