A Framework Classification of Heart Sound Signals in PhysioNet Challenge 2016 Using High Order Statistics and Adaptive Neuro-Fuzzy Inference System

Abstract

To investigate the performance of Adaptive Neuro-Fuzzy Inference System (ANFIS), activated by spectral analysis features, for detection of abnormal cardiac valves sound signals. A dataset of 1837 heart sound signals were acquired from international PhysioNet Challenge 2016 databases (classes A, B and E). This included 1369 normal and 468 abnormal signals. The signals were de-noised using Notch and Butterworth filtering, fed to Discrete Fourier Transform, and 5 features using High Order Spectral (HOS) analysis were extracted from the third Cumulant. Later, the ANFIS neural network was trained and tested to discern abnormal signals. The results showed that the selected features were statistically significant ( $\text{p} ). The proposed method was tested and achieved classification of 63-89% accuracy, 63-100% sensitivity, and 62-100% specificity, respectively. The results were compared with reports utilizing different neural network techniques, indicating competitive performance. The HOS spectral features can be reliable to participate in neural network systems to sort heart sound (HS) signals as normal or abnormal. The bispectral matrix is a new presentation of attributes describing signals. The ANFIS is a suggestive successful tool, which has not been attempted in Physio-net challenge 2016. The HOS attributes and ANFIS can participate successfully in PhysioNet Challenge 2016.