Non-contact apnea detection and vital-signs estimation based on radar signals

Continuous-wave (CW) radar is a promising modality for contactless cardiopulmonary monitoring. However, reliable apnea screening and vital-sign estimation remain challenging because respiratory and cardiac motions are weak, variable, and susceptible to motion artifacts and interference. This article presents a two-stage CW radar framework for protocol-labeled apnea screening and vital-sign estimation using a public clinically recorded dataset with synchronized medical reference signals. In the first stage, apnea screening is formulated as a one-class anomaly detection problem based on radar-derived respiratory-cycle dynamics, where fused local-density and reconstruction-error scores are used to detect apnea. In the second stage, a guarded physics-informed deep learning model, denoted GPIDL-CNN, is introduced for breathing-rate (BR) and heart-rate (HR) estimation by combining spectrogram representations, handcrafted signal descriptors, and confidence-aware physiological priors. Evaluated on data from 30 volunteers, the proposed apnea-screening stage achieved a balanced accuracy of 0.99 and an AUROC of 0.997. For vital-sign estimation, GPIDL-CNN achieved mean absolute errors of 1.45 breaths per minute for BR and 2.90 beats per minute for HR, outperforming reproduced deep-learning and signal-processing baselines under the same dataset and evaluation protocol. These results show that the proposed framework provides an effective approach for protocollabeled apnea detection and accurate radar-based BR and HR estimation from CW radar signals.