AI Heart Sound Denoising at the Edge

Cardiac auscultation with a digital stethoscope is an important method for early diagnosis of cardiovascular diseases (CVDs), especially in resource-constrained settings. However, its utility is often impaired by unwanted noise (e.g., lung sounds, ambient speech) obscuring critical heart signals in the recordings. This paper addresses the challenge of denoising heart sounds in noisy environments and performing this processing on-device for minimal delay and improved data privacy and security. We propose a deep learning-based denoising method using U-Nets based on the Spleeter architecture, which separates heart sound signals from noise in the time–frequency domain. The models are implemented on two edge computing platforms — a Raspberry Pi 4 B and a Qualcomm RB3 Gen 2 — demonstrating the feasibility of AI-driven heart sound denoising at the edge. We further optimize the models via quantization to Int8 precision to leverage the RB3 Gen 2’s neural processing unit (NPU) for acceleration. Experimental results show that the denoisers achieve significant enhancement in signal quality, with each yielding an average improvement in scale-invariant signal-to-distortion ratio (SI-SDR) exceeding 10 dB. The edge deployment meets real-time constraints, with the quantized model on the RB3 Gen 2 executing denoising substantially faster (at least 17 times faster than the Raspberry Pi implementation) and at lower power compared to the full-precision model, all without appreciable loss in denoising performance - 0.06 dB being the largest drop. These results demonstrate the potential of deploying advanced AI denoising on edge devices for immediate, on-device heart sound denoising.