Delineation of 12-Lead ECG Representative Beats Using Convolutional Encoder-Decoders with Residual and Recurrent...

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Accurate electrocardiogram (ECG) delineation is crucial for diagnosing cardiac conditions.
• Deep neural networks (DNNs) offer potential for automated ECG analysis.
• Challenges remain in developing robust and precise DNN-based ECG delineation algorithms.

Purpose of the Study:

• To compare the performance of four distinct encoder-decoder DNN architectures for 12-lead ECG delineation.
• To evaluate the accuracy and robustness of these models against a reference algorithm and standard tolerances.
• To assess the impact of noise on the delineation performance of different architectures.

Main Methods:

• Four encoder-decoder DNNs were implemented: CED-Net, CED-LSTM-Net, CED-U-Net, and CED-Res-Net.
• Models were trained on the PhysioNet ECG database (PTB-XL) using machine-based supervised learning with a reference algorithm (ETM).
• Performance was validated using the CSE database and subjected to noise tests (high-frequency, low-frequency, power-line noise).

Main Results:

• All DNN architectures demonstrated comparable accuracy in reproducing reference measurements, with average P-wave detection accuracy of 96.6%.
• The CED-Net model met all standard tolerances for P-duration, PQ-interval, QRS-duration, and QT-interval measurements.
• CED-Net, CED-Res-Net, and CED-LSTM-Net showed robustness to various noise types, maintaining low mean duration errors.

Conclusions:

• The developed DNN architectures, particularly CED-Net, are highly capable of accurate and reliable 12-lead ECG delineation.
• CED-Net exhibits superior robustness to noise compared to the CED-U-Net architecture.
• The study highlights the competitive performance of these DNNs in automated ECG analysis, achieving low time errors comparable to published studies.