A wavelet-based reduction of heart sound noise from lung sounds

doi:10.1016/S1386-5056(98)00137-3

International Journal of Medical Informatics

Volume 52, Issues 1–3, 1 October 1998, Pages 183-190

https://doi.org/10.1016/S1386-5056(98)00137-3 Get rights and content

Abstract

Heart sounds produce an incessant noise during lung sounds recordings. This noise severely contaminates the breath sounds signal and interferes in the analysis of lung sounds. In this paper, the use of a wavelet transform domain filtering technique as an adaptive de-noising tool, implemented in lung sounds analysis, is presented. The multiresolution representations of the signal, produced by wavelet transform, are used for signal structure extraction. In addition, the use of hard thresholding in the wavelet transform domain results in a separation of the nonstationary part of the input signal (heart sounds) from the stationary one (lung sounds). Thus, the location of the heart sound noise (1st and 2nd heart sound peaks) is automatically detected, without requiring any noise reference signal. Experimental results have shown that the implementation of this wavelet-based filter in lung sound analysis results in an efficient reduction of the superimposed heart sound noise, producing an almost noise-free output signal. Due to its simplicity and its fast implementation the method can easily be used in clinical medicine.

Introduction

Auscultation is the simplest diagnostic tool for pulmonary dysfunction. Although lung sounds are easily heard through a stethoscope, the incessant heart beating produces an intrusive, quasi-periodic, interference which influences the clinical auscultative interpretation of lung sounds. The introduction of pseudo-periodicities, the masking of the relevant signal and the modification of the energy distribution in the spectrum of lung sounds, due to heart sounds [1], require an effective reduction of heart sounds from the contaminated lung sound signal, to yield a successful lung sounds classification. High pass linear filtering (HPF) and adaptive filtering (AF) are the two basic approaches for heart noise reduction. Although HPF (with a cut-off frequency varying from 50 to 150 Hz) is effective in heart sound reduction 2, 3, degrades the respectively overlapped frequency region of breath sounds and fails to track the changing signal characteristics. AF technique overcomes these limitations, since it is based on a gradual reduction of the mean square error between the primary input signal (contaminated lung sounds) and a recorded or artificially produced reference signal, highly correlated to the noise component of the input signal (heart sounds) 4, 1, 5.

In this paper, a new type of adaptive filter for de-noising the contaminated lung sounds based on wavelet transform (WT) is presented. The WT sets a new perspective in lung sounds analysis, since it decomposes them into multiscale details, describing their power at each scale and position 6, 7. Applying a threshold-based criterion at each scale a filtering scheme can be composed, which weights WT coefficients according to signal structure. An adaptive separation of signal from ‘noise’, without requiring any reference signal, can be achieved through an iterative reconstruction–decomposition process, based on the derived weighted WT coefficients at each iteration.

Section snippets

Wavelet analysis

Wavelets are families of functions ψ_a,b(t) generated from a single base wavelet ψ(t) called the ‘mother wavelet’, by dilations and translations [6], i.e., $ψ_{a,b} t = 1 a ψ t−b a, a>0, b∈ †,$ where a is the dilation (scale) parameter and b is the translation parameter. The continuous wavelet transform of a 1-D function f(t)∈L²( $R$ ), where L²( $R$ ) denotes the vector space of measurable, square-integrable one-dimensional functions f(t), is defined in a Hilbert space, as the projection of the function onto the

Filtering algorithm

The proposed algorithm, initially proposed by Hadjileontiadis and Panas 9, 10, 11, is a wavelet domain filtering technique, based on the fact that explosive peaks in time domain (heart sound peaks) have large signal over many wavelet scales, while ‘noisy’ background (lung sounds) dies out swiftly with increasing scale. The definition of ‘noise’ is not always clear. Instead, it is better to view an N-sample signal as being noisy or incoherent relative to a basis of waveforms if it does not

Experiments and implementation

The study was conducted on four healthy volunteers, aged 23–50 years, with no known pulmonary or cardiac disorder. Several recordings took place on each subject from appropriate locations where heart sounds could be heard with the highest intensity [13], using a modified Littmann stethoscope. The whole analysis was implemented on an IBM-PC (Pentium/166 MHz) using the programming language ASYST 4.1 (Keithley, Taunton, MA 02780). After initialising filtering, a 12-Bit analogue-to-digital (A/D)

Results and discussion

Due to nonavailability of pure lung sounds, a qualitative rather than a quantitative procedure must be employed to evaluate the performance of WTST–NST algorithm. A rough estimation of noise reduction can be achieved, if it is assumed that non-breathing noises do not change significantly when breathing normally or when holding one’s breath [5]. With this assumption, close enough to reality, it can be possible to estimate efficiently the location and the shape of heart sound noise, especially

Conclusions

Summarising, a new adaptive noise reduction scheme (WTST–NST filter), which combines the efficiency of multiresolution analysis with hard thresholding, implemented in heart sound noise reduction of lung sounds, has been presented. Experiments have shown that the WTST–NST filter has resulted, in all cases, to a generally high de-noised signal quality without requiring any reference signal, with low computational cost and fast and easy implementation.

