Translation of Radar Signals into Latent Cardiac Event Space for Scalable, Annotation-free Heart Disease Diagnosis

Radar-based cardiac diagnosis faces significant scalability challenges due to its heavy reliance on expert-labeled data. We propose Radar-to-Event Net (R2E-Net), a novel cross-modal framework that recasts contactless diagnosis as a semantic translation task into a latent cardiac event space. Our method leverages ECG-supervised event representations as a structural prior, projecting radar signals into this shared space via a Transformer-based encoder. This unidirectional alignment enables fully annotation-free radar diagnosis, bypassing radar label dependency while preserving physiological semantics. Moreover, we introduce a robust spatiotemporal radar encoder, which eliminates point-selection heuristics and directly learns meaningful motion patterns from beamforming data. Evaluated on a large-scale real-world dataset of 7,336 clinical recordings, R2E-Net achieves an average F1 score of 82.4% without radar labels, which can even outperform radar models trained under full supervision. Moreover, it generalizes well across age, gender, and disease subtypes, and remains robust in the presence of interfering cardiac conditions. These results demonstrate that our approach offers a scalable, accurate, and annotation-free solution for real-world cardiac monitoring.