Advanced processing of lung sound (LS) recording is a significant means to separate heart sounds (HS) and combined low frequency noise from instruments (NI), with saving its characteristics. This paper proposes a new method of LS filtering which separates HS and NI simultaneously. It focuses on the application of least mean squares (LMS) algorithm with adaptive noise cancelling (ANC) technique. The second step of the new method is to modulate the reference input r1(n) of LMS-ANC to acquiesce combining HS and NI signals. The obtained signal is removed from primary signal (original lung sound recording-LS). The original signal is recorded from subjects and derived HS from it and it is modified by a band pass filter. NI is simulated by generating approximately periodic white gaussian noise (WGN) signal. The LMS-ANC designed algorithm is controlled in order to determine the optimum values of the order L and the coefficient convergence μ. The output results are measured using power special density (PSD), which has shown the effectiveness of our suggested method. The result also has shown visual difference PSD (to) normal and abnormal LS recording. The results show that the method is a good technique for heart sound and noise reduction from lung sounds recordings simultaneously with saving LS characteristics.
Biological signals are often noisy and non stationary. These factors tremendously complicate analysis of biosignals [
Auscultation is one of the most important non-invasive and simple diagnostic tools for detecting disorders in the respiratory tract like lung diseases. However, despite their effectiveness, these instruments only provide a limited and subjective perception of the respiratory sounds. The drawbacks of using stethoscopes and listening to the sounds are using the human ear area, and their inability to provide an objective study of the detected respiratory sounds. They lack sufficient sensitivity and the existence of imperfect system of nomenclature.
Besides the fact that normal and abnormal lung sounds are mixed in the airways and therefore pose a problem of classification of respiratory diseases, semi-periodic HS from heartbeat activity invariably interferes with the LS and therefore masks or inhibits clinical interpretation of LS particularly over low frequency components. The main frequency components of HS are in the range 20 - 100 Hz. This is the range in which LS has major components [
Generally, LSs are produced during inspiration and expiration cycles, and are found in the frequency range 20 - 1200 Hz. There are two types of lung sound, namely, normal and abnormal lung sounds. Normal breath sounds can be categorized into three classes: bronchial, bronchovesicular, and vesicular sounds. Each class of sounds is detected during auscultation according to the characteristics described in [
High-pass filtering of lung-sound recordings to reduce heart sounds would remove significant components of lung sounds. Filtering techniques are categorized as linear adaptive filters and filters employing time-frequency based methods. Several filtering schemes are outlined within these two categories. In [
Methods of heart sound localization are indicated in conjunction with the studies of heart-sound cancellation. Same researchers [11-13] confirm that the adaptive filter is more effective in reducing noise from time series data than linear filters, wavelet shrinkage, and chaosbased noise reduction schemes. The simplest method to reduce HS effects is to apply a high pass filter with cutoff frequency varying from 50 - 150 Hz [
This paper proposes a novel LS filtering method which combines:
• Separating heart sounds (HS) and combined low frequency noise from instruments (NI) simultaneously, while saving LS characteristics.
• Modulating the reference input r1(n) of LMS-ANC to acquiesce combining HS and NI signals.
To achieve the objectives of the algorithm the following tools are used:
• Accruing HS derived from LS original signal by suitable band pass filter with controlled cutoff frequency and order.
• Simulating NI by generating approximately periodic white gaussian noise (WGN).
• Combining HS and simulated WGN.
• Designing a suitable controlled algorithm LMS-ANC in order to determine the optimum values of L order and the coefficient convergence μ.
• Estimating the output results (obtained LS signal) and comparing it with combining signal used power special density (PSD).
PSD has been shown the effectiveness of the LMSANC filtering with suggested method.
The rest of this paper is organized as follows: Section 2 introduces the methodology, which describes the experimental procedure and the proposed modulation method of LMS-ANC. Section 3 presents results and discussion of output signals of ANC, and compares its PSD to normal and abnormal LS. Finally, conclusions are presented in Section 4.
The lung sound has recorded by Computerized Recording Channel of Lung Sound (CRCLS) on the chest and back, right and left [
Each subject has been instructed to breathe so that one breathing cycle occurred every approximately three seconds, (one cycle per ~3 s.). At every flow rate and had at least five breathing cycles at each target flow. Same recordings have been repeated for each subject and each recording consisted of approximately 20 s. For each target flow and concluded with an approximately 5 sec of breath hold (total of ~25 s).
In [4,7], researchers have used adaptive filtering with a pre-processing stage comprising a variable amplifier gain. Other groups used an adaptive filter based on the least mean square (LMS) algorithm to remove HS interferences. In both cases, researchers used HS recorded from close to the patient’s heart location as the reference signal for the adaptive system, which of course are not completely free of LS. The sounds signals have been pre-filtered to remove DC and to prevent aliasing, using CRCLS custom-built 8th order Butter worth band pass filters with pass band 7.5 to 2500 Hz, and amplified by a gain of 200. HS, in our case, has been derived from the original signal recorded from chest at different frequentcies range 20 to 100 Hz, 20 to 150 Hz and to 300 Hz. Other important parameter is the L order of band pass filters, which is applied in the range 2 ≤ L ≤ 120. The frequency range and L are applied depending on the properties of recoding signal and normal or abnormal. The obtained HS signal is used as a reference signal input of adaptive filter.
