A physiology-driven deep learning-based pipeline for ECG segmentation
Résumé
Knowing the durations and amplitudes of cardiac waves (P, QRS, and T) constituting the ElectroCardioGram (ECG) signal is crucial for diagnosing cardiac pathologies or predicting adverse events. Segmenting ECG waves visually by an expert can be laborious, highly time-consuming, and subjective. To cope with these limitations, an automatic segmentation of these waves emerges as the most sensible solution. However, accurate automatic segmentation of ECG waves is challenging due to the contamination of ECG signals by various types of noise and artefacts that can obscure or distort the waveforms. In this paper, a new supervised automatic ECG wave segmentation pipeline is proposed. It mainly relies on three stages. A shallow preprocessing stage that supports quasi-real-time data analysis. The second stage consists of a new multi-head attention-based CNN-LSTM model for ECG segmentation. The last stage includes a physiology-driven postprocessing algorithm aimed at addressing false positives in ECG segmentation, which significantly affects the evaluation of ECG segmentation methods. The effectiveness of the proposed three-stage pipeline compared to two standard unsupervised ECG segmentation pipelines and a recent deep learning-based approach is evaluated in this paper, using the well-known PhysioNet’s QT database. The obtained results demonstrate that the proposed pipeline outperforms the existing ones, particularly in the challenging segmentation of P and waves. The good behaviors of the proposed model also confirm the usefulness of both the use of a multi-head attention layer and the additional postprocessing algorithm.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |
![]()
Cite 10.13026/C24K53 Jeu de données Laguna, P., Mark, R. G., Goldberger, A. L., & Moody, G. B. (1997). The QT Database [Data set]. physionet.org. https://doi.org/10.13026/C24K53

