M40. PREDICTING DEPRESSION FROM ADOLESCENCE TO ADULTHOOD USING MACHINE LEARNING: INSIGHTS FROM A NATIONALLY REPRESENTATIVE COHORT

Early identification of individuals at risk for depression is essential for timely intervention. While genetic risk scores are increasingly used in psychiatric research, their predictive utility remains limited. This study assessed whether adolescent psychosocial and environmental data could predict depressive outcomes up to two decades later. We further evaluated sex-specific model performance and the incremental value of polygenic scores (PGS) for depression. Data were drawn from Wave I of the National Longitudinal Study of Adolescent to Adult Health (Add Health; N = 15,701). Six machine learning (ML) algorithms—logistic regression, support vector machines (linear and RBF kernels), random forest, XGBoost, and a fully connected neural network—were trained on adolescent psychosocial and environmental features. Outcomes included depressive symptoms in adolescence and adulthood, as well as clinical diagnoses of depression at both timepoints. Models were evaluated using the area under the receiver operating characteristic curve (ROC-AUC). Analyses were stratified by sex and repeated with the inclusion of depression PGS to assess their contribution to predictive performance. We also performed feature importance to verify the most relevant questions. XGBoost consistently yielded the highest predictive performance across outcomes. For adolescent depressive symptoms, the top AUC was 0.849 (females). Predictive accuracy declined for adult depressive symptoms (maximum AUC = 0.661) and adult clinical diagnoses (AUC = 0.636), likely reflecting increased influence of life-course exposures. Models trained on adolescent data predicted clinical diagnoses in adolescence with moderate accuracy (AUC = 0.788). Non-linear models consistently outperformed linear approaches, but linear models can still be used with some success. Feature importance analyses indicated shifts over time: emotional well-being, parental connectedness, and academic behaviors were significant predictors during adolescence, whereas general health, behavioral risks, and family history were more relevant to adult outcomes. Sex-stratified models revealed distinct predictive profiles. The addition of PGS did not significantly improve the prediction when used in combination with the environment. Models with and without PGS yielded nearly identical AUCs, suggesting that the current genetic risk scores have limited incremental utility in the presence of rich environmental data. Adolescent psychosocial and environmental variables allow for moderate-to-high prediction of short- and long-term depressive outcomes, particularly among females. While predictive performance declines over time, early-life features remain informative for adult diagnoses. The lack of improvement with PGS suggests that the added predictive value of current genetic predictors may be limited in the presence of comprehensive environmental and psychosocial data. These findings support adolescence as a critical window for identifying individuals at elevated risk for depression and suggest that ML models, especially XGBoost, offer promising tools for early risk stratification in psychiatric research.