The main types of linear adaptive filters are the adaptive noise cancellation and linear prediction. Many researchers have applied ANC to lung sound recordings to reduce heart sounds [
The linear adaptive filter with finite memory (finiteduration impulse response—FIR) has been applied for heart sound reduction and noise, which means that the internal structure of the adaptive filter contains only forward feed capability [7,14]. The most common form of a finite memory, or FIR filter is the FIR transversal filter and it adapts algorithm, which updates the tap weights ωk of the transversal filter. So the mean square error (MSE) is minimized and an estimate of the desired output results consists of unit-delay elements that delay each of the M-samples of the input (L is the filter order); elements that multiply weights by input samples; and adders. Each sample, k, of the L-samples of an input reference vector, r1(n), that is multiplied by the conjugate of a weight value, ωk, and these products are summed to form the filter output y(n).
a) Input of the primary signal x(n)b) Input of reference signal r1(n)c) Filter output y(n)d) Finally the estimation error e(n).
The primary input signal consists of the desirable signal L(n) and the noise signal r0(n) as shown in Equation (1),
The reference input r1(n) is modulated to a complex signal, which contains HS and generated signal NI simulated by WGN. Therefore, the generated signal NI accompanied with recording process of LS, which is created from instruments and surrounding environment of LS recording.
The generated NI is based on the methods presented in [19,20]. The methods proposed temporally to generate WGN on real quasi period. It is added to the synthetic HS in various magnitude scales in order to acquire synthetic biosignals corrupted by WGN at a wide range
of signal to noise ratios (SNR). Only SNR higher than 0 dB is considered, because significant noise level distorts ECG in such a degree that low magnitude complexes are not identifiable. However, this method is evaluated on both synthetic periodic signals of the known HS period combined with WGN on real quasi period. Equation (2) is the reference modulated HS signal, h(n), and generated WGN, d(n).
The output of the Adaptive filter, e(n), is the minimum MSE (MMSE) estimate of the component of the primary signal as shown in Equation (3).
where y(n) is the structure of the FIR, which can be represented in Equation (4),
The weights are updated with each iteration, n, based on feedback of the estimation error to the adaptive filter unit.
The LMS algorithm choice is based on its simplicity and it does not require measurements of the pertinent correlation function. It also requires matrix inversion.
It uses Equation (3) to get square error e2(n), which is the square error (MSE) between the output y(n) (equal to wx(n)) and the reference signal r1(n), as given Equation (5)
Shortly, it is possible to calculate the filter coefficient vector for each iteration k having information about the previous coefficients and gradient µ, multiplied by a constant, as shown in Equation (6)
The gradient vector is defined by Equation (7)
This equation with equation 6 Lead to get Equation (8)
The coefficient µ is constant, which must be chosen for quick adaptation without losing stability. The filter is stable if µ satisfies the following condition,
where L is the filter order and Pxx is the power of the input signal [
The LMS algorithm is initiated with an arbitrary value w(0) for the weight vector at n = 0. The successive corrections of the weight vector eventually leads to the minimum value of the mean squared error.
The results demonstrate the ability of the proposed method, where the modified signal r1(n) filtered in Matlab from LS (primary) as stated in the experimental procedure. The obtained results have been controlled by the following parameters:
• The length of signal or segment 25,000 to 32,000 samples at 11,025 Hz frequency sampling rate. This length should include 3 R peaks of ECG waveform as minimum in this segment.
• The convergence factor µ, which controlled using Equation (9).
• The order L of LMS-ANC is chosen in the range 4 ≤ L ≤ 64. The middle value of this range is the optimum value that worked with normal and abnormal LS signals.
The effect of changing the above parameters is monitored by PSD spectrum of LMS-ANC output signal e(n) and the graphic view of output signal e(n).
Figures 4 and 5 show the primary signal of an abnormal LS recording before and after applying method. In comparison, there is a significant difference between the views of both graphics, where frequency components of recording in
These parameters are very important; because it lead to appropriate filtering and to get adequate adaptation of designed filter.
This interprets that PSD of LS before filtering contains more frequencies which can’t be differentiated as illustrated PSD. PSD after filtering demonstrate different components of frequency. This explains that the HS frequency and NI have been removed successfully from LS. Calculated PSD, which is used in this work depends on FFT, explains that the frequency components before filtering are more in spectrum, which is blind and solid. On the other hand, when the signal contains less frequencies, spectrum shows distinguishable frequency values and bands after applying the method as shown in
Furthermore, the method has been applied to normal LS recording using the same steps stated previously.