Procedure for denoising dual-axis swallowing accelerometry signals

Abstract

Dual-axis swallowing accelerometry is an emerging tool for the assessment of dysphagia (swallowing difficulties). These signals however can be very noisy as a result of physiological and motion artifacts. A novel scheme for denoising those signals is proposed, i.e., a computationally efficient search for the optimal denoising threshold within a reduced wavelet subspace. To determine a viable subspace, the algorithm relies on the minimum value of the estimated upper bound for the reconstruction error. A numerical analysis of the proposed scheme using synthetic test signals demonstrated that the proposed scheme is computationally more efficient than minimum noiseless description length (MNDL) based de-noising. It also yields smaller reconstruction errors (i.e., higher signal-to-noise (SNR) ratio) than MNDL, SURE and Donoho denoising methods. When applied to dual-axis swallowing accelerometry signals, the proposed scheme improves the SNR values for dry, wet and wet chin tuck swallows. These results are important to the further development of medical devices based on dual-axis swallowing accelerometry signals.

Description

FIELD OF INVENTION

This invention relates in general to the field of dual-axis swallowing accelerometry signal analysis and more specifically to a method for denoising such signals.

BACKGROUND OF THE INVENTION

Swallowing accelerometry is a potentially informative adjunct to bedside screening for dysphagia. These measurements are minimally invasive, requiring only the superficial attachment of a sensor anterior to the thyroid notch. Even though single-axis accelerometers were traditionally used for swallowing accelerometry, recent studies have shown that dual-axis accelerometers can capture more of the clinically relevant information. Nevertheless, such measurements are inherently very noisy due to various physiological and motion artifacts. Denoising of dual-axis swallowing accelerometry signals is therefore essential for the development of a robust medical device based on these signals.

Estimation of unknown signals in white Gaussian noise has been dealt with by others. Wavelet denoising has previously been proposed as a valuable option. Wavelet denoising removes the additive white Gaussian noise from a signal by zeroing the wavelet coefficients with small absolute value. The suggested optimal threshold is equal to σ_ε√{square root over (2 log N)}

where σ_ε² is the variance of the additive noise and N is the length of the signal. This approach requires the knowledge of the noise variance, which can be estimated from the wavelet coefficients at the finest scale. However, wavelet denoising with the suggested optimal threshold does not necessarily produce the optimal results for signals that are not smooth. i.e., signals with noiseless coefficients being of very small amplitude for a large number of basis functions. Recent attempts to overcome this shortcoming have yielded methods that can suffer from high computational complexities for very long signals, and do not necessarily reach the optimal results.

It is an object of this invention to: (1) reduce high computational complexity; and, (2) reduce reconstruction error associated with denoising swallowing accelerometry signals.

SUMMARY OF THE INVENTION

This invention teaches a method for denoising of long duration dual-axis swallowing accelerometry signals using a computationally efficient algorithm. The algorithm achieves low computational complexity by performing a search for the optimal threshold in a reduced wavelet subspace. To find this reduced subspace, the proposed scheme uses the minimum value of the estimated reconstruction error. By finding this value, the proposed approach also achieves a smaller reconstruction error than previous approaches such as MNDL. SURE-based and Donoho's approaches. This finding has been confirmed for both, synthetic test signals and dual-axis swallowing accelerometry signals.

BRIEF DESCRIPTION OF THE DRAWINGS

In the drawings, embodiments of the invention are illustrated by way of example. It is to be expressly understood that the description and drawings are only for the purpose of illustration and as an aid to understanding, and are not intended as a definition of the limits of the invention.

FIG. 1 is a series of six graphs (a) to (f) comparing the denoising approach of the method of the present invention to the MDL-based and Donoho approaches.

FIG. 2 is a table, identified as Table 1, of SNRs (dB) between the Donoho approach and the method of the present invention.

FIG. 3 is a schematic view of a system for acquiring dual-axis swallowing accelerometry signals to be denoised by the denoising method of the present invention.

