READ: Reasoning Like a Cardiologist for Contactless Arrhythmia Diagnosis

Arrhythmia poses a severe health risk, necessitating long-term monitoring to prevent complications like stroke. However, existing radar monitoring solutions fail to detect concurrent arrhythmias and lack clinical interpretability. In this paper, we bridge the gap by proposing READ (Radar Evidence-based Arrhythmia Diagnosis). Unlike previous “black-box” approaches, READ integrates mmWave radar with Large Language Models (LLMs) to deliver transparent, evidence-based diagnoses. However, training such a reasoning-capable model is hindered by the scarcity of mmWave datasets with detailed diagnostic annotations. To tackle this challenge, we first pretrain on large-scale ECG data to learn robust physiological patterns and then transfer this knowledge to radar domain using our proposed Cross-Physics Knowledge Distillation (CPKD) framework. Furthermore, we design a Morphological Grouping Tokenizer and a Morphology-Anchored Clinical Reasoning module, which parse raw signals into semantic tokens to guide step-by-step diagnostic inference. Empirical evaluation on a real-world dataset of 205 patients, collected in collaboration with medical institutions, demonstrates READ 's superior performance, achieving a 90.12% exact match ratio and a 0.9625 F ₁ score. Our framework significantly outperforms state-of-the-art baselines while providing interpretable clinical reports.