Toward Robust Radar-Based Heart Sound: A Deep Transfer Learning Approach for Denoising and High-Frequency Recovery

Abstract

Radar-based heart sound (HS) monitoring is an emerging and valuable technique for noncontact health monitoring. However, raw radar HS data often suffer from uncertainty and poor quality due to signal attenuation in wireless propagation channels or motion artifacts from subjects. This inherent uncertainty and low signal fidelity result in limited usage for prescreening and proactive diagnosis in practical applications. This article presents a deep transfer learning (TL) framework tailored for radar HS enhancement to alleviate the measurement uncertainty. The architecture features dual encoder–decoder branches, in which each branch incorporates a pretrained ResNet50 as the encoder to efficiently process radar HS data. This strategy leverages the pretrained network’s ability to transfer knowledge learned from ImageNet to the radar HS domain, thereby improving training efficiency and performance. Visual comparisons of raw radar HS inputs, outputs from different pretrained networks, and ground-truth stethoscope signals validate the superiority of the ResNet50 backbone. This proposed TL strategy achieves a balance between recovering weak high-frequency components and suppressing background noisy artifacts. Experimental results demonstrate that the proposed model outperforms alternative architectures across multiple evaluation metrics. Specifically, it achieves a noise reduction of 10.04 dB, a distortion reduction of 5.43 dB, and a fivefold improvement in the high-frequency component recovery ratio compared to the input signal. The proposed HS denoising method yields more robust instantaneous heart rate (HR) estimation compared to the raw radar-based HS.