基于机器学习的缺血性脑卒中抗栓治疗后消化道出血预测模型研究  

Prediction of gastrointestinal bleeding after antithrombotic therapy for is-chemic stroke based on machine learning

作　　者：罗颢文 俞鹏飞 刘建模 韩梦琦 易应萍[1] LUO Hao-wen;YU Peng-fei;LIU Jian-mo;HAN Meng-qi;YI Ying-ping(Medical Big-Data Center,Second Affiliated Hospital of Nanchang University,Nanchang,Jiangxi 330006,China;不详)

机构地区：[1]南昌大学第二附属医院医疗大数据研究中心,江西南昌330006 [2]南昌大学医学部公共卫生学院,江西南昌330006

出　　处：《现代预防医学》2023年第2期199-204,221,共7页Modern Preventive Medicine

基　　金：国家重点研发计划(2020YFC2002901);国家自然科学基金(81960609);江西省研究生创新专项资金项目(YC2021-S199);南昌大学第二附属医院资助项目(2021efyB03)。

摘　　要：目的筛选出缺血性脑卒中抗栓治疗后发生消化道出血的关键变量,评价Catboost、支持向量机(SVM)、logistic回归(LR)三种机器学习算法对缺血性脑卒中抗栓治疗后消化道出血的预测效果。方法选取2018年1月1日-2020年1月1日南昌大学第二附属医院确诊为急性缺血性脑卒中并接受抗栓治疗的住院患者,根据单因素分析结果确定初始变量,综合多因素logistic回归、RFE、lasso回归三种特征选择方法筛选变量,比较Catboost、SVM、LR在缺血性脑卒中抗栓治疗后消化道出血预测模型中的效果。结果在1605名缺血性脑卒中患者中,消化道出血的患者84名,单因素分析初步确定了17个变量,根据三种特征选择方法确定年龄、GCS、谷草/谷丙、碱性磷酸酶、低密度脂蛋白、出血性转化为关键变量,构建机器学习模型后重复交叉验证结果显示,Catboost算法的综合性能较好,特异度、准确率、AUC、阳性似然比分别为0.851(95%CI:0.85~0.853)、0.84(95%CI:0.838~0.841)、0.848(95%CI:0.841~0.855)、4.463(95%CI:4.378~4.549),LR发现消化道出血的能力较强,灵敏度和阴性似然比分别为0.723(95%CI:0.711~0.734)和0.345(95%CI:0.33~0.36)。结论Catboost用于预测缺血性脑卒中抗栓治疗后消化道出血具有更强的优势,结合三种特征选择方法筛选的关键变量,为缺血性脑卒中抗栓治疗后的消化道出血的预防和干预提供一定参考。Objective To select the key variables based on feature selection to evaluate the predictive effect of Catboost, support vector machine(SVM) and Logistic regression(LR) machine learning algorithms on gastrointestinal bleeding after antithrombotic therapy for ischemic stroke. Methods From January 1, 2018 to January 1, 2020, inpatients diagnosed with acute ischemic stroke and receiving antithrombotic therapy in the Second Affiliated Hospital of Nanchang University were selected.Their basic information, score scale, past medical history, blood test results, and co-morbidity variables were collected. The initial variables were determined according to the results of univariate analysis, and the variables were then screened by mul-ti-factor Logistics regression, RFE, and Lasso regression. The predictive effects of Catboost, SVM, and LR in the model after antithrombotic therapy for ischemic stroke were compared. Results Among 1 605 patients with ischemic stroke, 84 patients had gastrointestinal bleeding. Seventeen variables with differences were initially identified by univariate analysis. Age, AST/ALT, alkaline phosphatase low-density lipoprotein, and hemorrhagic transformation were identified as key variables according to three feature selection methods. Repeated cross-validation results showed that the comprehensive performance of Catboost algorithm was better. The specificity, ACC, AUC, and positive likelihood ratio(PLR) were 0.851(95% CI: 0.85-0.853), 0.84(95%CI: 0.838-0.841), 0.848(95%CI: 0.841-0.855), and 4.463(95%CI: 4.378-4.549), respectively. LR had a strong ability to detect gastrointestinal bleeding, and its sensitivity and negative likelihood ratio(NLR) were 0.723(95% CI: 0.711-0.734)and 0.345(95% CI: 0.33-0.36), respectively. Conclusion Catboost has a stronger advantage in predicting gastrointestinal bleeding after antithrombotic therapy for ischemic stroke. The key variables selected by combining three methods provide some reference for the prevention and intervention of gastrointestinal bleeding after ant

关 键 词：缺血性脑卒中 抗栓治疗 消化道出血 机器学习 特征选择 Catboost 

分 类 号：R743.3[医药卫生—神经病学与精神病学]