Depressive Symptom Predictors in Older Mexican Adults: Interaction Structures and Nonlinear Effects from Machine Learning Explainability

Abstract

Background Depression among older adults remains a critical public health issue in Latin America, where aging populations face compounded social and health vulnerabilities. Objective To characterize psychosocial and clinical predictors of depressive symptoms among older Mexican adults using LASSO-guided logistic regression, Random Forest classification, and post-hoc explainability methods. Methods A cross-sectional analysis was conducted using baseline data from a cohort ( n = 1,252) of adults aged ≥60 affiliated to the Instituto Mexicano del Seguro Social (IMSS) in Mexico City. Depressive symptoms were assessed with the 35-item Center for Epidemiologic Studies Depression scale (CESD-R). A hybrid method combining LASSO-guided logistic regression and Random Forest (RF) classification was applied, followed by post-hoc explainability analysis using SHAP values, Friedman’s H-statistic, and Accumulated Local Effects (ALE) plots. Results Low perceived social support, social isolation risk, and male sex were consistently identified as influential predictors. Logistic regression showed moderate discrimination (AUC-ROC = 0.668) and satisfactory calibration; the RF model yielded significantly lower discrimination (AUC-ROC = 0.588; DeLong’s test p = 0.016). Sex and age operated largely through interactions with other predictors (H = 0.867 and 0.604, respectively). ALE plots identified a U-shaped association between age and depression risk and dose-response gradients for diabetes and hypertension complications. Conclusions Post-hoc explainability methods revealed interaction structures and nonlinear effects undetectable by conventional regression. The LSNS-6 and MOS-SSS can serve as first-line screening tools in routine geriatric consultations. Risk stratification is further informed by comorbidity burden, particularly the presence of type 2 diabetes or hypertension complications, rather than diagnosis alone.