ECGVEDNET: A Variational Encoder-Decoder Network for ECG Delineation in Morphology Variant ECGs.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Electrocardiogram (ECG) delineation is crucial for cardiovascular diagnosis, but deep learning models face challenges with data availability and ECG morphology variations.
• Existing deep delineation frameworks are limited by the generalizability of models trained on insufficient data and the inherent variability of ECG complexes.

Purpose of the Study:

• To introduce a novel deep delineation model, ECGVEDNET, and a large-scale 12-leads ECG dataset, ICDIRS, to overcome limitations in ECG delineation.
• To develop a model capable of addressing ECG morphology variations through a well-regularized latent space.
• To establish a transfer learning framework for applying knowledge from large datasets to smaller ones.

Main Methods:

• Development of ICDIRS, a large-scale ECG dataset with 156,145 QRS onset and 156,145 T peak annotations.
• Introduction of ECGVEDNET, a variational encoder-decoder network designed with a regularized latent space to minimize morphology variations.
• Implementation of a transfer learning framework to leverage knowledge from ICDIRS for smaller datasets.

Main Results:

• ECGVEDNET achieved high accuracy on ICDIRS: 86.28%/88.31% for QRS onset and 89.94%/91.16% for T peak within 5/10 ms tolerance.
• On the QTDB dataset, the model demonstrated state-of-the-art performance with average time errors of -1.86 ± 8.02 ms for QRS onset and -0.50 ± 12.96 ms for T peak.
• The proposed methods show effectiveness on both large-scale and smaller datasets.

Conclusions:

• ECGVEDNET, trained on the large-scale ICDIRS dataset, effectively addresses challenges in ECG delineation caused by data limitations and morphology variations.
• The model achieves state-of-the-art performance, improving accuracy in identifying fiducial points for cardiovascular diagnosis.
• The study will release source code and pre-trained models to facilitate further research and application.