Predicting six-year depression trajectories in adulthood: A super learner ensemble approach using the Add Health dataset.

BACKGROUND: Depression is a leading global cause of disability, and much focus has been directed toward identifying population-wide predictors of adult depression. The present study aimed to identify the strongest predictors of future depression severity using a comprehensive panel of biopsychosocial predictors and the Super Learner (a stacked ensemble algorithm based on cross-validation loss). METHOD: Participants were 3208 adults from the Add Health dataset, a nationally representative longitudinal cohort study of U.S. adults, split into training (n = 2247) and fully held-out test (n = 961) sets. Data came from two waves: 2016-2018 and 2022-2025. Depression severity was measured at both waves using the 5-item negative affect subscale of the Center for Epidemiologic Studies Depression Scale. The Super Learner ensemble comprised ten base learners, including elastic net and random forest. Shapley Additive Explanations (SHAP) values were calculated on the held-out test set. RESULTS: The Super Learner did not outperform all base learners on the test set (root mean squared error = 2.315). Bootstrap and equivalence testing indicated it was not reliably different from eight of ten learners. SHAP analysis revealed baseline depression score as the most influential predictor of future depression score (mean |SHAP| = 0.416), followed by perceived stress, prior depression diagnosis, everyday discrimination, and social gathering frequency. CONCLUSION: Depression history was the strongest predictor of depression severity after six years. No model explained more than a quarter of outcome variance. This ceiling and the ensemble's non-superiority compared to other constituent single algorithms tempers individual-level prediction expectations.