Machine learning-based predictive factor analysis of depression among Chinese adolescents.

INTRODUCTION: Adolescent depression has emerged as a critical global public health concern, with rising prevalence in China posing severe threats to psychological development and social adaptation. Traditional statistical methods face limitations in capturing complex non-linear relationships and interactions among influencing factors, while machine learning algorithms offer advantages in predictive modeling of mental health disorders.

OBJECTIVE: This study aimed to: (1) compare the performance of seven ML algorithms in classifying low and high depression risk groups among Chinese adolescents; (2) identify key predictive factors from demographic, personality, and PGI-related variables; (3) explore non-linear relationships and interactive effects between critical factors; and (4) explore preliminary threshold values for key factors as potential references for risk identification.

METHODS: A total of 559 Chinese adolescents completed assessments of demographic characteristics, Big Five personality traits, personal growth initiative, and depression symptoms. Model performance was compared using Friedman tests and Nemenyi post-hoc tests appropriate for correlated cross-validation data. Seven ML algorithms were trained and optimized using 5-fold cross-validation. Feature importance was analyzed via traditional metrics and SHAP values, and SHAP interaction effects were tested using permutation tests. Threshold analysis was conducted using the Youden's J statistic.

RESULTS: LightGBM outperformed other models with an AUC of 0.834, achieving balanced accuracy, sensitivity, and specificity. Neuroticism emerged as the most robust predictor across all models, followed by proactive change, agreeableness, extraversion, and growth resilience. Demographic factors showed minimal predictive power. SHAP permutation tests confirmed significant interactions between neuroticism and proactive change and between proactive change and agreeableness, whereas no significant interaction was found between neuroticism and agreeableness. Preliminary thresholds were identified for key factors within this sample.

CONCLUSION: ML algorithms, particularly lightGBM, effectively identify adolescent depression risk, with personality traits and PGI serving as core predictive factors. The findings highlight the value of integrating multi-dimensional variables in depression prediction and provide preliminary references for early intervention. Given the cross-sectional design and lack of external validation, conclusions regarding generalizable cutoffs and causal inference should be made with caution. Targeted strategies focusing on reducing neuroticism and enhancing proactive growth behaviors may mitigate depression vulnerability in Chinese adolescents.