非自杀性自伤行为是全球公共卫生领域备受关注的问题，会对青少年的身心健康产生严重消极影响。已有研究发现个体因素、家庭因素和学校因素都会对自伤行为产生影响，但当前的研究主要关注单个或少数几个变量对自伤行为的影响机制，而缺乏一个整合的模型来预测自伤行为的发生。机器学习是一种能够在大量数据中寻找规律，从而对新数据做出预测的方法。有研究者采用机器学习的方法对自伤行为进行预测，但纳入因素较为有限，多集中于个人因素，而对环境因素的关注较少。本研究的目的在于采用机器学习的方法，综合考虑个人、家庭和学校因素，构建一个能够有效预测青少年非自杀性自伤行为的模型，并进一步探讨预测模型中不同特征的重要性。

本研究以问卷调查的方式收集了1658份有效问卷，主要包括个人因素（如抑郁、焦虑、睡眠障碍）、家庭因素（如父母教养方式）和学校因素（如校园欺凌）等变量，基于以上变量构建逻辑回归和机器学习模型（随机森林、决策树、支持向量机和多层感知器）预测自伤行为和重复自伤行为。

研究发现：(1)初中生非自杀性自伤行为的发生率为44.0%，不同自伤方式的发生率存在差异；(2)性别、留守、独生对自伤行为存在影响，个人因素中的抑郁、焦虑、睡眠障碍等，家庭因素中的父母教养方式和学校因素中的校园欺凌、歧视都会影响自伤行为，但未发现年龄和年级对自伤行为的显著影响；(3)构建的逻辑回归和机器学习模型都能够有效预测自伤和重复自伤行为，其中逻辑回归和支持向量机的预测效果最好（预测自伤的准确性均为0.69，预测重复自伤的准确性分别为0.75和0.76）；(4) 所有模型中重要性排名前5位的特征，出现次数最多的是抑郁、焦虑、睡眠障碍、生理压力体验、校园欺凌、外化行为问题和父亲关爱，说明这些变量是自伤行为的重要预测因素。

本研究同时考察个人、家庭和学校因素对非自杀性自伤行为的影响，对于自伤行为的影响因素和功能作用有更加全面的探讨，发现学校因素中校园欺凌、家庭因素中父亲关爱以及个人因素中抑郁、焦虑、睡眠障碍等变量对自伤行为的重要作用；同时通过构建非自杀性自伤行为的预测模型，为自伤行为的预防和干预提供依据，减少青少年自伤行为的发生，从而促进青少年的身心健康发展。

Non-suicidal self-injury (NSSI) represents a significant worldwide public health issue, with the potential to exert adverse effects on the physical and mental health of adolescents. Previous studies have found that all of the individual, family and school factors can lead to the risk of NSSI, but current research mainly focuses on the relationship and mechanism between NSSI and a single or just a few variables, lacking an integrated model to predict the occurrence of self-injury. Machine learning makes it possible to make predictions by using large amounts of data. Some researchers have used machine learning methods to predict NSSI, but the predictors are relatively limited, paying most attention to individual factors and little to family and school factors. This study aims to use machine learning algorithms to develop an effective model which can accurately predict the NSSI of adolescents considering all of individual, family, and school factors.

A total of 1658 valid questionnaires were collected. Individual factors (such as depression, anxiety and sleeping disorder), family factors (such as parenting styles) and school factors (such as bullying) are measured. Based on the above variables, Logistic Regression and machine learning models (Random Forest, Decision Tree, Support Vector Machine, and Multilayer Perception) were developed to predict NSSI and repeated NSSI.

The main results of the study are as follows: (1) The prevalence of NSSI in junior high school students is 44.0%, and there are variations in the frequency of different self-harm behaviors; (2) Gender, left-behind and being the only child have an impact on NSSI. Individual factors such as depression, anxiety and sleep disorders, family factors such as parenting styles, and school factors such as bullying affect NSSI. No significant influence of age and grade on NSSI was found; (3) All of the Logistic Regression and machine learning models can effectively predict NSSI, with Logistic Regression and Support Vector Machines having better predictive effects than other models (the accuracy of predicting NSSI is 0.69, and the accuracy of predicting repeated NSSI is 0.75 and 0.76 respectively); (4) Among the top 5 most important features in all models, depression, anxiety, sleep disorders, physiological stress experience, bullying, externalizing behavior problems and father’s care appeared most frequently, which proved to be important predictors of NSSI.

This study examines the influence of individual, family, and school factors on NSSI at the same time, finding that bullying in school factors, father's care in family factors, and depression, anxiety, sleep disorders, physiological stress experience and externalizing behavior problems in individual factors play the most important roles in predicting NSSI. Meanwhile, by developing a predictive model of NSSI, we provide an effective way to identify the individuals at high risk of NSSI, thus NSSI can be prevented and reduced, promoting the physical and mental health of adolescents.