Predicting non-suicidal self-injury behavior among adolescents with depressive disorders: a comparative study based on different machine learning methods.

Predicting non-suicidal self-injury (NSSI) among adolescents is challenging due to its complex behavioral drivers and diverse predictors. Machine learning (ML) methods, though promising in various predictions, have been underexplored for NSSI among adolescents, particularly with large-scale cross-sectional datasets. In this study, we compiled 30 potential NSSI predictive variables from existing research and used psychological survey data from 2,343 participants across 14 Chinese psychiatry hospitals. The dataset was split into a 7:3 ratio for training and testing. Various ML models (logistic regression, Naive Bayes, SVM, decision tree, random forest, and AdaBoost) were applied, with performance evaluated using accuracy, precision, recall, F1 score, and AUC. All selected ML methods generally performed well in predicting adolescent NSSI. Random forest achieved the highest AUC (0.823), followed by AdaBoost (0.771), decision tree (0.760), logistic regression (0.758), SVM (0.714), and Naive Bayes (0.712). Compared to other methods, Naive Bayes achieved a significant lower accuracy value (0.68) and recall value (0.68) due to its prerequisite assumption of distribution of dataset. With the better explainable nature of RF model, we ranked the importance value of top 20 factors and found that psychological resilience showed the strongest influence in contributing to NSSI. The results suggest that ML methods applied in this study could hold significant value in predicting NSSI among Chinese adolescents with depressive disorders.