Artificial neural networks in computer-assisted classification of heart sounds in patients with porcine bioprosthetic valves

Development of artificial neural network-based algorithms for the classification of bileaflet mechanical heart valve sounds

OBJECTIVES

As is true for all mechanical prostheses, bileaflet heart valves are prone to thrombus formation; reduced hemodynamic performance and embolic events can occur as a result. Prosthetic valve thrombosis affects the power spectra calculated from the phonocardiographic signals corresponding to prosthetic closing events. Artificial neural network-based classifiers are proposed for automatically and noninvasively assessing valve functionality and detecting thrombotic formations. Further studies will be directed toward an enlarging data set, extending the investigated frequency range, and applying the presented approach to other bileaflet mechanical valves.

METHODS

Data were acquired for the normofunctioning St. Jude Regent valve mounted in the aortic position of a Sheffield Pulse Duplicator. Different pulsatile flow conditions were reproduced, changing heart rate and stroke volume. The case of a thrombus completely blocking 1 leaflet was also investigated. Power spectra were calculated from the phonocardiographic signals and used to train artificial neural networks of different topologies; neural networks were then tested with the spectra acquired in vivo from 33 patients, all recipients of the St. Jude Regent valve in the aortic position.

RESULTS

The proposed classifier showed 100% correct classification in vitro and 97% when applied to in vivo data: 31 spectra were assigned to the right class, 1 received a false positive classification, and 1 was "not classifiable."

CONCLUSION

Early malfunction detection is necessary to prevent thrombotic events in bileaflet mechanical heart valves. Following further clinical validation with an extended patient database, artificial neural network-based classifiers could be embedded in a portable device able to detect valvular thrombosis at early stages of formation: this would help clinicians make valvular dysfunction diagnoses before the appearance of critical symptoms.

A color spectrographic phonocardiography (CSP) applied to the detection and characterization of heart murmurs: preliminary results

Background

Although cardiac auscultation remains important to detect abnormal sounds and murmurs indicative of cardiac pathology, the application of electronic methods remains seldom used in everyday clinical practice. In this report we provide preliminary data showing how the phonocardiogram can be analyzed using color spectrographic techniques and discuss how such information may be of future value for noninvasive cardiac monitoring.

Methods

We digitally recorded the phonocardiogram using a high-speed USB interface and the program Gold Wave http://www.goldwave.com in 55 infants and adults with cardiac structural disease as well as from normal individuals and individuals with innocent murmurs. Color spectrographic analysis of the signal was performed using Spectrogram (Version 16) as a well as custom MATLAB code.

Results

Our preliminary data is presented as a series of seven cases.

Conclusions

We expect the application of spectrographic techniques to phonocardiography to grow substantially as ongoing research demonstrates its utility in various clinical settings. Our evaluation of a simple, low-cost phonocardiographic recording and analysis system to assist in determining the characteristic features of heart murmurs shows promise in helping distinguish innocent systolic murmurs from pathological murmurs in children and is expected to useful in other clinical settings as well.

Phonocardiographic signal analysis method using a modified hidden Markov model

Auscultation is an important diagnostic indicator for cardiovascular analysis. Heart sound classification and analysis play an important role in the auscultative diagnosis. This study uses a combination of Mel-frequency cepstral coefficient (MFCC) and hidden Markov model (HMM) to efficiently extract the features for pre-processed heart sound cycles for the purpose of classification. A system was developed for the interpretation of heart sounds acquired by phonocardiography using pattern recognition. The task of feature extraction was performed using three methods: time-domain feature, short-time Fourier transforms (STFT) and MFCC. The performances of these feature extraction methods were then compared. The results demonstrated that the proposed method using MFCC yielded improved interpretative information. Following the feature extraction, an automatic classification process was performed using HMM. Satisfactory classification results (sensitivity > or =0.952; specificity > or =0.953) were achieved for normal subjects and those with various murmur characteristics. These results were based on 1398 datasets obtained from 41 recruited subjects and downloaded from a public domain. Constituents characteristics of heart sounds were also evaluated using the proposed system. The findings herein suggest that the described system may have the potential to be used to assist doctors for a more objective diagnosis.

Diagnosing aortic valve stenosis by parameter extraction of heart sound signals

The objective of this study was to develop an automatic signal analysis system for heart sound diagnosis. This should support the general practitioner in discovering aortic valve stenoses at an early stage to avoid or decrease the number of surgical interventions. The applied analysis method is based on classification of heart sound signals utilising parameter extraction. From the wavelet decomposition of a representative heart cycle as well as from the Short Time Fourier Transform (STFT) and the Wavelet Transform (WT) spectra new time series were derived. In several segments, parameters were extracted and analysed. In addition, features of the Fast Fourier Transform (FFT) of the raw signal were examined. In this study, 206 patients were enrolled, 159 with no heart valve disease or any other heart valve disease but aortic valve stenosis and 47 suffering from aortic valve stenosis in a mild, moderate or severe stage. To separate the groups, a linear discriminant function analysis was applied leading to a reduced parameter set. The introduced two classification stage (CS) system for automatic detection of aortic valve stenoses achieves a high sensitivity of 100% for moderate and severe aortic valve stenosis and a sensitivity of 75% for mild aortic valve stenosis. A specificity of 93.7% for patients without aortic valve stenosis is provided. The developed method is robust, cost effective and easy to use, and could, therefore, be a suitable method to diagnose aortic valve stenosis by general practitioners.

