A CNN-LSTM Hybrid Architecture for Early Stroke Risk Prediction: A Comparative Study with Random Forest and XGBoost

Stroke remains the second-leading cause of global mortality, affecting approximately 15 million individuals annually. Early risk prediction through artificial intelligence offers a transformative pathway to preventive clinical intervention. This study investigates the performance of three machine learning and deep learning approaches: Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and a novel Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) Hybrid Architecture for binary stroke risk classification using the Kaggle Stroke Prediction Dataset and the CDC Behavioural Risk Factor Surveillance System (BRFSS) 2015 survey. The proposed CNN-LSTM hybrid achieved the highest AUC-ROC of 0.967, sensitivity of 0.940, and F1-score of 0.930 under stratified 10-fold cross-validation, significantly outperforming both RF (AUC 0.891) and XGBoost (AUC 0.913). Three structured experiments confirm that the proposed architecture outperforms classical ensemble methods with statistically significant margins (p < 0.05). The findings demonstrate the potential of integrating spatial convolutional feature extraction with temporal LSTM-based modelling to improve stroke prediction in clinical decision-support systems.