Deep learning based ECG segmentation for delineation of diverse arrhythmias

Area of Science:

• Biomedical Engineering
• Cardiology
• Artificial Intelligence in Medicine

Background:

• Accurate electrocardiogram (ECG) waveform delineation is crucial for diagnosing heart conditions.
• Deep learning segmentation models show promise but struggle with arrhythmias.
• The impact of diverse arrhythmias on ECG delineation quality requires detailed investigation.

Purpose of the Study:

• To investigate the effect of arrhythmias on ECG delineation accuracy.
• To develop and evaluate strategies for improving ECG delineation performance in the presence of arrhythmias.
• To address the limitations of current models in handling underrepresented arrhythmias.

Main Methods:

• Development of a U-Net-like segmentation model tailored for ECG delineation, focusing on diverse arrhythmias.
• Implementation of a post-processing algorithm for noise removal and automatic boundary determination of P, QRS, and T waves.
• Training and evaluation on diverse datasets, including LUDB and QTDB, with specific assessment across various arrhythmia types.

Main Results:

• The proposed model achieved high performance, with F1-scores exceeding 99% for QRS and T waves, and over 97% for P waves on the LUDB dataset.
• Observed that models performing well on standard benchmarks may falter on underrepresented arrhythmias like tachycardias.
• Identified specific challenges and proposed solutions for improving delineation in complex arrhythmic cases.

Conclusions:

• The developed U-Net based model with post-processing effectively delineates ECG waveforms, even with diverse arrhythmias.
• Performance disparities highlight the need for models robust to underrepresented arrhythmias.
• Further strategies are proposed to enhance ECG analysis accuracy across a wider spectrum of cardiac conditions.