A Novel Approach for Classification of Normal/Abnormal Phonocardiogram Recordings using Temporal Signal Analysis and Machine Learning

Abstract

This paper discusses a novel approach used for classification of phonocardiogram (PCG) excerpts into normal and abnormal classes as a part of Physionet 2016 challenge [10]. The dataset used for the competition comprises of cardiac abnormalities such as mitral valve prolapse (MVP), benign murmurs, aortic diseases, coronary artery disease, miscellaneous pathological conditions etc.[3], We present the approach used for classification from a general machine learning application standpoint, giving details on feature extraction, type of classifiers used comparing their performances individually and in combination. We propose a technique which leverages previous research on feature extraction with a novel approach to modeling temporal dynamics of the signal using Markov chain analysis [7, 9]. These newly introduced Markov features along with other statistical and frequency domain features, trained over an ensemble of artificial neural networks and gradient boosting trees, with bagging, gave us an accuracy of 82% on the validation dataset provided in the competition and was consistent with the test data with the best result of 78%. IntroductionThis work describes a novel approach designed for Physionet 2016 Challenge Classification of Normal/Abnormal Heart Sound Recordings. The objective here is to classify Phonocardiogram (PCG) recordings into normal and abnormal categories. A comprehensive detail on the database, explaining how the PCG signals were collected and the type of abnormalities found are discussed in the paper mentioned in the reference section [3].Usually, statistical features such as means, standard deviations of systole, diastole intervals, and signal complexity features are used as features for classification and these are enough to give decent results. But these features fail to completely capture the temporal information of the signal. This could be very important since it represents how each heart beat changes over time.To capture the temporal dynamics, we take PCG signal beat by beat and assign each beat a symbol/category based on different thresholds set on features (ratios of systole intervals to RR interval, diastole intervals to RR interval, beat energy, the power of frequency component above 200 Hz). Thus, a sequence of symbols for the entire signal is obtained. We then extract features out of this sequence for classification. One of the ways we employed is to create a Markov chain with symbols being the states of the matrix and the resulting transition probabilities are used as features. These features along with marginal probabilities of states and rest of the acoustic features like sample entropy, instantaneous frequency analysis etc. are used to train an ensemble/bag of 4 class boosted tree classifiers [8, 13] and 4 artificial neural networks [12]. PreprocessingThe training data consists of PCG signals of varying length, anywhere between 5s to just over 120s all sampled at 2000 Hz. For training, all the signals were re-sampled to 1000 Hz and features were extracted. Since PCG record...

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

“…The suggested framework achieved 63-89% accuracy, indicating suggestive promising outcomes in comparison with other techniques attempting the challenge but on lower number of samples. It is a preliminary first attempt to utilize ANFIS on 1837 samples in contrast to other investigations, in which researchers utilized all HS samples (3126), or utilized more features and other sophisticated classifiers [37].…”

“…less than 1000 samples) such as references [8,30,33,38], as seen in Table V. However, some techniques in Table V employed more than 3000 HS signals [14,31,[34][35][36][37]; they had attempted some of these difficult signals, but they needed to explore high number of input parameters in the range of 13-124 different signal attributes. Thus, they reported better reliable performance than ANFIS, but this would not affect the suggestive capability of ANFIS to classify signals after training, particularly if the number of features was increased (e.g.…”

A Stacking-Based Model for Non-Invasive Detection of Coronary Heart Disease

“…2 mainly contains two loops. In the first loop, there are 10 kinds of possible combinations, including C 1 10 , C 2 10 , C 3 10 , C 4 10 , C 5 10 , C 6 10 , C 7 10 , C 8 10 , C 9 10 , C 10 10 as the input of the second loop. All possible combinations (without repeating them) are enumerated rather than putting them all into the next loop.…”

“…Davari et al extracted features from ECGs by frequency and nonlinear domain methods to identify CHD symptoms with support vector classifier (SVC) classifier [6]. Vernekar et al extracted Markov features along with other statistical and frequency domain features from phonocardiogram (PCG) and used the set of artificial neural network and gradient enhancement tree for model training [7]. Kumar et al also used ECG signals but with flexible analytic wavelet transform to characterize the CHD [8].…”

Two-stage Artificial Intelligence Clinical Decision Support System for Cardiovascular Assessment using Convolutional Neural Networks and Decision Trees

“…AI-assisted methods applied to biological signals can be categorized into four main categories: neural network-based and deep learning classification (Zabihi, Rad, Kiranyaz, Gabbouj, & Katsaggelos, 2016), support vector machine-based classification (Barhatte, Ghongade, & Thakare, 2015), hidden Markov-model-based classification (Vernekar, Nair, Vijaysenan, & Ranjan, 2016), and clustering-based classification (Clifford, 2016). Advanced deep learning techniques can automatically extract salient patterns directly from the input (i.e., training set) and use the produced knowledge to classify unseen samples (Shao, Wu, & Li, 2014).…”