Discovering Interpretable Semantics from Radio Signals for Contactless Cardiac Monitoring

Radio signals have emerged as a promising modality for cardiac monitoring, enabling fully contactless and operation-free measurement. However, the lack of semantic understanding, i.e., the ability to interpret signal dynamics in clinically meaningful terms, remains a fundamental barrier to performance, interpretability, and clinical translation. Here, we introduce a semantic representation framework for radio-based cardiac monitoring, grounded in an information bottleneck formulation. Our approach leverages intrinsic semantic invariance in the signal by integrating intra-modal variability compression with cross-modal semantic alignment. This enables the transformation of radio measurements into a structured representation space where cardiac semantics are encoded in an interpretable and clinically meaningful manner. We validate the proposed framework on a large-scale cohort of 9518 outpatients. The learned semantic representations exhibit strong alignment with reference semantics. Building on these representations, our method achieves interpretable and clinical-grade cardiac monitoring, including heart rhythm monitoring with a median inter-beat interval error of 9.4 ms (95% CI: 9.3-9.6), and arrhythmia diagnosis for atrial fibrillation and premature beats, with F1 scores of 0.929 (95% CI: 0.906-0.948) and 0.867 (95% CI: 0.847-0.887), respectively. We also demonstrate the effectiveness of the proposed framework in practical long-term, daily-life deployment scenarios. These results highlight semantic representation as a key enabler for achieving reliable and transparent radio cardiac monitoring.