摘要

【摘要】
目的 本研究旨在探究执行功能和重度抑郁症（Major depressive disorder，MDD）自杀的关系及利用MDD患者的执行功能基于机器学习算法开发自杀风险预测模型。

方法 研究对象包括51名MDD自杀者，55名MDD非自杀者和100名健康对照者。执行功能采用N-back任务、情感转移任务（Affective shifting task，AST）、爱德华州赌博任务（Iowa gambling task，IGT）来测量。方差分析比较三组执行功能差异。利用极端梯度提升（Extreme gradient boosting ，XGBoost）分类算法构建预测模型，将预测性能及曲线下面积（Area under the curve ，AUC）进行比较，局部解释技术评估特征的相对重要性。

结果 三组在更新功能（P<0.001）、抑制功能（P=0.009）、转移功能(P<0.001)及决策功能(P<0.001)存在组间差异。MDD自杀组在抑制功能（P=0.013）和转移功能（P=0.023）比MDD非自杀组更差。利用综合数据构建了一个性能更优的自杀预测模型，执行功能可以提高预测模型的特异性、敏感性、准确性、F1分数及曲线下面积。

结论 本研究发现执行功能是自杀的重要危险因素。在预测自杀企图的模型中，利用综合的数据可以提高模型分类性能。可解释的模型能提高预测

结果可信度，有助于早期干预。

关键词：【关键词】：自杀;重度抑郁症;执行功能;机器学习

ABSTRACT

[Abstract]
Objective The aim of this study was to explore the relationship between executive function and suicide in severe major depressive disorder (MDD) and to develop a suicide risk prediction model based on machine learning algorithms using MDD patients' executive function.

Methods The study enrolled 51 MDD suicide cases, 55 MDD non-suicide cases, and 100 healthy controls. Executive function was measured using the N-back task, affective shifting task (AST), and Iowa Gambling Task (IGT). Variance analysis was used to compare differences in executive function among the three groups. The extreme gradient boosting (XGBoost) classification algorithm was used to construct the prediction model, which was compared based on predictive performance and area under the curve (AUC), and local interpretation techniques were used to evaluate the relative importance of features.

Results There were significant intergroup differences in updating (P<0.001), inhibition (P=0.009), shifting (P<0.001), and decision-making functions (P<0.001). The MDD suicide group showed poorer performance in inhibition (P=0.013) and shifting functions (P=0.023) compared to the MDD non-suicide group. A more effective suicide prediction model was developed using comprehensive data, and executive function improved the specificity, sensitivity, accuracy, F1 score, and AUC of the prediction model.

Conclusion   The findings suggest that executive function is an important risk factor for suicide. Using comprehensive data can improve the classification performance of the suicide attempt prediction model. An interpretable model can improve the credibility of the prediction results and aid in early intervention.

Key words: [Key words] Suicide; Major Depressive Disorder; Executive Function; Machine Learning