Predictionof Brain Stroke Using Federated Learning

Stroke often arises from an abrupt blockage in the blood vessels supplying the brain and heart. Detecting early warning signs of stroke can significantly reduce its impact. In this study, we propose an early prediction method for stroke using various Machine Learning (ML) techniques, considering factors such as hypertension, body mass index, heart disease, average glucose levels, smoking habits, prior stroke history, and age. These attributes, rich in information, were utilized to train three distinct classifiers: Logistic Regression, Decision Tree,and K-nearest neighbors for stroke prediction. In this study, Federated Learning (FL)has been applied to combine the ML models (Logistic Regression (LR), Decision Tree (DT), and K-Nearest Neighbors(KNN)) from distributed medical data sources while preserving patient privacy. By aggregating locally trained models from multiple hospitals or devices, FL ensures the robustness of the weighted voting classifier without requiring direct data sharing, thereby enhancingstroke prediction accuracy across diverse datasets. Subsequently, the results from these base classifiers were combined using a weighted voting approach to achieve the highest accuracy. Our study demonstrated an impressive accuracy rate of 97%, with the weighted voting classifier outperforming the individual base classifiers. This model proved to be the more accurate in predicting strokes. Additionally, the Area Under the Curve (AUC) value for the weighted voting classifier was notably high, and it exhibited the lowest false positive and false negative rates compared to other classifiers. Consequently, the weighted voting classifier emerged as an almost ideal tool for predicting strokes, offering valuable support to both physicians and patients in identifying and preventing potential stroke incidents.