A method and system for detecting epileptic EEG signals based on statistical channel selection

Abstract

The present invention discloses a method and system for detecting epileptic EEG signals based on statistical channel selection, the method comprising: performing data preprocessing on multi-channel EEG signals by a noise reduction and artifact removal algorithm; performing channel selection processing on the preprocessed multi-channel EEG signals based on the maximum variance principle and the maximum mutual information principle; performing three-dimensional trajectory reconstruction processing on the target EEG signal channel; identifying and storing the target EEG signal channel feature vector based on deterministic learning theory and in combination with a radial basis function neural network; designing a dynamic estimator to classify the dynamic pattern recognition library of the EEG signal channel characteristics to obtain a classification result. The embodiments of the present invention can improve the recognition efficiency and accuracy of the dynamic pattern of EEG signals of epileptic seizures. As a method and system for detecting epileptic EEG signals based on statistical channel selection, the present invention can be widely used in the field of deep learning technology for EEG signals.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning of electroencephalogram signals, and in particular to a method and system for detecting epileptic electroencephalogram signals based on statistical channel selection. Background Art

[0002] Epilepsy is a neurological disease caused by abnormal hypersynchronous discharges of brain neuron cell groups. The morphology of electroencephalogram (EEG) signals changes significantly during epileptic seizures. Traditional research methods for epileptic seizure detection based on machine learning mainly extract the features of EEG signals through linear dynamic feature analysis (time domain analysis, frequency domain analysis, time-frequency domain analysis) and nonlinear dynamic analysis. After the features are input into the classifier and relevant thresholds are set, epileptic seizure detection is achieved. However, these methods usually assume that EEG signals are linear and stable, which cannot accurately reflect the non-stationary characteristics of the signals. In addition, due to the high dimensionality and complexity of EEG signals, the difficulty of model training is increased, which may lead to information redundancy and model overfitting. In addition, compared with traditional machine learning methods, deep learning models use Feature extraction is integrated with classifier design, and the features in the data are automatically learned, without the need for manual selection and feature extraction of epileptic seizure feature types. However, it requires a large amount of labeled data to train the network and has a strong dependence on data. The deep learning model is an end-to-end black box model, and its black box nature makes the internal decision-making process lack interpretability. Since EEG signals are generally multi-channel and long-term recorded data, this requires the equipment to have powerful computing power and high hardware support. It cannot be ignored that patients will inevitably move their eyeballs, limbs, and muscles during the collection process. These movements will generate artifact signals such as electrooculography and electromyography, which will cause certain interference to the interpretation of the EEG. Summary of the invention

[0003] In order to solve the above technical problems, the purpose of the present invention is to provide an epileptic EEG signal detection method and system based on statistical channel selection, which can improve the recognition efficiency and accuracy of the dynamic pattern of epileptic EEG signals.

[0004] The first technical solution adopted by the present invention is: a method for detecting epileptic EEG signals based on statistical channel selection, comprising the following steps:

[0005] Preprocessing the multi-channel EEG signal data by using a noise reduction and artifact removal algorithm to obtain a preprocessed multi-channel EEG signal, wherein the noise reduction and artifact removal algorithm includes an independent component analysis algorithm and an improved adaptive noise complete set empirical mode decomposition algorithm;

[0006] Based on the maximum variance principle and the maximum mutual information principle, the preprocessed multi-channel EEG signal is subjected to channel selection processing to obtain the target EEG signal channel;

[0007] Perform three-dimensional trajectory reconstruction on the target EEG signal channel to obtain the target EEG signal channel feature vector;

[0008] Based on the deterministic learning theory and combined with the radial basis function neural network, the target EEG signal channel feature vector is identified and stored, and a dynamic pattern recognition library of EEG signal channel features is constructed;

[0009] A dynamic estimator is designed to classify and process the dynamic pattern recognition library of EEG signal channel characteristics, and the interictal EEG signals and pre-ictal EEG signals of epileptic patients are obtained.

[0010] Furthermore, the step of performing data preprocessing on the multi-channel EEG signal by using a noise reduction and artifact removal algorithm to obtain a preprocessed multi-channel EEG signal specifically includes:

[0011] Acquire multi-channel EEG signals of epilepsy patients and perform format matching processing to obtain multi-channel EEG signals with time series;

[0012] The artifact signal is filtered out of the multi-channel EEG signal with time series by using the independent component analysis algorithm to obtain the multi-channel EEG signal after preliminary filtering;

[0013] The multi-channel EEG signal after preliminary filtering is subjected to noise filtering by using an improved adaptive noise complete set empirical mode decomposition algorithm to obtain a multi-channel EEG signal after filtering.

