Predicting The Risk Of Recurrence Of Ischemic Stroke Within 1 Year Based On Machine Learning And MRI Radiomics

Purpose:The aim of this study was to construct different machine learning ischemic stroke(IS)recurrence prediction models based on magnetic resonance imaging(MRI)radiomics and clinical data of IS patients and four machine learning algorithms to predict the risk of recurrence in IS patients within one year.Methods:MRI images and clinical data were collected from March 1,2019,to March 5,2021,from patients admitted to the Second Affiliated Hospital of Nanchang University with a diagnosis of IS,and patients were followed up for recurrence within1 year.Brain infarction lesions were labeled on diffusion-weighted imaging(DWI)by imaging physicians in MRI examinations,and consistency tests were performed on the labeled images to extract DWI brain radiomics features,and feature selection was performed using least absolute shrinkage and selection operator(LASSO)regression.The data were randomly divided into training and validation sets in the ratio of 7:3,and four machine learning algorithms,logistic regression(LR),support vector classification(SVC),light gradient boosting machine(Light GBM),and random forest(RF),were used to construct recurrent prediction models.For each algorithm,three models were constructed based on radiomics features,clinical data,and a combination of radiomics features and clinical data,respectively.The sensitivity,specificity and area under the operating characteristic(ROC)curve(AUC)of the models were compared to evaluate the predictive efficacy of each model.Results:A total of 612 patients with IS were included in the study,of whom 337(55.1%)were men and 275(44.9%)were women,with a mean patient age of 63.9±11.16 years,and 67(10.95%)patients had an endpoint event within 1 year of follow-up.Smoking history( χ~2=3.868,P=0.049),stroke history( χ~2=5.646,P=0.017),ischemic heart disease history(χ~2=4.695,P=0.030),m RS(χ~2=5.008,P=0.025),age(Z=-2.349,P=0.019),white blood cell count(t=2.244,P=0.025),absolute neutrophil value(Z=-1.967,P=0.049),total protein(t=-2.391,P=0.017),albumin(t=-2.725,P=0.007),alkaline phosphatase(Z=-1.984,P=0.047),creatinine(Z=-3.177,P=0.001),and prothrombin time(Z=-2.079,P= 0.038),international standard ratio(Z=-2.042,P=0.041),fibrinogen concentration(Z=-2.689,P= 0.007)and other variables were statistically significant between the recurrent and non-recurrent groups.The MDice coefficient,MIo U and Hausdorff distance labeled by the imaging physicians were 0.87,0.77 and 4.36 mm,respectively,and the intraclass correlation coefficient(ICC)mean value was 0.98.1037 radiomics features were extracted from DWI images,and a total of 20 radiomic features were selected after feature screening,including 10 first-order features,1 shape feature and 9 texture features.Prediction models based on three different data combinations with four different algorithms were constructed,and the model constructed by the Light GBM algorithm achieved the best prediction precision in the validation set.The sensitivity,specificity and AUC of the Light GBM model based only on radiomics features were 65.0%,67.1% and 0.647,respectively,while based olny on clinical data were 70.0%,79.9%and 0.735,respectively.The best performance was achieved by the Light GBM model based on a combination of radiomics features and clinical data,with sensitivity,specificity and AUC of 85.0%,80.5% and 0.789,respectively.The model showed that clinical variables with large platelet ratio,smoking history,age,fibrinogen concentration,hemoglobin,albumin,creatinine,white blood cell count,absolute neutrophil value,and discharge m RS could be used to predict IS recurrence,and radiomics with first-order features such as kurtosis,skewness,range,and texture features such as High Gray Level Zone Emphasis Gray Level Non Uniformity and Size Zone Non Uniformity could be used to predict IS recurrence.Conclusions:The predictive performance of the Light GBM-based ischemic stroke recurrence prediction model was the best.The combination of MRI radiomics features and clinical data could improve the predictive performance of the model.Smoking history,age,discharge m RS,fibrinogen concentration,white blood cell count,absolute neutrophil value,creatinine,albumin,and hemoglobin were significant for IS recurrence risk prediction.Kurtosis,skewness and range in the first-order features and High Gray Level Zone Emphasis,Gray Level Non Uniformity and Size Zone Non Uniformity in the texture features are important factors in predicting recurrence.