Deep learning-based approach for denoising heart vibration signals

Salman ALUHAMMAD ALALI1, Amar Kachenoura1, Lotfi Senhadji2, Alfredo Hernandez3, Cindy Michel4, Laurent Albera5, Ahmad KARFOUL6
1Univ Rennes, Inserm, LTSI - UMR 1099, 2LTSI - Inserm - Univ Rennes, 3INSERM - LTSI U 1099, 4CardiaMetrics, 5LTSI, UMR 1099, université de Rennes, 6Université de Rennes


Abstract

Heart Failure (HF) is one of the leading causes of death worldwide. Daily monitoring of cardiac biomarkers offers an effective solution for patients at risk. Unlike existing non-invasive systems, limited to long-term monitoring, Implantable Devices (IDs) provide promising alternatives. Our group has contributed to the development of a mini-invasive ID placed at the gastric fundus to acquire heart's mechanical activities. This innovative implant is deliverable via gastroscopy techniques, providing a less invasive alternative to conventional cardiac implants. However, gastric site-related artifacts limit the ACCelerometry (ACC) signal analysis. Indeed, these artifacts distort mechanical heart signals, obscuring key heart sounds S1 and S2. Here, we develop a Deep Learning (DL)-based approach for denoising ACC signals acquired from this ID, to highlight events of interest, S1 and S2.

The proposed method takes advantage of the ability of Convolutional Neural Network (CNN) to filter out noise and extract relevant features. This enables CNNs to capture the contextual information and the inherent patterns within the signal and artifacts. Drawing from this principle, the new denoising method consists of two steps: i) pretraining a low-cost CNN for a classification task (HF vs. HF-free subject), and ii) selecting, from the pre-trained kernels, the one that best enhances the Signal-to-Noise-Ratio (SNR) for S1 and S2 waves.

A preclinical dataset issued from 7 pigs (3 HF and 4 HF-free) is used to evaluate the performance of the method. A comparison with Wavelet-based method, Empirical Mode Decomposition, Principal Component Analysis, Canonical Correlation Analysis, and Independent Component Analysis is conducted. The CNN-based strategy achieved an SNR improvement of 20% for S1 and 16% for S2, surpassing traditional methods.

The obtained results demonstrated that the proposed method improved the reliability of ACC segmentation into key events, S1 and S2, enhancing the capabilities of the new implant for long-term monitoring of cardiac conditions.