An electroencephalogram signal classification method based on physical information residual polynomial network

Abstract

The present application relates to a kind of electroencephalogram classification method based on physical information residual polynomial network, design to be trained network, first with polynomial feature extraction layer carries out nonlinear feature extraction, then through LSTM network capture long-term time series dependence of electroencephalogram, then realize the neural network architecture of E / I path separation, respectively to interface excitatory pathway and inhibitory pathway, then by the feature correlation analysis layer of embedding Wilson-Cowan neural population dynamics equation carries out physical constraint, finally in succession fusion layer, classification layer, classification layer completes the construction of to-be-trained network, then based on each sample formed by each multi-channel electroencephalogram, for to-be-trained network training, obtains electroencephalogram classification model, while guaranteeing electroencephalogram classification prediction accuracy, significantly enhance biological explainability.

Description

Technical Field

[0001] This invention relates to a method for classifying electroencephalogram (EEG) signals based on physical information residual multinomial networks, belonging to the field of signal classification technology. Background Technology

[0002] Because epileptic seizures are sudden and unpredictable, electroencephalography (EEG) is used as the primary means of epilepsy monitoring. It is used to record the electrical activity of brain neuron populations in real time to capture abnormal brain electrical patterns during the preictal and ictal phases.

[0003] In existing EEG recognition research, traditional machine learning methods struggle to accurately capture the temporal characteristics of EEG signals due to the strong subjectivity of feature selection and the limited expressive power of the models, resulting in low model classification accuracy. While deep learning algorithms improve prediction accuracy, they lack the support of physical mechanisms, limiting model interpretability. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a method for classifying EEG signals based on a physical information residual multinomial network. The method uses a physical information unfolding layer for nonlinear feature extraction, captures the long-term temporal dependencies of EEG signals through an LSTM network, implements a neural network architecture to achieve E / I pathway separation, and embeds the Wilson-Cowan neurodynamic equation as a physical constraint. This method significantly enhances biological interpretability while ensuring prediction accuracy.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention designs an EEG signal classification method based on physical information residual multinomial network, and performs the following steps A to C to obtain an EEG signal classification model, which is used to classify the detected multi-channel EEG signals according to preset categories;

[0006] Step A. Based on the preset categories, obtain a preset number of real categories corresponding to each preset category of multi-channel EEG signals to form each sample, and proceed to Step B;

[0007] Step B. Construct the network to be trained, including a multinomial feature extraction layer, an LSTM network, an excitatory pathway, an inhibitory pathway, a feature correlation analysis layer, a fusion layer, and a classification layer. The input of the multinomial feature extraction layer constitutes the input of the network to be trained. The output of the multinomial feature extraction layer is connected to the input of the LSTM network. The output of the LSTM network is connected to the input of the excitatory pathway and the input of the inhibitory pathway, respectively. The outputs of the excitatory pathway and the inhibitory pathway are connected to the input of the feature correlation analysis layer. The output of the feature correlation analysis layer is sequentially connected in series with the fusion layer and the classification layer. The output of the classification layer constitutes the output of the network to be trained. Then proceed to step C.

[0008] Step C. Based on each sample, using the multi-channel EEG signals in the sample as input and the predicted category corresponding to the multi-channel EEG signals in the sample as output, the network to be trained is trained in combination with the target loss function to obtain the EEG signal classification model.

[0009] As a preferred technical solution of the present invention: the polynomial feature extraction layer receives multi-channel EEG signals and performs polynomial expansion on the multi-channel feature data of each time step in the multi-channel EEG signals according to the preset maximum expansion order, so as to obtain the feature sequence corresponding to each time step and output it to the LSTM network.

