Explainable AI decision model for ECG data of cardiac disorders

Explainable AI decision model for ECG data of cardiac disorders

Dikshant Sagar , Atul Anand, Tushar Kadian, Manu K Shetty, Anubha Gupta
SBILab, IIIT Delhi, Maulana Azad Medical College, New Delhi, India

Need of an innovation: Diagnosing heart diseases via an efficient AI model on ECG signals which can save considerable time for cardiologists resulting in effective triage at peripheral levels, and effective referral service that can save patients’ lives.

Novelty: Our study is among the first to develop AI models to diagnose cardiac disorders using ECG along with the interpretability of the AI model. We have benchmarked the performance of the recent Deep Learning architectures for the detection of cardiac disorders on a large ECG dataset, PTB-XL, that is publicly available. We have proposed our custom-designed CNN architecture (DL model) and compared its performance with these state-of-the-art methods. In order to test the generalizability of our proposed best performing DL model, we have assessed its performance on another ECG dataset of arrhythmia patients.

Utility and Impact: Our AI model can help non-cardiologists in easy diagnosis and triage of patients with chest pain and other symptoms as a screening system for cardiovascular disorders. Moreover, our model exhibits the explainability/interpretability of the disease class prediction on the ECG waveforms that are characteristic of those cardiac diseases, helping medical doctors and caregivers trust the decisions made by the AI model. Interpretability of the automated ECG AI model augments the effectiveness of cardiologists in diagnosing heart disease accurately with less human error, especially in overloaded healthcare setups in low/middle-income countries such as India. Furthermore, it helps to gain confidence and improve the trust of the cardiologist toward an automated ECG AI model. Therefore, it can be implemented in a clinical setup, where low/middle-income countries struggle with a high burden of heart disease and poor healthcare delivery infrastructure.

This AI model can be used to diagnose cardiac disorders using the web app (https://ecgdetect.sbilab.iiitd.edu.in/). The decision taken by the AI model for any particular patient’s ECG can be interpreted using the ShAP plot, highlighting the abnormal ECG waves. We made a web app as open source for anyone to access. This interpretability also works as the validation of the efficacy of the model proposed and trained in this work.