Peak Attention U-Net: Enhancing ECG delineation with attention

Cardiovascular diseases are one of the leading causes of death worldwide, making accurate analysis of electrocardiograms (ECGs) critical for early diagnosis and effective treatment. ECG delineation – the precise identification of waveform boundaries and peaks – is essential for clinical assessment but remains a challenging and time-consuming task when performed manually. While previous automated approaches have achieved reasonable performance, they are often limited by their inability to detect waveform peaks, reliance on pre-segmented heartbeats, and lack of generalizability across full-length 12-lead ECGs. This paper presents Peak Attention U-Net , a novel deep learning model for automated ECG delineation with enhanced peak detection capabilities. Building upon the U-Net encoder–decoder architecture, our approach integrates attention gates to selectively focus on salient waveform features, enabling precise identification of P, QRS, and T waves, as well as their respective peaks, across full-length 12-lead ECG signals. We evaluate the model on the LUDB dataset and demonstrate that Peak Attention U-Net achieves state-of-the-art performance in fiducial point delineation, with final F1-scores of 89.26 for P peak, 99.72 for R peak, and 97.51 for T peak, representing significant improvements over the second-best model in the literature (+13.67%, +12.19%, and +11.45%, respectively). The model is lightweight, efficient, and generalizes well to diverse cardiac conditions, supporting real-time clinical applications and deployment in wearable devices. These results demonstrate the effectiveness of the proposed model and its potential to advance automated ECG analysis in biomedical signal processing.