[0014] The ideal frequency range of the EEG signal is set, and the EEG frequency is screened and processed on the filtered multi-channel EEG signal to obtain the preprocessed multi-channel EEG signal.

[0015] Furthermore, the step of performing channel selection processing on the preprocessed multi-channel EEG signal based on the maximum variance principle and the maximum mutual information principle to obtain the target EEG signal channel specifically includes:

[0016] Perform channel mean calculation on the preprocessed multi-channel EEG signal to obtain the EEG signal channel mean;

[0017] Determine the variance value of the EEG signal channel according to the mean value of the EEG signal channel, and select the EEG signal channel corresponding to the maximum variance value according to the maximum variance principle to obtain the first EEG signal channel;

[0018] According to the maximum mutual information principle, the Pearson correlation coefficients of several EEG signal channels with the first EEG signal channel are obtained and arranged in descending order, and the EEG signal channels corresponding to the first n Pearson correlation coefficients are selected as the second EEG signal channels;

[0019] The first EEG signal channel and the second EEG signal channel are combined to obtain a target EEG signal channel.

[0020] Furthermore, the step of performing three-dimensional trajectory reconstruction processing on the target EEG signal channel to obtain the target EEG signal channel feature vector specifically includes:

[0021] Through the phase space reconstruction method, the three-dimensional trajectory of the target EEG signal channel is reconstructed to obtain the phase space of the reconstructed time series of the EEG signal;

[0022] Setting feature extraction calculation indicators, wherein the feature extraction calculation indicators include a Euclidean distance feature indicator, an average energy feature indicator, and a maximum Lyapunov exponent feature indicator;

[0023] According to the feature extraction calculation index, the phase space of the reconstructed time series of the EEG signal is subjected to feature extraction processing to obtain the target EEG signal channel feature vector.

[0024] Furthermore, the step of identifying and storing the target EEG signal channel feature vector based on the deterministic learning theory and in combination with the radial basis function neural network to construct a dynamic pattern recognition library of EEG signal channel features specifically includes:

[0025] Based on the deterministic learning theory, the local dynamics identification processing is performed on the target EEG signal channel feature vector to obtain the dynamic pattern data of the EEG signal channel feature;

[0026] The dynamic pattern data of EEG signal channel characteristics are converted into constant weight storage based on radial basis function neural network, and a dynamic pattern recognition library of EEG signal channel characteristics is constructed.

[0027] Furthermore, the step of converting the dynamic pattern data of the EEG signal channel characteristics into constant weight storage based on the radial basis function neural network to construct a dynamic pattern recognition library of the EEG signal channel characteristics specifically includes:

[0028] Considering the differences between EEG signal channel characteristics in different dynamic modes, the dynamic mode data of EEG signal channel characteristics are modeled and processed through dynamic functions to construct a dynamic model of EEG signal channel characteristics.

[0029] According to the dynamic model of EEG signal channel characteristics, an identifier based on radial basis function neural network is designed;

[0030] The state error and weight estimation error of the dynamic model of the EEG signal channel characteristics are obtained, and the weights of the identifier based on the radial basis function neural network are adaptively updated;

[0031] According to the weight convergence of radial basis function neural network, the weight of the identifier corresponding to the mean of the final steady-state process is selected as the constant weight to construct a dynamic pattern recognition library of EEG signal channel characteristics.

[0032] Further, the step of designing a dynamic estimator to classify and process the dynamic pattern recognition library of EEG signal channel features to obtain the interictal EEG signals of epileptic patients and the preictal EEG signals of epileptic patients specifically includes:

[0033] A dynamic estimator is designed, a constant weight is embedded into the dynamic estimator, and state estimation processing is performed on the dynamic pattern recognition library of EEG signal channel characteristics and the dynamic pattern data of EEG signal channel characteristics respectively, so as to obtain a state estimation value of the dynamic pattern recognition library and a state estimation value of the dynamic pattern data;

[0034] Compare the state estimation value of the dynamic pattern recognition library with the state estimation value of the dynamic pattern data to obtain the state estimation error;

[0035] The state estimation error is calculated and classified by L1 norm to obtain the interictal EEG signals and preictal EEG signals of epileptic patients.

