Explainable AI driven machine learning and deep learning for heart disease classification and prediction

Abstract

This thesis presents the development of Explainable Artificial Intelligence (XAI)-driven newlineMachine Learning (ML) and Deep Learning (DL) approaches for the accurate classification and newlineprediction of Heart Disease (HD), specifically targeting Coronary Artery Disease (CAD), Heart newlineFailure (HF), and rare congenital disorders such as Anomalous Left Coronary Artery from newlinethe Pulmonary Artery (ALCAPA). Effective prediction in these domains is often hindered by newlinechallenges including class imbalance, lack of annotated medical images, limited interpretability newlineof models, and the need for extensive manual intervention in model selection and tuning. To newlineaddress these challenges, the research integrates XAI frameworks with optimized ML, DL, and newlineAutoML pipelines to ensure high performance, transparency, and clinical relevance. newlineThe first objective focuses on CAD prediction using an Extension of the Z-Alizadeh Sani newlinedataset (EZAS) dataset. Resampling techniques such as ADASYN, SMOTE, and Borderline- newlineSMOTE were utilized to mitigate class imbalance. Feature selection methods like ANOVA newlineand Chi-square tests helped to reduce redundancy, while a suite of ML classifiers, including newlineAdaBoost and CatBoost, were trained and evaluated. The AdaBoost classifier achieved a peak newlineaccuracy of 98.9%, and its decision-making process was made interpretable using SHapley newlineAdditive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanation (LIME). newlineThe second objective introduces an Automated Machine Learning (AutoML) framework newlineto overcome the high dependency on expert-driven tuning in traditional ML workflows. For newlineHF prediction, AutoGluon was employed for automated model selection and hyperparameter newlineoptimization. This approach demonstrated superior performance, with the CatBoost model newlineachieving 99.39% accuracy and a ROC-AUC of 99.89%. SHAP values were used to enhance newlineinterpretability at both global and local levels, ensuring transparency in clinical decision support. newlineThe third objective addresses CAD detection from imaging data using a novel ConvM