Objective:A machine learning algorithm was used to develop a predictive model of self-injury among college students and to explore the high-risk factors for self-injury among college students.Methods:From November to December 2022, a convenience sample of 791 college students from a university in Hebei Province was selected.Whether the self-injurious behavior occurred or not was regarded as an outcome variable.The basic demographics data were collected for statistical analysis.The adolescent self-harm questionnaire, the acquired helplessness scale, the Chinese version of the interpersonal needs questionnaire, the adolescent life events scale, and the childhood traumatic experiences questionnaire were used for assessment.The predictor variables were statistically analyzed by SPSS 26.0 software, and the performance of the model was evaluated by random forest, support vector machine and logistic regression so as to predict the self-injury behavior of college students.The model performance was evaluated by the accuracy, F1 score, sensitivity, specificity, and AUC value of the model, and the optimal model was selected.Finally, the optimal model was used to analyze the high-risk factors of college students' self-injury behaviors.Results:(1) The results of one-way ANOVA showed that the detection rate of self-injury behavior among college students was 42.4%(335/791), and the detection rate of male students was significantly higher than that of female students ( χ2=14.139, P<0.05). Individuals with lower-middle monthly household income(RMB 3 000-5 999) had a significantly higher detection rate of self-injury behavior than those with other monthly household income( P<0.05). (2) The accuracy of random forest, support vector machine, and logistic regression models were 85.53%, 85.96%, and 68.86%, F1 scores were 0.853, 0.864, and 0.676, and sensitivities were 83.91%, 89.04%, and 64.91%, respectively.The AUCs of support vector machine, logistic regression models and random forest were 0.89, 0.73 and 0.92.(3) The top ten characteristic variables of high risk factors for college students' self-injury behaviors based on the random forest algorithm with better predictive efficacy were emotional abuse, frustration of belonging, helplessness, interpersonal relationship factor, despair, emotional neglect, academic stress factor, monthly family income, perception of tiredness, and health adaptation factor, in that order. Conclusions:Random forest is optimal for predicting self-injury behavior among college students compared to support vector machine and logistic regression.Factors influencing self-injury behavior among college students originate from environmental factors, individual factors and interpersonal factors.