Logistic回归模型和XGBoost模型对急性缺血性脑卒中患者发生吞咽障碍的预测价值  

Predictive value of Logistic regression and XGBoost models for dysphagia in patients with acute ischemic stroke

作　　者：周升霞 张佳 王祖萍[3] 付丽萍[3,5] 李萍 ZHOU Shengxia;ZHANG Jia;WANG Zuping;FU Liping;LI Ping(School of Nursing,Xinjiang Medical University,Urumqi 830011,China;Department of Neurology,the Second Affiliated Hospital of Xinjiang Medical University,Urumqi 830063,China;Department of Outpatient,the Second Affiliated Hospital of Xinjiang Medical University,Urumqi 830063,China;Intensive Care Unit,the Second Affiliated Hospital of Xinjiang Medical University,Urumqi 830063,China;Xinjiang Regional Population-based Disease and Health Care Research Center,Urumqi 830011,China)

机构地区：[1]新疆医科大学护理学院,乌鲁木齐830011 [2]新疆医科大学第二附属医院神经内科,乌鲁木齐830063 [3]新疆医科大学第二附属医院门诊部,乌鲁木齐830063 [4]新疆医科大学第二附属医院重症监护室,乌鲁木齐830063 [5]新疆区域人群疾病与健康照护研究中心,乌鲁木齐830011

出　　处：《新疆医科大学学报》2024年第8期1179-1185,共7页Journal of Xinjiang Medical University

基　　金：新疆神经系统疾病重点实验室开放课题项目(XJDX1711-2227)。

摘　　要：目的筛选危险因素构建急性缺血性脑卒中后吞咽障碍风险预测模型,对比XGBoost模型和Logistic回归模型的优劣性。方法选取2022年1-12月新疆医科大学第二附属医院神经内科573例急性缺血性脑卒中患者,按7∶3比例随机分为建模组(n=401)和验证组(n=172)。筛选发生吞咽障碍的危险因素,以单因素分析有统计学意义的变量分别建立Logistic回归模型和XGBoost模型。在验证组数据集上使用十折交叉验证法进行内部验证,采用校准曲线、受试者工作特征曲线(ROC曲线)和决策曲线评价两种模型的预测效能。结果多因素Logistic回归分析结果显示,年龄、NIHSS评分、GCS评分、BI指数、脑干病变、构音障碍、失语症、咽反射(正常)是急性缺血性脑卒中后吞咽障碍的影响因素。XGBoost模型特征重要性排序前8位分别为年龄、BI指数、NIHSS评分、咽反射、TOAST分型、白蛋白、文化程度、营养评分。对比两种模型结果显示,XGBoost模型的准确性、精确度、敏感度、F1分值分别为0.849、0.830、0.754、0.790,表现优于Logistic回归模型。Logistic回归、XGBoost模型预测吞咽障碍的AUC值分别是0.894、0.925,两者AUC值比较,差异无统计学意义(P>0.05)。模型的校准曲线和临床决策曲线均显示XGBoost模型准确度和临床实用价值优于Logistic回归模型。结论XGBoost模型和Logistic回归模型均能有效预测急性缺血性脑卒中后吞咽障碍风险,XGBoost模型表现更优,可为临床早期预防急性缺血性脑卒中吞咽障碍提供参考。Objective:To screen risk factors to construct a risk prediction model for dysphagia after acute ischaemic stroke,and to compare the advantages and disadvantages of XGBoost model and Logistic regression model.Methods:573 patients with acute ischemic stroke in the hospital were selected from January to December 2022,and randomly divided into modelling group(n=401)and validation group(n=172).According to a 7∶3 ratio risk factors for the occurrence of dysphagia were screened,and Logistic regression models and XGBoost models were established with variables that were statistically significant by univariate analysis,respectively.Internal validation was performed on the validation group dataset using the ten-fold cross-validation method,and the predictive efficacy of the 2 models was evaluated using calibration curves,subject work characteristic curves(ROC curves)and decision curves.Results:The results of multifactorial Logistic regression analysis showed that age,NIHSS score,GCS score,BI index,brainstem lesions,dysarthria,aphasiaand pharyngeal reflexes(normal)were the influencing factors of dysphagia after acute ischemic stroke.The top 8 rankings of importance of features in the XGBoost model were age,BI index,NIHSS score,pharyngeal reflexes,TOAST typing,albumin,education level and nutritional score.Comparison of the results of the 2 models showed that the accuracy,precision,sensitivity and F1 score of the XGBoost model were 0.849,0.830,0.754 and 0.790,respectively,which outperformed the Logistic regression model.The AUC values of Logistic regression and the XGBoost model for the prediction of dysphagia were 0.894 and 0.925,respectively,and the difference was not statistically significant(P>0.05).The calibration curve and clinical decision curve of the model showed that the accuracy and clinical utility value of the XGBoost model were better than that of Logistic regression model.Conclusion:Both XGBoost model and Logistic regression model can effectively predict the risk of dysphagia after acute ischaemic stroke and t

关 键 词：急性缺血性脑卒中 吞咽障碍 LOGISTIC回归 XGBoost模型 

分 类 号：R47[医药卫生—护理学]