Early identification of pediatric depression in western China: A multicenter, citywide evaluation of nine machine learning models.

BACKGROUND: Childhood depression is an emerging global concern, yet knowledge and early detection tools remain limited in low- and middle-income regions (LMIRs). METHODS: To address this gap, a school-based prospective cohort study was conducted in Nanchong, Sichuan Province, China, from June 2023 to May 2024. A representative sample of 136,699 children aged 9-13 years was surveyed to develop and validate a prediction model for depressive symptoms. Depression was assessed using the Center for Epidemiologic Studies Depression Scale (CESD). Chi-square tests and multivariate logistic regression identified fifteen significant predictors (p < 0.001). Nine machine learning models including Support Vector Machine, Naive Bayes, Decision Trees, Logistic Regression, Random Forest, LightGBM, AdaBoost, XGBoost, and Multilayer Perceptron were trained and validated. RESULTS: A training-testing set (n = 112,742) and an independent validation set (n = 23,957) were identified a depression rate of 8.8 %. All models demonstrated good performance (AUC = 0.796-0.830). Logistic regression (LR) exhibited the most balanced overall performance (AUC = 0.827; recall = 0.721;specificity = 0.785). Misunderstandings and disputes with peers in past year, subjective learning experiences, gender, and sleep duration on school days were identified as the most important predictors in the Logistic model with SHAP. LIMITATIONS: This single-city study may limit generalizability to other regions and cultural contexts. Additionally, reliance on a self-report scale (CESD) may introduce subjective bias in symptom assessment. CONCLUSIONS: The models showed satisfactory performances in early screening for pediatric depression. These algorithms can assist in early identification and allow timely interventions cost-effectively.