Longitudinal machine learning prediction of non-suicidal self-injury among Chinese adolescents: A prospective...

Non-suicidal self-injury (NSSI) is an important public health problem among adolescents, yet traditional prediction approaches yield limited accuracy. This study aims to develop and evaluate machine learning models for predicting NSSI and to examine the patterns of feature importance using longitudinal data.

Data were obtained from the Chengdu Positive Child Development Cohort, which covers students from five primary and middle schools in China. The analysis utilized four waves of longitudinal data collected from 3,483 students over a period of 2.5 years. A progressive prediction framework with three prediction windows was constructed, and seven machine learning algorithms were compared. Five-fold cross-validation was used to ensure the robustness of performance evaluation. SMOTE-NC (Synthetic Minority Over-sampling Technique-Nominal Continuous) was used to address the class imbalance in the training set. Model performance was evaluated by area under the receiver operating characteristic curve (AUROC), accuracy, precision, recall, and F1 score. SHapley Additive exPlanations (SHAP) analysis was conducted to assess the models' interpretability.

Random Forests showed superior performance in all windows (AUROC = 0.843, 0.855, and 0.853, respectively). The top predictive risk factors included suicide-related behaviors, depression, delinquent behaviors, and anxiety, while spirituality, emotional competence, life satisfaction, and empathy emerged as protective factors.

The progressive prediction framework achieved robust longitudinal NSSI prediction. Positive child development factors, particularly spirituality and emotional competence, emerged as key protective factors against NSSI risk. These model-interpretable results provide an evidence-based foundation for developing targeted prevention strategies and early intervention programs.