Diagnosing Patient Stroke Status Using Modern AI After Dataset Balancing: A Comprehensive Comparative Study

Accurate prediction of stroke patient status is vital for early intervention and clinical decision-making. This study systematically evaluates the latest artificial intelligence (AI) methods for stroke diagnosis on a balanced real-world dataset. Using Synthetic Minority Oversampling Technique (SMOTE) for class balancing, we compare traditional and modern AI models, including XGBoost, Random Forest, Deep Neural Networks, TabTransformer, and TabPFN (a transformer-based model for tabular data). Stacked ensemble models were also tested. Results demonstrate that transformer-based and ensemble approaches outperform classic methods, with stacked models and TabPFN achieving AUC scores up to 0.96 and accuracy exceeding 97%. Model interpretability was assessed using SHAP. The findings confirm that combining SMOTE with state-of-the-art AI yields robust and generalizable tools for stroke prediction, supporting safe and transparent clinical implementation.