A classifier based on the artificial neural network approach for cardiologic auscultation in pediatrics

OBJECTIVE

This research work was aimed at developing a reliable screening device for diagnosis of heart murmurs in pediatrics. This is a significant problem in pediatric cardiology because of the high rate of incidence of heart murmurs in this population (reportedly 77-95%), of which only a small fraction arises from congenital heart disease. The screening devices currently available (e.g. chest X-ray, electrocardiogram, etc.) suffer from poor sensitivity and specificity in detecting congenital heart disease. Thus, patients with heart murmurs today are frequently assessed by consultation as well with advanced imaging techniques. The most prominent among these is echocardiography. However, echocardiography is expensive and is usually only available in healthcare centers in major cities. Thus, for patients being evaluated with a heart murmur, developing a more accurate screening device is vital to efforts in reducing health care costs.

METHODS AND MATERIAL

The data set was collected from incoming pediatrics at the cardiology clinic of The Children's Hospital (Denver, Colorado), on whom echocardiography had been performed to identify congenital heart disease. Recordings of approximately 10-15s duration were made at 44,100Hz and the average record length was approximately 60,000 points. The best three cycles with respect to signal quality sounds were extracted from the original recording. The resulting data comprised 241 examples, of which 88 were examples of innocent murmurs and 153 were examples of pathological murmurs. The selected phonocardiograms were subject to the digital signal processing (DSP) technique of fast Fourier transform (FFT) to extract the energy spectrum in frequency domain. The spectral range was 0-300Hz at a resolution of 1Hz. The processed signals were used to develop statistical classifiers and a classifier based on our in-house artificial neural network (ANN) software. For the latter, we also tried enhancements to the basic ANN scheme. These included a method for setting the decision-threshold and a scheme for consensus-based decision by a committee of experts.

RESULTS

Of the different classifiers tested, the ANN-based classifier performed the best. With this classifier, we were able to achieve classification accuracy of 83% sensitivity and 90% specificity in discriminating between innocent and pathological heart murmurs. For the problem of discrimination between innocent murmurs and murmurs of the ventricular septal defect (VSD), the accuracy was higher, with sensitivity of 90% and specificity of 93%.

CONCLUSIONS

An ANN-based approach for detection and identification of congenital heart disease in pediatrics from heart murmurs can result in an accurate screening device. Considering that only a simple feature set was used for classification, the results are very encouraging and point out the need for further development using improved feature set with more potent diagnostic variables.

A virtual instrument for acquisition and analysis of the phonocardiogram and its internet-based application

The objective of this study is to develop a phonocardiogram (PCG) acquisition and analysis instrument using virtual instrumentation technology and investigate its Internet-based application. The PCG instrument was developed using a Pentium 200 computer, a data acquisition board, and a two-channel custom designed bio-signal preamplifier. LabVIEW was used to create the instrument's front panels. Spectral and joint time-frequency analyses were implemented into the instrument. This instrument can be used to display the PCG and to analyze the individual heart sound and murmur for the detection of heart valve diseases. Using a test-bed, the PCG data acquisition and analysis were performed remotely over the Internet. Through the main PCG panel, an operator can control the acquisition and analysis of PCG signals. In the remote test, real-time transmission of the PCG signal over the Internet was possible. Remote operators were able to view smoothly scrolling PCG waveforms and could control all the acquisition parameters and perform spectral and time-frequency analyses on the acquired heart sound. This study demonstrated that a LabVIEW-based medical virtual instrument provides a low-cost and flexible solution for data acquisition and analysis of PCG. It also showed that the current Internet supports the transmission of real-time PCG signals. Compared with other telemedicine systems, this application transfers not only the medical data, but also the virtual instrument and its signal processing capability through the Internet.

Artificial neural network-based method of screening heart murmurs in children

BACKGROUND

Early recognition of heart disease is an important goal in pediatrics. Efforts in developing an inexpensive screening device that can assist in the differentiation between innocent and pathological heart murmurs have met with limited success. Artificial neural networks (ANNs) are valuable tools used in complex pattern recognition and classification tasks. The aim of the present study was to train an ANN to distinguish between innocent and pathological murmurs effectively.

METHODS AND RESULTS

Using an electronic stethoscope, heart sounds were recorded from 69 patients (37 pathological and 32 innocent murmurs). Sound samples were processed using digital signal analysis and fed into a custom ANN. With optimal settings, sensitivities and specificities of 100% were obtained on the data collected with the ANN classification system developed. For future unknowns, our results suggest the generalization would improve with better representation of all classes in the training data.

