Supraventricular ectopic beats and ventricular ectopic beats detection based on improved U-net.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Supraventricular ectopic beats (SVEB) and ventricular ectopic beats (VEB) are common arrhythmias with diverse presentations.
• Accurate automatic localization of SVEB and VEB is crucial for effective clinical diagnosis and management.
• Existing methods face challenges due to the uncertain occurrence and morphological variability of ectopic beats.

Purpose of the Study:

• To develop a deep learning model for simultaneous automatic localization of SVEB and VEB.
• To enhance the U-net architecture for improved feature extraction and localization accuracy.
• To evaluate the proposed method's performance against established benchmarks using clinical datasets.

Main Methods:

• A modified U-net architecture, termed U-net, was developed incorporating group convolution and a multi-scale 2D deformable convolution module.
• The model utilizes a dual-output strategy, processing both high and low-resolution features via Dice loss.
• Patient-specific training was performed on the MIT-BIH arrhythmia database, with evaluation metrics including Sensitivity, Positive Prediction Rate, and F1-scores.

Main Results:

• The U-net model achieved state-of-the-art F1-scores for both SVEB (81.3%) and VEB (95.4%) across 24 testing records.
• The method demonstrated leading performance in Sensitivity and Positive Prediction Rate compared to existing studies.
• The dual-output approach effectively integrated learning from different feature resolutions.

Conclusions:

• The proposed U-net deep learning method shows significant potential for accurate and efficient detection of SVEB and VEB.
• This advancement could lead to improved clinical diagnostic capabilities for ectopic beats.
• The refined deep learning architecture offers a promising tool for arrhythmia analysis.