EEG signal classification method based on wavelet scattering transform and differential entropy feature fusion

Abstract

The present invention discloses an EEG signal classification method that fuses wavelet scattering transform and differential entropy features, belongs to the technical field of EEG signal classification, and is used for EEG signal classification, including importing collected EEG epilepsy signals, inputting samples into a wavelet scattering transform network, performing scattering transform on channel signals, extracting second-order decomposition scattering coefficients, and obtaining time-frequency domain features of channel signals; performing differential entropy calculation on channel signals, obtaining nonlinear dynamic features of channel signals, performing discriminant correlation analysis on time-frequency domain features and nonlinear dynamic features, and constructing multimodal fusion features; inputting fusion features into a long short-term memory neural network for training and learning, and evaluating the results using average classification accuracy. The present invention reduces feature dimension through feature fusion methods by analyzing the time-frequency domain features and nonlinear dynamic features of signals, thereby improving classification accuracy, better distinguishing EEG signals of epileptic patients from those of normal people, and providing effective assistance for diagnosis and treatment.

Description

Technical Field

[0001] The invention discloses an electroencephalogram (EEG) signal classification method integrating wavelet scattering transform and differential entropy features, and belongs to the technical field of EEG signal classification. Background Art

[0002] Epilepsy is a chronic non-communicable disease caused by abnormal paroxysmal hypersynchronous electrical activity of brain neurons. Epileptic seizures are often sudden and uncontrollable, and it is one of the most common neurological diseases in the world. The World Health Organization reports that there are approximately 50 million epilepsy patients worldwide, and epilepsy is also listed as one of the five most important neurological and psychiatric diseases in the world. Currently, visual inspection of the electroencephalogram by neurologists is the main method for clinical diagnosis of epilepsy. However, this traditional method is not only time-consuming and inefficient, but also has a strong subjective judgment factor. Therefore, the development of an effective and reliable automatic detection method for epilepsy has become an urgent problem to be solved, and it also has strong clinical significance.

[0003] Electroencephalogram (EEG) is a representation of the electrical activity on the surface of the brain relative to time. Abnormal discharge of brain neurons causes epileptic seizures. Therefore, the period of epileptic seizures can be recorded by recording scalp EEG signals, thereby obtaining a large amount of data for the study of EEG signal recognition during the epileptic seizure stage. The development of deep learning technology has provided new methods and new ideas for epilepsy diagnosis. Deep learning technology has more advantages than traditional methods in dealing with complex nonlinear problems, which can greatly reduce manual participation and improve recognition accuracy. It is generally believed that the brain's advanced cognitive functions are usually related to the time-frequency and nonlinear dynamic characteristics of EEG signals. In order to make full use of the above characteristics, the wavelet scattering transform method is used to extract the time-frequency domain features, and the differential entropy of the EEG signal is calculated to extract the nonlinear dynamic features. The wavelet scattering transform is an extension of the wavelet transform. Inspired by the convolutional neural network, the signal is convolved using a series of low-pass filter groups, and a series of wavelet scattering coefficients of different orders are obtained by taking the modulus and local averaging. Then, the differential entropy feature of the signal is extracted. The entropy feature is a measure of the average uncertainty and complexity of the signal. For typical continuous non-stationary signals such as EEG, the entropy feature can well characterize the signal uncertainty, thereby increasing the algorithm's ability to distinguish different types of epileptic EEG signals.

[0004] After extracting the time-frequency domain features and nonlinear dynamic features of epileptic signals, the discriminant correlation analysis feature fusion algorithm is used to fuse the two features. A large number of studies have shown that fusion features are more helpful in improving the classification accuracy of the classification model than single features. Finally, the multimodal fusion feature vector is input as an input feature into the long short-term memory recurrent neural network (LSTM) for training, and the trained model is tested using test data to obtain the classification accuracy of the signal. Summary of the invention

[0005] The purpose of the present invention is to provide an EEG signal classification method integrating wavelet scattering transform and differential entropy features, so as to solve the problem of inaccurate EEG signal classification in the prior art.

