Explainable AI-driven machine learning for heart disease detection using ECG signal

Abstract

The majority of countries globally find that cardiovascular diseases have the greatest death rate. Early detection of heart disease is necessary to prevent deaths from it. Big data for medical diagnostics has made it easier to construct sophisticated machine learning (ML) and deep learning (DL)-based models for the early automatic identification of heart disease. Using three common ECG datasets, i.e., the Physionet Challenge 2016, the PASCAL Challenge Competition, and the MIT-BIH - random forest (RF) and extreme gradient boosting (XGBoost), two tree explainer classifiers are proposed for the detection of heart disease in this study. During pre-processing variety of techniques like bandpass filtering, denoising, butterworth filtering and zero-phase filtering of the signal are carried out. After the signals are restored to the time domain form using IDWT, DWT and EWT are utilized to extract features. SHapley Additive exPlanations (SHAP) is used for the feature importance which identifies important features that have more impact on the prediction output of the models. It is demonstrated from the simulation results that EWT with XGB performs well in the three datasets considered. RF explainer algorithm with EWT feature performs best and results in an AUC of 95.36 % for the PASCAL Challenge Competition dataset. However, with AUC values of 97.44 % and 98.25 % for the Physionet Challenge 2016 and MIT-BIH datasets, respectively, the XGB explainer algorithm with EWT features performs better than (DWT+XGB), (DWT+RF), and (EWT+RF). When all three datasets are compared to the current models, overall the suggested model performs better and has a higher AUC.