Open-world electrocardiogram classification via domain knowledge-driven contrastive learning.

Area of Science:

• Cardiology
• Artificial Intelligence
• Machine Learning

Background:

• Automatic electrocardiogram (ECG) classification aids disease diagnosis but struggles with limited training data covering only specific ECG types.
• Conventional models fail to recognize unseen or unknown ECG types present in real-world data.
• This limitation hinders the reliability of automated diagnostic tools in diverse clinical scenarios.

Purpose of the Study:

• To address the challenge of classifying limited known ECG types and identifying unknown ECG types in an open-world setting.
• To develop a novel method for open-world ECG classification that improves the identification of both observed and unobserved ECG classes.
• To enhance the robustness and reliability of automated ECG diagnostic systems.

Main Methods:

• Proposed a customized method integrating clinical knowledge into contrastive learning.
• Generated "hard negative" samples to guide the learning of distinguishable diagnostic ECG features.
• Employed multi-hypersphere learning for compact ECG representation and classification.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art approaches on 12-lead ECG datasets (CPSC2018, PTB-XL, Georgia).
• Achieved higher accuracy in identifying unseen ECG classes and certain seen classes.
• Outperformed comparative methods, highlighting its effectiveness in open-world scenarios.

Conclusions:

• Open-world ECG classification is crucial for improving the reliability of automatic ECG diagnosis.
• The proposed method effectively tackles the challenge of classifying known and unknown ECG types.
• The findings contribute to more robust and clinically applicable automated ECG analysis tools.