A method for removing blink artifacts in single-channel electroencephalogram signals

Abstract

A method for removing blink artifacts in single-channel EEG signals, which relates to a method for processing EEG signals. In order to solve the problems that the existing methods for removing blink artifacts in EEG signals require manual marking and additional electrodes to record reference signals, resulting in incomplete artifact removal and large loss of EEG signals. The present invention inputs a single-channel EEG signal into a low-pass filter, extracts an approximate blink artifact signal and identifies the interval where the approximate blink artifact signal is located; calculates the similarity of EEG signals in multiple intervals, determines the typical blink artifacts in the EEG signals of multiple intervals, and constructs the specific form of the corresponding biorthogonal filter bank; automatically determines the decomposition depth of wavelet transform; and removes the blink artifacts in the single-channel EEG signal based on the specific form of the filter bank and the decomposition depth of discrete wavelet transform. The beneficial effects are that blink artifacts can be removed more effectively, and at the same time, the original EEG information can be better retained.

Description

Technical Field

[0001] The present invention relates to a method for processing electroencephalogram (EEG) signals. Background Art

[0002] The signals collected by EEG devices often contain short-duration and high-amplitude spikes caused by involuntary blinking activities; since the frequency of such spike signals is mainly concentrated in the 0-12 Hz range, the low-frequency components of EEG signals will be contaminated by blinking artifacts; therefore, directly analyzing the contaminated EEG signals may lead to incorrect understanding of brain activities, especially when analyzing the signals collected by electrodes close to the eyes.

[0003] It is not advisable to directly discard the signal segments containing blinking artifacts, because some key EEG information may be lost; moreover, it is also impossible to avoid the interference caused by blinking artifacts by asking the subjects not to blink, because this will not only cause discomfort to the users, but may also affect normal brain activities. The existing blinking artifact removal schemes can generally be divided into several categories based on regression, adaptive filters, independent component analysis, discrete wavelet transform (DWT), and deep learning. Although these methods can be used to remove blinking artifacts, their disadvantages in processing single-channel EEG signals cannot be ignored: the regression and adaptive filter methods require additional channels to record reference signals, which will increase the hardware complexity of single-channel devices; the independent component analysis-based methods usually require manual marking, which has high professional requirements for users; the existing DWT-based methods do not consider the individual differences between subjects, that is, the wavelet functions used when processing different EEG signals are fixed; in addition, when there is not enough data, the performance of deep learning will also decline.

[0004] Therefore, in order to promote the application of single-channel EEG devices in real life, there is an urgent need to develop a more effective, human-unattended, reference-signal-free blinking artifact removal algorithm for single-channel EEG signals. Summary of the Invention

[0005] The purpose of the present invention is to solve the problem that the existing methods for removing blinking artifacts in EEG signals require manual marking and additional electrodes to record reference signals, resulting in incomplete artifact removal and large loss of EEG signals, and a method for removing blinking artifacts in single-channel EEG signals is proposed.

[0006] A method for removing blinking artifacts in single-channel EEG signals according to the present invention, the removal method includes the following steps;

[0007] Step 1: Input a single-channel EEG signal into a low-pass filter to extract an approximate signal of blinking artifacts;

[0008] Step 2: Identify the approximate signal of the blink artifact extracted in Step 1, identify the interval where the approximate signal of the blink artifact is located, and obtain multiple interval EEG signals;

[0009] Step 3: Calculate the similarity of the multiple interval EEG signals obtained in Step 2, and determine the typical blink artifacts among the multiple interval EEG signals;

[0010] Step 4: Construct the specific form of the corresponding biorthogonal filter bank based on the typical blink artifacts determined in Step 3;

[0011] Step 5: Perform a morphology-based similarity judgment on each interval signal identified in Step 2, and automatically determine the corresponding discrete wavelet transform decomposition depth according to the similarity threshold;

[0012] Step 6: Remove the blink artifacts in the single-channel EEG signal based on the specific form of the filter bank determined in Step 4 and the discrete wavelet transform decomposition depth automatically obtained in Step 5.

