Development and validation of a simplified time-dependent interpretable machine learning-based survival model for older adults with multimorbidity

Abstract

Multimorbidity elevates late-life mortality, yet existing tools remain complex. Using two nationally representative Chinese cohorts—the Chinese Longitudinal Healthy Longevity and Happiness Family Study (CLHLS-HF; n = 8675) and the China Health and Retirement Longitudinal Study (CHARLS, n = 4171)—we developed and externally validated a simplified, time-dependent, interpretable survival model. A four-stage feature-selection pipeline (univariate Cox, L1-penalized Cox, multi-model importance with 100 bootstraps, and cumulative performance) identified four routinely available predictors: age, BMI, and cooking and toileting abilities. Among five algorithms, a parsimonious Cox model performed best (C-index 0.7524 internal; 0.7104 external) with a favorable time-Brier Score (0.1417; 0.1157), good calibration, decision-curve net benefit, and subgroup fairness. Time-dependent permutation importance confirmed age as dominant, toileting ability as short-term, and cooking ability as mid- to long-term contributors, while BMI showed modest, stable effects. Implemented as the M-SAGE online tool, this four-item model enables rapid, interpretable mortality risk stratification and supports individualized interventions for older adults with multimorbidity.