Stroke is the leading cause of disability and death globally, emphasising the necessity for precise and timely prediction. This study used a variety of machine learning (ML) algorithms to predict the risk of a stroke, including Logistic Regression (LR), Random Forest (RF), Gradient Boosting (GB), and the Stacking Model. Imputation, scaling, one-hot encoding, and SMOTE were used to correct class imbalance in a stroke-related factor dataset. Despite a low precision of 5.41%, the LR model identified positive instances with 74.62% accuracy, 77.02% recall, and 82.96% AUC-ROC. The (RF) model has excellent accuracy (95.68%) and AUC-ROC (75.21%), but recall was poor at 6.21%. GB balanced performance metrics with 84.86% accuracy, 50.93% recall, and 80.20% AUC-ROC. The Stacking Model, which combined the strengths of various models, had the best accuracy at 96.30% but poorer precision and recall at 4.02 and 4.35%. These data show that while ensemble approaches like RF and Stacking Models give improved accuracy, LR’s strong recall is critical in medical diagnostics where true positives (TP) are crucial. To improve stroke prediction accuracy and clinical usefulness, future research should incorporate new characteristics, better preprocessing methods, and deep learning models.