A New Direct Heart Sound Segmentation Approach using Bi-directional GRU

Abstract

Heart sound segmentation is a key step in automatic analysis of phonocardiogram (PCG) for early pathology detection. In this paper, we propose a novel method inspired by the Part of Speech (POS) tagging problem for heart sound segmentation. We use a Bi-directional Gated Recurrent Unit (GRU) to predict the state of the heart sound cycles directly, steering away from the traditionally used envelopes and time-frequency based features. Our method is evaluated on a large dataset using a 10-fold cross-validation. The proposed method has achieved overall 96.86% accuracy and the F1 score is 98.40% on the test sets. The proposed method has outperformed other existing state of the art methods by 1–3 percentage in terms of accuracy and F1.