[0010] The LSTM network receives the feature sequences corresponding to each time step in the multi-channel EEG signal, and updates the hidden state of the LSTM network for each time step according to the following equation:

[0011] ;

[0012] ;

[0013] ;

[0014] ;

[0015] ;

[0016] ;

[0017] Obtain the time-series feature sequence output by the LSTM network. These are respectively output to the excitatory pathway and the inhibitory pathway; among them, This indicates the number of time steps in a multichannel EEG signal. Indicates the first of the multichannel EEG signals Feature sequences at each time step Representing time-series feature sequences The Middle Features corresponding to each time step Indicates the output of the candidate memory unit. Indicates the output of the current memory unit. , , , These represent the weight matrices for the forget gate, input gate, output gate, and candidate memory units, respectively. , , , These represent the biases of the forget gate, input gate, output gate, and candidate memory unit, respectively. , , These represent the outputs of the forget gate, the input gate, and the output gate, respectively. express Activation function Indicates the preset initial value. This indicates element-wise multiplication. This indicates the sigmod activation function.

[0018] As a preferred embodiment of the present invention: the excitatory pathway receives a time-series feature sequence output from the LSTM network. For each time step, firstly, for each time step... Corresponding features Perform polynomial expansion to obtain the corresponding features Then follow the formula below:

[0019] ;

[0020] ;

[0021] ;

[0022] ;

[0023] Obtain the excitatory characteristic sequence output by the excitatory pathway The output is sent to the feature correlation analysis layer, where... This indicates the number of time steps in a multichannel EEG signal. Represents excitatory characteristic sequences The Middle Features corresponding to each time step This represents the weight matrix in the excitatory pathway. This indicates a bias in the excitatory pathway. Represents the normalization function. express Activation function This represents the weighting parameter in the excitatory pathway. This represents the state vector of an excitatory neuron. This represents the normalized result of the excitatory state. This indicates an excitatory pre-activated state.

[0024] As a preferred embodiment of the present invention: the inhibitory path receives the temporal feature sequence output from the LSTM network. For each time step, firstly, for each time step... Corresponding features Perform polynomial expansion to obtain the corresponding features Then follow the formula below:

[0025] ;

[0026] ;

[0027] ;

[0028] ;

[0029] ;

[0030] ;

[0031] Obtain the inhibitory characteristic sequence output by the inhibitory pathway The output is sent to the feature correlation analysis layer, where... This indicates the number of time steps in a multichannel EEG signal. Indicating inhibitory characteristic sequences The Middle Features corresponding to each time step , These represent the first and second weight matrices in the inhibitory pathway, respectively. , These represent the first and second biases in the inhibitory pathway, respectively. Represents the normalization function. express Activation function This represents the weighting parameter in the inhibitory pathway. Indicates the first The activation coefficient of each neuron in the first layer. Indicating the inhibitory pathway The output of the first layer of neurons, Indicates the first The layer normalization result output by the layer. Indicating the inhibitory pathway The output of the first layer of neurons, This indicates an excitatory pre-activated state.

[0032] As a preferred embodiment of the present invention: the feature correlation analysis layer embeds the Wilson-Cowan neural population dynamics equation, and the feature correlation analysis layer receives excitatory feature sequences output from excitatory pathways. Inhibitory signature sequences from inhibitory pathway outputs ;

[0033] Next, the feature correlation analysis layer uses the pre-trained state transition neural network corresponding to the excitatory pathway. and pre-trained state transition neural networks corresponding to the inhibitory pathways. Execute the following formula:

[0034] ;

[0035] ;

[0036] Obtain the predicted value of the rate of change of excitability state Predicted value of rate of change of inhibitory state ,in, Represents excitatory characteristic sequences and inhibitory characteristic sequences The splicing representation, and These represent the training parameters for the corresponding excitatory pathway and the training parameters for the corresponding inhibitory pathway, respectively.

[0037] Then, the feature correlation analysis layer is based on the step size in the dynamic evolution process. and the first step characteristics in the excitatory dynamics evolution process. Preset initial values, characteristics of the first step in the inhibitory dynamic evolution process The preset initial value is used to update the state using the Euler method according to the following formula;

[0038] ;

[0039] ;

[0040] Obtain the temporal characteristic sequence of excitatory dynamics evolution Inhibitory dynamics evolution time sequence ,in, This represents the number of steps in the dynamic evolution process. This represents the predicted rate of change of excitability state corresponding to the step size in the dynamic evolution process. This represents the predicted rate of change of the state of inhibition corresponding to the step size of the dynamic evolution process.

