Elsevier

Signal Processing

Volume 86, Issue 12, December 2006, Pages 3787-3795
Signal Processing

An iterative kurtosis-based technique for the detection of nonstationary bioacoustic signals

https://doi.org/10.1016/j.sigpro.2006.03.020Get rights and content

Abstract

A novel technique for the detection of nonstationary bioacoustic signals, such as explosive lung and bowel sounds, in clinical auscultative recordings is presented. The technique uses kurtosis (zero-lag fourth-order statistics) to form an iterative kurtosis-based detector (IKD). The IKD iteratively detects important peaks of the kurtosis, estimated within a sliding window along the signal under investigation, which indicate the presence of non-Gaussianity in the raw signal. The efficiency of the IKD has been illustrated by experimental results that demonstrate its ability to clearly detect bioacoustic signals of diagnostic interest in the presence of background signal with high amplitude.

Introduction

The aim of this paper is to propose a novel kurtosis-based technique for the detection of nonstationary bioacoustic signals, such as lung sounds and bowel sounds, in auscultation recordings. In general, we can assume that an auscultation recording is composed of a signal of diagnostic interest, which is associated with the relevant pathology, and a superimposed signal of no particular interest, which is considered as background signal (BGS). This assumption is valid especially in the cases of discontinuous adventitious lung sounds (DALS), i.e., crackles and squawks, and explosive bowel sounds (EBS). Their acoustic energy of DALS and EBS, generated by the breathing pattern and the motility of the bowel, respectively, is highly correlated with the corresponding pulmonary [1], [2], and bowel dysfunction [3], [4]. Hence, their detection provides a way of diagnosing the relevant pathology. However, the separation of DALS or EBS (signal of interest (SI)) from the BGS is not always an easy and practical task for the physician to perform, due to the nonstationary character of the SI and its variation in duration and amplitude. Previous research efforts, which address the DALS and EBS detection problem, include higher-order statistics-based autoregressive modeling, wavelet transform, and neuro-fuzzy modeling. A thorough review of these methods can be found in [5]. A fractal dimension-based detector has also been proposed to address the same problem [6].

An alternative method is introduced, based on the estimation of the kurtosis (zero-lag fourth-order statistics) [7] within a sliding window along the recordings under investigation. Higher-order statistics have also been applied recently to the detection of voice activity [8]. Since kurtosis is (theoretically) zero for Gaussian signals [7], like BGSs, significant deviations from this value can be attributed to the presence of non-Gaussian signals, such as the SIs. This deviation from zero value can be utilized in forming a criterion for identifying the presence of nonstationary transient signals. This kurtosis-based criterion has been adopted by the proposed iterative kurtosis-based detector (IKD) that gradually separates the SI from the BGS. It should be noted that SI refers to the portions, in the time domain, of the recorded signal that includes the DALS or EBS. Needless to say that these portions include also BS. Hence, the IKD segments the signal into two regions, one that consists of SI (DALS or EBS, plus BS) and one that consists of BS only. Unlike IKD, a detector based on variance changes fails in cases where Gaussian noise is present. The performance of the IKD scheme is evaluated through experimental results that show its ability to clearly detect bioacoustic signals of diagnostic interest, even when the amplitude of the BGS is high. The simplicity and the moderate computational complexity of the IKD scheme make it favorable for real-time implementation, such as continuous DALS or EBS screening.

The rest of the paper is organized as follows. Section 2 deals with the mathematical background related to the kurtosis definition, properties, and estimation. The proposed IKD technique is described in Section 3. Some experimental results, which demonstrate the IKD performance, are given in Section 4. Finally, Section 5 concludes the paper.

Section snippets

Mathematical background

Let {x(k)} be a real random zero-mean process that is fourth-order stationary. The second and fourth-order moments of {x(k)}, are defined as [7]:R2x(τ1)Ex(k)x(k+τ1),R4x(τ1,τ2,τ3)Ex(k)x(k+τ1)x(k+τ2)x(k+τ3),respectively, where E[·] denotes the expectation operator and τi (i=1,2,3) correspond to time lags. The fourth-order cumulants sequence C4x(τ1,τ2,τ3) of {x(k)} is defined as [7]:C4x(τ1,τ2,τ3)R4x(τ1,τ2,τ3)-3[R2x(τ1)]2.The fourth-order cumulant for zero lags, i.e., τ1=τ2=τ3=0, is the kurtosis

