Applying Multiple Machine Learning Methods to Identify Correlates of Non-Suicidal Self-Injury Behaviors in Chinese Middle School Students: A Cross-Sectional Study

Background: Adolescent non-suicidal self-injury (NSSI) has become a major public health concern, with its rising prevalence driven by complex and interacting factors. Traditional analytical methods have limited capacity to capture non-linear relationships among high-dimensional variables. This study aimed to compare the performance of five machine learning algorithms in identifying NSSI and to determine its key correlates among Chinese adolescents. Participants and Setting: A total of 2,295 middle school students from Hebei and Shandong provinces in China, aged between 11 and 16 years, participated in the study. Methods: Five machine learning algorithms—Random Forest, AdaBoost, XGBoost, LightGBM, and Naive Bayes—were applied to identify NSSI using 43 features encompassing psychological traits, emotional states, and environmental factors. SMOTE was used to address class imbalance, and hyperparameters were tuned via 5-fold cross-validated grid search. Model performance was evaluated using ROC AUC, PR AUC, accuracy, sensitivity, specificity, PPV, and NPV. SHAP analysis was employed to determine feature effect directionality. Results: The prevalence of NSSI was 21.7%. Individuals with NSSI reported significantly higher levels of Academic Pressure, Self-criticism, Negative self-schema, Hopelessness, and Depression ( P < 0.001). XGBoost achieved the highest ROC AUC (0.884), PR AUC (0.741), accuracy (85.62%), and specificity (92.50%), while Naive Bayes achieved the highest sensitivity (71.72%). Self-criticism ranked first in four of the five models. SHAP analysis confirmed that higher Self-criticism, Negative self-schema, Academic Pressure, Depression, and Hopelessness were associated with increased NSSI probability. Conclusion: Self-criticism, Negative self-schema, Academic Pressure, Depression, and Hopelessness were consistently identified as the most important correlates of NSSI across models. These results highlight the potential of machine learning approaches for identifying key correlates of adolescent NSSI, although external validation is needed before clinical application.