联合经典统计学与机器学习方法构建和检验青少年自杀危险性动态预测模型：前瞻性队列研究

In China, there are more than 10,000 adolescent die from suicide every year, and 20%-30% adolescent seriously considered ending their lives. It brings tremendous and deeply harms to the family and society. Predicting suicide risk accurately is a fundamental method to identify the individual with high risk of suicide, and reduce the cases of suicide. While, the current methods for prediction are far more than accurate. Many of them consider the suicide action simply comes from aggregating with suicide idea, and do not consider the factor named “suicide capacity” which is the main factor transferring the ideation to action. As well, lacking the dynamical assessment, important information as time effects are missing from the previous prediction methods. This study aims to develop an adolescent cohort in schools, define the suicide risk combining the “suicide desire” and “suicide capacity”, and bring their related factors for assessment. In order to facilitate the dynamic assessment, we will select four presentative time points within a school year to assess suicide risk and its risk factors, and monitoring factors that could be collecting from others every month. In every time point, we will adopt Generalized Linear Mixed Models and machine learning to develop prediction models. To explore the time effects of important risk factors and select the best model through comparing different models developed in different time point. Based on this study, we aim to develop a prediction model for adolescent suicide risk which would match the current mental health work in the school system, and help schools to predict individuals with high risk of suicide and to reduce the suicide cases.

我国每年有逾万名青少年死于自杀，更有20%-30%的青少年认真考虑过结束生命，这给家庭和社会造成沉重而深远的危害。准确预测自杀危险性，是发现自杀高危个体，减少自杀死亡的重要手段。目前的预测方法，没有认识到自杀行动并不是自杀意念的简单增加，未考虑“自杀能力”这个导致自杀行动的重要因素；也因无法动态评估造成重要信息的遗漏，从而预测不准确。研究拟在学校建立青少年队列， 引入“自杀能力”，结合“自杀意念”与“自杀能力”确定自杀危险性；在一学年中有代表性的四个时点评估自杀危险性和相关因素，且每月通过家长、老师和第三方数据监测可观察的情境性因素以实现动态评估，在每个时点均采用广义线性混合效应模型和非回归机器学习模型探索预测模型。通过比较不同时点的模型探索因素及其时间效应，并选择最优模型。希望通过本研究探索一套能与我国学校心理健康工作相结合，准确识别自杀高危个体的自杀危险性动态预测模型。

青少年自杀是重要的公共卫生问题。青少年自杀不仅给家庭造成无可复加的悲痛，也给社会造成巨大的损失。准确预测自杀危险性，对及时准确的发现自杀高危个体和减少自杀死亡有重要意义。本项目在此背景下，构建了一套符合青少年、且信效度良好的的评估工具，由学生、家长和老师共同参与，静态结合动态评估的方式，以有代表性中学生样本展开研究，取得了有价值的研究成果。中学生过去一学期自杀意念、自杀计划、自杀准备和自杀未遂的发生率分别为12.7%、5.9%、2.9%、9.4%，其中女性的自杀意念、自杀计划和自杀未遂发生率高于男性，居住在城市的学生的自杀意念发生率高于城镇，城镇又高于农村。多因素logistic回归发现，有精神障碍家族史、与父母联系紧密度差、有受虐待经历、孤独、社会支持差、有被欺侮、上学路上不安全感、有睡眠障碍、自杀能力高、和有情绪问题的学生，过去一学期自杀意念发生较多。与自杀意念相比，具有冲动特性的学生，过去一学期的自杀未遂发生较多。采用多因素Logistic回归构建自杀预测模型，模型由与父母联系紧密度、孤独、生活满意度、游戏障碍、睡眠障碍、情绪问题、家长报告自杀风险、心理问题组成，初步检验模型结果较好。本研究结果丰富了青少年自杀理论领域，在国内外杂志上发表学术论文3篇，参加会议论文3篇，获得相关子项目4项。本研究结果还具有卓越的现实意义。以本研究结果为基础构建预警系统和工作模式收到省内教育部门认可，获得基础教育成果奖-特等奖，且工作模式在全省范围内试点和推广。以项目为基础构建的培训课程覆盖全省12,230名老师。