An electroencephalogram signal pattern recognition method based on state-dependent convolution sparse model

Abstract

The application discloses a state-related convolution sparse model-based electroencephalogram signal pattern recognition method, aiming at the problem that existing convolution sparse coding is difficult to recognize state-related biomarkers, state-shared and state-specific waveforms are used for modeling, and the waveform characteristics of electroencephalogram signals in different states are clearly distinguished. Based on this, the application adopts waveform inconsistency constraints to efficiently identify potential biomarkers related to certain states, and the identified potential biomarkers have good interpretability and can be used as an effective tool for auxiliary medical diagnosis.

Description

Technical Field

[0001] This invention relates to the field of neural signal decoding, and in particular to a method for EEG signal pattern recognition based on a state-related convolutional sparse model. Background Technology

[0002] Typical waveforms in electroencephalogram (EEG) signals can provide effective biomarkers for understanding complex cognitive processes in the brain. For example, variations in sleep spindle waves (spontaneous periodic oscillations of 12-14 Hz) can indicate multiple diseases such as schizophrenia; sharp waves and high-frequency oscillations in epilepsy are also highly correlated with epileptic seizures. Therefore, discovering and identifying these typical waveforms can help detect biomarkers related to neurological diseases.

[0003] To identify these typical waveforms, both knowledge-based and data-driven approaches have been employed. Traditional knowledge-based methods typically presuppose characteristics of the target waveform, such as frequency range and time-domain shape. For example, some methods define a typical waveform as neural activity in time-frequency representation where the power exceeds a preset threshold and falls within a predefined frequency range. Data-driven methods aim to reduce reliance on predefined assumptions by learning typical waveforms directly from neural signals. Typical methods include empirical pattern decomposition, cyclic periodic analysis, and short amplitude fluctuation detection, which can automatically identify typical waveforms based on time-domain characteristics.

[0004] For example, Chinese patent document CN109480834A discloses a method for classifying EEG signals based on fast multidimensional empirical mode decomposition, including: (1) collecting several sets of EEG signals and preprocessing them; (2) obtaining all intrinsic mode function signals by fast multidimensional empirical mode decomposition of the preprocessed signals; (3) performing spectral analysis on each layer of each intrinsic mode function signal, selecting signal layers with average power spectra concentrated in the 8-12Hz and 18-26Hz frequency bands as new multidimensional signals; (4) extracting the features of the EEG signals by passing the new multidimensional signals through a spatial filter; (5) inputting the features into a classifier for classification, selecting the optimal parameters in CSP according to the classification accuracy, and using the EEG features under the optimal parameters to classify EEG signals under different motor imagery tasks.

[0005] In recent years, dictionary-based learning methods, a data-driven approach, have demonstrated excellent performance in learning typical waveforms. Convolutional sparse coding, a type of convolutional dictionary learning, decomposes neural signals into the convolution of translation-invariant waveforms and sparse impulse responses. A limitation of existing convolutional sparse coding methods is that most do not explicitly consider the relationship between the learned waveform and the EEG signal state. Furthermore, state specificity is a crucial characteristic of biomarkers, meaning that a signal only frequently appears in a specific state. Therefore, directly applying existing convolutional sparse coding methods for biomarker identification is ineffective, as it is difficult to distinguish state-specific waveforms as potential biomarkers from unrelated, shared waveforms.

[0006] To address this issue, it is necessary to construct a pattern recognition method for EEG signals based on a state-related convolutional sparse model, which can automatically identify typical waveforms specific to a given state, thereby more accurately identifying relevant potential biomarkers. Summary of the Invention

[0007] This invention discloses a method for EEG signal pattern recognition based on a state-related convolutional sparse model. It addresses the problem that existing convolutional sparse coding methods have difficulty in recognizing state-related biomarkers. By clearly distinguishing between state-shared and state-specific waveforms, it improves the accuracy of existing methods in recognizing potential biomarkers. Furthermore, the identified potential biomarkers have good interpretability and can serve as an effective tool for assisting medical diagnosis.