DETAILED DESCRIPTION

Methodology of the Invention

Consider N noisy discrete-time observations:x(n)=f(n)+ε(n)  (1)where n=0, . . . , N−1, f(n) is a sampled version of a noiseless continuous signal, and ε(n) is the additive white Gaussian noise drawn from N (0, σ_ε²)

Assume that f(n) can be expanded using

basis functions, b_k(n), on the observation space, B_N:f(n)=Σ_k=1^Nc_kb_k(n)  (2)

wherec_k=b_k(n),f(n)  (3)and (p,q) denotes the inner product of vectors p and q. However, given the noisy observations, the coefficients, c_k, can only be approximated as follows:ĉ_k=b_k(n),x(n)=c_k+b_k(n),ε(n)  (4)Denoising and Reconstruction Error

If f(n) can be described with M nonzero coefficients, where M<<N, then many estimated coefficients, ĉ_k, represent samples of a zero mean Gaussian random variable with variance σ_ε². A classical approach known as wavelet denoising diminishes the effects of noise by first expanding the noisy signal in terms of orthonormal bases of compactly supported wavelets. The estimated coefficients below some threshold, τ, are disregarded either by hard or soft thresholding. The value of τ is always chosen based on an attempt to minimize the so-called reconstruction error,

r_e:

$\begin{array}{cc}{r}_e=\frac{1}{N}{f\left(n\right)-\hat{f}\left(n\right)}^2& \left(5\right)\end{array}$ where ∥·∥ denotes the Euclidean norm and {circumflex over (f)}(ii) represents the estimated noiseless signal. r_e is a sample of random variable R_e that has the following expected value:

$\begin{array}{cc}E\left\{R_e\right\}=\frac{m}{N}{\sigma }^2+\frac{1}{N}{\Delta m}^2& \left(6\right)\end{array}$ where m represents the number of coefficients describing f(n) in some subspace of B_N and Δm is a vector of length N-m, representing the coefficients of bases that are not selected to describe the unknown signal. In reality, r_e is not available and only the number of coefficients not disregarded by the thresholding operation, {circumflex over (m)}, is known. In a recent contribution, probabilistic upper and lower bounds for r_e were derived based on the available data error:

$\begin{array}{cc}{d}_e=\frac{1}{N}{x\left(n\right)-\hat{f}\left(n\right)}^2& \left(7\right)\end{array}$

Therefore, it, has been shown that the upper bound for r_e is equal to

$\begin{array}{cc}{r}_{\mathrm{eub}}\left(\hat{m}\left(\tau \right),{\sigma }^2,\alpha ,\beta \right)=\frac{{\sigma }_{\varepsilon }^2\sqrt{2\hat{m}\left(\tau \right)}}{N}\left(\sqrt{2\hat{m}\left(\tau \right)}+\beta \right)+d_e-{\sigma }_{\varepsilon }^2+\frac{2{\mathrm{\alpha \sigma }}_{\varepsilon }}{\sqrt{N}}\left(\frac{{\mathrm{\alpha \sigma }}_{\varepsilon }}{\sqrt{N}}+\sqrt{\frac{{\alpha }^2{\sigma }_{\varepsilon }^2}{N}+d_e-\left(1-\frac{\hat{m}\left(\tau \right)}{N}\right)\frac{{\sigma }_{\varepsilon }^2}{2}}\right),& \left(8\right)\end{array}$ where α and β represent the parameters for validation probability (P_v=Q(α)) and confidence probability (P_c=Q(β)), with Q(·) for an argument λ being, defined as

$Q\left(\lambda \right)={\int }_{-\lambda }^{+\lambda }\left(\frac{1}{\sqrt{2\pi }}\right)e^{-x^2/2_{\mathrm{dx}}}.$

In addition, {circumflex over (m)}(r) denotes the number of bases whose expansion coefficients are greater than τ in some subspace of B_N.

It should be note that for some values of {circumflex over (m)} the reconstruction error given by eqn. (5) and its upper bound given by eqn. (8) achieve a minimum due to the bias-to-variance trade-off. The principle of MDL has been borrowed from coding theory to find such a minimum value. Also, it has been demonstrated that, smaller reconstruction errors can be achieved with MDL-derived thresholds.

Algorithm for Determining Optimal Threshold

The MNDL-based approach can be computationally expensive for very long data sets since the bases are incrementally added to the subspace describing the unknown signal. Considering the length of acquired dual-axis accelerometry signals (>>10⁵ points}, an attempt should be made to minimize the search space, while choosing a threshold that minimizes the reconstruction error. In some cases the MNDL-based approach can yield higher reconstruction errors than Donoho's approach.

