Heart sound analysis for symptom detection and computer-aided diagnosis

https://doi.org/10.1016/j.simpat.2003.11.005Get rights and content

Abstract

Heart auscultation (the interpretation by a physician of heart sounds) is a fundamental component of cardiac diagnosis. It is, however, a difficult skill to acquire. In this work, we develop a simple model for the production of heart sounds, and demonstrate its utility in identifying features useful in diagnosis. We then present a prototype system intended to aid in heart sound analysis. Based on a wavelet decomposition of the sounds and a neural network-based classifier, heart sounds are associated with likely underlying pathologies. Preliminary results promise a system that is both accurate and robust, while remaining simple enough to be implemented at low cost.

Introduction

Heart auscultation (the monitoring of sounds produced by the heart and blood circulation) is a fundamental tool in the diagnosis of heart disease. While somewhat eclipsed in the research literature due to the advent of electrocardiographic (ECG) and echocardiographic methods, there are heart defects that are difficult to detect using ECG (e.g., structural abnormalities in natural or implanted heart valves, and defects characterized by heart murmurs and abnormal sounds). Furthermore, auscultation remains the primary tool for screening and diagnosis in primary health care, due in part to the higher cost and relatively limited availability of the equipment, and to the special skills necessary to administer and interpret the results of ECG and echocardiography. In some circumstances, particularly in remote areas or developing countries, auscultation may be the only means available.

However, forming a diagnosis based on sounds heard through either a conventional acoustic or an electronic stethoscope is itself a very special skill, one that can take years to acquire. Because this skill is also very difficult to teach in a structured way, the majority of internal medicine and cardiology programs offer little or no such instruction. Despite its obvious utility, primary care physicians are documented to have poor auscultatory skills [1], [2], [3]. It would be very useful if the benefits of auscultation could be obtained with a reduced learning curve, using equipment that is low-cost, robust, and easy to use.

The complex and highly nonstationary nature of heart sound signals can make them challenging to analyze in an automated way. However, recent technological developments have made extremely powerful digital signal-processing techniques both widely accessible and practical. Local frequency analysis and wavelet (local scale analysis) approaches are particularly applicable to problems of this type. Some of these methods have been applied to study the fundamental mechanisms underlying the production of sound by the heart, and the correlation between these sounds and various heart defects (e.g., [4], [5], [6], [7], [8], [9], [10], [11], [12], [13]). For an excellent survey and discussion of work in this area see [14].

Despite the clear success of digital signal-processing (DSP) based techniques, many heart sounds associated with defects remain subtle, and difficult to detect and discriminate from similar sounds with no underlying pathology.

Some of the same technological advances that have supported DSP oriented methods have also facilitated the development of increasingly sophisticated models of the heart, from relatively simple models focusing on sound formation such as [15], to full 3-D structural models.

The goal of this work is to combine local analysis methods, information provided via modeling, and classification techniques to detect, characterize and interpret sounds corresponding to symptoms and signs important for cardiac diagnosis. It is hoped that the results of this analysis may prove valuable in themselves as a diagnostic aid, and as input to machine diagnostic systems, as for example the Fallot computational model [16]. The modeling and analysis techniques described are general and could be applied to a broad range of signal-processing problems.

Section snippets

The structure of heart sounds

Heart sounds are complex and highly nonstationary signals. The “beats” associated with these sounds are reflected in the signal by periods of relatively high activity, alternating with comparatively long intervals of low activity.

The cardiac cycle consists of two periods, systole and diastole [17]. Four classes of sound components may be audible on heart auscultation: the primary components (S1, S2, S3, and S4) are short “beats”, the other classes of sounds are murmurs (longer duration),

A simple heart sound model

The heart–thorax acoustic system, like the heart sounds themselves, is extremely complex. An approach to modeling complex systems that has proven useful in a number of domains is to approximate them as linear systems, and to use the tools developed for the study of such systems. Usually, this approach includes the assumption that the system of interest is time-invariant. Clearly, this is not the case here.

Durand and Pibarot [14] have proposed a linear model with both time-varying and

Determination of the inputs and transfer functions

To determine the relative locations and amplitudes of the input impulses, we first observe that for stable, damped linear systems, the response to an impulse is maximum in magnitude at the time corresponding to the application of the impulse. When the input to the system is a sequence of impulses, there will be peaks in energy in the output (which may or may not coexist with amplitude extrema in the output) for each impulse.

Locating such peaks is an ideal application for time–frequency or

The characterization of abnormal heart sounds

We will consider two types of abnormal heart sounds in this section: those resulting from coarctation of the aorta, and those exhibiting so-called “splits.”