[0008] A method for EEG signal pattern recognition based on a state-correlated convolutional sparse model includes the following steps:

[0009] (1) Obtain the raw multi-state EEG signals, preprocess them, and divide them into training set and test set according to the proportion;

[0010] (2) Construct a state-sensitive convolutional sparse coding model. In this model, the input signal of a certain state is composed of the sparse impulse response of the waveform common to all states, the sparse impulse response of the waveform specific to that state, and the sum of independent Gaussian noise. Among them, the waveform common to all states constitutes the state common dictionary, and the waveform specific to that state constitutes the state specific dictionary.

[0011] (3) Input multi-state EEG signals from the training set and learn each waveform and sparse impulse response using the following alternating optimization strategy:

[0012] (3-1) Fix each waveform, use the matching chase algorithm to optimize the sparse impulse response of each waveform in the input signal, and update the current sparse impulse response;

[0013] (3-2) Fix the sparse impulse response and state-specific waveforms of all waveforms, use the gradient descent method to optimize the common waveforms of all states, and update the common dictionary of the current state.

[0014] (3-3) Fix the sparse impulse response and state-specific waveforms of all waveforms. For each state EEG signal, use the gradient method to optimize the state-specific waveform and update the current state-specific dictionary.

[0015] (3-4) Iterate through steps (3-1) to (3-3) until all waveforms and sparse impulse responses converge or the number of iterations exceeds a preset value;

[0016] (4) Using a dictionary of attentional states as potential markers, the frequency of occurrence of potential markers is estimated based on the EEG signals of each test sample, and the frequency is used to predict whether the test sample is in an attentional state.

[0017] In step (1), a 5th-order Butterworth filter is used for preprocessing.

[0018] In step (2), the constructed convolutional sparse coding model is as follows:

[0019] Let the EEG signal of state c be x. c (t), the state has a total waveform. The waveform unique to state c is The duration of all waveforms is L, and the corresponding sparse impulse response uses the time point at which the waveform begins to appear. With the amplitude of the occurrence Characterization, x c The independent Gaussian noise contained in (t) is ∈ c If (t), then the formula for the model is:

[0020]

[0021] in, and They are respectively and In x c The number of times it appears in (t).

[0022] In step (3-1), the objective function for optimizing the sparse impulse response is as follows:

[0023]

[0024] Among them, the state has a dictionary The corresponding impulse response set State-specific dictionary The corresponding impulse response set The function g represents the convolution of the dictionary with the corresponding sparse impulse response:

[0025]

[0026]

[0027] Regular terms To control the sparsity of the impulse response, λ w This represents the weight of the regularization term.

[0028] In step (3-2), the optimization objective function for the state dictionary D0 is as follows:

[0029]

[0030] y c (t) represents x c (t) The error remaining after removing a specific state component: y c (t)=x c (t)-g(S c D c ,t), function corr metric d i With d j Similarities between them:

[0031]

[0032] in, The time-reversed version of d j To ensure that D0 and D c Distinctiveness, regularity term λ c corr(D0, D) -0 Controlling D0 with other dictionaries D -0 The similarity between them, λ c λ is the weight of this term; s corr(D0, D0) controls the similarity within D0 to avoid duplicate solutions, λ s Let be the weight of this term. To avoid trivial solutions, l2-norm constraints are added to the elements in D0.

[0033] In step (3-3), the optimization objective function for the state-specific dictionary Dc is as follows:

[0034]

[0035] Among them, z c (t) represents x c (t) is derived from D c Components: Regular term λ c corr(D c D -c ) and λ s corr(Dc D c The same logic applies to optimizing the objective function D0 mentioned above.

[0036] In step (4), for the input test sample, the attention state-specific dictionary D is calculated. c The number of times each element appears in D is used to obtain D. c The ratio f of elements in the middle c The formula is as follows:

[0037]

[0038] in, D c The total number of times the element appears in the test sample. Let f be the total number of times each element in D0 appears in the test sample; if f c Greater than the dividing threshold γ learned during training c If the test sample is in state c, then the test sample is considered to be in state c.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. This invention uses state-specific waveforms and state-independent common waveforms to model EEG signals, clearly distinguishing the different components in the state signal.