[0013] Furthermore, the specific method for constructing the specific form of the corresponding biorthogonal filter bank is as follows:

[0014] Based on the typical blink artifacts determined in Step 3, and using the sine-cosine optimization algorithm to optimize the objective function value R loss to obtain the minimum value of the objective function value R loss and uniquely determine the biorthogonal filter bank coefficients λ by using this optimization process; obtain the specific form of the biorthogonal filter bank according to the obtained biorthogonal filter bank coefficients λ;

[0015] The specific expression form of the objective function value R loss is as follows:

[0016]

[0017] where p r is the typical blink artifact determined in Step 3; is the first low-frequency component after the discrete wavelet transform decomposition; L is the discrete wavelet transform decomposition depth, and in formula (1), L = 4; is the covariance between the first low-frequency component of the discrete wavelet transform and the typical blink artifact determined in Step 3; is the standard deviation of the typical blink artifact determined in Step 3; is the standard deviation of the first low-frequency component of the discrete wavelet transform.

[0018] Further, the specific form of the biorthogonal filter bank includes a low-pass filter H0(z), a high-pass filter H1(z), a low-pass reconstruction filter G0(z), and a high-pass reconstruction filter G1(z); the specific method for obtaining the low-pass filter H0(z), the high-pass filter H1(z), the low-pass reconstruction filter G0(z), and the high-pass reconstruction filter G1(z) is as follows:

[0019] G0 = [1, 2λ + 4, 8λ + 7, 12λ + 8, 8λ + 7, 2λ + 4, 1] (2)

[0020]

[0021] where G0 is the filter coefficient of the low-pass reconstruction filter G0(z); H0 is the filter coefficient of the low-pass filter H0(z); A is the first intermediate coefficient, A = -(3 + λ); B is the second intermediate coefficient, C is the third intermediate coefficient,

[0022] Perform a Z-transform on the low-pass reconstruction filter G0(z) to obtain the high-pass filter H1(z); use the low-pass filter H0(z) to perform a Z-transform to obtain the high-pass reconstruction filter G1(z);

[0023] H1(z) = z -1 G0(-z) (4)

[0024] G1(z) = zH0(-z) (5)

[0025] where z is the Z-transform symbol.

[0026] Further, the specific method for extracting the blink artifact approximation signal in step one is as follows:

[0027] Input the single-channel EEG signal into the low-pass filter for low-pass filtering to generate a filtered signal, which is the extracted blink artifact approximation signal;

[0028] The cut-off frequency of the low-pass filter is 12 Hz.

[0029] In this embodiment, since the frequency of the blink artifact signal is mainly concentrated in 0 - 12 Hz, therefore, the cut-off frequency of the low-pass filter is selected as 12 Hz.

[0030] Further, the specific method for obtaining the multi-interval EEG signals in step two is as follows:

[0031] The local maximum value of the extracted approximate signal of the blink artifact is compared with the amplitude threshold Th. When the local maximum value of the approximate signal of the blink artifact is greater than the amplitude threshold Th, the local maximum value at this time is considered to be a peak of the blink signal. Taking the peak of the blink signal as a reference, 100 milliseconds are pushed forward and 500 milliseconds are pushed backward respectively to obtain the interval where the blink artifact is located, and then multiple intervals of EEG signals are obtained.

[0032] Further, the calculation formula of the amplitude threshold Th is:

[0033]

[0034] where n is the discrete signal points in the input single-channel EEG signal; f(n) is the input single-channel EEG signal; e(n) is the generated filtered signal; N ef is the signal length of the input single-channel EEG signal; median(·) is the median function; |·| is the absolute value function.

[0035] Further, the specific method for determining the typical blink artifact in the multiple intervals of EEG signals in step three is:

[0036] The similarity calculation is performed on the multiple intervals of EEG signals obtained in step two to obtain the similarity matrix M; the sum of each column element of the similarity matrix M is obtained to obtain the row vector matrix M*; the elements of the row vector matrix M* are compared with each other to obtain the maximum value m r , the maximum value m r The corresponding r-th interval of EEG signal is the typical blink artifact;

[0037] The m r is the sum value of the r-th column element of the similarity matrix M.