[0006] The EEG signal classification method based on wavelet scattering transform and differential entropy feature fusion includes:

[0007] S1 imports the collected EEG epilepsy signal and performs filtering preprocessing operations;

[0008] S2 performs segmentation operations on the preprocessed EEG epilepsy signals and constructs a sample library through data enhancement;

[0009] S3 inputs the sample into the wavelet scattering transform network, performs scattering transform on the channel signal, extracts the second-order decomposition scattering coefficient, and obtains the time-frequency domain characteristics of the channel signal;

[0010] S4 performs differential entropy calculation on the channel signal to obtain the nonlinear dynamic characteristics of the channel signal;

[0011] S5 performs discriminant correlation analysis on time-frequency domain features and nonlinear dynamic features to construct multimodal fusion features;

[0012] S6 inputs the fusion features into the long short-term memory neural network for training and learning, and uses the average classification accuracy to evaluate the results.

[0013] The filtering includes using a bandpass filter to filter the EEG epileptic signal.

[0014] S2 includes dividing the entire EEG epilepsy signal into several h-second segments with h seconds as a window according to the epilepsy detection task, and each segment of data has 50% data overlap when cutting the segments.

[0015] S3 includes using wavelet scattering transform to filter and analyze the input signal by cascading iterative multi-scale complex wavelet transform, modular operation and local mean operation to obtain multi-level scattering transform coefficients of the signal;

[0016] For the one-dimensional input signal , using complex analytical complex bandpass filter Perform convolution, use the modulus algorithm and take the modulus value, and obtain the first-order scattering coefficient through averaging :

[0017] ;

[0018] In the formula, is the complex wavelet function, is a low pass filter.

[0019] S3 includes using a second high-frequency complex wavelet to recover high-frequency energy, convolving the wavelet modulus with bandpass filters of different scales, and passing the high-frequency coefficients through Find the local average and get the second-order scattering coefficient :

[0020] ;

[0021] In the formula, yes The complex wavelet function of the adjacent fine scale.

[0022] S4 includes differential entropy for:

[0023] ;

[0024] In the formula, represents the probability density function of continuous information, Indicates the range of information values.

[0025] S4 includes a Gaussian distribution for a specific length The differential entropy of the EEG epileptic signal is expressed as:

[0026] ;

[0027] In the formula, is the mean of the Gaussian distribution, is the variance of Gaussian distribution, and the differential entropy is equal to the logarithm of the energy spectrum in a specific frequency band. Represents sample input.

[0028] S5 includes reducing feature dimensions through discriminant correlation analysis, analyzing the correlation between variables to enhance the correlation within the class and reduce the correlation between classes;

[0029] Assume that the characteristic matrix of wavelet scattering transform coefficients is , the differential entropy feature matrix is , respectively calculate the matrix , The matrix after dimensionality reduction ,matrix :

[0030] ; ;

[0031] In the formula, , , , is the transformation matrix, for The spatial projection of for The final transformation matrix is for The spatial projection of for The final transformation matrix.

[0032] S5 includes obtaining the final input feature vector by adding and fusion :

[0033] .

[0034] Long short-term memory neural network is used as the classifier, the fusion features are used as input variables for training, and the classification accuracy is used as the evaluation criterion to verify the classifier performance.