In light of the computational and reconstruction limitations or the MNDL-based approach, a new denoising strategy is proposed here. The goal of this new approach is twofold. First, it should be computationally efficient. Second, it should attain a minimum reconstruction error. Minimization of the search space can be achieved by exploiting the fact that the optimal threshold is usually larger than the actual threshold which minimizes the reconstruction error. The algorithm for determining the optimal threshold is defined through the following steps:

• • 1. Estimate the variance of the noise ε from the median, MED_x, of N/2 wavelet coefficients at the finest scale:

$\begin{array}{cc}{\hat{\sigma }}_{\varepsilon }=\frac{{\mathrm{MED}}_x}{0.6745}& \left(9\right)\end{array}$

• • 2. Based on the estimated noise variance, and for each τ selected from a set 0<τ≦{circumflex over (σ)}_ε√{square root over (2 log(N))}, evaluate the upper bound given by equation (8). Use the soft thresholding procedure to compute the data error required for the evaluation of the upper bound.
  • 3. Determine the optimal threshold for wavelet denoising as:

$\begin{array}{cc}{\tau }_{\mathrm{opt}}=\underset{\tau }{\arg \min }{r}_{\mathrm{eub}}\left(\hat{m}\left(\tau \right),{\sigma }^2,\alpha ,\beta \right)& \left(10\right)\end{array}$

• • 4. Denoise a recording using the optimal value of threshold, τ_opt, and the soft thresholding procedure.

The above procedure is repeated independently for signals acquired from each axis of a dual-axis accelerometer. Unlike the MNDL-based approach, soft thresholding is applied in the above steps, since it yields an estimated signal as smooth as the original signal with high probability. Hard thresholding can produce abrupt artifacts in the recovering signal leading to a higher reconstruction signal.

Numerical Analysis

The results of a two-step numerical analysis are presented in this section. First, the performance of the proposed algorithm is examined using two test signals. The goal of this analysis is to compare the performance of the proposed scheme against that of other well-established techniques under well-controlled conditions. In the second step, the proposed denoising algorithm is applied to the acquired dual-axis swallowing accelerometry signals. The goal is to understand the benefits of the proposed approach in the context of a real biomedical application.

Performance Analysis Using Synthetic Test Signals

Referring to FIG. 1, the first test signal is the so-called Blocks signal, which is a standard signal used in the analysis of various denoising schemes. Assuming that the length of the signal is N=1024 points, the reconstruction error is evaluated for four methods: the proposed method, and the MNDL-based method and a new SURE-based approach. The first test is to numerically examine which of the four schemes provides the lowest reconstruction error for 18 mother wavelets (Haar wavelet, Daubechies wavelets with the number of vanishing moments varying between two and six, Meyer wavelet, Coiflet wavelets with the order varying, between one and five, and Symlet wavelets with the order varying between two and seven). The signal is contaminated with zero-mean additive white Gaussian noise, and SNR=10 dB. For each mother wavelet, 1000 realizations are used. α=10 and β=40 are used for both the MNDL-based approach and the proposed method. The reconstruction errors for the proposed method (circles), the MNDL-based denoising (x'S), the SURE-based approach (diamonds) and Donoho's approach (squares) are shown in FIG. 1(a). Amongst the 18 wavelet functions, considered, the Haar wavelet (the wavelet indexed as 1 on the x-axis of FIG. 1(a)) provides the smallest reconstruction error, since, the structure of the wavelet closely resembles the structure of the signal.

The next task is to examine the reconstruction error under various SNR values with the Haar wavelet. One thousand realizations are used for each SNR value yielding the results depicted in FIG. 1(b). From the graph, it is clear that the proposed method (solid line) provides the smallest error for various SNR levels with the MNDL-based (dotted line) and SURE-based (dashdotted line) methods also providing a small error. Donoho's approach (dashed line) consistently yields the highest reconstruction error. Despite the small reconstruction error over different SNL levels, the MNDL-based method suffers from high computational complexity. To further understand the computational bottlenecks, the SNR value is kept constant at 10 dB, but the length of the Blocks signal is varied between N=2¹⁰ and N=2¹⁵ points. The durations required to execute the specific algorithms are tracked using built-in MATLAB functions. The time to complete the denoising task, averaged over ten realizations of the Block signal at each signal length is reported in FIG. 1(c). As expected, as N increases, there is an obvious upward trend for all for algorithms. Donoho's approach (dashed line) is the least computationally expensive. However, for the MNDL-based approach (dotted line) the time required to complete the task increases significantly with signal length. For example, the average duration required for the MNDL-based approach to denoise a signal with length of N=2¹⁵ points is 157 seconds. On the other hand, the time required by the proposed algorithm (solid line) to denoise the same signal is 0.74 seconds. In fact, computation time of the proposed method increases logarithmically with signal length (the duration is approximately equal to log₁₀ (N^0.35)).