[0038] Further, the specific form of the similarity matrix M is:

[0039]

[0040]

[0041] where N is the total number of interval EEG signals; p1 is the first interval EEG signal; p i is the i-th interval EEG signal; p j is the j-th interval EEG signal; p N is the N-th interval EEG signal; is the similarity between the i-th interval EEG signal and the j-th interval EEG signal; is the standard deviation of the i-th interval EEG signal; is the standard deviation of the j-th interval EEG signal; cov(pi , p j ) is the covariance between the i-th interval EEG signal and the j-th interval EEG signal.

[0042] Furthermore, the specific form of the row vector matrix M* is:

[0043]

[0044] where m1 is the sum value of the elements in the first column of the similarity matrix M; m N is the sum value of the elements in the N-th column of the similarity matrix M.

[0045] Furthermore, the similarity threshold in step five is set to 0.85.

[0046] The beneficial effects of the present invention are as follows: This removal method is used to remove blink artifacts from EEG signals, especially suitable for EEG signals collected by portable single-channel EEG devices; this removal method does not require additional electrodes to record reference signals, reducing hardware complexity, and does not require manual marking or parameter adjustment. It will automatically determine relevant parameters according to the characteristics of the input signal, reducing the professional requirements for users; the experimental results of simulated EEG signals and real EEG signals show that this removal method not only more effectively removes blink artifacts, but also better retains the original EEG information, which is of great benefit to fields such as BCI. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flowchart of a method for removing blink artifacts from a single-channel EEG signal described in Embodiment 1;

[0048] Figure 2 is a comparison chart of simulation data results based on mean absolute error and Pearson correlation coefficient in Embodiment 1;

[0049] Figure 3 is a comparison chart of simulation data results of mean absolute error and Pearson correlation coefficient under different signal-to-noise ratio conditions in Embodiment 1;

[0050] Figure 4 is a comparison chart of visualization results of artifact removal from simulated EEG signals by different methods in Embodiment 1;

[0051] Figure 5 is a comparison chart of visualization results of artifact removal from real EEG signals by different methods in Embodiment 1. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] Combined with Figures 1 to 5 This embodiment is described. A method for removing blink artifacts from a single-channel EEG signal described in this embodiment includes the following steps;

[0053] Step 1: Input a single-channel EEG signal into a low-pass filter to extract an approximate signal of blink artifacts;

[0054] Step 2: Identify the approximate signal of blink artifacts extracted in Step 1, identify the intervals where the approximate signal of blink artifacts is located, and obtain multiple interval EEG signals;

[0055] Step 3: Calculate the similarity of the multiple interval EEG signals obtained in Step 2 to determine the typical blink artifacts among the multiple interval EEG signals;

[0056] Step 4: Construct the specific form of the corresponding biorthogonal filter bank based on the typical blink artifacts determined in Step 3;

[0057] Step 5: Perform a morphology-based similarity judgment on each interval signal identified in Step 2, and automatically determine the corresponding discrete wavelet transform decomposition depth according to the similarity threshold;

[0058] Step 6: Remove the blink artifacts in the single-channel EEG signal based on the specific form of the filter bank determined in Step 4 and the discrete wavelet transform decomposition depth automatically obtained in Step 5.

[0059] In the preferred embodiment, the specific method for constructing the specific form of the corresponding biorthogonal filter bank is as follows:

[0060] Based on the typical blink artifacts determined in Step 3, and using the sine-cosine optimization algorithm to optimize the objective function value R loss to obtain the minimum value of the objective function value R loss and uniquely determine the biorthogonal filter bank coefficients λ using this optimization process; obtain the specific form of the biorthogonal filter bank according to the obtained biorthogonal filter bank coefficients λ;

[0061] The specific expression form of the objective function value R loss is as follows:

[0062]

[0063] where p r is the typical blink artifact determined in Step 3; is the first low-frequency component after discrete wavelet transform decomposition; L is the discrete wavelet transform decomposition depth, and in formula (1), L = 4; is the covariance between the first low-frequency component of the discrete wavelet transform and the typical blink artifact determined in Step 3; is the standard deviation of the typical blink artifact determined in Step 3; is the standard deviation of the first low-frequency component of the discrete wavelet transform.