[0035] Compared with the prior art, the present invention has the following beneficial effects: the present invention analyzes the time-frequency domain characteristics and nonlinear dynamic characteristics of EEG epilepsy signals, reduces the feature dimension through feature fusion method, improves the classification accuracy, and thus better distinguishes the EEG signals of epilepsy patients from those of normal people, providing effective assistance for diagnosis and treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a technical flow chart of the present invention;

[0037] Figure 2 This is the result graph before the data set signal is filtered;

[0038] Figure 3 It is the result image after the data set signal is filtered;

[0039] Figure 4 It is the result diagram of the embodiment. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] The EEG signal classification method based on wavelet scattering transform and differential entropy feature fusion includes:

[0042] S1 imports the collected EEG epilepsy signal and performs filtering preprocessing operations;

[0043] S2 performs segmentation operations on the preprocessed EEG epilepsy signals and constructs a sample library through data enhancement;

[0044] S3 inputs the sample into the wavelet scattering transform network, performs scattering transform on the channel signal, extracts the second-order decomposition scattering coefficient, and obtains the time-frequency domain characteristics of the channel signal;

[0045] S4 performs differential entropy calculation on the channel signal to obtain the nonlinear dynamic characteristics of the channel signal;

[0046] S5 performs discriminant correlation analysis on time-frequency domain features and nonlinear dynamic features to construct multimodal fusion features;

[0047] S6 inputs the fusion features into the long short-term memory neural network for training and learning, and uses the average classification accuracy to evaluate the results.

[0048] The filtering includes using a bandpass filter to filter the EEG epileptic signal.

[0049] S2 includes dividing the entire EEG epilepsy signal into several h-second segments with h seconds as a window according to the epilepsy detection task, and each segment of data has 50% data overlap when cutting the segments.

[0050] S3 includes using wavelet scattering transform to filter and analyze the input signal by cascading iterative multi-scale complex wavelet transform, modular operation and local mean operation to obtain multi-level scattering transform coefficients of the signal;

[0051] For the one-dimensional input signal , using complex analytical complex bandpass filter Perform convolution, use the modulus algorithm and take the modulus value, and obtain the first-order scattering coefficient through averaging :

[0052] ;

[0053] In the formula, is the complex wavelet function, is a low pass filter.

[0054] S3 includes using a second high-frequency complex wavelet to recover high-frequency energy, convolving the wavelet modulus with bandpass filters of different scales, and passing the high-frequency coefficients through Find the local average and get the second-order scattering coefficient :

[0055] ;

[0056] In the formula, yes The complex wavelet function of the adjacent fine scale.

[0057] S4 includes differential entropy for:

[0058] ;

[0059] In the formula, represents the probability density function of continuous information, Indicates the range of information values.

[0060] S4 includes a Gaussian distribution for a specific length The differential entropy of the EEG epileptic signal is expressed as:

[0061] ;

[0062] In the formula, is the mean of the Gaussian distribution, is the variance of Gaussian distribution, and the differential entropy is equal to the logarithm of the energy spectrum in a specific frequency band. Represents sample input.

[0063] S5 includes reducing feature dimensions through discriminant correlation analysis, analyzing the correlation between variables to enhance the correlation within the class and reduce the correlation between classes;

[0064] Assume that the characteristic matrix of wavelet scattering transform coefficients is , the differential entropy feature matrix is , respectively calculate the matrix , The matrix after dimensionality reduction ,matrix :

[0065] ; ;

[0066] In the formula, , , , is the transformation matrix, for The spatial projection of for The final transformation matrix is for The spatial projection of for The final transformation matrix.

[0067] S5 includes obtaining the final input feature vector by adding and fusion :

[0068] .

[0069] Long short-term memory neural network is used as the classifier, the fusion features are used as input variables for training, and the classification accuracy is used as the evaluation criterion to verify the classifier performance.