To more closely mimic a real swallowing scenario, the test signal shown in FIG. 1(d), is used in the analysis. The signal is defined as:

$\begin{array}{cc}f\left(n\right)=\{\begin{array}{cc}{f}_o\left(n\right)+0.6\cos \left(210\pi \mathrm{nT}\right)& 8100\le n\le 16430\\ f_o\left(n\right)+0.5\cos \left(140\pi \mathrm{nT}\right)& 11400\le n\le 18330\\ f_o\left(n\right)+0.2\cos \left(120\pi \mathrm{nT}\right)& 13200\le n\le 25230\\ f_o\left(n\right)+0.4\cos \left(160\pi \mathrm{nT}\right)& 12250\le n\le 23400\\ f\left(n\right)w\left(n\right)& 8100\le n\le 25230\end{array}& \left(11\right)\end{array}$ where w(n) is Gaussian window with standard deviation σ_g=1.9 andf₀(n)=0.1 sin(8πnT)+0.2 sin(2πnT)+0.15 sin(20πnT)+0.15 sin(6πnT)+0.12 sin(14πnT)+0.1 sin(4πnT)  (12)with 0≦n≦N−1, N=35000 and T=10⁻⁴ seconds. The duration of the signal is based on previously reported swallow durations. It should be mentioned that this signal only mimics a realistic signal, and does not represent a model of a swallow. The same group of wavelets as in the Blocks signal analysis are used to examine the reconstruction error. It is assumed again that the signal is contaminated with additive zero-mean Gaussian noise and SNR=0.10 dB. For this particular signal, the Meyer wavelet (indexed by number 7 in FIG. 1(e)) achieved the smallest reconstruction error since the structure of the wavelet resembles the structure of the signal. It should be pointed out that the MNDL-based method consistently provides the highest error for all considered wavelets. Given that the method is sensitive to the choice of α and β we varied the two parameters to further examine the obtained error. The MNDL method still maintained the highest reconstruction error for this particular signal. The main reason for these results is the hard-thresholding procedure used in this method. Consequently, the better results are indeed expected with an approach implementing a soft-thresholding procedure. As the next step, the reconstruction error is evaluated using the Meyer wavelet for various SNR values for all four approaches. From the results shown in FIG. 1 (f), it is obvious that the proposed method (solid line) achieves a significantly smaller reconstruction error than the other three methods.Denoising Dual-Axis Swallowing Accelerometry SignalsExperimental Protocol

During a three month period, four hundred and eight participants (aged 18-65) were recruited at a public science centre. All participants provided written consent. The study protocol was approved by the research ethics boards of the Toronto Rehabilitation Institute and Bloorview Kids Rehab, both located in Toronto, Ontario, Canada. As seen in FIG. 3, a dual-axis accelerometer 102 (ADXL322, Analog Devices) was attached to the participant's neck (anterior to the cricoid cartilage) using double-sided tape, and an output signal of the accelerometer 102 was communicated to a signal processor 104 of a computer 106. The axes of acceleration were aligned to the anterior-posterior (A-P) and superior-inferior (S-I) directions. Data were band-pass filtered in hardware with a pass band of 0.1-3000 Hz and sampled at 10 kHz using a custom LabVIEW program running on a laptop computer.

With the accelerometer attached, each participant was cued to perform 5 saliva swallows (denoted as dry in Table 1). After each swallow, there was a brief rest to allow for saliva production. Subsequently, the participant completed 5 water swallows (denoted as wet in Table 1) by cup with their chin in the natural position (i.e., perpendicular to the floor) and water swallows in the chin-tucked position (denoted as WTC in Table 1). The entire data collection session lasted 15 minutes per participant.

Results of Denoising

The acquired dual-axis swallowing accelerometry signals were denoised using Donoho's approach, the MNDL-based approach, the SURE-based approach and the proposed approach. In particular, a 10-level discrete wavelet transform using the Meyer wavelet with soft thresholding was implemented. Before denoising, the signals were pre-processed using inverse filters to annul effects of the data collection system on the acquired data. In order to compare the performance of the aforementioned denoising schemes, SNR values were evaluated before and after denoising using the following formula:

