Denoising of Heart Sound Recordings A self-supervised approach for real non-stationary noise conditions

In the context of assisting patients with heart failure, implantable devices measuring cardiac mechanical activity offer promising solutions for remote autonomous monitoring of individuals at risk. However, cardiac vibration signals acquired by these implants are often subject to various specific noises associated with sensor placement and patient activity. In this paper, we propose a novel supervised method for cleaning noise-corrupted sequences and improving their pathophysiological interpretation. Our approach involves modeling the characteristics of both clean cardiac components and noise contributions through a qualitative classification of recordings. These models are then utilized to generate a controllable database, which trains a deep-learning-based denoising autoencoder. We have tested this procedure on a preclinical prototype implanted subcutaneously in four healthy pigs. We evaluate the denoising performance of our method using various quantitative and qualitative metrics. Our cleaning process, by reducing the mean relative error by up to a factor of five and enhancing the contrast of heart sounds to levels similar to their clean counterparts, improves the accuracy of physiological estimates such as heart rate or systole and diastole durations.Clinical relevance- Our approach consistently increases the number of signals suitable for further medical analysis. The proposed combination of a data-driven training database and a machine-learning denoising algorithm can significantly improve the quality and reliability of heart sound signals.