Machine Learning Approaches for Stroke Risk Prediction Using Healthcare Data

Abstract

Stroke is a leading cause of mortality and disability globally, and precise risk prediction models are required. Herein, machine learning classifiers are utilized to predict and estimate stroke incidence based on a publicly accessible healthcare dataset of demographic, lifestyle, and clinical variables. Logistic Regression and Random Forest were experimented and trained in a class imbalance setting following preprocessing steps included missing values imputation and encoding categorical variables. Random Forest model had 95.4% accuracy, precision of 79.2%, recall of 62.7%, F1-score of 70.1%, and ROC-AUC of 91.6%, higher than Logistic Regression model with 92.8% accuracy, precision of 64.8%, recall of 55.1%, F1-score of 59.6%, and ROC-AUC of 87.9%. Confusion matrices and ROC/Precision-Recall curves also showed the discriminative power of the models. The importances of the features plot indicated that the primary predictors were hypertension, body mass index, mean glucose value, and age. The results demonstrate the potential of machine learning to facilitate early risk prediction of stroke, hence facilitating timely clinical intervention and resource planning. The study augments the evidence for the use of predictive analytics in healthcare decision support systems.