Who’s at risk for emergent depression years later? Predictive modeling in a nine-year longitudinal cohort

Overview

This longitudinal cohort study applied explainable artificial intelligence methods to identify predictors of emergent major depressive disorder (MDD) in adulthood across a nine-year interval. A community-based sample of 931 adults without baseline MDD diagnosis was followed from 2004–2006 to 2013–2014, during which 6.23% developed incident MDD. The investigation leveraged 46 validated composite variables spanning inflammatory, developmental trauma, coping, emotion regulation, personality, and social support domains to generate predictive models.

Methods and approach

Six machine-learning algorithms were evaluated with nested five-fold cross-validation, incorporating multiple configurations of predictor set length and missing data handling strategies. Shapley additive explanations (SHAP) analysis was used to characterize the direction and magnitude of multivariable predictor contributions. Elastic net regression emerged as the optimal classifier, with performance evaluated using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and calibration indices.

Key Findings

Elastic net regression achieved an AUC of 0.724 (95% confidence intervals 0.657–0.792), with acceptable sensitivity and high negative predictive value. Calibration indices indicated moderate-to-good alignment between predicted and observed probabilities. Key psychosocial predictors of elevated emergent MDD risk included greater perceived stress, early life minimization and stress exposure, and family/spousal strain. Protective or risk-reducing factors encompassed problem-focused coping strategies and personality attributes including self-acceptance, sense of control, and self-directedness. Behavioral disengagement showed elevated risk associations. Demographic risk factors included younger age and racial minority status. Comorbid generalized anxiety disorder, panic disorder, and substance use disorder symptom severity at baseline further stratified risk.

Implications

Explainable artificial intelligence approaches demonstrate feasibility for distal risk prediction of emergent MDD using theoretically grounded, readily obtainable biopsychosocial variables. The identified predictor profile suggests targets for preventive intervention including stress management, coping skill enhancement, and emotion regulation optimization. If externally validated, such multivariable predictive models could facilitate integration into healthcare delivery systems to enable stratified prevention and early intervention protocols. Future implementation research should address validation in diverse populations, cost-effectiveness of risk stratification approaches, and mechanisms by which identified risk factors mechanistically contribute to MDD emergence.