I-BEAT: Interpretable Transformer Model for Intra-Beat Wave Detection on Ambulatory ECG

Accurate detection of ECG waves is essential for diagnosing heart conditions, but existing methods often struggle with noise and morphological variability in ambulatory settings. Additionally, the lack of interpretability in current deep learning approaches limits their clinical adoption. To address these challenges, we propose I-BEAT, a novel interpretable transformer-based model capable of detecting specific intra-beat waves - P-waves, QRS complexes, and T-waves - directly, eliminating the need for prior heartbeat identification, a key limitation of current approaches. Our method improves detection accuracy while providing insights into the model's decision-making process through attention and saliency maps, enhancing its clinical relevance. Using manually annotated datasets (QTDB and LUDB) and employing strict patient separation, I-BEAT achieves F1-scores of 94.59%, 98.76%, and 97.53% for P, QRS, and T wave detection, respectively. Our innovative proximity-based labeling method addresses class imbalance, ensuring robust performance in continuous ambulatory detection. Furthermore, by incorporating processing techniques to deal with patient movement artifacts and variable signal quality, our approach effectively overcomes the challenges of long-term ECG recordings in real-world scenarios. Combining high accuracy, interpretability, and robustness, I-BEAT represents a significant step toward reliable and explainable ECG analysis in ambulatory and clinical settings.