Learning Seismocardiogram Beat Denoising Without Clean Data

Mohammad Nikbakht; David Lin; Omer T Inan

Abstract

Noninvasive monitoring of cardiovascular health plays a crucial role in predicting risks and reducing mortality rates, especially in the context of trauma care. The seismocardiogram (SCG) in particular is a noninvasive signal that has been shown to monitor key health parameters related to blood volume loss estimation, suggesting its ability to guide trauma care intervention. Robust extraction of features from SCG signals in noisy environments is challenging due to low signal-to-noise (SNR) ratios. In addition, lack of access to clean ground truth signals makes developing denoising algorithms even more difficult. In this work, we propose a novel deep learning-based approach for denoising SCG signals without requiring access to clean ground truth signals. Experimental results showed (1) enhancement in the signal-to-noise ratio (SNR) (approximately 10 dB and 9 dB increase for AO and AC regions of -10 dB SNR beats respectively), and (2) improvements in feature extraction accuracy (approximately 3x and 1.5x for AO and AC features of -10 dB SNR beats respectively) using the denoising model. Thus, the model effectively reduces noise, and improves the quality of SCG signals, leading to improved accuracy in feature extraction in noisy environments. This is a promising step forward in improving the quality and utility of SCG signals for clinical and research purposes.

Keywords: Seismocardiogram; Deep Learning; Noise Reduction; Signal to Noise Ratio

Links
[Full text PDF][Bibtex]