Discovering Interpretable Semantics from Radio Signals for Contactless Cardiac Monitoring.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Contactless cardiac monitoring using radio signals is promising but lacks clinical interpretability.
• Interpreting radio signal dynamics in clinically meaningful ways is a barrier to adoption.

Purpose of the Study:

• To develop a semantic representation framework for radio-based cardiac monitoring.
• To enable interpretable and clinically meaningful analysis of cardiac signals from radio measurements.

Main Methods:

• Utilized an information bottleneck formulation for semantic representation.
• Integrated intra-modal variability compression and cross-modal semantic alignment.
• Transformed radio measurements into a structured semantic space.

Main Results:

• Validated on 9518 outpatients, showing strong alignment with reference semantics.
• Achieved clinical-grade heart rhythm monitoring with a median inter-beat interval error of 9.4 ms.
• Demonstrated high accuracy in diagnosing atrial fibrillation (F1=0.929) and premature beats (F1=0.867).

Conclusions:

• Semantic representation is key for reliable and transparent radio cardiac monitoring.
• The proposed framework enables interpretable, clinical-grade cardiac monitoring in daily life.
• This approach overcomes limitations in signal interpretation for radio-based healthcare technologies.