$\begin{array}{cc}\mathrm{SNR}=10{\log }_{10}\left(\frac{E_f}{E_{\hat{\varepsilon }}}\right)& \left(13\right)\end{array}$ where E_f represents the approximate energy of the noise-free signal, and E_{{circumflex over (ε)}} represents an approximate variance of the white Gaussian noise. The approximate energy is calculated as E_f={circumflex over (σ)}_x²={circumflex over (σ)}_{{circumflex over (ε)}}², where {circumflex over (σ)}_x² is the variance of the observed signal, and {circumflex over (σ)}_{{circumflex over (ε)}}² represents the variance of the noise calculated by (9). Similarly, E_{{circumflex over (ε)}}={circumflex over (σ)}_x² for the noisy signals, and for the denoised signals E_{{circumflex over (ε)}}=r_eub({circumflex over (m)}(τ),{circumflex over (σ)}_{{circumflex over (ε)}}²,α,β) for the threshold estimated by (10).

Using the SNR metric given by (13), the results of the analysis are summarized in Table 1. Donoho's approach provides the least amount of improvement in SNR as expected, followed by the MNDL-based approach. The SURE-based approach achieves greater improvement in the SNR values in comparison to the other two aforementioned approaches. Nevertheless, as demonstrated by the results in Table 1, the SURE approach exhibits strong variations in performance. The proposed approach provides the greatest improvement in SNR values. On average, the greatest gain in SNR is over Donoho's approach (3.8 dB and 4.0 dB in the A-P and S-I directions, respectively), while smaller improvements were obtained over the SURE-based approach (2.0 dB and 1.3 dB in the A-P and S-I directions, respectively). Nevertheless, the proposed approach still provides a statistically significant improvement over SURE-based approach in denoising the dual-axis swallowing accelerometry signals (Wilcoxon rank-sum test, p<<10⁻¹⁰ for both directions). This improvement was achieved regardless of whether or not the different swallowing types were considered individually or as a group. As a last remark, it should be noted that these SNR values were estimated using eqn. (13), which from our experience with swallowing signals, provides a conservative approximation. In reality, we expect the gains in SNR to be even greater.

CONCLUSION

A denoising algorithm is proposed for dual-axis swallowing accelerometry signals, which have potential utility in the non-invasive diagnosis of swallowing difficulties. This algorithm searches for the optimal threshold value in order to achieve the minimum reconstruction error for a signal. To avoid the high computational complexity associated with competing algorithms, the proposed scheme conducts the threshold search in a reduced wavelet subspace. Numerical analysis showed that the algorithm achieves a smaller reconstruction error than Donoho, MNDL- and SURE-based approaches. Furthermore, the computational complexity of the proposed algorithm increases logarithmically with signal length. The application of the proposed algorithm to dual-axis swallowing accelerometry signals demonstrated statistically significant improvements in SNR over the other three considered methods.

Claims

1. A computer executed procedure for denoising swallowing accelerometry signals, the procedure executed by a processor coupled to memory of the computer on a signal output from an accelerometer attached to a person's neck during swallowing, the procedure comprising: defining a set of N discrete time observations from said signal to have a noiseless component f(n) and a noise component ε(n);estimating a set of wavelet coefficients ck for said noiseless component f(n);estimating a variance {circumflex over (σ)}ε of the noise component ε(n) from a median MEDx of a finest scale subset of said wavelet coefficients;defining a set of wavelet coefficient threshold values τ within a range thereof defined as function of said estimated variance, wherein the wavelet coefficient threshold values τ are each selected from a set 0<τ≦{circumflex over (σ)}ε√(2 log(N));for each of said threshold values τ, calculating an upper bound for a reconstruction error re;selecting an optimal threshold value τopt from said set of threshold values that minimizes said upper bound; anddenoising the signal by suppressing those of said wavelet coefficients below said optimal threshold value,wherein the upper bound is determined by

2. The computer executed procedure of claim 1, wherein said suppressing comprises suppressing via one of a hard and a soft thresholding process.

3. The computer executed procedure of claim 2, wherein said suppressing comprises suppressing via a soft thresholding process.

4. The computer executed procedure of claim 1, wherein the signal is output from a dual-axis accelerometer.

5. The computer executed procedure of claim 4, wherein the procedure is executed independently for signals acquired from each axis of the dual-axis accelerometer.

6. The computer executed procedure of claim 1, wherein said finest scale subset consists of N/2 finest scale wavelet coefficients.

7. The computer executed procedure of claim 6, wherein the variance {circumflex over (σ)}ε of the noise component ε(n) is estimated from the median MEDx of the finest scale subset consisting of N/2 finest scale wavelet coefficients by dividing MEDx by 0.6745.