Decision Support System for Early Detection of Cardiac Health Status

Abstract

Cardiovascular diseases have surpassed cancer as the leading cause of death on the planet today. Numerous decision-making systems with computer-assisted support have been developed to assist cardiologists to detect heart disease, and thus, lowering the mortality rate. The purpose of this research is to classify audio signals received from the heart as normal or abnormal. The PhysioNet Computing in Cardiology (CinC) 2016 benchmark dataset, popularly known as PhysioNet 2016, has been used to validate the proposed methodology presented here. PhysioNet 2016 contains a total of 3,200 PCG recordings divided into sub-datasets A-F. In this work, researchers have proposed three different techniques with respect to decision support system. In first technique, textural features such as Linear Binary Pattern (LBP), Adaptive-LBP, and Ring-LBP have been extracted from the existing spectrogram and combined with the features extracted from the chromagram. It has been observed that the combination of features extracted from both the image variants has resulted in a greater accuracy as compared to the scenario where researchers were using only the spectrogram. The experiment yielded the mean accuracy, precision, and F1-score as 94.87, 93.11, and 95.273, respectively. In second technique, authors suggest a unique methodology for the detection of important events in an audio signal using a biologically-inspired depiction of the audio stream through a picture known as Gammatonegram which correlates to the processing of audio in the cochlea membrane of the inner human auditory system. In this study, texture-related features which include Linear Ternary Pattern (LTP), Local Directional Pattern (LDP), Geometric Local Textural Pattern (GLTP), and Local Phase Quantization (LPQ) have been extracted from a visual representation of PCG signal such as Spectrogram, Scalogram, Mel-spectrogram, and Gammatonegram. As compared to the case when researchers were employing other images, it has been noticed that the fusion of the features retrieved from the Gammatonegram has led to an increased overall classification performance metrics. The experiment resulted an overall accuracy of 94.00 % with precision and F1 scores of 91.77 and 93.61 respectively. In third technique, conversion of PCG signals into 2-D Time-Frequency images, viz. Tempogram, Chromagram, and Spectrogram has been performed. Further, data augmentation iv methods have been used to improve the imbalanced dataset. The present study utilizes a U-Net architecture-based Convolutional Neural Network (CNN), incorporating CNN, Residual Neural Network (ResNet), Visual Geometry Group (VGG), and Inception V3 blocks as encoders and decoders, to make a comparative evaluation of PCG signal classification models. The model under consideration achieved a validation accuracy of 95.89% with F1-score as 95.90%.