Automatic diagnosis of arrhythmia with electrocardiogram using multiple instance learning: From rhythm annotation to...

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Electrocardiograms (ECGs) are crucial for diagnosing cardiac conditions like arrhythmias.
• Deep learning (DL) has significantly advanced automatic ECG diagnosis.
• Previous studies often focused separately on rhythm-level or heartbeat-level ECG analysis.

Purpose of the Study:

• To investigate if an abnormal heartbeat detection model can be trained using only rhythm-level ECG data.
• To explore the relationship between rhythm-level annotations and individual heartbeat classifications.
• To develop a high-performing automatic diagnostic model for arrhythmias.

Main Methods:

• Utilized multiple instance learning (MIL) to link ECG records (rhythm-level labels) with individual heartbeats (to be predicted).
• Sequentially trained a rhythm model for arrhythmia detection and a heartbeat model for normal/arrhythmia classification.
• Trained and validated models on a large dataset of 61,853 ECG records with rhythm annotations.

Main Results:

• The heartbeat model achieved a macro-average F1 score of 0.807 in classifying normal and four types of arrhythmias.
• This rhythm-annotation-based model significantly outperformed a model trained directly on heartbeat-annotated data.
• Demonstrated the feasibility of training effective heartbeat-level diagnostic models with less granular data.

Conclusions:

• Training heartbeat-level diagnostic models using solely rhythm-level ECG annotations is a viable and effective strategy.
• This approach offers a practical solution for leveraging large, rhythm-annotated ECG datasets.
• The findings pave the way for more efficient and scalable automatic ECG diagnostic systems.