[0041] As a preferred embodiment of the present invention: the fusion layer receives the excitatory dynamics evolution time-series feature sequence output from the feature correlation analysis layer. Inhibitory dynamics evolution time series characteristics sequence ,choose The final step of the feature composition , The final step of the feature composition Then press Perform the splicing, and follow the formula below:

[0042] ;

[0043] ;

[0044] Obtain parameters The output is sent to the classification layer, where, , These represent the first and second weights in the fusion layer, respectively. , These represent the first and second biases in the fusion layer, respectively. Represents the ReLU activation function. Represents the temporal characteristic sequence of excitatory dynamics evolution. middle The final feature, Represents the temporal characteristic sequence of inhibitory dynamics evolution. middle The final feature, Represents the concatenated E / I feature vector. This represents the first layer fused feature vector;

[0045] The classification layer receives parameters from the output of the fusion layer. For each preset category, the following formula is used:

[0046] ;

[0047] ;

[0048] Obtaining multi-channel EEG signals regarding the first The probability of each category ,in, Indicates the number of categories. Indicates the first Weight matrices corresponding to each category Indicates the first The bias corresponding to each category This indicates that the multi-channel EEG signal corresponds to the first Data values ​​for each category, Indicates the original category The output of the final classifier.

[0049] As a preferred embodiment of the present invention: the target loss function in step C includes the cross-entropy loss function, the physical residual loss function, the E / I balance loss function, and the L2 regularization loss function, wherein the cross-entropy loss function is as follows:

[0050] ;

[0051] Obtain the cross-entropy loss result ,in, Indicates the number of samples. Indicates the number of categories. For the first The sample belongs to the first The true category of each category, This represents the predicted probability of the model. , They represent the first Multichannel EEG signals in each sample, and the predicted categories corresponding to the multichannel EEG signals;

[0052] The physical residual loss function is as follows:

[0053] ;

[0054] Obtain physical residual loss results ,in, This represents the number of steps in the dynamic evolution process. This represents the rate of change during the dynamic evolution of excitability. This represents the rate of change during the inhibitory dynamics evolution process. Represents the first step in the dynamic evolution process. The theoretical rate of change of the excitatory state. Represents the first step in the dynamic evolution process. The theoretical rate of change of the inhibited state;

[0055] The E / I balance loss function is as follows:

[0056] ;

[0057] Obtain E / I balance loss results ,in, Represents the temporal characteristic sequence of excitatory dynamics evolution. middle The final feature, Represents the temporal characteristic sequence of inhibitory dynamics evolution. middle The final feature, This represents the function for finding the average.

[0058] The L2 regularization loss function is as follows:

[0059] ;

[0060] Obtain L2 regularization loss results ,in, This represents the parameters to be trained in the network. This represents the set of all parameters to be trained in the network to be trained. This represents the squared L2 norm of the parameter;

[0061] Then, according to the following formula:

[0062] ;

[0063] Weighted total loss ,in, , , These represent the weights of the physical residual loss, the E / I balance loss, and the L2 regularization loss, respectively.

[0064] The EEG signal classification method based on physical information residual multinomial network described in this invention has the following technical advantages compared with the prior art:

[0065] This invention designs a method for classifying electroencephalogram (EEG) signals based on a physical information residual multinomial network. The method involves designing a training network, firstly using a multinomial feature extraction layer for nonlinear feature extraction, then capturing the long-term temporal dependencies of EEG signals through an LSTM network, followed by a neural network architecture that separates the excitatory and inhibitory pathways, respectively. Next, a feature correlation analysis layer embedding the Wilson-Cowan neural population dynamics equation provides physical constraints. Finally, a fusion layer and a classification layer are sequentially connected, with the classification layer completing the construction of the training network. Based on samples composed of multi-channel EEG signals, the training network is then trained to obtain an EEG signal classification model. This method significantly enhances biological interpretability while maintaining the accuracy of EEG signal classification prediction. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating the framework of the EEG signal classification method based on physical information residual multinomial networks designed in this invention.

