A study on several critical problems on arrhythmia detection using varying-dimensional electrocardiography.

Area of Science:

• Cardiology
• Artificial Intelligence
• Signal Processing

Background:

• The PhysioNet/Computing in Cardiology Challenge 2021 (CinC2021) questioned the necessity of all 12 standard electrocardiography (ECG) leads for accurate abnormality classification.
• Developing effective models for classifying a wide spectrum of ECG abnormalities remains a challenge.

Purpose of the Study:

• To investigate if subsets of ECG leads can provide sufficient information for comparable classification performance.
• To design deep learning models capable of classifying diverse ECG abnormalities using reduced lead configurations.
• To enhance model efficiency and interpretability in ECG analysis.

Main Methods:

• Development of novel convolutional recurrent neural network architectures.
• Implementation of a 'lead-wise' mechanism for parameter reuse in ECG neural networks, reducing model size.
• Manual design of auxiliary detectors based on clinical diagnostic rules to improve performance and interpretability.

Main Results:

• Achieved competitive challenge scores on reduced-lead ECG subsets (2- to 12-lead) in the CinC2021 post-challenge session.
• Demonstrated comparable performance even with an extreme reduction to 2-lead ECGs.
• The 'lead-wise' mechanism significantly reduced model size while maintaining performance.

Conclusions:

• Reduced-lead ECG configurations can yield performance comparable to the standard 12-lead ECG for abnormality classification.
• Proposed deep learning models and mechanisms offer effective and efficient solutions for ECG analysis.
• The findings provide a foundation for future research in reduced-lead ECG analysis and AI-driven diagnostics.