References (14)

V.K. Iyer et al.
Reduction of heart sounds from lung sounds by adaptive filtering
IEEE Trans. Biomed. Eng.
(1986)
N. Gavriely et al.
Spectral characteristics of normal breath sounds
J. Appl. Physiol.
(1981)
G. Charbonneau et al.
An accurate recording system and its use in breath sounds spectral analysis
J. Appl. Physiol.
(1983)
B. Widrow et al.
Adaptive noise cancelling: principles and applications
Proc. IEEE
(1975)
M. Kompis, E. Russi, Adaptive heart-noise reduction of lung sounds recorded by a single microphone, IEEE Eng. Med....
I. Daubechies
Orthonormal bases of compactly supported waveless
Commun. Pure Appl. Math.
(1988)
G. Mallat
A theory for multiresolution signal decomposition: The wavelet representation
IEEE Trans. Patt. Anal. Machine Intell.
(1989)

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

Cited by (74)

Pre-processing techniques to enhance the classification of lung sounds based on deep learning
2024, Biomedical Signal Processing and Control
Deep learning has recently proved a huge potential in the classification of lung sounds. Most studies rely on publicly available data sets that are usually well-cleaned and annotated by expert physicians. The result of annotation is subjective by definition and, above all, large and public data sets are not collected in the scope of a very specific clinical investigation. Other works rely on private and suitably collected data sets that either may or may not stem from clinical studies. The main issue in these cases is represented by the reliability and noisiness of auscultations.
This paper delves into the significant impact of quantitative, systematic and reproducible cleaning of data sets of lung sounds. For “cleaning a data set” we mean discarding the records that carry mostly noise and interfering signals, since machine learning can be significantly impaired by outliers.
The developed pre-processing techniques are tested on several data sets of lung sounds. We designed a deep neural network (DNN) for the diagnosis of interstitial lung diseases (ILD) in patients affected by connective tissue diseases (CTD). The devised DNN can provide significant performance on the clean data set with impressive accuracy, F1-score, and F2-score of 97% with respect to the high-resolution computer tomography. Considering that the screening of ILD in patients affected by chronic autoimmune diseases is still an open issue, the proposed pipeline represents the enabling technology for the early, safe, reliable and cheap diagnosis of CTD-ILD.
An efficient new adaptive variational mode decomposition algorithm for extracting adventitious lung sounds
2024, Biomedical Signal Processing and Control
The rapid and accurate extraction of adventitious lung sounds (ALSs) from original lung sounds is greatly significant for diagnosing lung diseases and assessing lung condition. Due to interference from heart sounds and various noises, the empirical mode decomposition (EMD), variational mode decomposition (VMD) and some other classical algorithms have poor performance. Here, an efficient new adaptive variational mode decomposition (AVMD) algorithm is proposed to improve the efficiency for extracting ALSs. For minimizing the bandwidth of each extracted mode and reducing frequency aliasing between modes, the proposed algorithm constructs a frequency-constrained model based on VMD. In order to further improve its performance, the algorithm adopts a strategy of gradually extracting each mode to determine the optimal number of modes for avoiding the under-decomposition or over-decomposition, and uses Pool Adjacent Violators (PAV) to automatically initialize the center frequency of each mode to reduce algorithm iteration steps. Based on the above improved characteristics, the experiments presented in this paper, which conducted on the artificial simulation signal and 102 groups of lung sounds from 20 patients with severe pulmonary diseases, highlight the proposed AVMD's superiority over the other five algorithms when comprehensively considering both decomposition quality and computational efficiency. Furthermore, the above results of 102 groups of lung sounds are statistically analyzed to obtain the characteristic frequency distribution of ALSs, which can provide important references for clinicians and demonstrate the potential application of the proposed algorithm in the diagnosis of pulmonary diseases.
Extracting cervical spine popping sound during neck movement and analyzing its frequency using wavelet transform
2022, Computers in Biology and Medicine
Citation Excerpt :
In this study, a denoising algorithm denoted Wavelet Transform-Based Stationary-Nonstationary (WTST-NST) is adopted to remove the noise that originates from the cervical spine. The algorithm doesn't require a reference signal a does an adaptive filter, and it has been used to differentiate heart sounds from lung sounds [18]. In particular, according to the WTST-NST filter algorithm, noise is removed where it is present, preserving the original signal unchanged as much as possible.
