自杀行为预测中的机器学习方法:一项系统综述

Machine Learning Approaches to the Prediction of Suicidal Behavior:A Systematic Review

目的:全面总结机器学习预测自杀行为的算法特点、应用场景和模型效能。方法:系统搜索三个数据库中(PubMed、EMBASE、Web of Science)截至2022年11月应用机器学习算法预测自杀意念、自杀尝试和自杀死亡的文献。结果:研究共纳入71篇文献,提取81个模型。机器学习预测自杀行为的模型效果良好(自杀意念AUC区间0.70~0.97;自杀尝试AUC区间:0.63~0.96;自杀死亡AUC区间:0.69~0.93),但阳性预测值偏低。当前研究多采用临床或社区样本,缺少学校样本。模型纳入的特征多为人口统计学和精神病理学变量,缺乏心理、生理和社会变量。结论:机器学习提升了自杀预测能力,有望作为辅助筛查工具应用于临床实践。

Objective:To summarize the characteristics,application scenarios and model performance of machine learning in predicting suicidal behavior.Methods:Three databases(PubMed,EMBASE,Web of Science)were systematically searched for articles using machine learning algorithms to predict suicidal ideation,suicidal attempts,and suicidal death up to November 2022.Results:A total of 71 articles were included and 81 models were extracted.Predicting suicidal behavior models of machine learning was good(AUC interval of suicidal ideation was 0.70~0.97;AUC interval of suicidal attempt:0.63~0.96;AUC interval for suicide:0.69~0.93),but the positive predictive value was low.Current studies mostly used clinical or community samples,but lacked school samples.The model included mostly demographic and psychopathological variables and lacked psychological,physiological,and social variables.Conclusion:Machine learning improves the ability to predict suicide,and is expected to be as auxiliary screening tool used in clinical practice.