[0041] 2. By introducing inconsistency constraints between dictionaries, this invention explicitly considers the state-specificity of waveforms during waveform recognition, which is beneficial for more efficient extraction of potential biomarkers related to specific states. The results show that this invention can achieve higher accuracy than existing methods. Attached Figure Description

[0042] Figure 1 This is a flowchart of the EEG signal pattern recognition method based on a state-related convolutional sparse model according to the present invention.

[0043] Figure 2 This is a comparison chart showing the effects of the method of this invention and the comparative method on a real dataset. Detailed Implementation

[0044] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not constitute any limitation thereof.

[0045] This example uses an internal electroencephalogram (iEEG) dataset containing 21 patients with refractory epilepsy. Since sufficient seizure samples are required for training, this example selects 9 of these patients, each with 5 or more seizure samples. The example categorizes EEG signals into the following states: interictal period, preictal period (15 minutes before a seizure), and ictal period.

[0046] This example uses leave-one-out (LOO) to conduct a single-patient epilepsy prediction experiment (classifying pre-ictal and interictal epilepsy). Electrode placement varies significantly among patients due to the heterogeneity of lesion areas; therefore, data from the same patient were used simultaneously for training and testing to maintain consistency in electrode placement and recorded waveforms.

[0047] By trying different parameter values ​​on the dataset, the hyperparameter selection of this invention is as follows: based on the selected filter band (γ) and sampling frequency, the length L of the target waveform is set to 50 sampling time points, and the weight coefficient λ of the regularization term is... c With λ s Choose values ​​of 0.03 and 0.01 respectively, and let w be the size of the dictionary specific to all states. c (c∈{1,...,C}) is set to 2, the size of the state dictionary w0 is set to 5, and the classifier threshold γ used for epilepsy prediction is set to 5. c f is set to be calculated during training. c 90th percentile, maximum number of iterations N iter The selection is 50.

[0048] like Figure 1 As shown, a method for EEG signal pattern recognition based on a state-related convolutional sparse model includes the following steps:

[0049] 1. EEG signal preprocessing and dataset partitioning:

[0050] The raw multi-state EEG signals were acquired and filtered to the γ band using a 5th-order Butterworth filter. The signals were then divided into training and testing sets using a leave-one-out method (leaving out one EEG signal related to an epileptic seizure as the test set). The preprocessed signals were divided into non-overlapping segments with a duration of 15 seconds.

[0051] 2. Construct a state-sensitive convolutional sparse coding model:

[0052] Let the EEG signal of state c be x. c (t), the state has a total waveform. The waveform unique to state c is The duration of all waveforms is L, and the corresponding sparse impulse response uses the time point at which the waveform begins to appear. With the amplitude of the occurrence Characterization, x c The independent Gaussian noise contained in (t) is ∈ c If (t), then the formula for the model is:

[0053]

[0054] in, and They are respectively and In x c The number of times it appears in (t).

[0055] In this example, x c (t) includes two states: preictal and interictal. The total number of waveforms for each state is w0, which is 5, and the number of state-specific waveforms is w c (c∈{1,...,2}) is 2, and the waveform duration is L with 50 sampling points.

[0056] 3. Alternately optimize various waveforms and sparse impulse responses:

[0057] (3-1) Fix each waveform, optimize the sparse impulse response of each waveform in the input signal, and update the current sparse impulse response.

[0058] The objective function for optimizing the sparse impulse response is as follows:

[0059]

[0060] Among them, the state has a dictionary The corresponding impulse response set State-specific dictionary The corresponding impulse response set The function g represents the convolution of the dictionary with the corresponding sparse impulse response:

[0061]

[0062]

[0063] Regular terms To control the sparsity of the impulse response, λ w This represents the weight of the regularization term.

[0064] In this example, the matching chase algorithm is used for optimization.

[0065] (3-2) Fix the sparse impulse response and state-specific waveforms of all waveforms, use gradient descent to optimize the common waveforms of all states, and update the common dictionary of the current state.

[0066] The optimization objective function for the state dictionary D0 is as follows:

[0067]

[0068] y c (t) represents x c (t) The error remaining after removing a specific state component: y c (t)=x c (t)-g(S c D c ,t), function corr metric d i With d j Similarities between them:

[0069]

[0070] in, The time-reversed version of d j .

