Stacked Bi-LSTM Network and Dual Signal Transformation for Heart Sound Denoising

Abstract

This work introduces a dual signal transformation Bidirectional LSTM network for real-time heart sound denoising. The proposed methodology employs a stacked network (ST) approach, which combines a synthesis basis and learned analysis using short-time Fourier transform (STFT). The model is trained on the noisy data, generated by the mixer of White noise and clean heart sound collected from the Michigan heart sound database. The network is capable of processing frames in real-time (one in, one out) and has achieved promising results. The dual signal transformation Bidirectional LSTM network can reliably use the magnitude spectra to get information and uses a learned feature-base to add phase information by combining these two types of signal transformations. The experimental results show that the proposed method improves the signal-to-noise ratio (SNR) of heart sounds under various SNR levels. The algorithm is implemented on Raspberry Pi 4 device for real-time implementation. The real-time experimental results using Raspberry Pi 4 show that the proposed network outperforms the adaptive overlapping group sparse denoising for heart sound signals (adaOGS) technique and conventional wavelet-based approaches for denoising heart sounds. © 2023, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.