中学生非自杀性自伤风险预测模型的构建与评价

Construction and evaluation of risk prediction model for non-suicidal self-injury of middle school students

目的基于不同机器学习算法构建中学生非自杀性自伤(NSSI)风险预测模型,并对模型的效果进行评价,为校园NSSI的防控提供指导。方法于2023年3月,采用分层整群随机抽样与方便抽样结合的方法抽取江西省南昌市、抚州市和上饶市共3372名初、高中学生为研究对象,采用一般情况调查表、自尊量表、渥太华自伤量表、社会支持评定量表、中文版Olweus欺负问卷、事件归因方式量表、青少年心理韧性量表及青少年生活事件量表进行问卷调查。将数据按照7∶3分为训练集(n=2361)和测试集(n=1011),基于单因素及LASSO回归结果筛选变量,使用随机森林、支持向量机、Logistic回归及极端梯度提升树(XGBoost)4种机器学习算法分别构建NSSI风险预测模型,使用曲线下面积(AUC)、灵敏度、特异度、阳性预测值、阴性预测值、F1指数对模型效果进行评价和比较。结果中学生NSSI的检出率为34.4%,单因素分析显示,不同学段、性别、户籍所在地、是否担任班干部及4种不同被欺凌类型(身体、言语、关系、网络欺凌人)的中学生NSSI检出率差异均有统计学意义(χ^(2)值分别为27.17,15.81,11.54,4.63;68.22,140.63,77.81,13.95,P值均<0.05)。NSSI为因变量纳入LASSO回归模型中进行变量筛选,结果显示,学段、自尊、主观支持、支持利用度、被言语欺凌、情绪控制、人际关系、受惩罚、亲友和财产丧失及健康与适应问题10个变量为预测变量。随机森林、支持向量机、Logistic回归、XGBoost算法的AUC值依次为0.76,0.76,0.76,0.77,两两比较差异均无统计学意义(Z=-0.59~0.82,P值均>0.05);灵敏度依次为0.62,0.61,0.62,0.61;特异度依次为0.74,0.78,0.78,0.78;阳性预测值依次为0.56,0.59,0.60,0.59;阴性预测值依次为0.79,0.79,0.80,0.79;F1指数依次为0.59,0.60,0.61,0.60。结论4种NSSI的风险预测模型效果均较好,Logistic回归模型效果略优于其余算法。学校及家长应关注NSSI对应的预测因素,以减少中学生NSSI的发生。

Objective To construct a non-suicidal self-injury(NSSI)risk prediction model for middle school students using different machine learning algorithms and evaluate the model's effectiveness,so as to provide guidance for the prevention and control of NSSI in campus.Methods In March 2023,a total of 3372 middle and high school students from schools in Nanchang,Fuzhou and Shangrao cities in Jiangxi Province were selected by combining stratified random cluster sampling and convenient sampling methods.Questionnaire surveys were conducted using various instruments including general information questionnaire,Self-esteem Scale,Ottawa Self-injury Scale,Social Support Assessment Scale,Chinese Version of the Olweus Bullying Questionnaire,Event Attribution Style Scale,Adolescent Resilience Scale,and Adolescent Life Events Scale.Data were divided into training set(n=2361)and test set(n=1011)at a ratio of 7∶3,and variables were selected based on univariate and LASSO regression results.Four machine learning algorithms including namely random forest,support vector machine,Logistic regression and XGBoost,were used to construct NSSI risk prediction models,and the models'performance was evaluated and compared using metrics including area under curve(AUC),sensitivity,specificity,positive predictive value,negative predictive value and F1 score.Results The detection rate of NSSI among middle school students was 34.4%.Univariate analysis showed that there were statistically significant differences in NSSI detection rates among middle school students of different grades,genders,registered residence locations,whether they were class cadres and four types of bullying(physical,verbal,relational bullying and cyberbullying)(χ^(2)=27.17,15.81,11.54,4.63;68.22,140.63,77.81,13.95,P<0.05).NSSI was included as the dependent variable in the LASSO regression model for variable screening,and the results regression identified 10 predictive variables including grade level,self-esteem,subjective support,support utilization,verbal bullying,emotional control,interpersonal relationships,punishment,loss of relatives and property,and health and adaptation issues.The AUC values of random forest,support vector machine,Logistic regression,and XGBoost algorithms were 0.76,0.76,0.76 and 0.77,respectively,with no statistically significant differences between pairwise comparisons(Z=-0.59-0.82,P>0.05).Sensitivity values were 0.62,0.61,0.62 and 0.61,respectively.Specificity values were 0.74,0.78,0.78 and 0.78,respectively.Positive predictive values were 0.56,0.59,0.60 and 0.59,respectively.Negative predictive values were 0.79,0.79,0.80 and 0.79,respectively.F1 scores were 0.59,0.60,0.61 and 0.60,respectively.Conclusions All four non-suicidal self-injury risk prediction models perform well,with the Logistic regression model slightly outperforming the others.Schools and parents should pay attention to the predictive factors corresponding to NSSI,so as to reduce the occurrence of NSSI among middle school students.