基于机器学习的急性穿支动脉闭塞性脑梗死预后预测模型研究

Study of a prediction model for acute penetrating artery territory infarction based on machine learning

目的基于机器学习算法评估急性穿支动脉闭塞性脑梗死预测模型并筛选优势模型,为临床管理急性穿支动脉梗死患者提供依据。方法选取441例急性穿支动脉闭塞性脑梗死患者为研究对象,排除临床信息不完整10例,多次脑梗死患者28例,共纳入403例。将结果变量分为预后良好组[改良Rankin量表(mRS)评分0~2分]和预后不良组(mRS评分>2分)。采用单、多因素Logistic回归(LR)以逐步回归法分析筛选预测变量。使用LR、随机森林(RF)、支持向量机(SVM)3种机器学习算法构建功能预后预测模型,在测试集中通过受试者操作特征(ROC)曲线的曲线下面积(AUC)、准确度、灵敏度、特异度等指标比较预测模型对患者发病90 d功能预后的预测价值。结果403例患者中男性占68.73%,年龄(60.4±11.4)岁。从44个变量中选出7个变量作为预测变量,分别为白细胞计数、血小板计数、就诊时血糖、胆固醇、既往糖尿病病史、既往服用降糖药物史、既往吸烟史(P均<0.05)。LR、RF、SVM预测预后的AUC分别为0.610、0.690、0.780。结论机器学习算法在预测急性穿支动脉闭塞性脑梗死中有一定的预判能力。RF、SVM(非线性模型)在预测模型中的表现优于传统LR模型(线性模型)。

Objective To evaluate the performance of prediction models for acute penetrating artery territory occlusive cerebral infarction based on machine learning algorithms and select the optimal model,aiming to provide evidence for clinical management of acute penetrating artery territory infarction.Methods A total of 441 patients diagnosed with acute perforator artery territory infarction were enrolled in this study.Patients with incomplete clinical information(n=10)and multiple cerebral infarctions(n=28)were excluded,resulting in a final sample size of 403 patients.The outcome variables were divided into two groups:good prognosis(mRS scores of 0-2)and poor prognosis(mRS scores>2).Univariate and multi-variate Logistic regression(LR)using the stepwise regression method were employed to identify prediction variables.LR,random forest(RF)and support vector machine(SVM)models were utilized to develop a prognostic prediction model.The dataset was further divided randomly into a training set and a test set in a 7:3 ratio.In the test set,the predictive performance of the model for 90-day functional prognosis in patients with BAD(with poor prognosis defined as mRS scores>2)was evaluated using metric such as the area under the receiver operating characteristic(ROC)curve(AUC),accuracy,sensitivity and specificity,etc.Results Among 403 patients with BAD,68.73%of them were male,with an average age of(60.4±11.4)years.Using the stepwise regression method,7 prediction variables were selected from a pool of 44 variables:white blood cell count,platelet count,blood glucose,cholesterol,history of diabetes mellitus,history of taking hypoglycemic drugs,and history of smoking(all P<0.05).The AUC of LR,RF and SVN for predicting clinical prognosis was 0.610,0.690,and 0.780,respectively.Conclusions Machine learning algorithms have demonstrated certain predictive ability for acute penetrating artery territory infarction.The performance of RF and SVM models(nonlinear models)is superior to traditional logistic regression model(linear model).