Deep-Learning Based Segmentation of In-Ear Cardiac Sounds

Abstract

Cardiovascular disease is a leading cause of death worldwide. Auscultation, a common diagnostic method, involves listening to heart sounds to detect valve contraction irregularities, with heart sound segmentation (i.e., identifying heartbeat phases) being a crucial first step. Due to the high level of expertise required for traditional auscultation, previous work has automated segmentation using digital sounds from stethoscopes or phonocardiographs; however, these methods depend on specialist medical devices, limiting continuous and wide usage. Earable devices now offer continuous heart sound capture via in-ear microphones (IEM) on Active Noise Cancellation (ANC) earphones, opening new possibilities for portable, continuous, out-of-hospital heart sound segmentation. However, techniques developed for phonocardiogram (PCG) signals are not directly applicable because of the distinct differences in signal characteristics between IEM and PCG signals. In this work, we demonstrate the unique potential of using in-ear cardiac sounds for heart sound segmentation. We begin by analysing the temporal and frequency differences between IEM and PCG signals, then introduce a U-Net deep learning model tailored for in-ear heart sound segmentation. Additionally, we propose a more stringent evaluation method for segmentation accuracy and use this to evaluate our method and the baselines. We collected data from 11 participants, and our model achieved 84% accuracy, outperforming established baselines.