[0064] In a preferred embodiment, the specific form of the biorthogonal filter bank includes a low-pass filter H0(z), a high-pass filter H1(z), a low-pass reconstruction filter G0(z), and a high-pass reconstruction filter G1(z); the specific methods for obtaining the low-pass filter H0(z), the high-pass filter H1(z), the low-pass reconstruction filter G0(z), and the high-pass reconstruction filter G1(z) are as follows:

[0065] G0 = [1, 2λ + 4, 8λ + 7, 12λ + 8, 8λ + 7, 2λ + 4, 1] (2)

[0066]

[0067] where G0 is the filter coefficient of the low-pass reconstruction filter G0(z); H0 is the filter coefficient of the low-pass filter H0(z); A is the first intermediate coefficient, A = -(3 + λ); B is the second intermediate coefficient, C is the third intermediate coefficient,

[0068] Performing a Z-transform using the filter coefficient of the low-pass reconstruction filter G0(z) to obtain the high-pass filter H1(z); performing a Z-transform using the low-pass filter H0(z) to obtain the high-pass reconstruction filter G1(z);

[0069] H1(z) = z -1 G0(-z) (4)

[0070] G1(z) = zH0(-z) (5)

[0071] where z is the Z-transform symbol.

[0072] In a preferred embodiment, the specific method for extracting the blink artifact approximation signal in step one is as follows:

[0073] Inputting the single-channel EEG signal into the low-pass filter for low-pass filtering to generate a filtered signal, and this filtered signal is the extracted blink artifact approximation signal;

[0074] The cut-off frequency of the low-pass filter is 12 Hz.

[0075] In a preferred embodiment, the specific method for obtaining the multi-interval EEG signals in step two is as follows:

[0076] Comparing the local maximum value of the extracted blink artifact approximation signal with the amplitude threshold Th. When the local maximum value of the blink artifact approximation signal is greater than the amplitude threshold Th, this local maximum value is considered a blink signal peak; using the blink signal peak as a reference, respectively pushing forward 100 milliseconds and backward 500 milliseconds to obtain the interval where the blink artifact is located, and thus obtaining the multi-interval EEG signals.

[0077] In a preferred embodiment, the calculation formula of the amplitude threshold Th is:

[0078]

[0079] where n is the discrete signal points in the input single-channel EEG signal; f(n) is the input single-channel EEG signal; e(n) is the generated filter signal; N ef is the signal length of the input single-channel EEG signal; median(·) is the median function; |·| is the absolute value function.

[0080] In a preferred embodiment, the specific method for determining the typical blink artifacts in the multiple interval EEG signals in step three is as follows:

[0081] Calculate the similarity of each of the multiple interval EEG signals obtained in step two to obtain a similarity matrix M; sum the elements of each column of the similarity matrix M to obtain a row vector matrix M*; compare the elements of the row vector matrix M* with each other to obtain the maximum value m r , the maximum value m r The corresponding r-th interval EEG signal is the typical blink artifact;

[0082] The m r is the sum value of the elements in the r-th column of the similarity matrix M.

[0083] In a preferred embodiment, the specific form of the similarity matrix M is:

[0084]

[0085]

[0086] where N is the total number of interval EEG signals; p1 is the first interval EEG signal; p i is the i-th interval EEG signal; p j is the j-th interval EEG signal; p N is the N-th interval EEG signal; is the similarity between the i-th interval EEG signal and the j-th interval EEG signal; is the standard deviation of the i-th interval EEG signal; is the standard deviation of the j-th interval EEG signal; cov(p i ,p j ) is the covariance between the i-th interval EEG signal and the j-th interval EEG signal.

[0087] In a preferred embodiment, the specific form of the row vector matrix M* is:

[0088]

[0089] Among them, m1 is the sum value of the elements in the first column of the similarity matrix M; m N is the sum value of the elements in the Nth column of the similarity matrix M.

[0090] In the preferred embodiment, the similarity threshold in step five is set to 0.85.

[0091] In this embodiment, since the discrete wavelet transform decomposition is not only affected by the adaptive biorthogonal wavelet function, but also related to the discrete wavelet transform decomposition depth L; therefore, after determining the adaptive wavelet function, L also needs to be determined; when L is small, a relatively complete blink signal is retained, and at the same time, more electroencephalogram signals are included; when L is large, in addition to the higher computational cost, the blink artifacts will be scattered in multiple DWT decomposition components; an ideal value of L should make include relatively complete blink artifacts, and at the same time contain as few electroencephalogram signals as possible, so that the blink artifacts can be removed by directly discarding to achieve the purpose of removing blink artifacts; in addition, it is also found that the similarity between j and p decreases with the increase of the value of L; in order to automatically determine the value of L corresponding to the signal, the present invention sets the similarity threshold to 0.85.

