Study on the use of standard 12-lead ECG data for rhythm-type ECG classification problems

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Current deep learning ECG classification primarily focuses on optimizing model architecture.
• A novel approach is proposed to fuse diverse single-lead ECG data for improved classification.

Purpose of the Study:

• To evaluate the efficacy of fusing single-lead ECG data for classification.
• To compare the performance of a fused-data model against a single-lead data model.

Main Methods:

• A 152-layer Squeeze-and-Excitation Residual Network (SE-ResNet) was employed as the baseline model.
• Performance was compared between SE-ResNet trained on fused multi-lead ECG data versus single-lead ECG data.
• Experiments utilized five types of rhythm-specific single-lead ECG data from Konkuk University Hospital.

Main Results:

• Lead -aVR and Lead II demonstrated superior classification performance when data was fused.
• The -aVR lead achieved high F1 scores: normal (98.7%), atrial fibrillation (98.2%), atrial premature contractions (95.1%), and ventricular premature contractions (97.4%).
• These results indicate significant potential for clinical application in medical diagnostics.

Conclusions:

• Fusion of single-lead ECGs using a 152-layer SE-ResNet enhances classification performance compared to single-lead training.
• The proposed fusion methodology proves effective across various single-lead ECG signal types.
• Lead -aVR and Lead II are identified as optimal leads for single-lead ECG classification via data fusion.