A system for heart sound classification

A block diagram representing a simple system for heart sound classification is shown in Fig. 11. Heart sounds (sampled at an 8 kHz sample rate, 16 bits/sample) are first hand segmented into 4096 sample segments, each consisting of a single heartbeat cycle. Each segment is transformed using a seven level wavelet decomposition, based on a Coifman fourth order wavelet kernel. The resulting transform vectors, 4096 values in length, are reduced to 256 element feature vectors by discarding the four

Results and discussion

The system was evaluated using heart sounds corresponding to five different heart conditions: normal, mitral valve prolapse (MVP), coarctation of the aorta (CA), ventricular septal defect (VSD), and pulmonary stenosis (PS). CA often leads to an increased A2 sound, while VSD often produces an increased P2. In PS, the P2 is soft and may not be audible, making S2 appear single. A click may also be present.

The classifier was trained using 10 shifted versions (over a range of 100 samples) of a

Conclusions and future work

In this paper, we have introduced a model for the generation of heart sounds, and demonstrated its usefulness as a source of relevant features for cardiac diagnosis. Establishing a correlation between different pathologies and specific features in the transfer functions, and the evaluation of the utility of this approach as part of a complete computer-based diagnostic aid (e.g., in conjunction with the system described in [16]) are areas of future work.

We have also presented a preliminary study

Acknowledgments

This work was supported in part by the Programming Environments Laboratory, Department of Computer and Information Science, Linköping University, Sweden, and by the United States National Science Foundation under grant # 9870454. The authors wish to thank Dr. D. Roy and Dr. C.B. Mahnke for their comments.

References (18)

  • R.L. Donnerstein et al.

    Hemodynamic and anatomic factors affecting the frequency content of Still 's innocent murmur

    The American Journal of Cardiology

    (1994)
  • N.E. Reed et al.

    Diagnosing congenital heart defects using the Fallot computational model

    Artificial Intelligence in Medicine

    (1997)
  • D. Roy et al.

    Helping family physicians improve their cardiac auscultation skills with an interactive cd-rom

    Journal of Continuing Education in the Health Professions

    (2002)
  • D.L. Roy

    The paediatrician and cardiac auscultation

    Paediatric Child Health

    (2003)
  • C.B. Mahnke, A. Norwalk, D. Hofkosh, J. Zuberbuhler, Y.M. Law, Comparison of two educational interventions on pediatric...
  • D. Barschdor, U. Femmer, E. Trowitzsch, Automatic phonocardiogram signal analysis in infants based on wavelet...
  • B. El-Asir, L. Khadra, A. Al-Abbasi, M. Mohammed, Time–frequency analysis of heart sounds, in: Proc. of the 1996 IEEE...
  • B. El-Asir, L. Khadra, A. Al-Abbasi, M. Mohammed, Multiresolution analysis of heart sounds, in: Proc. of the Third IEEE...
  • J. Ritola et al.

    Comparison of time–frequency distributions in the heart sound analysis

    Medical and Biological Engineering and Computing

    (1996)
There are more references available in the full text version of this article.

Cited by (152)

  • Input reduction of convolutional neural networks with global sensitivity analysis as a data-centric approach

    2022, Neurocomputing
    Citation Excerpt :

    AI has recently been used to develop and improve upon many previously human-only jobs. Among these are high frequency trading, portfolio management and prediction [1], military and health analysis [2], law analysis [3] and vast numbers of other occupations embracing a broad spectrum of disciplines. Convolutional Neural Networks (CNNs), the outcome of the rapid development of deep learning techniques, are variants of feed-forward neural networks.

  • Automated detection of heart valve disorders with time-frequency and deep features on PCG signals

    2022, Biomedical Signal Processing and Control
    Citation Excerpt :

    The S3 and S4 components indicate murmur sounds, and the PCG signal carrying the murmurs is categorized as an abnormal heart sound. The murmur is called systolic, diastolic and continuous according to its position in the heart cycle [6,7]. Various HVDs such as aortic stenosis (AS), mitral regurgitation (MR), mitral stenosis (MS), and mitral valve prolapse (MVP) can be diagnosed and detected using PCG signals [8].

  • Recent Advances in PCG Signal Analysis using AI: A Review

    2024, International Journal on Smart Sensing and Intelligent Systems
  • A Pre-Screening Technique for Coronary Artery Disease with Multi-Channel Phonocardiography and Electrocardiography

    2024, Non-Invasive Health Systems based on Advanced Biomedical Signal and Image Processing
View all citing articles on Scopus
View full text