Deep Learning Radar for High-Fidelity Heart Sound Recovery in Real-World Scenarios

Radar vibrometry using millimeter waves to recover acoustics is an exciting new topic for biometrics by remote sensing but signal integrity is challenged due to wireless channel propagation. In this article, a dual-autoencoder (DAE) network is developed to achieve high-quality radar heart sound (HS) measurement. The input to the deep learning (DL) network is the spectrogram of the initial radar HS. In the dual architecture, the two autoencoders (AEs) take in the real and imaginary components of the spectrogram. In addition, a hybrid training strategy by mixing radar data and digital stethoscope data is proposed to denoise distorted HS signal and recover high-frequency HS components. The major advantage of the hybrid training strategy compared with the standard training strategy, only using radar data as training input, is the signal distortion performance. Generalization experiments in real-world scenarios are carried out to verify the effectiveness of the proposed network. Results show satisfactory noise reduction and high-frequency recovery performance in three quantitative metrics: noise reduction, signal distortion, and maximum spectral support.