Novel Approaches to Heart Disease Detection with TabNet and FT Transformer Models

Heart disease has been one of the leading causes of mortality worldwide, necessitating accurate predictive models to support early diagnosis and intervention. Traditional ML models have been widely used for heart disease detection due to their interpretability and efficiency on tabular datasets. Recent advancements in deep learning introduced novel architectures, such as TabNet and FT-Transformer, which leveraged attention mechanisms and feature tokenization to capture complex, high-order feature interactions directly from tabular data. This paper applied and compared traditional ML models, TabNet, and FT-Transformer on a heart disease dataset to evaluate their effectiveness in predicting heart disease outcomes. The approach combined the strengths of traditional ML models with the enhanced feature representation and interpretability offered by deep learning-based methods, particularly TabNet's sequential attention mechanism and FT-Transformer's selfattention architecture. Model performance was analyzed across various metrics, including accuracy, F1-score, and AUC-ROC, and feature importance was evaluated to gain insights into the key predictors of heart disease. Results demonstrated that while traditional ML models remained competitive, TabNet and FT-Transformer provided promising alternatives, capturing complex feature dependencies that might otherwise have gone unnoticed.