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

BACKGROUND

Major depressive disorder (MDD) is prevalent in adulthood, but there remains a dearth of studies identifying predictors of emergent MDD in adulthood with high-dimensional biopsychosocial predictor sets. Our study thus examined how explainable artificial intelligence (XAI) models might accurately detect predictors of emergent MDD nine years later.

METHODS

Community adults who did not meet diagnostic criteria for MDD at Wave 1 (W1; 2004–2006) participated in the current study (N = 931). Forty-six W1 validated composite variables, including inflammation, childhood maltreatment, coping, emotion regulation, personality, and social support, were used to predict emergent MDD at Wave 2 (W2; 2013–2014). Six machine-learning models, each with four varying configurations of predictor set length and missingness handling strategies, were tested using five-fold nested cross-validation to determine which model had the best multivariable predictive performance. Shapley additive explanations (SHAP) analysis informed the sign and strength of each multivariable predictor.

RESULTS

Elastic net regression achieved the best classification accuracy (AUC = 0.724; 95% confidence intervals = 0.657–0.792), with moderate sensitivity and a high negative predictive value in predicting W2 emergent MDD, observed in 6.23% of the sample. Moderate-to-good calibration values were also observed, highlighting acceptable alignment between predicted probabilities and observed prevalences. Key psychosocial correlates of higher W2 emergent MDD risk included greater perceived stress, early life minimization and stress, as well as family and spousal strain. Other key correlates included fewer problem-focused coping strategies, lower self-acceptance, sense of control, and self-directedness, as well as greater tendencies for behavioral disengagement. Demographic correlates included younger age and racial minority identity. Comorbid mental health symptoms, especially higher W1 generalized anxiety disorder, panic disorder, and substance use disorder symptom severity, were also clinical correlates of greater W2 emergent MDD risk.

CONCLUSIONS

XAI may inform clinically actionable distal risk modeling for emergent MDD using easily measurable, theory-driven variables. If externally validated, scalable multivariable predictive models could be integrated into healthcare systems to inform prevention strategies. These predictive models might inform tailored treatment strategies, including enhancing approach-focused coping, emotion regulation, and appraisals, as well as objective indicators of social support and stress. CLINICAL TRIAL NUMBER: Not applicable.