With the excessive use of smartphones, cervical spine pain is becoming increasingly prevalent. A denoised cervical spine popping sound can aid in monitoring and estimating the state of the cervical spine. However, cervical spine popping sounds that are collected when a subject performs neck movements is contaminated by constant noise. Therefore, a denoising algorithm called Wavelet Transform-Based Stationary-Nonstationary (WTST-NST) is adopted to remove the noise. The input signal is decomposed using wavelet transform to obtain the wavelet coefficients. The wavelet coefficients are then separated into two parts, the nonstationary part and the stationary part, using stationary-nonstationary filtering technology. Finally, the wavelet coefficients of the nonstationary part are reconstructed to obtain the denoised cervical spine popping sound. In addition, the frequency components of the sound are analyzed using the multiresolution analysis of the wavelet transform. The experimental results demonstrate that the implementation of the WTST-NST algorithm in the sound analysis of cervical spine facet joints efficiently reduces the overlapped noise, producing an almost pure cervical spine popping sound. Furthermore, the frequency components of cervical spine popping sounds during the smartphone use period are significantly higher than that in the non-use period and are significantly associated with self-reported neck and upper back pain during the smartphone use period. Therefore, the WTST-NST algorithm preserved almost all the features of the sampled input signal. The denoised cervical spine popping sound can be used to quickly and conveniently monitor the status of the cervical spine during the smartphone use period.
Lung sound signal denoising using discrete wavelet transform and artificial neural network
2022, Biomedical Signal Processing and Control
Citation Excerpt :
The line graph also shows that the rate of SNR improvement at lower SNRs has been better, and as the input SNR increases, this improvement has gradually diminished. The concept of the discrete wavelet transform denoising method, assuming that the non-static wavelet coefficients are much lower than the static part, was first proposed by Hadjileontiadis [25,26], and Panas [27]. In the following, various methods for selecting the optimal threshold value are presented.
Computerized analysis of Lung Sound (LS) is a promising method for assessing pulmonary function. However, the LS signal is severely contaminated by background noise from various sources. Conventional denoising methods may not be practical due to the noisy nature of the LS as well as its spectral overlap with different noise sources. This paper proposes an adaptive technique based on Discrete Wavelet Transform and Artificial Neural Network (DWT-ANN) to filtrate LS signals in a noisy environment. This new method mixes the multi-resolution property of DWT with ANN as a nonlinear adaptive filter. In this research, separate models for signal denoising with different SNRs (0, 5, 10, and 15 dB) were designed. Then a single model was introduced as a combined model to eliminate any information about input signals SNR before the denoising process. The results showed that the combined model, in addition to having a close performance to the individual models, could perform the denoising process well in the range of −2 to 20 dB, which is outside the range that the model has been trained with. In addition, comparing the results of our proposed method with the DWT method, it was observed that the SNR of the denoised signal was significantly enhanced. At SNR = 0 dB, this improvement is 9.18 compared to only 3.90 using the DWT method.
Recent Advances in PCG Signal Analysis using AI: A Review
2024, International Journal on Smart Sensing and Intelligent Systems
Noisy Neonatal Chest Sound Separation for High-Quality Heart and Lung Sounds
2023, IEEE Journal of Biomedical and Health Informatics

View all citing articles on Scopus

View full text

A wavelet-based reduction of heart sound noise from lung sounds

Abstract

Introduction

Section snippets

Wavelet analysis

Filtering algorithm

Experiments and implementation

Results and discussion

Conclusions

Reduction of heart sounds from lung sounds by adaptive filtering

IEEE Trans. Biomed. Eng.

Spectral characteristics of normal breath sounds

J. Appl. Physiol.

An accurate recording system and its use in breath sounds spectral analysis

J. Appl. Physiol.

Adaptive noise cancelling: principles and applications

Proc. IEEE

Orthonormal bases of compactly supported waveless

Commun. Pure Appl. Math.

A theory for multiresolution signal decomposition: The wavelet representation

IEEE Trans. Patt. Anal. Machine Intell.