[0071] To ensure that D0 and D c Distinctiveness, regularity term λ c corr(D0, D) -0 Control D0 and other dictionaries D besides D0 -0 The similarity between them, λ c λ is the weight of this term; s corr(D0, D0) controls the similarity within D0 to avoid duplicate solutions, λ s This is the weight of the term. Furthermore, to avoid trivial solutions, the elements in D0 are constrained by the l2-norm.

[0072] In this example, the waveform of each state in D0 is randomly updated while keeping the other elements in D0 unchanged. Specifically, Defined as only by Composed of y c The components of (t):

[0073]

[0074] Furthermore, in order to use differentiable... Replacing the original corr, this example first uses the LogSumExp(LSE) function to smoothly approximate the maximum value function max; furthermore, d i and The discrete convolution operation between them is rewritten as d i Toeplitz matrix Toe(d) i )and Matrix multiplication between them. Therefore, It can be written in the following form:

[0075]

[0076] Therefore, the optimized state has a common waveform. The objective function can be written in the following form:

[0077]

[0078] This example uses stochastic gradient descent (SGD) to update. because To improve computational efficiency, sparse activation can be addressed by optimizing only sparse activation. Existing signal segments.

[0079] (3-3) Fix the sparse impulse response and state-specific waveforms of all waveforms. For each state EEG signal, use the gradient method to optimize the state-specific waveform and update the current state-specific dictionary.

[0080] State-specific dictionary D c The optimization objective function is as follows:

[0081]

[0082] Among them, z c (t) represents x c (t) is derived from D c Components: Regular term λ c corr(D c D -c ) and λ s corr(D c D c The same logic applies to optimizing the objective function D0 mentioned above.

[0083] In this example, while keeping D c Randomly update D while keeping the other elements unchanged. c Each state in the waveform has its own unique waveform. The specific update method is similar to step (3-2).

[0084] (3-4) Iterate through steps (3-1) to (3-3) until all waveforms and sparse impulse responses converge or the number of iterations exceeds a preset value.

[0085] In this example, the maximum number of iterations N iter The preset value is 50, and the convergence criterion is that the difference between the waveform calculated in two adjacent iterations and the sparse impulse response is less than the preset value 1e-6.

[0086] 4. Verify the learned state-specific dictionary:

[0087] Using a dictionary of attentional states as potential biomarkers, the frequency of occurrence of potential biomarkers is estimated based on the EEG signals of each test sample, and the frequency is used to predict whether the test sample is in an attentional state.

[0088] For the input test sample, calculate the attention state-specific dictionary D. c The number of times each element appears in D is used to obtain D. c The ratio f of elements in the middle c The formula is as follows:

[0089]

[0090] in, D c The total number of times the element appears in the test sample. Let f be the total number of times each element in D0 appears in the test sample; if f c Greater than the dividing threshold γ learned during training c If the test sample is in state c, then the test sample is considered to be in state c.

[0091] In this example, attention state c is the pre-ictal state, and D... c A dictionary specific to the pre-seizure phase of epilepsy; a classifier threshold γ used for epilepsy prediction. c f is set to be calculated during training. c The 90th percentile.

[0092] 5. Performance evaluation of latent marker identification methods:

[0093] Pre-epileptic samples were designated as positive samples, while interictal samples were designated as negative samples. Since each test sample contained only one seizure, the focus of the results was on precision, specifically the accuracy in classifying the sample as pre-epileptic. Precision was calculated as follows: Where tp represents the number of true positive results and fp represents the number of false positive results; the precision given in this example is the average of 5 random leave-one-out cross-validations.

[0094] like Figure 2 As shown, using an internal electroencephalogram (iEEG) dataset from patients with refractory epilepsy, the f-values ​​of test samples during the preictal and interictal periods were plotted. c Distribution. It can be seen that, compared with existing methods based on convolutional sparse coding, this invention achieves better results in the pre-seizure and inter-seizure phases of certain patients. c The distribution showed more significant differences, thus demonstrating higher accuracy.