The IKD technique

We consider that the N-sample sound recording x={x(k):k=1,2,,N} can be written in the form ofx=s+b,where s and b stand for the SI and the BGS, respectively. To inspect the variations of the kurtosis along the given sound recording, x, an M-sample sliding window (MN) is defined, with one sample difference between successive windows. The latter is selected to obtain point-to-point values of the estimated kurtosis. Hence, the onset time and end of the SI can be accurately captured. The value of M

Implementation issues

The IKD was implemented on a PC (Pentium IV/2.4 GHz) using Matlab 6.1 (The Mathworks, Inc., Natick, MA) and tested on a subset of the real DALS and EBS signals used in [10], [11], respectively. The DALS signals have been recorded by means of a modified stethoscope (with an added microphone) [10], whereas the EBS have been acquired using an electronic stethoscope (Audioscope, Medivisio OY, Helsinki, Finland) [11]. The obtained signals were digitized with a 12-bit analog-to-digital converter at a

Conclusions

The IKD, an iterative kurtosis-based detector proposed in this paper, proved to be a very promising tool for the efficient detection of explosive lung and bowel sounds in pulmonary and abdominal auscultation recordings, respectively. Deviations from Gaussianity in the recordings are reflected in the variations of the corresponding kurtosis time series thus, the detection of the nonstationary SI becomes feasible through the iterative selection of those signal portions that correspond to high

Acknowledgments

The authors would like to thank all anonymous reviewers for their valuable comments. The financial support provided by the Research Committee of the Technological Educational Institute (T.E.I.) of Serres, Greece, is acknowledged.

References (11)

  • N. Gavriely

    Breath Sounds Methodology

    (1995)
  • L.J. Hadjileontiadis, S.M. Panas, Autoregressive modeling of lung sounds using higher-order statistics: estimation of...
  • W.B. Cannon

    Auscultation of the rhythmic sounds produced by the stomach and intestine

    Amer. J. Physiol.

    (1905)
  • S. Michael et al.

    Computerised phonoenterography: the clinical investigation of a new system

    J. Clin. Gastroenterol.

    (1994)
  • L.J. Hadjileontiadis et al.

    Intelligent system modeling of bioacoustic signals using advanced signal processing techniques

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

Cited by (36)

  • A nonparametric test for slowly-varying nonstationarities

    2018, Signal Processing
    Citation Excerpt :

    Some parametric methods that have been proposed in the past years assume the underlying signal can be modeled by a time-varying autoregressive process (TVAR) [3,4,9,10]. Among the nonparametric methods, the work in [11] has presented a technique for detecting changes in high-order statistics, whereas [12] has proposed a test for second-order stationarity in the TF domain. A common drawback of TF-based methods, however, is the computational load required to estimate full TF representations [13].

  • Recovery of gastrointestinal tract motility detection using Naive Bayesian and minimum statistics

    2014, Computers in Biology and Medicine
    Citation Excerpt :

    Various researchers have investigated automatic detection and classification of BSs. The most promising techniques used for short-term BS analysis are fractal dimension [6], higher order statistics [7], wavelet decomposition [8] and artificial neural network [9]. Studies performed by Dimoulas et.al [10], Kim et.al. [11],

  • Estimation algorithm of the bowel motility based on regression analysis of the jitter and shimmer of bowel sounds

    2011, Computer Methods and Programs in Biomedicine
    Citation Excerpt :

    Also, the conventional iterative stationary/nonstationary signal separation techniques (i.e. wavelet transform-based or fractal dimension-based filters) can provide erroneous BS detection and segmentation due to misinterpretation of regularly sustained (RS) signals (i.e. crepitating, musical/whistling BS) frequently presented during auscultation [16,23]. Recently, Rekanos and Hadjileontiadis [24] proposed an iterative kurtosis-based detector (IKD) algorithm for the detection of BS and demonstrated its effectiveness. However, unfortunately, in this algorithm, since the threshold used for the detection of the BS portions is adjusted by varying the power of the overall BGS signals rather than the kurtosis of the inherent BS signals at a particular time, significant numbers of relatively subtle BS segments could be missed or incorrectly detected.

  • BSMonitor: Noise-Resistant Bowel Sound Monitoring via Earphones

    2024, IEEE Transactions on Mobile Computing
View all citing articles on Scopus
View full text