[0070] At present, the signals collected by EEG equipment are usually subject to many noise interferences, such as power frequency interference, eye movement artifacts, etc. The epilepsy signals collected by EEG equipment are also typical nonlinear and non-stationary signals. The quality problem of EEG signals undoubtedly increases the difficulty of improving the accuracy of the signal classification network. Therefore, in order to make the classification network obtain higher classification accuracy, it is necessary to perform pre-processing operations such as filtering on the EEG data. First, the frequency range where the power frequency interference is located is eliminated. According to different acquisition equipment, foreign countries generally have 60hz interference, and my country generally has 50hz interference. The bandpass filter finally retains the data in the frequency range of 0.5-42Hz. After that, the scalp EEG signal is divided into time periods according to the task detection requirements. Since epilepsy EEG data belongs to class imbalanced data, in order to solve this problem and increase the availability of data, each task data has 50% data overlap. The processed and divided EEG signal data is simultaneously subjected to wavelet scattering transform and differential entropy calculation to obtain the decomposition scattering coefficients and differential entropy coefficients of the signal at each level, and the time-frequency domain features and nonlinear dynamic features are obtained for discriminant correlation analysis to construct a multimodal fusion feature vector. Finally, the fusion feature vector is input as training data into the long short-term memory recurrent neural network (LSTM) for classification. This method has excellent results and can effectively distinguish the EEG signals of epileptic patients during the attack period, the interictal period and healthy people.

[0071] The present invention is obtained by collecting electrodes placed on the surface of the patient's scalp to record the changes in electrical activity caused by brain area activity. The experimental database comes from the epilepsy EEG database of the University of Bonn in Germany, and contains 5 subsets, named A, B, C, D, and E. The data is recorded using a standard 10-20 electrode placement system. The data acquisition frequency is 173.61Hz, and each subset contains 100 single-channel segments, divided into 23.6s time periods. Each EEG time series consists of 4096 samples.

[0072] The EEG signal was filtered using a bandpass filter, where the cutoff frequency of the low-pass filter was 42 Hz and the pass frequency of the high-pass filter was 0.5 Hz. According to the epilepsy detection task, the entire EEG data was divided into several 6-second segments with a window of 6 seconds; the signal after scattering transformation has translation invariance and local stability.

[0073] In order to recover the high-frequency coefficients, a second high-frequency complex wavelet is used to recover the high-frequency energy. The specific method is to convolve the wavelet modulus with bandpass filters of different scales again, but the above process will lead to a decrease in the invariance when recovering the high-frequency components. In order to ensure the stability of the high frequency, the output coefficients with local translation invariance are obtained, and the high-frequency coefficients are locally averaged to obtain the second-order scattering coefficients.

[0074] In the embodiment, the epilepsy EEG database of the University of Bonn in Germany contains EEG data of epilepsy patients and normal people, which are divided into 5 subsets, named A, B, C, D, and E. Among them, the A and B subsets are normal people's data, and the C, D, and E subsets are epilepsy patients' data. The data is recorded using a standard 10-20 electrode placement system, with a data acquisition frequency of 173.61Hz and an acquisition time of 23.6s. Each subset contains 100 single-channel segments, a total of 409,600 samples. This embodiment divides 4 different classification tasks according to the epilepsy detection task, namely normal-epileptic seizure period (AE, BE, AB-E), epileptic seizure interval-epileptic seizure period (CE, DE, CD-E), normal-epilepsy (AB-CDE), and non-epileptic seizure-epileptic seizure (ABCD-E).

[0075] Since the original EEG signal acquisition process is affected by the 50Hz or 60Hz power frequency signal and is easily disturbed by uncontrollable external factors, the scalp EEG often has a certain randomness. In order to conduct subsequent research and analysis on the characteristics of the EEG signal, it is necessary to perform a series of preprocessing operations on the original EEG signal without losing effective information to reduce the classification error rate. Most of the effective information of the EEG signal is concentrated between 1-60Hz, so bandpass filtering is a common EEG signal preprocessing method. The present invention uses a bandpass filter to perform preprocessing operations on the EEG signal, wherein the low-pass filter cutoff frequency is 42hz, and the high-pass filter pass frequency is 0.5hz, and finally obtains an EEG signal with a frequency range of 0.5-42Hz.