[0092]

[0093] In this embodiment, in order to reduce the interference of blink artifacts on electroencephalogram signals, an adaptive single-channel blink artifact removal method (FBLFP-ABOW) proposed in this embodiment, based on the discrete wavelet transform (DWT), takes into account the individual differences between subjects, constructs an adaptive wavelet function according to the signal characteristics and automatically determines the discrete wavelet transform decomposition depth; effectively removes blink artifacts while reducing the loss of electroencephalogram signals, does not require additional electrodes to record reference signals, and also does not require manual marking, which greatly promotes the application of single-channel electroencephalogram devices in brain-computer interfaces (BCIs) in non-clinical environments;

[0094] Since the blink signal and the non-blink electroencephalogram signal in the simulated data are known, the removal method described in this embodiment first comprehensively evaluates the performance of FBLAP-ABOW using the simulated data; the evaluation content involves the mean absolute error, Pearson correlation coefficient, and power spectral density absolute error between the non-blink electroencephalogram signal and the filtered signal. Figure 2The experimental results in [the specific context] show that, compared with three recently proposed methods, the average absolute error value between the EEG signals removed by FBLAP-ABOW and the blink-artifact-free signals is the smallest, and the Pearson correlation coefficient is the largest. Figure 3 It shows that the larger the signal-to-noise ratio, the better the performance of FBLAP-ABOW. In practice, due to the interference brought by various environments, it is difficult to achieve a large signal-to-noise ratio. It can be seen that FBLAP-ABOW still performs the best under low signal-to-noise ratio conditions. Figure 4 Examples of simulated EEG signals filtered by different methods are shown. In addition, the results of the sum of the absolute errors of the power spectral density show that, compared with the comparison methods (25.79, 21.97, 22.90), the spectral loss corresponding to FBLAP-ABOW is the smallest (15.87).

[0095] The removal method described in this embodiment also qualitatively tested the performance of FBLAP-ABOW using the signals collected under three different real application scenarios. Figure 5 Examples of real EEG signals filtered by different methods are shown. It is found that the performance of FBLAP-ABOW on real data is also better than other methods, and FBLFP-ABOW has stronger adaptability to data. That is to say, there is no situation where FBLFP-ABOW performs well in one dataset but poorly in another dataset.

[0096] As described above, only the specific embodiments of the present invention are provided, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for removing blink artifacts in single-channel EEG signals, characterized in that, The removal method comprises the following steps; Step 1: Input a single-channel EEG signal into a low-pass filter to extract an approximate blink artifact signal; Step 2: Identify the approximate blink artifact signal extracted in Step 1, identify the intervals where the approximate blink artifact signal is located, and obtain multiple interval EEG signals; Step 3: Calculate the similarity of the multiple interval EEG signals obtained in Step 2 to determine the typical blink artifacts among the multiple interval EEG signals; The specific method for determining the typical blink artifacts among the multiple interval EEG signals is: Calculate the similarity for each of the multiple interval EEG signals obtained in step two to obtain a similarity matrix M; sum the elements of each column of the similarity matrix M to obtain a row vector matrix M*; compare the elements of the row vector matrix M* with each other to obtain the maximum value m in the row vector matrix M* r , the maximum value m r The corresponding r-th interval EEG signal is the typical blink artifact; wherein m r is the sum value of the r-th column elements of the similarity matrix M; Step 4: Construct the specific form of the corresponding bi-orthogonal filter bank based on the typical blink artifacts determined in Step 3; The specific method for constructing the specific form of the corresponding bi-orthogonal filter bank is; Based on the typical blink artifacts determined in step three, and using the sine-cosine optimization algorithm to optimize the objective function value R loss to obtain the minimum value of the objective function value R loss ; the biorthogonal filter bank coefficients λ are uniquely determined using this optimization process; the specific form of the biorthogonal filter bank is obtained according to the obtained biorthogonal filter bank coefficients λ; The objective function value R loss has the following specific form: where n is the discrete signal points in the input single-channel EEG signal; p r is the typical blink artifact determined in step 3; is the first low-frequency component after discrete wavelet transform decomposition; L is the decomposition depth of discrete wavelet transform, and in formula (1), L = 4; is the covariance between the first low-frequency component of discrete wavelet transform and the typical blink artifact determined in step 3; is the standard deviation of the typical blink artifact determined in step 3; is the standard deviation of the first low-frequency component of discrete wavelet transform; Step 5: Perform a morphology-based similarity judgment on each interval EEG signal identified in Step 2, and automatically determine the corresponding discrete wavelet transform decomposition depth according to the similarity threshold; Step 6: Remove the blink artifacts in the single-channel EEG signal based on the specific form of the filter bank determined in Step 4 and the discrete wavelet transform decomposition depth automatically determined in Step 5.

