Deep Multi-instance Networks for Bundle Branch Block Detection from Multi-lead ECG.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Bundle branch block (BBB) is a common cardiac disorder diagnosed using electrocardiogram (ECG) signals.
• Conventional diagnostic methods rely on handcrafted features with limited discriminative power and require costly, precise heartbeat annotations for supervised learning.
• Existing approaches face challenges in accuracy and efficiency due to the limitations of feature engineering and annotation requirements.

Purpose of the Study:

• To propose a novel end-to-end deep network for classifying three types of heartbeats: right bundle branch block (RBBB), left bundle branch block (LBBB), and others.
• To implement a multiple instance learning-based training strategy to overcome the need for detailed heartbeat annotations.
• To evaluate the proposed method's performance on established ECG databases.

Main Methods:

• Development of a novel end-to-end deep neural network architecture.
• Application of a multiple instance learning (MIL) training strategy to utilize record-level ECG labels.
• Training the model on the China Physiological Signal Challenge 2018 (CPSC) database and testing on the MIT-BIH Arrhythmia (AR) database.

Main Results:

• The proposed deep network achieved an overall accuracy of 78.58%.
• High sensitivity was reported for specific BBB types: 99.72% for RBBB and 84.78% for LBBB.
• The method demonstrated superior performance compared to baseline approaches, particularly in classifying RBBB and LBBB.

Conclusions:

• The developed deep learning model provides an effective solution for classifying BBB on ECG datasets.
• The multiple instance learning strategy successfully addresses the challenge of requiring only record-level labels, reducing annotation costs.
• This approach represents a promising advancement for automated BBB detection in clinical practice.