Stroke Prediction Base on Logistic Regression Model

Abstract

The prediction of stroke using risk factors and basic demographic information can be valuable for primary prevention and community healthcare workers. Machine learning models are becoming more prevalent in clinical prediction due to their high accuracy. This study investigates stroke prediction using four machine learning models including logistic regression, random forest, Lasso Regression (Least Absolute Shrinkage and Selection Operator), and Extreme Gradient Boosting (XGBoost). Six key variables—age, gender, work type, heart disease history, marital status, residency type, smoking status, Body Mass Index (BMI), and average glucose level—are used for prediction. Among the dataset of 3,255 observations, 180 individuals experienced stroke, indicating an extremely imbalanced dataset. Therefore, balanced accuracy is the key metric used to compare model performance. The balanced accuracy for logistic regression, Lasso Regression, random forest, and XGBoost are 78.4%, 61.7%, 50%, and 67.7%, respectively. Logistic regression demonstrated the strongest performance, while also highlighting the significant role of age, particularly for individuals over 45.