AI-Driven Predictive Models for Early Stroke Risk Assessment Using Machine Learning and Deep Learning

Stroke remains one of the foremost causes of mortality and long-term disability globally, demanding robust and timely predictive mechanisms for clinical intervention. This paper proposes an AI-driven hybrid predictive framework that combines traditional machine learning (ML) algorithms — including Random Forest and XGBoost — with deep learning (DL) architectures such as Bidirectional LSTM (Bi-LSTM) and CNN-LSTM for early stroke risk stratification. The model is trained and evaluated on publicly available datasets including MIMIC-III and BRFSS, after rigorous preprocessing, feature engineering, and class-imbalance correction via SMOTE. The proposed hybrid ensemble achieves an accuracy of 95.7%, AUC-ROC of 0.98, and F1-score of 0.94, outperforming existing standalone approaches. Explainability modules incorporating SHAP and LIME are integrated to ensure clinical interpretability, enabling deployment in decision-support environments. The study addresses key challenges including data imbalance, missing values, feature importance, and model transparency, making a meaningful contribution toward intelligent, scalable, and clinically viable stroke prediction systems.