A Tensor Approach to Heart Sound Classification

Abstract

In the context of the PhysioNet/CinC 2016 Challenge, where a relatively large, labeled data set of phonocardiograms (PCGs) was made available, this work presents a mixed approach to the problem of its binary classification. Instead of laboriously selecting a set of PCG signal features that capture the fundamental differences between healthy and unhealthy heart sounds, a rather exhaustive set of features is generated for each heart beat segment, which is then represented in a 4-way tensor. In a second stage, such tensor representation is decomposed and compressed, to determine only a few of the most discriminating parameters, which are then fed to an otherwise standard classifier. This results in an accurate, compact and fast algorithm, that can effectively classify noisy PCG signals of different duration, achieving a balanced accuracy of 91.9% in 10-fold cross-validation, and 84.54% on the Challenge hidden test data (the 4 th highest score).

A benchmark of heart sound classification systems based on sparse decompositions

“…[5][6][7] Linear predictive coding coefficients have been also used. 7 In terms of classifiers, competitors mostly used Neural Networks, 5,8,9 Support Vector Machines 10,11 and Random Forest techniques. [12][13][14] In parallel, sparse representations of the heart sounds (Matching Pursuit decomposition and Linear Predictive Coding of the residual) have been shown to provide a compact and meaningful representation of PCG signals.…”

Towards Domain Invariant Heart Sound Abnormality Detection Using Learnable Filterbanks

“…Notable features reported for this dataset includes, time, frequency and statistical features [8], Mel-frequency Cepstral Coefficients (MFCC) [9], and Continuous Wavelet Transform (CWT) [10]. Classifiers like SVM [11], k-Nearest Neighbor (k-NN) [9], Multilayer Perceptron (MLP) [10], [12] and Random Forest [8], deep learning approaches with 1D & 2D CNNs [13], [14], and Recurrent Neural Network (RNN) [15] based architectures were employed in the challenge submissions. The winning algorithm, similar to a good number of other submissions, proposed an ensemble; a static filter front-end 1D-CNN model combined with an Adaboost-Abstain classifier using a threshold-based voting algorithm.…”

Learning Front-end Filter-bank Parameters using Convolutional Neural Networks for Abnormal Heart Sound Detection

“…The 2016 Physionet/CinC Challenge released an archive of 4430 PCG recordings, which is the most extensive open-source heart sound dataset to date. Time, frequency and statistical features [6], Melfrequency Cepstral Coefficients (MFCC) [7], and Continuous Wavelet Transform (CWT), were some of the commonly used features by the PhysioNet challenge entrants. Among the top scoring systems, Maknickas et al [8] extracted Melfrequency Spectral Coefficients (MFSC) from unsegmented signals and used a 2D CNN.…”

“…Plesinger et al [9] proposed a novel segmentation method, a histogram based feature selection method and parameterized sigmoid functions per feature, to discriminate between classes. Various machine learning algorithms including SVM [10], k-Nearest Neighbor (k-NN) [7], Multilayer Perceptron (MLP) [11], [12], Random Forest [6], 1D [13] and 2D CNNs [8], and Recurrent Neural Network (RNN) [14] were employed in the challenge. A good number of submissions used an ensemble of classifiers with a voting algorithm [6], [11], [12], [13].…”