Enhancing single-lead ECG arrhythmia classification via multi-teacher decomposed feature distillation

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Arrhythmias manifest across all electrocardiogram (ECG) leads, but with varying prominence.
• Leveraging information from prominent ECG leads can potentially improve the analysis of less prominent ones.

Purpose of the Study:

• To propose a novel knowledge distillation (KD) model for enhancing weak ECG leads using information from stronger leads.
• To improve the accuracy of arrhythmia detection through signal enhancement.

Main Methods:

• A knowledge distillation (KD) model was developed, utilizing single-lead signals for the student network and twelve-lead signals for the teacher network.
• Tucker decomposition was employed to decompose the teacher network's feature maps, facilitating information transfer.
• The model was evaluated on the Chapman ECG dataset for classification tasks.

Main Results:

• The proposed KD model achieved a high accuracy of 96.48% on the Chapman ECG dataset classification task.
• The method effectively enhanced the information content of weaker ECG leads by learning from stronger leads.

Conclusions:

• The KD model demonstrates a promising approach for improving ECG signal quality and diagnostic accuracy.
• This technique offers a potential advancement in non-invasive cardiac arrhythmia detection and analysis.