[0076] According to the requirements of the epilepsy detection task, 6s of data is defined as a window, and a 23.6s EEG epilepsy signal data is divided into 4 6-second segments, each data set is divided into 400 data segments, and the number of samples in each segment is 1024; EEG signals belong to class imbalanced data, with non-epileptic seizure time accounting for the majority and seizure time accounting for the minority. The imbalance of the number of samples will lead to a decrease in classification effect. In order to improve the efficiency of data utilization, data expansion is required. On the basis of the segmentation in step 2-1, each segment of data is overlapped by 50%. 3s of the data in every 6s of data will be repeated with the previous segment. The imbalanced data is expanded and reconstructed, and the sample size is increased, so that each 23.6s data can be divided into 7 valid segments, each containing 1024 samples.

[0077] By improving the basic wavelet transform, the wavelet scattering transform can obtain the multi-level scattering transform coefficient of the signal after filtering and analyzing the input signal in a cascade manner of iterative multi-scale complex wavelet transform, modular operation and local mean operation. The signal after scattering transformation has translation invariance and local stability. In this embodiment, the wavelet scattering transform is set to include 8 wavelet filters per octave for the first-order wavelet convolution, 1 wavelet filter per octave for the second-order wavelet convolution, and the invariance scale is set to 64.

[0078] Discriminant correlation analysis (DCA) is performed on time-frequency domain features and nonlinear dynamic features to construct multimodal fusion features. Feature fusion methods can integrate low-level or mid-level features extracted from different modules into a single representation, thereby capturing the comprehensive and discriminative information provided by multiple features to enhance the overall representation and improve subsequent processing tasks.

[0079] In the feature fusion stage, more representative features are obtained according to the DCA algorithm and combined together by addition. The DCA method uses a diagonalized scatter matrix to separate features of different categories, and uses singular value decomposition to transform two groups of features to increase the correlation between features of the same category. The resulting fusion features not only enhance the connection between features of the same category, but also highlight the differences between features of different categories.

[0080] Use LSTM network as classifier, train fusion features as input variables, and verify classifier performance with classification accuracy as evaluation criteria. LSTM network uses time series to analyze input, which can solve the long-term dependency problem prevalent in general recursive neural networks. Using LSTM can effectively transmit information in time series and will not ignore the valid information before the long time series, and can also solve the gradient disappearance and gradient explosion problems in RNN. In this embodiment, the fusion features obtained by step 5 are used as input samples, divided into test sets and training sets according to the ratio of 2:8, the hidden layer nodes of LSTM network are set to 100, the learning rate is set to 0.01, and five-fold cross validation is used. At the same time, the influence of wavelet scattering transforms of different orders on the classification results is compared, and Table 1 lists the classification results under wavelet scattering transforms of different orders.

[0081] Table 1 Statistics of test set classification results

[0082] ;

[0083] In Table 1, Acc is the classification accuracy. It can be seen from Table 1 that with the increase of the order of wavelet scattering coefficients, the classification accuracy of each task also increases, especially for AE, CE, DE, AB-E, ABCD-E tasks, which can achieve 100% classification accuracy. Based on the above results, it can be seen that the order selection of wavelet scattering transform is very important for the experimental results. The results obtained by selecting the second-order coefficient combined with the differential entropy feature and performing discriminant correlation analysis have achieved very good results, indicating that the model has a good distinction between epileptic signals and normal signals, which meets the expected goals. It should be noted that the above classification results can be used to assist in determining the category of EEG signals, but are not directly used for the diagnosis and treatment of diseases. The classification results can be used in scientific research, model testing and other fields.

[0084] To solve the problem of unbalanced sample quantity, the present invention uses the fragment overlapping method to expand the data, which is conducive to improving the model training and classification capabilities; the wavelet scattering transform and the differential entropy algorithm can effectively extract the features of complex nonlinear signals, and the fused features have good robustness; the long short-term memory recursive neural network (LSTM) is used as a classifier, which can capture the long-term dependencies of temporal signals and effectively improve the accuracy of EEG epilepsy signal classification.

