Advancing Stroke Prediction: A Comparative Study of Hybrid Ensemble and Deep Learning Models

Stroke remains a major contributor to global mortality and long-term disability which places a strain on global healthcare systems. This study introduces two advanced hybrid ensemble learning models-the Stacking Classifier and Voting Classifier-for stroke prediction. To ensure model robustness, preprocessing techniques such as SMOTE for addressing class imbalance and feature standardization for numerical stability were implemented. The Stacking Classifier, which combines Random Forest, XGBoost, and LightGBM as base learners with a Logistic Regression meta-learner, achieved exceptional performance, including an accuracy of 96.97 %, precision of 97.40 %, recall of $\mathbf{9 6. 4 9 \%}$, and a ROC-AUC of $\mathbf{9 9. 5 7 \%}$. Similarly, the Voting Classifier, employing Random Forest, XGBoost, CatBoost, and LightGBM with soft voting, delivered competitive results with an accuracy of $\mathbf{9 7. 0 2 \%}$ and a ROC-AUC of $\mathbf{9 9. 6 0 \%}$. Both models consistently outperformed traditional machine learning techniques, standalone deep learning models, and hybrid deep learning approaches. These results establish the proposed models as effective, reliable, and clinically relevant tools for precise and early stroke risk assessment, enhancing patient outcomes and facilitating timely medical interventions.