Machine learning and network analysis of psychosocial factors associated with non-suicidal self-injury severity among college students.

Background

Non-suicidal self-injury (NSSI) is a prevalent and clinically significant behavior. Although multiple psychosocial factors are linked to NSSI, their relative predictive importance and conditional structure remain unclear. This study applied complementary machine learning and network analysis approaches to examine psychosocial correlates of NSSI severity.

Methods

Participants were 2436 college students in China (27.1% male; mean age = 19.77 years). Psychosocial variables included emotion dysregulation, impulsivity, parenting style, negative life events, perceived social support, and emotion regulation strategies. Five supervised machine learning models were compared for predicting NSSI severity, and Shapley Additive Explanations (SHAP) analysis characterized predictor contributions. EBICglasso network analysis characterized psychosocial network structure and subgroup differences by NSSI status and sex.

Results

Regularized linear models performed most consistently across validation and test sets, and Elastic Net was selected for stable predictive performance and robustness to correlated predictors (RMSE = 5.33, R² = 0.18). Feature importance and SHAP analyses identified emotion dysregulation, maltreatment experience, suicide exposure, negative life events, and parenting-related factors as important features associated with NSSI severity, whereas perceived social support showed an inverse predictive pattern. Network analysis revealed a broadly modular partial-correlation structure, with emotion dysregulation showing relatively high strength centrality. Network Comparison Tests indicated significant differences by NSSI status and exploratory sex-related differences in overall network structure.

Conclusions

Both approaches highlighted emotion dysregulation and negative life events as psychosocial correlates of NSSI severity. These findings support the complementary value of predictive and network approaches for characterizing NSSI-related psychosocial vulnerability and generating hypotheses for longitudinal research.