XGBoost Model For Diagnosis Of Cognitive Impairment In Acute Stroke

Background and Objectives:Cognitive impairment is prevalent after ischmic stroke incidence.Timely and effective diagnosis is urgent for post-stroke cognitive impairment.Our study's aim was to construct and compare models for diagnosing cognitive impairment among ischemic stroke inpatients,using machine learning method and conventional logistic regression.Methods:334 patients with acute ischemic stroke were admitted consecutively at Neurology Department of the first affiliated Hospital of Zhejiang University between December 2013 and Febrary 2015.Standard stroke workup and treatment were carried out,and demographic,clinical and risk factor data were collected.According to the diagnostic criteria,the patients were classified into vascular cognitive impairment and cognitive normality.Then,all of the patients were randomly divided into a training group and a validation group at the ratio of 2:1.Conventional logistic regression model and newer machine learning model were derived from the training dataset based on several determinants of the cognitive impairment.Both models were tested on the validation group and compared with each other by diagnostic capacity for vascular cognitive impairment.Discrimination of the two models was accessed with receiver operating characteristic curve analysis,respectively,and diagnostic accuracy was accessed using net reclassification improvement(NRI).Results:The training group comprised 220 participants:141 with cognitive impairment,and the validation group comprised 114 patients:68 with cognitive impairment.Variables which were included in both of the two models comprised:age,education,BMI,TC,TG,GHB,mRS scores,The logistic regression model had a c-statistic of 0.77(95%CI:0.71-0.84),whereas the machine learning algorithm had a c-statistic of 0.85(95%CI 0.79-0.90)based on the training dataset.Moreover,the machine learning model had a significantly better diagnostic capacity as accessed by NRI(P=0.020).In the validation dataset,the conventional regression method had a c-statistic of 0.74(95%CI:0.65-0.83),and the machine learning model had a c-statistic of 0.73(95%CI:0.63-0.82).Also,the results didn't show statistical significant for diagnostic accuracy measured by NRI(P=0.830).Conclusions:The machine learning algorithm outperformed conventional regression model in the training dataset.