[0036] The second technical solution adopted by the present invention is: an epileptic EEG signal detection system based on statistical channel selection, comprising:

[0037] A preprocessing module is used to perform data preprocessing on the multi-channel EEG signal by using a noise reduction and artifact removal algorithm to obtain a preprocessed multi-channel EEG signal, wherein the noise reduction and artifact removal algorithm includes an independent component analysis algorithm and an improved adaptive noise complete set empirical mode decomposition algorithm;

[0038] A channel selection module is used to perform channel selection processing on the preprocessed multi-channel EEG signals based on the maximum variance principle and the maximum mutual information principle to obtain the target EEG signal channel;

[0039] A trajectory reconstruction module is used to perform three-dimensional trajectory reconstruction processing on the target EEG signal channel to obtain a target EEG signal channel feature vector;

[0040] The feature recognition module is used to identify and store the target EEG signal channel feature vector based on deterministic learning theory and combined with the radial basis function neural network, and to build a dynamic pattern recognition library of EEG signal channel features;

[0041] The feature classification module is used to design a dynamic estimator to classify and process the dynamic pattern recognition library of the EEG signal channel features to obtain the interictal EEG signals and preictal EEG signals of epileptic patients.

[0042] The method and system of the present invention have the following beneficial effects: the present invention performs data preprocessing on multi-channel EEG signals through a noise reduction and artifact removal algorithm, performs noise reduction and artifact removal processing on EEG signals after channel dimension reduction based on independent component analysis and an improved fully adaptive noise set empirical mode decomposition algorithm, and obtains EEG signals with smooth waveform and low noise, and further performs channel selection processing on the preprocessed multi-channel EEG signals based on the maximum variance principle and the maximum mutual information principle, selects the EEG channels most closely related to epileptic seizures, and performs segmentation processing on the EEG data to divide the interictal period, preictal period and seizure period for subsequent training stages, performs three-dimensional trajectory reconstruction processing on the target EEG signal channel and identifies and stores the target EEG signal channel feature vector based on deterministic learning theory and combined with a radial basis function neural network, performs dynamic identification on the features of the interictal period and preictal period of each epileptic sample, stores the identified results with constant weights, constructs an epilepsy dynamic pattern recognition library, and realizes the functions of fast dynamic pattern recognition and epilepsy detection according to the dynamic matching inside the system, which can improve the recognition efficiency and recognition accuracy of the dynamic pattern of EEG signals of epileptic seizures. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flowchart of the steps of a method for detecting epileptic EEG signals based on statistical channel selection according to the present invention;

[0044] Figure 2 It is a structural block diagram of an epileptic EEG signal detection system based on statistical channel selection according to the present invention;

[0045] Figure 3 It is a flowchart of the steps of epileptic EEG signal detection provided by a specific embodiment of the present invention;

[0046] Figure 4 It is a schematic diagram of a signal result obtained by decomposing the improved adaptive noise complete set empirical mode decomposition algorithm provided by a specific embodiment of the present invention;

[0047] Figure 5 is a schematic diagram of a three-dimensional grid arrangement of neurons provided by a specific embodiment of the present invention;

[0048] Figure 6 It is a three-dimensional visualization schematic diagram of the pre-ictal signal of the EEG signal provided by a specific embodiment of the present invention. DETAILED DESCRIPTION

[0049] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only provided for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0050] Reference Figure 1 The present invention provides a method for detecting epileptic EEG signals based on statistical channel selection, the method comprising the following steps:

[0051] S100, performing data preprocessing on the multi-channel EEG signal by a noise reduction and artifact removal algorithm to obtain a preprocessed multi-channel EEG signal, wherein the noise reduction and artifact removal algorithm includes an independent component analysis algorithm and an improved adaptive noise complete set empirical mode decomposition algorithm;

[0052] Specifically, the multi-channel EEG signals of epilepsy patients are acquired and format-matched to obtain a multi-channel EEG signal with a time series; the multi-channel EEG signal with a time series is subjected to artifact signal filtering through an independent component analysis algorithm to obtain a multi-channel EEG signal after preliminary filtering; the multi-channel EEG signal after preliminary filtering is subjected to noise filtering through an improved adaptive noise complete set empirical mode decomposition algorithm to obtain a filtered multi-channel EEG signal; the ideal frequency range of the EEG signal is set, and the EEG frequency screening process is performed on the filtered multi-channel EEG signal to obtain a preprocessed multi-channel EEG signal.

[0053] In this embodiment, the CHB-MIT dataset is an EEG data set from Boston Children's Hospital collected in the MIT EEG database. The dataset consists of scalp EEG records of children with intractable epilepsy. The collected raw EEG data is imported into the training model, and the raw EEG signals are represented as a set of time series, namely ,in , is the number of training samples, Indicates training samples, . is the number of channels of EEG data, here we take .

[0054] Independent component analysis (ICA) denoising is an effective blind source separation technique and a powerful tool for reducing artifacts. Epileptic seizure detection uses limited electrodes to record EEG data. In the case of limited channels, ICA can better distinguish artifacts such as blinking or muscle activity than other artifact removal methods, retain the original EEG signal information, have less loss to each blind source, and have a more convenient calculation process. Therefore, in terms of EEG signal denoising, ICA fast algorithm denoising technology has an outstanding performance.