CONCLUSION

We demonstrated that ANNs show significant potential in their use as an accurate diagnostic tool for the classification of heart sound data into innocent and pathological classes. This technology offers great promise for the development of a device for high-volume screening of children for heart disease.

Evaluation of Karhunen-Loève expansion for feature selection in computer-assisted classification of bioprosthetic heart-valve status

This paper analyses the performance of four different feature-selection approaches of the Karhunen-Loève expansion (KLE) method to select the most discriminant set of features for computer-assisted classification of bioprosthetic heart-valve status. First, an evaluation test reducing the number of initial features while maintaining the performance of the original classifier is developed. Secondly, the effectiveness of the classification in a simulated practical situation where a new sample has to be classified is estimated with a validation test. Results from both tests applied to a reference database show that the most efficient feature selection and classification (> or = 97% of correct classifications (CCs)) are performed by the Kittler and Young approach. For the clinical databases, this approach provides poor classification results for simulated 'new samples' (between 50 and 69% of CCs). For both the evaluation and the validation tests, only the Heydorn and Tou approach provides classification results comparable with those of the original classifier (a difference always < or = 7%). However, the degree of feature reduction is particularly variable. The study demonstrates that the KLE feature-selection approaches are highly population-dependent. It also shows that the validation method proposed is advantageous in clinical applications where the data collection is difficult to perform.

Automated neural network detection of wavelet preprocessed electrocardiogram late potentials

The aim of the study is to investigate the potential of a feedforward neural network for detecting wavelet preprocessed late potentials. The terminal parts of a simulated QRS complex are processed with a continuous wavelet transform, which leads to a time-frequency representation of the QRS complex. Then, diagnostic feature vectors are obtained by subdividing the representations into several regions and by processing the sum of the decomposition coefficients belonging to each region. The neural network is trained with these feature vectors. Simulated ECGs with varying signal-to-noise ratios are used to train and test the classifier. Results show that correct classification ranges from 79% (high-level noise) to 99% (no noise). The study shows the potential of neural networks for the classification of late potentials that have been preprocessed by a wavelet transform. However, clinical use of this method still requires further investigation.

Use of an artificial neural network to analyse an ECG with QS complex in V1-2 leads

A feed-forward neural network with back-propagation algorithm is used to distinguish anterior wall myocardial infarction (AI) and non-infarction based on analysis of computerised electrocardiograms. Data used in the study are from 132 patients diagnosed as having AI by automated electrocardiograph analysis. Their ECGs show an abnormal Q-wave (or QS complex) or small R progression in leads V1 and V2. However, 66 of them are diagnosed as old AI from the history, physical examination, echocardiogram and other laboratory data, whereas the other 66 are not. The network is trained with the data from half of the AI and non-infarction patients; respectively. The diagnostic accuracy rate is then tested with the remaining 66 patients (33 infarction, 33 non-infarction) who have not been exposed to the network. The neural network correctly identifies 90.2% of the patients with AI and 93.3% of the patients without infarction. The neural network is capable of diagnosing anterior wall myocardial infarction better than a computer electrocardiograph.

Spectral characterization and classification of Carpentier-Edwards heart valves implanted in the aortic position

This paper demonstrates an improvement in the performance of spectral phonocardiography, combined with pattern recognition techniques for monitoring the condition of bioprosthetic heart valves. The analysis of the heart sounds is performed using a modified forward-backward overdetermined Prony's method. Results show that the condition of the bioprosthesis affects mostly the higher part of the spectrum (i.e., above 250 Hz) where no frequency components were found for malfunctioning cases. Therefore, the amplitudes of the three highest frequency components are used as the input vector of an adaptive single layer perceptron-based classifier to identify normal and malfunctioning classes. For the sample set examined, this method gives 100% correct discrimination between normal and malfunctioning Carpentier-Edwards (C-E) valves.

Modelling techniques and their application for monitoring in high dependency environments--learning models

This paper reviews the use of learning models including Bayesian classifiers and artificial neural networks in monitoring and interpreting biosignals. Generally learning models applied for analysis of biosignals are "black-box' types trained on the basis of measured signals. It is illustrated that the training and application of learning models more or less follow the same sequences. The main focus is the interpretation of electrical signals from the brain (electroencephalogram (EEG) and evoked potentials (EP)). Current analysis of these signals often reveals sudden changes in the EEG or evoked potentials to be the earliest discernible signs of inadequate perfusion of the brain. They may reflect problems such as systemic arterial oxygen desaturation or hypotension arising from other body system failures during critical illness. It is suggested that these brain signals should be recorded in the critical care unit, and that they should form part of the annotated database of biosignals established during the IMPROVE project. This would allow for the development of new methods for on-line warning of impending damage to the central nervous system, such that corrective actions could be taken before permanent damage occurred.