[0095] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for EEG signal pattern recognition based on a state-related convolutional sparse model, characterized in that, Includes the following steps: (1) Obtain the raw multi-state EEG signals, preprocess them, and divide them into training set and test set according to the proportion; (2) Construct a state-sensitive convolutional sparse coding model. In this model, the input signal of a certain state is composed of the sparse impulse response of the waveform common to all states, the sparse impulse response of the waveform specific to that state, and the sum of independent Gaussian noise. Among them, the waveform common to all states constitutes the state common dictionary, and the waveform specific to that state constitutes the state specific dictionary. (3) Input multi-state EEG signals from the training set and learn each waveform and sparse impulse response using the following alternating optimization strategy: (3-1) Fix each waveform, use the matching chase algorithm to optimize the sparse impulse response of each waveform in the input signal, and update the current sparse impulse response; (3-2) Fix the sparse impulse response and state-specific waveforms of all waveforms, use the gradient descent method to optimize the common waveforms of all states, and update the common dictionary of the current state. (3-3) Fix the sparse impulse response and state-specific waveforms of all waveforms. For each state EEG signal, use the gradient method to optimize the state-specific waveform and update the current state-specific dictionary. (3-4) Iterate through steps (3-1) to (3-3) until all waveforms and sparse impulse responses converge or the number of iterations exceeds a preset value; (4) Using a dictionary of attentional states as potential markers, the frequency of occurrence of potential markers is estimated based on the EEG signals of each test sample, and the frequency is used to predict whether the test sample is in an attentional state.

2. The EEG signal pattern recognition method based on a state-related convolutional sparse model according to claim 1, characterized in that, In step (1), a 5th-order Butterworth filter is used for preprocessing.

3. The EEG signal pattern recognition method based on a state-related convolutional sparse model according to claim 1, characterized in that, In step (2), the constructed convolutional sparse coding model is as follows: Let the EEG signal of state c be x. c (t), the state has a total waveform. The waveform unique to state c is The duration of all waveforms is L, and the corresponding sparse impulse response uses the time point at which the waveform begins to appear. With the amplitude of the occurrence Characterization, x c The independent Gaussian noise contained in (t) is ∈ c If (t), then the formula for the model is: in, and They are respectively and In x c The number of times it appears in (t).

4. The EEG signal pattern recognition method based on a state-related convolutional sparse model according to claim 1, characterized in that, In step (3-1), the objective function for optimizing the sparse impulse response is as follows: Among them, the state has a dictionary The corresponding impulse response set State-specific dictionary The corresponding impulse response set The function g represents the convolution of the dictionary with the corresponding sparse impulse response: Regular terms To control the sparsity of the impulse response, λ w This represents the weight of the regularization term.

5. The EEG signal pattern recognition method based on a state-related convolutional sparse model according to claim 1, characterized in that, In step (3-2), the optimization objective function for the state dictionary D0 is as follows: y c (t) represents x c (t) The error remaining after removing a specific state component: y c (t)=x c (t)-g(S c D c ,t), function corr metric d i With d j Similarities between them: in, The time-reversed version of d j To ensure that D0 and D c Distinctiveness, regularity term λ c corr(D0,D -0 Control D0 and other dictionaries D -0 The similarity between them, λ c λ is the weight of this term. s corr(D0,D0) controls the similarity within D0 to avoid duplicate solutions, λ s This is the weight of that item.

6. The EEG signal pattern recognition method based on a state-related convolutional sparse model according to claim 5, characterized in that, To avoid trivial solutions, the elements in D0 are constrained by the l2 norm.

7. The EEG signal pattern recognition method based on a state-related convolutional sparse model according to claim 6, characterized in that, In step (3-3), the state-specific dictionary D c The optimization objective function is as follows: Among them, z c (t) represents x c (t) is derived from D c Components: Regularization term λ c corr(D c D -c ) and λ s corr(D c D c The same logic applies to optimizing the objective function D0 mentioned above.

8. The EEG signal pattern recognition method based on a state-related convolutional sparse model according to claim 1, characterized in that, In step (4), for the input test sample, the attention state-specific dictionary D is calculated. c The number of times each element appears in D is used to obtain D. c The ratio f of elements in the middle c The formula is as follows: in, D c The total number of times the element appears in the test sample. Let f be the total number of times each element in D0 appears in the test sample; if f c Greater than the splitting threshold γ learned during training c If the test sample is in state c, then the test sample is considered to be in state c.

Images