[0055] ICA processing is performed on the original EEG signal:

[0056] First, the sample data is centered and whitened. Using the whitening matrix , remove the correlation between observed variables, and its expression is:

[0057] ;

[0058] In the above formula, represents the EEG signal after centering and whitening. represents the original EEG observation signal, is the whitening matrix, represents the mixing matrix, Independent signal source.

[0059] make have to:

[0060] ;

[0061] In the above formula, is an orthogonal matrix, and the estimation of independent components can be expressed as:

[0062] ;

[0063] In the above formula, is the mixing matrix The inverse matrix of .

[0064] Further By solving the problem, the artifact signal can be filtered out according to the proportion of each component to obtain the required EEG component.

[0065] After removing the artifacts of the EEG signal, the improved adaptive noise complete ensemble empirical mode decomposition algorithm (ICEEMDAN) is used to filter out the power frequency interference and noise in the EEG. The expression is:

[0066] ;

[0067] in For decomposition modal components, is the number of modal components, To decompose the residual.

[0068] Due to the modal component and Both are high-frequency, low-amplitude noise components, and the EEG frequency range retained by the embodiment of the present invention is 0.5Hz~50Hz, so these two components need to be removed. At this time, a multi-channel EEG signal with relatively smooth waveform and low noise is obtained, as shown in the following example. Figure 4 shown.

[0069] S200, based on the maximum variance principle and the maximum mutual information principle, performing channel selection processing on the preprocessed multi-channel EEG signal to obtain a target EEG signal channel;

[0070] Specifically, the preprocessed multi-channel EEG signal is subjected to channel mean calculation processing to obtain the EEG signal channel mean; the variance value of the EEG signal channel is determined according to the EEG signal channel mean, and the EEG signal channel corresponding to the maximum variance value is selected according to the maximum variance principle to obtain the first EEG signal channel; according to the maximum mutual information principle, the Pearson correlation coefficients of several EEG signal channels with the first EEG signal channel are obtained and arranged in descending order, and the EEG signal channels corresponding to the first n Pearson correlation coefficients are selected as the second EEG signal channel; the first EEG signal channel and the second EEG signal channel are combined to obtain the target EEG signal channel.

[0071] In this embodiment, after ICA-ICEEMDAN processing, the data preprocessing part has been completed. Next, from a statistical point of view, according to the maximum variance and maximum mutual information principles, the preprocessed multi-channel EEG data channels are selected to select the five EEG channels most closely related to epileptic seizures, remove irrelevant channels, and reduce the complexity of the model.

[0072] First calculate the Training samples The mean of each channel: .

[0073] According to the calculation formula of variance: ,have to .

[0074] According to the maximum variance principle, the channel with the maximum variance of this sample is selected as .

[0075] According to the maximum mutual information principle, the remaining 22 channels are calculated and The Pearson correlation coefficient is expressed as:

[0076] ;

[0077] And arrange them from high to low, and select the first four as the channels of the training sample. The training samples are represented as: To further simplify the notation, let , the preprocessed EEG training samples are expressed as: .

[0078] S300, performing three-dimensional trajectory reconstruction processing on the target EEG signal channel to obtain a target EEG signal channel feature vector;

[0079] Specifically, a three-dimensional trajectory reconstruction process is performed on the target EEG signal channel through a phase space reconstruction method to obtain the phase space of the reconstructed time series of the EEG signal; feature extraction calculation indicators are set, and the feature extraction calculation indicators include a Euclidean distance feature indicator, an average energy feature indicator, and a maximum Lyapunov exponent feature indicator; feature extraction processing is performed on the phase space of the reconstructed time series of the EEG signal according to the feature extraction calculation indicators to obtain a feature vector of the target EEG signal channel.

[0080] In this embodiment, the phase space reconstruction technology is used to reconstruct the training samples. The three-dimensional trajectory of the five EEG channels is reconstructed, and the phase space of the reconstructed time series is expressed as: ;

[0081] in, is the subscript of the corresponding channel, , is the embedding dimension of the phase space, is the delay time in phase space. In this experiment, in order to visualize the topological structure of the system's internal dynamics through three-dimensional visualization, , , at this time, the reconstructed phase space is expressed as .

[0082] The Euclidean distance, average energy and maximum Lyapunov index are selected as the phase space The extracted features are calculated as follows:

[0083] Euclidean distance:

[0084] ;

[0085] Average Energy:

[0086] ;

[0087] Maximum Lyapunov exponent:

[0088] ;

[0089] In the above formula, represents the initial state of the system, Represents the final state of the system.

[0090] Combine the three features extracted above into a feature vector .

