Deep Recurrent Learning for Heart Sounds Segmentation based on Instantaneous Frequency Features

TL;DR

This paper introduces a deep recurrent neural network utilizing instantaneous frequency features for automatic segmentation of heart sounds in phonocardiograms, achieving near state-of-the-art accuracy with a compact model.

Contribution

It presents a novel combination of LSTM neural networks and Fourier Synchrosqueezed Transform for effective heart sound segmentation using specific time-frequency features.

Findings

Achieved an average sensitivity of 89.5%

Attained an average positive predictive value of 89.3%

Reached an average accuracy of 91.3%

Abstract

In this work, a novel stack of well-known technologies is presented to determine an automatic method to segment the heart sounds in a phonocardiogram (PCG). We will show a deep recurrent neural network (DRNN) capable of segmenting a PCG into its main components and a very specific way of extracting instantaneous frequency that will play an important role in the training and testing of the proposed model. More specifically, it involves a Long Short-Term Memory (LSTM) neural network accompanied by the Fourier Synchrosqueezed Transform (FSST) used to extract instantaneous time-frequency features from a PCG. The present approach was tested on heart sound signals longer than 5 seconds and shorter than 35 seconds from freely-available databases. This approach proved that, with a relatively small architecture, a small set of data, and the right features, this method achieved an almost state-of-the-art performance, showing an average sensitivity of 89.5%, an average positive predictive value of 89.3\% and an average accuracy of 91.3%.