[0085] The technical process of the present invention is as follows Figure 1 , use a bandpass filter to filter the collected data and divide the task segments, and then conduct two branches. One branch uses wavelet scattering transform to obtain time-frequency domain features, and the other branch calculates the differential entropy of the signal to obtain nonlinear dynamic features. Then both branches use the discriminant correlation analysis algorithm for feature fusion. Finally, the obtained fused feature vector is input into the long short-term memory network as training data, and the epilepsy data signal is classified.

[0086] Figure 2 It is the result diagram before the data set signal is filtered, corresponding to the epileptic signal attack segment of data set A. The horizontal axis is the number of data acquisitions, and the vertical axis is the signal amplitude; Figure 3 It is the result diagram after the data set signal is filtered. The horizontal axis of the epileptic signal seizure segment corresponding to data set A is the number of data acquisitions, and the vertical axis is the corresponding signal amplitude after the filtering operation; Figure 4 3 is a result diagram of the embodiment, the horizontal axis is the classification task, the vertical axis is the classification accuracy corresponding to the eight classification tasks, and the classification task accuracy under different wavelet coefficients is statistically analyzed.

[0087] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features may be replaced by equivalents, and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The EEG signal classification method based on wavelet scattering transform and differential entropy feature fusion is characterized by: include: S1 imports the collected EEG epilepsy signal and performs filtering preprocessing operations; S2 performs segmentation operations on the preprocessed EEG epilepsy signals and constructs a sample library through data enhancement; S3 inputs the sample into the wavelet scattering transform network, performs scattering transform on the channel signal, extracts the second-order decomposition scattering coefficient, and obtains the time-frequency domain characteristics of the channel signal; S4 performs differential entropy calculation on the channel signal to obtain the nonlinear dynamic characteristics of the channel signal; S5 performs discriminant correlation analysis on time-frequency domain features and nonlinear dynamic features to construct multimodal fusion features; S6 inputs the fusion features into the long short-term memory neural network for training and learning, and uses the average classification accuracy to evaluate the results; The filtering includes using a bandpass filter to filter the EEG epileptic signal; S2 includes dividing the entire EEG epilepsy signal into several h-second segments with h seconds as a window according to the epilepsy detection task, and each segment of data has 50% data overlap when intercepting the segments; S3 includes using wavelet scattering transform to filter and analyze the input signal by cascading iterative multi-scale complex wavelet transform, modular operation and local mean operation to obtain multi-level scattering transform coefficients of the signal; For the one-dimensional input signal , using complex analytical complex bandpass filter Perform convolution, use the modulus algorithm and take the modulus value, and obtain the first-order scattering coefficient through averaging : ； In the formula, is the complex wavelet function, is a low-pass filter; S3 includes using a second high-frequency complex wavelet to recover high-frequency energy, convolving the wavelet modulus with bandpass filters of different scales, and passing the high-frequency coefficients through Find the local average and get the second-order scattering coefficient : ； In the formula, yes The complex wavelet function of the adjacent fine scale; S4 includes differential entropy for: ； In the formula, represents the probability density function of continuous information, Indicates the interval of information value; S4 includes a Gaussian distribution for a specific length The differential entropy of the EEG epileptic signal is expressed as: ； In the formula, is the mean of the Gaussian distribution, is the Gaussian distribution variance, and the differential entropy is equal to the logarithm of the energy spectrum in a specific frequency band. represents sample input; S5 includes reducing feature dimensions through discriminant correlation analysis, analyzing the correlation between variables to enhance the correlation within the class and reduce the correlation between classes; Assume that the characteristic matrix of wavelet scattering transform coefficients is , the differential entropy feature matrix is , respectively calculate the matrix , The matrix after dimensionality reduction ,matrix : ； ； In the formula, , , , is the transformation matrix, for The spatial projection of for The final transformation matrix is for The spatial projection of for The final transformation matrix of S5 includes obtaining the final input feature vector by adding and fusion : ； Long short-term memory neural network is used as the classifier, the fusion features are used as input variables for training, and the classification accuracy is used as the evaluation criterion to verify the classifier performance.

Images