[0091] S400, based on the deterministic learning theory and in combination with the radial basis function neural network, the target EEG signal channel feature vector is identified and stored, and a dynamic pattern recognition library of EEG signal channel features is constructed;

[0092] Specifically, based on the deterministic learning theory, local dynamic identification processing is performed on the target EEG signal channel feature vector to obtain the dynamic pattern data of the EEG signal channel characteristics; the dynamic pattern data of the EEG signal channel characteristics are converted into constant weight storage based on the radial basis function neural network to construct a dynamic pattern recognition library of the EEG signal channel characteristics.

[0093] Furthermore, considering the differences between different dynamic modes of EEG signal channel characteristics, the dynamic mode data of EEG signal channel characteristics are modeled and processed through dynamic functions to construct a dynamic model of EEG signal channel characteristics; according to the dynamic model of EEG signal channel characteristics, an identifier based on radial basis function neural network is designed; the state error and weight estimation error of the dynamic model of EEG signal channel characteristics are obtained, and the weights of the identifier based on radial basis function neural network are adaptively updated; according to the weight convergence of the radial basis function neural network, the weight of the identifier corresponding to the mean of the final steady-state process is selected as the constant weight to construct a dynamic pattern recognition library of EEG signal channel characteristics.

[0094] First of all, it should be noted that in adaptive control theory, the stability of system parameters and the exponential convergence of tracking errors are crucial, which usually depends on the input signal satisfying the continuous excitation condition. However, in the field of system identification based on radial basis function (RBF) neural networks, the traditional continuous excitation condition is often difficult to achieve because it requires that the input trajectory must correspond exactly to the activation function of the hidden layer nodes. Deterministic learning proves that under the local radial basis function (RBFN) network, almost any neural network input near a periodic or quasi-periodic trajectory can satisfy the continuous excitation condition. This allows the unknown system dynamic trajectory to be approximated by the local RBFN network with high accuracy. Due to the convergence of the neural network weights, the learned system dynamic information can be effectively stored and represented in the form of a time-invariant, constant weight vector. In addition, the weight update algorithm based on Lyapunov stability theory provides a solid theoretical basis for ensuring the convergence of the neural network weights and the global stability of the system. The algorithm not only ensures the convergence of the weights, but also maintains the stability of the system, thereby significantly improving the robustness and reliability of the adaptive control system.

[0095] In this embodiment, the purpose of knowledge acquisition is to be able to reuse the acquired knowledge in subsequent tasks. In order to quickly identify dynamic patterns, an efficient and intelligent method is proposed - deterministic learning theory. The features extracted in the previous step are locally and accurately identified dynamically, and the time-varying dynamic patterns are stored with time-invariant constant weights using a constant RBF neural network as a learned training pattern. The experiment is repeated for each case sample of epilepsy patients, and the corresponding dynamic pattern recognition library is constructed to obtain multiple training patterns.

[0096] Consider the following general form of nonlinear EEG system dynamics:

[0097] ;

[0098] in, represents the characteristics of the EEG dynamic system, and has , represents the parameter vector of the system, and represents the dynamics function of each EEG channel, each component are all continuous and smooth nonlinear functions. Represents the modeling error of the EEG system.

[0099] The following dynamic RBF neural network identifier is used, and its expression is:

[0100] ;

[0101] Let state error , then:

[0102] ;

[0103] Then let the weight estimation error have to:

[0104] ;

[0105] in, is the weight estimator and is the ideal weight that converges finally, represents the gain constant to be designed, and . represents the ideal approximation error, is a dynamic RBF neural network, is a Gaussian function, and the weight update of the RBF neural network adopts the following adaptive learning law:

[0106] ;

[0107] In the above formula, Represents the learning law for weight estimation. represents the learning law for the weight estimation error, and represents the tuning parameter of the learning law, Indicates the state error, express The first dimension of EEG feature variables Dimensional variables.

[0108] According to the weight convergence of the RBF neural network, the mean of the final steady-state process can be selected as the constant weight:

[0109] ;

[0110] In the construction stage of the dynamic pattern library, the acquired EEG signals of the interictal period and the preictal period are processed in the same way as the training stage to obtain multiple test patterns of the interictal period and the preictal period.

[0111] The RBF neural network is evenly arranged in a three-dimensional grid, and the number of neurons , the receptive field width of the neuron is set , the center points are evenly arranged in In the region of ,like Figure 5 and Figure 6 shown.

[0112] S500, design a dynamic estimator, classify and process the dynamic pattern recognition library of EEG signal channel characteristics, and obtain the interictal EEG signals and pre-ictal EEG signals of epileptic patients.