2. The method for removing blink artifacts in single-channel EEG signals according to claim 1, wherein, The specific form of the bi-orthogonal filter bank includes a low-pass filter H0(z), a high-pass filter H1(z), a low-pass reconstruction filter G0(z), and a high-pass reconstruction filter G1(z); the specific method for obtaining the low-pass filter H0(z), the high-pass filter H1(z), the low-pass reconstruction filter G0(z), and the high-pass reconstruction filter G1(z) is: Among them, G0 is the filtering coefficient of the low-pass reconstruction filter G0(z); H0 is the filtering coefficient of the low-pass filter H0(z); A is the first intermediate coefficient, A = -(3 + λ); B is the second intermediate coefficient, C is the third intermediate coefficient, Perform a Z-transform on the low-pass reconstruction filter G0(z) to obtain the high-pass filter H1(z); use the low-pass filter H0(z) to perform a Z-transform to obtain the high-pass reconstruction filter G1(z); H1(z) = z -1 G0(-z) (4) G1(z) = zH0(-z) (5) where z is the Z-transform symbol.

3. A method for removing blink artifacts in single-channel EEG signals according to claim 1, characterized in that, The specific method for extracting the approximate blink artifact signal in Step 1 is: Perform low-pass filtering on the single-channel EEG signal input into the low-pass filter to generate a filtered signal, and this filtered signal is the extracted approximate blink artifact signal; The cut-off frequency of the low-pass filter is 12 Hz.

4. A method for removing blink artifacts in single-channel EEG signals according to claim 1, characterized in that, The specific method for obtaining multiple interval EEG signals in Step 2 is: Compare the local maximum value of the extracted approximate blink artifact signal with the amplitude threshold Th. When the local maximum value of the approximate blink artifact signal is greater than the amplitude threshold Th, this local maximum value is considered a blink signal peak; using the blink signal peak as a reference, respectively push forward 100 milliseconds and backward 500 milliseconds to serve as the intervals where the blink artifacts are located, and then obtain multiple interval EEG signals.

5. The method for removing blink artifacts in single-channel EEG signals according to claim 4, wherein The calculation formula for the amplitude threshold Th is: Among them, f(n) is the input single-channel EEG signal; e(n) is the generated filtered signal; N ef is the signal length of the input single-channel EEG signal; median(·) is the median function; |·| is the absolute value function.

6. The method for removing blink artifacts in single-channel EEG signals according to claim 1, characterized in that, The specific form of the similarity matrix M is: Among them, N is the total number of interval EEG signals; p1 is the first interval EEG signal; p i is the i-th interval EEG signal; p j is the j-th interval EEG signal; p N is the N-th interval EEG signal; is the similarity between the i-th interval EEG signal and the j-th interval EEG signal; is the standard deviation of the i-th interval EEG signal; is the standard deviation of the j-th interval EEG signal; cov(p i , p j ) is the covariance between the i-th interval EEG signal and the j-th interval EEG signal.

7. A method for removing blink artifacts in single-channel EEG signals according to claim 6, characterized in that The specific form of the row vector matrix M* is: Among them, m1 is the sum value of the elements in the first column of the similarity matrix M; m N is the sum value of the elements in the Nth column of the similarity matrix M.

8. A method for removing blink artifacts in single-channel EEG signals according to claim 1, characterized in that The similarity threshold in Step 5 is set to 0.85.

Images