X-ECG: Explainable Foundation model for Electrocardiogram

Huynh Dang Nguyen, Trong-Thang Pham, Đỗ Anh Thuận, DINH PHUONG NHI, Brijesh Patel, Ngan Le, Van Nguyen

Keywords: Electrocardiogram, Foundation model, Abnormal, Explainable

Abstract: Electrocardiography (ECG) is a cost-effective and widely accessible tool for evaluating cardiac health. While numerous machine learning methods have been developed to assist cardiologists in diagnosis, many suffer from lacking explainability, making it difficult to understand why a particular disease is classified. To address this limitation, we introduce X-ECG, an explainable ECG foundation model. To train this model, we first curate wave-level anomalies annotations on public datasets, using a rule-based algorithm that finds abnormal waves, intervals or segments in ECG signal according to established clinical knowledge. To help models learn where to focus, we propose an attention-guided training approach that enables the model to highlight relevant regions. To the best of our knowledge, X-ECG is the first ECG foundation model with built-in explainability. Our experiments show that using our dataset to guide the model not only adds explainability but also improves performance in arrhythmia classification and report generation tasks.

Primary Area: interpretability and explainable AI

Submission Number: 8537