[0113] Specifically, a dynamic estimator is designed, constant weights are embedded into the dynamic estimator, and state estimation processing is performed on the dynamic pattern recognition library of EEG signal channel characteristics and the dynamic pattern data of EEG signal channel characteristics, respectively, to obtain the state estimation value of the dynamic pattern recognition library and the state estimation value of the dynamic pattern data; a difference comparison is made between the state estimation value of the dynamic pattern recognition library and the state estimation value of the dynamic pattern data to obtain the state estimation error; the state estimation error is calculated and classified by the L1 norm to obtain the interictal EEG signals of epileptic patients and the pre-ictal EEG signals of epileptic patients.

[0114] In this embodiment, in the classification stage, a dynamic estimator is designed to generate a series of metric errors according to the L1 norm principle, and the test pattern is quickly identified based on the minimum error principle. The classification mechanism of the dynamic estimator to achieve interictal and preictal periods is now described. The constant weights obtained by training the RBF network based on the EEG data of epileptic patients are embedded in the dynamic estimator as the learned nonlinear EEG dynamics knowledge, and its expression is:

[0115] ;

[0116] in Indicates An estimator, represents the state of the estimator, are the setting parameters of the dynamic estimator, is the EEG feature to be classified in the input dynamics estimator.

[0117] The multiple dynamic patterns stored in the dynamic pattern library are compared with the EEG pattern to be classified in the input dynamic estimator, and the following error detection system is obtained, which is expressed as:

[0118] ;

[0119] in is the state estimation error, and The average L1 norm of is:

[0120] ;

[0121] in is the period of the EEG signal.

[0122] Furthermore, when a test pattern of a dynamic system is given, an error estimation system can be obtained by subtracting the dynamic estimation system from the test pattern. Based on this, the basic idea of ​​the dynamic estimator to achieve the classification of interictal and preictal EEG data is: since the dynamic estimation system embeds the internal dynamics of the dynamic pattern obtained in the training phase, if the dynamic pattern of the EEG to be classified (test set) is different from the training pattern The constant RBF network can quickly recall the learned system dynamic information and accurately approximate the dynamic mode trained along the trajectory, making the average L1 norm error The state synchronization error is proved to be approximately proportional to the internal dynamic difference between the dynamic estimation system and the test pattern, that is, the synchronization error can be used to measure the similarity between the test pattern and the training pattern.

[0123] In summary, if Figure 3As shown, in the embodiment of the present invention, first, in the training stage, the collected multi-channel EEG data is imported into the training network after format matching to obtain the original electroencephalogram (EEG) signal. Secondly, the original EEG signal is preprocessed based on the independent component analysis and improved completely adaptive noise ensemble empirical mode decomposition (ICA-ICEEMDAN) algorithm, wherein ICA is used to remove physiological artifacts such as electromyography and electrooculography in EEG, and the required EEG components are retained; ICEEMDAN is then used to filter out the power frequency interference and noise in EEG, and the required EEG frequency of 0.5Hz~50Hz is retained. At this point, a multi-channel EEG signal with relatively smooth waveform and low noise is obtained.

[0124] Next, from a statistical point of view, according to the principles of maximum variance and maximum mutual information, channel selection is performed on the preprocessed multi-channel EEG data channels, and the five EEG channels most closely related to epileptic seizures are selected, and irrelevant channels are removed to reduce the complexity of the model.

[0125] Next, the phase space reconstruction technique is used to reconstruct the three-dimensional trajectory of the five EEG channels of the training sample, and the Euclidean distance, average energy and maximum Lyapunov exponent are selected as the extracted features. Finally, based on the deterministic learning theory, the features extracted in the previous step are locally and accurately identified dynamically, and the time-varying dynamic patterns are stored with time-invariant constant weights as a learned training pattern. The experiment is then repeated for each case sample of epilepsy patients, and the corresponding dynamic pattern recognition library is constructed to obtain multiple training patterns.

[0126] In the classification stage, the acquired EEG signal is subjected to the same operation as the training stage to obtain multiple test patterns. A dynamic estimator is designed to generate a series of metric errors based on the L1 norm principle, and the test pattern is quickly identified based on the minimum error principle.

[0127] Therefore, the embodiments of the present invention and related technologies have the following beneficial effects:

[0128] 1) The embodiment of the present invention combines independent component analysis and improved fully adaptive noise ensemble empirical mode decomposition (ICA-ICEEMDAN) algorithm to remove power frequency interference and physiological artifacts of the original EEG signal, and obtains an EEG signal with smooth waveform and low noise.

[0129] 2) The embodiment of the present invention proposes a new channel selection method, which selects channels of multi-channel EEG data from a statistical perspective based on the maximum variance and maximum mutual information principles, and selects several EEG channels that are most closely related to epileptic seizures. This solves the problem of too many channels and large data volume in EEG data, which leads to long model training time, information redundancy, and easy overfitting, and reduces the amount of model calculation.

