Explainable Machine Learning Reveals Pattern-Specific Drivers of Depression in Middle-Aged and Older Adults with Multimorbidity: A Nationwide Cross-Sectional Study from CHARLS

Background/Objectives: Depression is a leading cause of disability among middle-aged and older adults, with its burden markedly amplified by multimorbidity. In rapidly aging China, distinct multimorbidity phenotypes interact with depressive symptoms through complex, non-linear pathways that traditional models struggle to capture. Interpretable machine learning offers high predictive accuracy alongside clinically transparent explanations, yet few studies have applied this approach to large national Chinese cohorts while stratifying by empirically derived multimorbidity patterns. The aim of this study was to identify the pattern-specific determinants of depression among middle-aged and older adults with multimorbidity using explainable machine learning techniques based on data from the China Health and Retirement Longitudinal Study (CHARLS). Methods: In this cross-sectional analysis of the China Health and Retirement Longitudinal Study (CHARLS) 2020 data, 14,981 participants were divided into a multimorbidity group (≥2 of 12 physician-diagnosed chronic conditions; n = 8265) and a matched control group (0 or 1 condition; n = 6716). Depressive symptoms were defined as a CES-D-10 score ≥ 10 (binary outcome). Forty-four candidate predictors were screened via LASSO regression, the Boruta algorithm, and univariate logistic regression. Fifteen supervised machine learning models were trained separately for each group using SMOTE for class imbalance, nested 5-fold cross-validation for hyperparameter tuning, and an independent 20% test set for external validation. Model performance was ranked by the composite AB Score (equal weighting of AUROC and Brier score). The champion models were interpreted globally via SHAP summary plots, directional waterfall diagrams, and novel impact-intensity network diagrams. The same pipeline was repeated across four literature-derived multimorbidity patterns within the multimorbidity group. Results: Depressive-symptom prevalence was 42.96% (3551/8265) in the multimorbidity group versus 27.01% (1814/6716) in the control group. The Gaussian process classifier achieved the highest AB Score (0.7782) in the multimorbidity group (external AUROC of 0.7545, Brier score of 0.1981); the generalized additive model was optimal in the control group (AB Score 0.7856, external AUROC 0.7387, Brier score 0.1676). Across both groups, the top contributors by mean absolute SHAP were IADL disability, abnormal sleep duration, ADL disability, rural residence, and lower education level. Pattern-specific SHAP networks revealed clinically meaningful heterogeneity: cardio-metabolic clusters were driven by marital status and headache; digestive–joint clusters by IADL disability and shoulder–knee pain synergy; respiratory clusters by gender and multisite pain; and cardiovascular–digestive clusters by sleep duration and vigorous activity. Conclusions: Interpretable machine learning models demonstrated clinically useful discrimination and excellent calibration while revealing universal drivers (functional disability and sleep disturbance) and cluster-specific synergistic interactions underlying depressive symptoms in multimorbid Chinese adults. These phenotype-tailored insights provide a practical roadmap for precision screening and targeted interventions in aging populations.