Prospective prediction of first onset of major depressive disorder in midlife using machine learning.

Purpose: In this paper we leverage machine learning (ML) models to prospectively predict the first onset of Major Depressive Disorder (MDD), one of the most common and disabling mental health conditions. While such prediction models hold potential for enabling early interventions, few studies have applied ML approaches to this task, and those that have are heterogeneous in nature. Moreover, the clinical utility of these predictive models remains largely unexamined.

Methods: Data stemmed from CoLaus|PsyCoLaus, a population-based cohort study. In total, 1350 participants, age 35-66 years without lifetime MDD at baseline participated in the physical and psychiatric baseline and at least one psychiatric follow-up evaluation. Models based on logistic regression, elastic net, random forests, and XGBoost were trained using an extensive array of psychosocial, environmental, biological, and genetic predictors. Discriminative performance, calibration, clinical utility, and individual predictor contributions were assessed using nested cross-validation.

Results: Discriminative performance was comparable between models (areas under the precision-recall curve between 0.36 and 0.38; areas under the receiver operating characteristic curve between 0.65 and 0.68). Decision curve analysis suggested clinical utility of logistic regression, elastic net, and random forests for threshold probabilities between 10% and 40%. Across all models, neuroticism, sex, and age were the most important predictors.

Conclusions: Although the prediction models achieved discriminative performance levels above chance, further refinement is necessary. The addition of biological and genetic predictors did not elevate performance markedly. Additional research seems warranted given the limited number and heterogeneous nature of existing studies, the burden associated with MDD, and the potential to improve overall outcomes for people at risk for MDD.