[0130] 3) Based on deterministic learning theory, RBF neural network is used to dynamically identify the characteristics of each epileptic sample during the interictal and preictal periods, and the results after identification are stored with constant weights. An epilepsy dynamic pattern recognition library is constructed, and the functions of fast dynamic pattern recognition and epileptic seizure detection are realized based on the dynamic matching inside the dynamic estimator, and the results are interpretable.

[0131] Reference Figure 2 , an epileptic EEG signal detection system based on statistical channel selection, comprising:

[0132] A preprocessing module 201 is used to perform data preprocessing on the multi-channel EEG signal by using a noise reduction and artifact removal algorithm to obtain a preprocessed multi-channel EEG signal, wherein the noise reduction and artifact removal algorithm includes an independent component analysis algorithm and an improved adaptive noise complete set empirical mode decomposition algorithm;

[0133] The channel selection module 202 is used to perform channel selection processing on the preprocessed multi-channel EEG signal based on the maximum variance principle and the maximum mutual information principle to obtain the target EEG signal channel;

[0134] The trajectory reconstruction module 203 is used to perform three-dimensional trajectory reconstruction processing on the target EEG signal channel to obtain a target EEG signal channel feature vector;

[0135] The feature recognition module 204 is used to identify and store the target EEG signal channel feature vector based on the deterministic learning theory and in combination with the radial basis function neural network, and to build a dynamic pattern recognition library of EEG signal channel features;

[0136] The feature classification module 205 is used to design a dynamic estimator to classify the dynamic pattern recognition library of the EEG signal channel features to obtain the interictal EEG signals and preictal EEG signals of the epileptic patients.

[0137] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0138] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A method for detecting epileptic EEG signals based on statistical channel selection, characterized in that: The following steps are involved: Preprocessing the multi-channel EEG signal data by using a noise reduction and artifact removal algorithm to obtain a preprocessed multi-channel EEG signal, wherein the noise reduction and artifact removal algorithm includes an independent component analysis algorithm and an improved adaptive noise complete set empirical mode decomposition algorithm; Based on the maximum variance principle and the maximum mutual information principle, the preprocessed multi-channel EEG signal is processed by channel selection to obtain the target EEG signal channel, including: Perform channel mean calculation on the preprocessed multi-channel EEG signal to obtain the EEG signal channel mean; Determine the variance value of the EEG signal channel according to the mean value of the EEG signal channel, and select the EEG signal channel corresponding to the maximum variance value according to the maximum variance principle to obtain the first EEG signal channel; According to the maximum mutual information principle, the Pearson correlation coefficients of several EEG signal channels with the first EEG signal channel are obtained and arranged in descending order, and the EEG signal channels corresponding to the first n Pearson correlation coefficients are selected as the second EEG signal channels; Combining the first EEG signal channel with the second EEG signal channel to obtain a target EEG signal channel; Perform three-dimensional trajectory reconstruction on the target EEG signal channel to obtain the target EEG signal channel feature vector; Based on the deterministic learning theory and combined with the radial basis function neural network, the target EEG signal channel feature vector is identified and stored, and a dynamic pattern recognition library of EEG signal channel features is constructed; A dynamic estimator is designed to classify and process the dynamic pattern recognition library of EEG signal channel characteristics, and the interictal EEG signals and pre-ictal EEG signals of epileptic patients are obtained.

2. The epileptic EEG signal detection method based on statistical channel selection according to claim 1, characterized in that: The step of performing data preprocessing on the multi-channel EEG signal by using a noise reduction and artifact removal algorithm to obtain the preprocessed multi-channel EEG signal specifically includes: Acquire multi-channel EEG signals of epilepsy patients and perform format matching processing to obtain multi-channel EEG signals with time series; The artifact signal is filtered out of the multi-channel EEG signal with time series by using the independent component analysis algorithm to obtain the multi-channel EEG signal after preliminary filtering; The multi-channel EEG signal after preliminary filtering is subjected to noise filtering by using an improved adaptive noise complete set empirical mode decomposition algorithm to obtain a multi-channel EEG signal after filtering. The ideal frequency range of the EEG signal is set, and the EEG frequency is screened and processed on the filtered multi-channel EEG signal to obtain the preprocessed multi-channel EEG signal.

