Enhancing dynamic ECG heartbeat classification with lightweight transformer model.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Arrhythmia, a major cardiovascular disease, causes significant global mortality.
• Wearable ECG devices offer continuous monitoring but face challenges with signal interference.
• Existing heartbeat classification models are often parameter-heavy and perform poorly on dynamic ECG data.

Purpose of the Study:

• To develop a novel, lightweight model for accurate arrhythmia detection using wearable ECG data.
• To address the limitations of traditional models in handling noisy signals and large parameter sizes.
• To improve the detection of premature ventricular contractions (PVCs) and supraventricular premature beats (SPBs).

Main Methods:

• Proposed a lightweight model, Lightweight Fussing Transformer, incorporating a novel LightConv Attention (LCA) mechanism.
• Replaced the self-attention component of the Fussing Transformer with the more efficient LCA.
• Developed an enhanced embedding structure using Convolutional Neural Networks with attention to better capture heartbeat morphology.

Main Results:

• The LCA mechanism achieved performance comparable to or exceeding self-attention with significantly fewer parameters.
• The enhanced embedding structure effectively improved the weighting of internal heartbeat features.
• Experimental validation on real datasets demonstrated outstanding accuracy in detecting PVCs and SPBs.

Conclusions:

• The proposed Lightweight Fussing Transformer with LCA offers an efficient and accurate solution for arrhythmia detection from wearable ECGs.
• This model overcomes the limitations of traditional methods, particularly in noisy environments.
• The findings support the potential of this lightweight model for early warning systems in wearable health devices.