Self-supervised representation learning reveals explainable physiological structure in high-dimensional...

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Current AI in cardiology often limited by sensing modalities.
• Magnetocardiography (MCG) offers higher fidelity cardiac data than surface potentials.
• Exploring MCG's potential for advanced AI-driven cardiac analysis.

Purpose of the Study:

• Investigate if combining MCG with self-supervised learning yields meaningful cardiac representations.
• Develop and evaluate an AI model (MCG2Vec) for cardiac condition detection.
• Assess the physiological interpretability of learned cardiac embeddings.

Main Methods:

• Developed MCG2Vec, a contrastive encoder trained on raw 64-channel MCG recordings.
• Utilized recordings from 1732 consecutive patients.
• Evaluated learned embeddings using task-specific probes for coronary artery disease, ejection fraction, and atrial fibrillation risk.

Main Results:

• MCG2Vec achieved high accuracy in discriminating multivessel coronary artery disease (AUC 0.89), reduced ejection fraction (AUC 0.81), and atrial fibrillation risk (AUC 0.77).
• Attribution analyses confirmed physiological relevance of spatial and temporal patterns.
• Demonstrated explainable AI embeddings from non-invasive cardiac magnetic field data.

Conclusions:

• Combining high-fidelity MCG sensing with self-supervised learning generates structured, interpretable cardiac representations.
• This approach enhances AI's capability in non-invasive cardiac diagnostics.
• Measurement physics is crucial for advancing medical AI.