Predicting longitudinal depressive symptom trajectories among older Chinese adults with chronic health conditions: An interpretable machine learning study.

Objective

This study leveraged interpretable machine learning (ML) to map heterogeneous trajectories of depressive symptoms in Chinese older adults with chronic diseases, aiming to develop an interpretable, prediction-oriented framework for personalized mental health interventions.

Methods

We analyzed four-wave longitudinal data from 5492 participants in the China Health and Retirement Longitudinal Study. Following trajectory identification, 10 ML algorithms were compared. A 50-iteration bootstrap Recursive Feature Elimination (RFE) distilled 10 core predictors from 39 baseline features. Models were evaluated using an 80:20 stratified split, with MICE imputation strictly preventing data leakage. Model interpretability was extracted via SHapley Additive exPlanations (SHAP).

Results

Three distinct trajectories emerged: Consistently Low, Chronically High, and Rapidly Escalating Risk. The Elastic Net model demonstrated optimal discriminative power (ROC-AUC = 0.710, PR-AUC = 0.468) and a Brier score of 0.337. SHAP analysis indicated that extremely low life satisfaction, severe instrumental functional limitations, and poor self-rated health were strongly associated with higher predicted probabilities of high-risk trajectories, whereas high household income robustly protected the low-risk group. We translated these insights into an interactive web-based risk calculator.

Conclusion

This study identified critical depressive trajectory classes and established an interpretable predictive model. By translating complex analytics into an interpretable predictive framework, this approach provides a foundation for individualized risk profiling and highlights distinct patterns of risk that may inform future targeted mental health interventions for older adults with chronic conditions.