基于机器学习的血管内治疗急性缺血性卒中患者7 d内病死风险预测

Machine learning-based prediction of death within seven days in patients with acute ischemic stroke after endovascular treatment

摘要:

目的 基于机器学习方法构建接受血管内治疗的急性缺血性卒中患者7 d内病死率预测模型，探索关键预测因素。方法 纳入2021年1月至2023年6月在山西省人民医院神经内科接受血管内治疗的急性缺血性卒中患者293例。收集术前人口统计学、疾病史、辅助检查结果等33个变量。划分训练集与验证集，采用随机森林和极端梯度提升决策树（XGBoost）算法构建预测模型。通过准确率、灵敏度、特异度和受试者操作特征（ROC）曲线下面积（AUC）等指标评估模型的预测性能。结果 训练集中，随机森林模型在预测7 d病死率方面表现最佳，其AUC值为0.986，灵敏度为95.8%，特异度为91.1%，优于XGBoost模型和Logistic回归模型。验证集中，XGBoost模型在预测7 d病死率方面的AUC值（0.908）和特异度（98.0%）优于随机森林模型（AUC为0.860）和特异度为97.9%，但灵敏度（26.7%）低于随机森林模型（66.7%）。随机森林模型和XGBoost模型的关键预测因素包括美国国立卫生研究院卒中量表评分、格拉斯哥昏迷评分和阿尔伯塔卒中项目早期CT评分。结论 基于机器学习的模型能有效预测急性缺血性卒中患者的7 d内的病死率，为临床决策提供了有价值的工具。

Abstract:

Objective To establish a predictive model for death within 7 days in patients with acute ischemic stroke undergoing endovascular treatment based on the machine learning method， and to explore key predictive factors.Methods A total of 293 patients with acute ischemic stroke who received endovascular treatment in the Department of Neurology， Shanxi People's Hospital， from January 2021 to June 2023 were enrolled. A total of 33 preoperative variables were collected， including demographics， disease history， and auxiliary examination results. The patients were divided into a training set and a validation set， and the random forest algorithm and the Extreme Gradient Boosting （XGBoost） algorithm were used to establish predictive models. The performance of the models was assessed based on accuracy， sensitivity， specificity， and the area under the ROC curve （AUC）.Results In the training set， the random forest model showed the best performance in predicting 7-day mortality， with an AUC of 0.986， a sensitivity of 95.8%， and a specificity of 91.1%， with a better performance than the XGBoost model and the Logistic regression model. In the training set， the XGBoost model had better AUC and specificity than the random forest model in predicting 7-day mortality （AUC： 0.908 vs 0.860； specificity： 98.0% vs 97.9%）， but with a poorer sensitivity than the random forest model （26.7% vs 66.7%）. The key predictive factors in the random forest model and XGBoost model included National Institutes of Health Stroke Scale score， Glasgow coma score， and Alberta Stroke Program Early CT Score.Conclusions Machine learning-based models can effectively predict death within 7 days in patients with acute ischemic stroke， which provides a valuable tool for clinical decision-making.