X-ECG: Self-Supervised Explainable Foundation Model for Electrocardiogram

ACL ARR 2026 January Submission8583 Authors

Keywords: Electrocardiogram, Explainable, Foundation model

Abstract: Electrocardiography (ECG) is widely used for cardiac health evaluation, yet many machine learning methods for ECG analysis lack explainability. We introduce X-ECG, a self-supervised explainable ECG foundation model that learns to focus on clinically relevant regions without manual attention annotations. Our key component is Clinically-Guided Attention Localization, which generates attention pseudo-labels on-the-fly using rule-based clinical knowledge and supervises model attention toward these regions via KL divergence loss. This enables the model to highlight abnormal regions contributing to predictions without manual annotations, while improving arrhythmia classification and report generation performance. Experiments show that X-ECG achieves state-of-the-art anomaly localization while achieving state-of-the-art arrhythmia classification and report generation performance.

Paper Type: Long

Research Area: Clinical and Biomedical Applications

Research Area Keywords: Clinical and biomedical language models,clinical decision support

Contribution Types: Model analysis & interpretability

Languages Studied: English

Submission Number: 8583