Towards Interpretable Arrhythmia Classification With Human-Machine Collaborative Knowledge Representation.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Arrhythmia detection and classification are vital for diagnosing cardiovascular diseases.
• Current deep learning models lack interpretability, hindering clinical trust and understanding.
• There is a need for interpretable methods in automated cardiac arrhythmia analysis.

Purpose of the Study:

• To propose a novel interpretable arrhythmia classification approach using human-machine collaborative knowledge representation.
• To address the interpretability deficiency in current deep learning-based arrhythmia detection.
• To enhance classification accuracy through a human-in-the-loop mechanism.

Main Methods:

• Utilized an AutoEncoder to encode electrocardiogram (ECG) signals into distinct hand-encoded and machine-encoded knowledge components.
• Developed a classifier that processes this encoded knowledge for arrhythmia heartbeat classification.
• Implemented a human-in-the-loop (HIL) mechanism to refine hand-encoded knowledge.

Main Results:

• The proposed approach effectively classifies arrhythmias while providing interpretability.
• The human-in-the-loop mechanism demonstrated an improvement in classification accuracy.
• Experiments on the MIT-BIH Arrhythmia Database validated the method's efficacy.

Conclusions:

• The novel human-machine collaborative approach offers an interpretable solution for arrhythmia classification.
• Integrating human expertise via HIL can enhance the performance of AI models in cardiovascular diagnostics.
• This method represents a significant advancement in explainable AI for cardiac arrhythmia analysis.