Census-based adolescent depression risk assessment: A novel method and reflections

Diagnosing psychological issues in adolescents is generally more challenging than in adults. The need to enhance early identification and risk assessment for adolescent mental health problems has become particularly urgent, especially given the surge in such issues brought about by the COVID-19 pandemic and its isolation policies. Traditional approaches relying on scales or interviews can be difficult to administer at scale and may miss at-risk youths, while recent attempts at artificial intelligence-assisted clinical diagnosis have not yet achieved consistent performance for routine clinical use. Using youth census survey data from the National Survey of Children's Health (NSCH) from 2020 to 2022 (147,772 samples), we propose a novel data synthesis and training management method tailored for imbalanced health data, and develop a heuristic decision model based on daily-life indicators to predict survey-reported adolescent depression risk. Experimental results demonstrate that our data management approach improves the average accuracy of conventional models by 14%. Compared with general artificial intelligence methods, the proposed decision model achieves higher accuracy, precision, and sensitivity. Built on AutoML, the model can adapt to factor replacement and incremental data updates, supporting scalable screening workflows and model maintenance. Using Shapley values, we identify several key daily-life factors associated with adolescent depression risk, such as social interactions and sleep duration. The interpretability analysis further supports the utility of our approach for risk stratification and practical decision support, rather than replacing clinician-administered diagnosis.