GLF-Net: A quasi-periodic prior-guided waveform segmentation network for ECG delineation.

Abstract

Electrocardiogram (ECG) signals capture cardiac electrical activity and serve as a fundamental tool for diagnosing cardiovascular diseases. A standard ECG waveform is composed of P waves, QRS complexes, and T waves, each reflecting different aspects of cardiac function through their morphological characteristics. However, precise delineation of these waveform boundaries is often complicated due to overlapping signals between adjacent heartbeats and inherent variability in waveform morphology. Therefore, this study systematically investigates the critical role of R-peak prior information in guiding ECG delineation and proposes a stepwise segmentation strategy that leverages the quasi-periodic structure of cardiac cycles. Specifically, each wave point is classified into P, QRS, T, or background categories, with heartbeat segmentation anchored by R-peak prior to enhance waveform boundary localization accuracy. To further improve performance, we develop an encoder-decoder model designed to capture associations between ECG sub-waves by fusing fine-grained local morphological features with global temporal dependencies via a dedicated integration mechanism. Experimental results on the LUDB and QTDB datasets demonstrate that the proposed method achieves superior delineation performance, with average F1-scores of 97.60% and 94.59%, respectively, surpassing those of existing state-of-the-art approaches. These results highlight the promise of incorporating structural cardiac priors to achieve more reliable and robust automated ECG analysis in clinical practice.