Explainable Artificial Intelligence for Cardiovascular Risk Assessment: A Comparative Study of Interpretable Machine Learning Models

Reliable and interpretable artificial intelligence models are critical for effective healthcare decision-making. The current study evaluates the use of a variety of machine learning methodologies using 18 models, with the top five Stacked Ensemble, Clinical Extreme Gradient Boosting (XGBoost) Time-Aware Hybrid using Long Short-Term Memory (LSTM), Stacked Generalisation (XGBoost + Random Forest + support vector machine), and Tabular Neural Network (TabNet) selected for prediction of heart failure disease by analysis of the Explainable AI (XAI) technique assessment due to their high predictive accuracies. Each model was evaluated for predictive power and explainability using XAI approaches such as Local Interpretable Model-agnostic Explanations (LIME), SHapley Additive explanation (SHAP), Anchors, Attention Maps, and Permutation Importance. These strategies improved the transparency of the model and ensured that the explanations were clinically relevant. The results demonstrate that more intricate models like Stacked Ensemble and TabNet had higher predictive accuracy, and less complicated models like Clinical XGBoost and Stacked Generalisation (XGBoost + Random Forest + support vector machine) produced more interpretable decision routes. The explainable artificial intelligence (XAI) methodologies ensured that both simple and complicated models produced clinically interpretable explanations of how and why a model makes its predictions. The Time-Aware Hybrid model’s use of attention maps and integrated gradients allowed for detecting crucial time points within electrocardiogram (ECG) sequences. The Stacked Generalisation model (XGBoost + Random Forest + Support Vector Machine) and Anchors generated interpretable decision-making pathways consistent with clinical practices.