Explainable AI decision model for ECG data of cardiac disorders

Abstract

Electrocardiogram (ECG) data is used to monitor the electrical activity of the heart. It is known that ECG data could help in detecting cardiac (heart) abnormalities. AI-enabled automated analysis of ECG waves has many applications in the medical domain, such as diagnostic of heart diseases, prediction of stress level, etc. In this study, we implemented a number of deep neural networks on a publicly available dataset of PTB-XL of ECG signals for the detection of cardiac disorders. Our proposed ST-CNN-GAP-5 model produced better results compared to the existing state-of-the-art results on this dataset, achieving an AUC of 93.41%. The same network architecture is tested on another ECG dataset of arrhythmia patients to assess the generalizability of our DL model for ECG datasets, yielding an accuracy of 95.8% and an AUC of 99.46%, which is competitive in performance to the state-of-the-art models. Finally, we analyzed the ECG data using SHapley Additive exPlanations (SHAP) on the trained ST-CNN-GAP-5 to assess the explainability or interpretability of the decisions of this deep convolution network model. Results indicate that the model is able to highlight relevant alterations of the ECG waves as required by clinicians, making it explainable for diagnostic purposes. Deployment of such models can help in easing the burden on medical infrastructure in low- and middle-income populous countries.