Objective: To systematically evaluate the performance of machine learning (ML) models in predicting non suicidal self injury (NSSI) behavior among adolescents, providing an evidence based foundation for the development of clinically applicable risk assessment tools. Methods: A comprehensive search was conducted in PubMed, Embase, Web of Science, CNKI, and Wanfang databases for relevant studies from their inception to July 21, 2025. The Prediction Model Risk of Bias Assessment Tool (PROBAST) was used to evaluate the methodological quality of the included studies. Stata 18 software was used to calculate the area under the receiver operating characteristic curve (AUC) of the models, and publication bias assessment, sensitivity analysis, and Egger s test were performed. Results: The total of 12 studies (42 prediction models) involving 58 070 adolescents were included. There were 15 machine learning algorithms in total, among which Random Forest, Logistic Regression, XGBoost, and Support Vector Machines were the most frequently utilized. The most common predictors were gender (female), family function, depression, emotion regulation, and age. All 12 included studies exhibited a high risk of bias. The pooled AUC was 0.80 (95% CI =0.78-0.82), although heterogeneity was substantial ( I 2=95.8%, P <0.01). Sensitivity analysis confirmed the robustness of these findings (no overlap in 95% CI ), while Egger s test indicated the presence of publication bias ( P <0.05). Conclusions Machine learning demonstrates potential in the risk prediction of adolescent NSSI, but existing models have a high risk of bias. Future research should focus on improving methodological quality and optimizing model reliability through rigorous external validation.