Improving stroke prediction with deep learning models using tabular medical data

Stroke remains a leading cause of adult disability, posing significant healthcare and socio-economic burdens. This study explores advanced machine learning and deep learning models to enhance stroke prediction accuracy. We focus on a hybrid Convolutional neural networks- long short-term memory (CNN-LSTM) architecture and compare it against LSTM, recurrent neural networks (RNN), and Gradient Boosting models, incorporating fuzzy logic-based feature engineering. A comprehensive dataset encompassing key risk factors—such as age, hypertension, and heart disease—served as the basis for experimentation. Our methods included hyperparameter tuning, stratified K-fold cross-validation, and SMOTE to address class imbalance. The LSTM model achieved 98% accuracy, effectively capturing temporal trends in medical records. The CNN-LSTM model performed well with 95% accuracy, though its performance dropped to 78% when SMOTE was applied, highlighting the challenge of synthetic oversampling. RNN and Gradient Boosting models also showed strong results (~94%) when enhanced with fuzzy logic. These results emphasize the performance trade-offs of ensemble and hybrid learning strategies while demonstrating their potential in stroke risk prediction. This work contributes to the development of robust, AI-driven diagnostic tools in preventive healthcare.