Abstract
BACKGROUND: Depressive symptoms in adolescents can significantly affect their daily lives and pose risks to their future development. These symptoms may be linked to various factors experienced during both childhood and adolescence. Machine learning (ML) has attracted substantial attention in the field of adolescent depression; however, studies establishing prediction models have primarily considered childhood or adolescent features separately, resulting in a lack of analyses that incorporate factors from both stages. METHODS: We collected 39 features related to childhood and adolescence. Using the maximum relevance-minimum redundancy method and four ML algorithms, we determined the optimal feature subset for identifying depressive symptoms and constructed child-adolescent models. Stepwise logistic regression and four ML methods were employed to create demographic and combined models, respectively. The performance of each model was evaluated using a test set, and SHapley Additive exPlanations (SHAP) were utilized to interpret the models' prediction results. RESULTS: The proposed child-adolescent models exhibited superior performance on the test set than the demographic and combined models (AUC: 0.835-0.879 versus 0.530 and 0.840-0.876, respectively). The optimal feature subset included two childhood features (relationship quality with peers and parental absence) and four adolescence features (social trust, academic pressure, importance of the internet for entertainment, and positive parenting behaviour). These features were found to be more effective than demographic characteristics in distinguishing depressive symptoms in adolescents. CONCLUSIONS: This study demonstrates the correlation between adolescence depressive symptoms and specific factors from both childhood and adolescence, as well as the potential of ML to predict it. These findings may serve as a reference for future intervention studies.