Heart sound signal recovery based on time series signal prediction using a recurrent neural network in the long short-term memory model

In this research, we propose a method for recovering heart sound signals that comprises the long short-term memory prediction model based on the recurrent neuron network architecture. The complete heart sound signal is used to implement a prediction model to recover damaged or incomplete heart sound signals. Root mean square errors (RMSEs) and Pearson’s correlation coefficients are used for numerical evaluation. The signals of 13 out of 15 subjects are considerably improved, with the RMSE being as low as 0.03 ± 0.04. Using the Pearson correlation coefficients to estimate the degree of signal recovery, the highest coefficient of correlation between the original and recovered time-domain waveforms is 0.93, and that between the corresponding spectra is 0.967. Waveforms and spectra are used to compare the results graphically. The recovered signal more closely fits the original signal than the interfered signal. Additionally, excess frequency components in the recovered spectra are found to be filtered out and important features retained. Thus, the proposed method not only recovers incomplete or disturbed signals but also has the effect of a filter.