3. The epileptic EEG signal detection method based on statistical channel selection according to claim 2, characterized in that: The step of performing three-dimensional trajectory reconstruction processing on the target EEG signal channel to obtain the target EEG signal channel feature vector specifically includes: Through the phase space reconstruction method, the three-dimensional trajectory of the target EEG signal channel is reconstructed to obtain the phase space of the reconstructed time series of the EEG signal; Setting feature extraction calculation indicators, wherein the feature extraction calculation indicators include a Euclidean distance feature indicator, an average energy feature indicator, and a maximum Lyapunov exponent feature indicator; According to the feature extraction calculation index, the phase space of the reconstructed time series of the EEG signal is subjected to feature extraction processing to obtain the target EEG signal channel feature vector.

4. The epileptic EEG signal detection method based on statistical channel selection according to claim 3, characterized in that: The step of identifying and storing the target EEG signal channel feature vector based on the deterministic learning theory and in combination with the radial basis function neural network to construct a dynamic pattern recognition library of EEG signal channel features specifically includes: Based on the deterministic learning theory, the local dynamics identification processing is performed on the target EEG signal channel feature vector to obtain the dynamic pattern data of the EEG signal channel feature; The dynamic pattern data of EEG signal channel characteristics are converted into constant weight storage based on radial basis function neural network, and a dynamic pattern recognition library of EEG signal channel characteristics is constructed.

5. The epileptic EEG signal detection method based on statistical channel selection according to claim 4, characterized in that: The step of converting the dynamic pattern data of EEG signal channel characteristics into constant weight storage based on the radial basis function neural network to construct a dynamic pattern recognition library of EEG signal channel characteristics specifically includes: Considering the differences between EEG signal channel characteristics in different dynamic modes, the dynamic mode data of EEG signal channel characteristics are modeled and processed through dynamic functions to construct a dynamic model of EEG signal channel characteristics. According to the dynamic model of EEG signal channel characteristics, an identifier based on radial basis function neural network is designed; The state error and weight estimation error of the dynamic model of the EEG signal channel characteristics are obtained, and the weights of the identifier based on the radial basis function neural network are adaptively updated; According to the weight convergence of radial basis function neural network, the weight of the identifier corresponding to the mean of the final steady-state process is selected as the constant weight to construct a dynamic pattern recognition library of EEG signal channel characteristics.

6. The epileptic EEG signal detection method based on statistical channel selection according to claim 5, characterized in that: The step of designing a dynamic estimator to classify and process the dynamic pattern recognition library of EEG signal channel features to obtain the interictal EEG signals and preictal EEG signals of epileptic patients specifically includes: A dynamic estimator is designed, a constant weight is embedded into the dynamic estimator, and state estimation processing is performed on the dynamic pattern recognition library of EEG signal channel characteristics and the dynamic pattern data of EEG signal channel characteristics respectively, so as to obtain a state estimation value of the dynamic pattern recognition library and a state estimation value of the dynamic pattern data; Compare the state estimation value of the dynamic pattern recognition library with the state estimation value of the dynamic pattern data to obtain the state estimation error; The state estimation error is calculated and classified by L1 norm to obtain the interictal EEG signals and preictal EEG signals of epileptic patients.

7. An epileptic EEG signal detection system based on statistical channel selection, characterized in that: Includes the following modules: A preprocessing module, used for performing data preprocessing on the multi-channel EEG signal by a noise reduction and artifact removal algorithm to obtain a preprocessed multi-channel EEG signal, wherein the noise reduction and artifact removal algorithm includes an independent component analysis algorithm and an improved adaptive noise complete set empirical mode decomposition algorithm; The channel selection module is used to perform channel selection processing on the preprocessed multi-channel EEG signals based on the maximum variance principle and the maximum mutual information principle to obtain the target EEG signal channel, including: Perform channel mean calculation on the preprocessed multi-channel EEG signal to obtain the EEG signal channel mean; Determine the variance value of the EEG signal channel according to the mean value of the EEG signal channel, and select the EEG signal channel corresponding to the maximum variance value according to the maximum variance principle to obtain the first EEG signal channel; According to the maximum mutual information principle, the Pearson correlation coefficients of several EEG signal channels with the first EEG signal channel are obtained and arranged in descending order, and the EEG signal channels corresponding to the first n Pearson correlation coefficients are selected as the second EEG signal channels; Combining the first EEG signal channel with the second EEG signal channel to obtain a target EEG signal channel; A trajectory reconstruction module is used to perform three-dimensional trajectory reconstruction processing on the target EEG signal channel to obtain a target EEG signal channel feature vector; The feature recognition module is used to identify and store the target EEG signal channel feature vector based on deterministic learning theory and combined with the radial basis function neural network, and to build a dynamic pattern recognition library of EEG signal channel features; The feature classification module is used to design a dynamic estimator to classify and process the dynamic pattern recognition library of the EEG signal channel features to obtain the interictal EEG signals and preictal EEG signals of epileptic patients.

Images