Radar-Based Heart Sound Reconstruction Using Deep Learning

Cardiovascular diseases remain one of the major diseases in modern medicine, necessitating continuous, non-invasive monitoring solutions. In this context, frequency-modulated continuous wave radar sensors use electromagnetic waves to detect thoracic micro vibrations, offering a promising non-contact alternative to traditional auscultation. However, isolating high-fidelity heart sounds (corresponding to the two beats S1 and S2) from radar signals is difficult due to respiratory harmonics and the complex non-linear relationship between thoracic displacement and acoustic sound. Furthermore, the presence of people close to the patient could also alter the radar measurements. This Thesis presents a new hybrid signal processing and deep learning pipeline to reconstruct audible heart sounds from radar data. Experimental data were collected from healthy volunteers using a IWR6843ISK radar and a digital stethoscope as a reference. The proposed methodology moves from the time domain to the time-frequency domain: raw radar signals undergo beam- forming and phase extraction, followed by conversion to amplitude spectrograms via short-term Fourier transform. A convolutional neural network with encoder-decoder architecture was de- signed and trained to map noisy radar spectrograms into clean cardiac spectrograms. Finally, the waveform in the time domain is reconstructed using the Griffin-Lim algorithm to estimate the missing phase information. Quantitative evaluation using the log-spectral distance metric yielded an average error of 10.33 dB on the inference set, consisting of six 120-second recordings recorded specifically for performance evaluation. Qualitative listening tests confirm the effective recovery of fundamental heart tones(S1andS2). These results demonstrate the feasibility of the deep learning-enhanced radar stethoscope, paving the way for discreet, long-term monitoring of vital signs in home care and clinical settings.