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

Abstract

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）.

Keywords

• cerebral infarction|prognosis|branch atheromatous disease|penetrating artery|machine learning|predictive model