Federated learning-based prediction of depression among adolescents across multiple districts in China

Depression in adolescents is a serious mental health condition that can affect their emotional and social well-being. Detailed understanding of depression patterns and status of depressive symptoms in adolescents could help identify early intervention targets. Despite the growing use of artificial intelligence for diagnosis and prediction of mental health conditions, the traditional centralized machine learning methods require aggregating adolescents' data; this raises concerns about confidentiality and privacy, which hampers the clinical application of machine learning algorithms. In this study, we use federated learning to solve those problems. We included 583,405 middle and high school adolescents from 20 districts in Chengdu China, and collected from three aspects: individuals, families, and schools, containing 11 psychological phenomena to evaluate the status of depressive symptoms. We compared federated and local training frameworks; the results showed the area under the receiver operating characteristic curve for depression increased by up to 20 % (from 0.7544 with local training to 0.9064 with federated training). Moreover, based on the best-performing model, the XGBoost model, we explore the data heterogeneity in federated learning and found that stress, student burnout, and social connection were the three most important predictors of depression symptoms. We then assessed the impact of each subdimension of depression symptoms, results show that sleep was the most impact one which may provide clues to predict depression symptoms in early stage and improve control and prevention efforts.