A benchmark of heart sound classification systems based on sparse decompositions

Abstract

Background: Nowadays, cardiovascular diseases (CVD) remain the main cause of death worldwide. A heart sound signal or phonocardiogram (PCG) is the most simple, economical and non-invasive tool to detect CVDs. Advances in technology and signal processing allow the design of computer-aided systems for heart illnesses detection from PCG signals. Purpose: The paper proposes a pipeline and benchmark for binary heart sounds classification. The features extraction architecture is focused on the use of Matching Pursuit time-frequency decomposition using Gabor dictionaries and the Linear Predictive Coding method of a residual. We compare seven classifiers with two different approaches: feature averaging and cycle averaging. Methods: We test our proposal on the PhysioNet/CinC challenge 2016 database, which comprises a wide variety of heart sounds recorded from patients with normal and different pathological heart conditions. We conduct a 10-fold stratified cross-validation method to evaluate the performance of different classification algorithms. The feature sets were also tested when using an oversampling method for balancing. Results: The benchmark identified systems showing a satisfying performance in terms of accuracy, sensitivity, and Matthews correlation coefficient. Results can be improved when using feature averaging and an oversampling strategy.

A Comprehensive Survey on Heart Sound Analysis in the Deep Learning Era

“…, where P (X|y) is the likelihood probability distribution. The Naïve Bayes Classifier [41], [91], [102], [103] was widely used for heart sound classification due to its advantage of being easy-touse. Gaussian Mixture Models (GMMs) [53], [95] were used to estimate the data distribution by optimising the weights of Gaussian mixture components and mean and variance in each component.…”

“…4 presents a statistic of recent works from 2017 to 2022 that employ classic machine learning models for heart sound classification. SVMs have been very widely used for heart sound classification by learning a supporting hyperplane between classes [6], [33], [40], [41], [43], [46]- [48], [55], [60], [64], [75], [79], [85], [91]- [93], [95]- [97], [99], [100], [103]. Apart from linear projection between data samples and labels, SVMs can learn separating hyperplanes on non-linear data via non-linear kernels, such as radial basis function.…”

“…Furthermore, k-nearest neighbours (KNNs) has shown good performance on heart sound classification with the idea of classifying a data sample according to the classes of its knearest neighbours [40], [41], [43], [46], [60], [75], [85], [90], [91], [93], [103], [104]. Also, decision trees were successfully used for heart sound classification [7], [40], [41], [75], [88], [99], [103].…”

“…Furthermore, k-nearest neighbours (KNNs) has shown good performance on heart sound classification with the idea of classifying a data sample according to the classes of its knearest neighbours [40], [41], [43], [46], [60], [75], [85], [90], [91], [93], [103], [104]. Also, decision trees were successfully used for heart sound classification [7], [40], [41], [75], [88], [99], [103]. One reason is that limiting the number of decision nodes can help avoid overfitting [7], and another reason is that the structure of a decision tree can show the internal logic for classification.…”

“…Bagged trees [40], [99] assemble multiple decision trees for more complex model architectures, therefore it is possible to achieve better performance. Random forests [35], [79], [88], [90], [93], [103] further improve bagged trees with less features to be used when splitting each node.…”