[0067] Figure 2This is a radar comparison chart of the performance of the network model used in this invention and existing models;

[0068] Figure 3 This is a schematic diagram of the average convergence curve of the network model used in this invention. Detailed Implementation

[0069] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0070] This invention designs a brainwave signal classification method based on physical information residual multinomial network. In practical applications, the following steps A to C are specifically designed and executed to obtain a brainwave signal classification model, which is used to classify the detected multi-channel brainwave signals according to preset categories.

[0071] Step A. Based on the preset categories, obtain a preset number of real categories corresponding to each preset category of multi-channel EEG signals to form each sample, and proceed to Step B.

[0072] Step B. As Figure 1 As shown, a network to be trained is constructed, including a multinomial feature extraction layer, an LSTM network, an excitatory pathway, an inhibitory pathway, a feature correlation analysis layer, a fusion layer, and a classification layer. The input of the multinomial feature extraction layer constitutes the input of the network to be trained. The output of the multinomial feature extraction layer is connected to the input of the LSTM network. The output of the LSTM network is connected to the input of the excitatory pathway and the inhibitory pathway, respectively. The outputs of the excitatory pathway and the inhibitory pathway are connected to the input of the feature correlation analysis layer. The output of the feature correlation analysis layer is sequentially connected in series with the fusion layer and the classification layer. The output of the classification layer constitutes the output of the network to be trained. Then, proceed to step C.

[0073] In the actual data processing of the network under the above structural design, the multinomial feature extraction layer receives multi-channel EEG signals and performs multinomial expansion on the multi-channel feature data of each time step in the multi-channel EEG signals according to the preset maximum expansion order, so as to obtain the feature sequence corresponding to each time step and output it to the LSTM network.

[0074] In this application, the multinomial feature extraction layer maps the input multichannel EEG signals from a linear space to a nonlinear space, capturing the nonlinear dynamic characteristics of the EEG signals and enhancing the model's expressive power. To avoid dimensionality explosion, a lightweight multinomial expansion is used to generate multinomial features. The 0th order corresponds to the original input; the 1st order term corresponds to linear features, with an output dimension of 16; and the 2nd order term corresponds to squared features, i.e., for each channel... calculate Then it is concatenated with the original features, and the output dimension is 32.

[0075] The LSTM network receives the feature sequences corresponding to each time step in the multi-channel EEG signal, and updates the hidden state of the LSTM network for each time step according to the following equation:

[0076] ;

[0077] ;

[0078] ;

[0079] ;

[0080] ;

[0081] ;

[0082] Obtain the time-series feature sequence output by the LSTM network. These are respectively output to the excitatory pathway and the inhibitory pathway; among them, This indicates the number of time steps in a multichannel EEG signal. Indicates the first of the multichannel EEG signals Feature sequences at each time step Representing time-series feature sequences The Middle Features corresponding to each time step Indicates the output of the candidate memory unit. Indicates the output of the current memory unit. , , , These represent the weight matrices for the forget gate, input gate, output gate, and candidate memory units, respectively. , , , These represent the biases of the forget gate, input gate, output gate, and candidate memory unit, respectively. , , These represent the outputs of the forget gate, the input gate, and the output gate, respectively. express Activation function Indicates the preset initial value. This indicates element-wise multiplication. This indicates the sigmod activation function.

[0083] Next, the temporal feature sequence output by the LSTM network. The signals will be delivered to the excitatory and inhibitory pathways, respectively. The excitatory pathway receives the temporal feature sequence output from the LSTM network. For each time step, firstly, for each time step... Corresponding features Perform polynomial expansion to obtain the corresponding features Then follow the formula below:

[0084] ;

[0085] ;

[0086] ;

[0087] ;

[0088] Obtain the excitatory characteristic sequence output by the excitatory pathway The output is sent to the feature correlation analysis layer, where... This indicates the number of time steps in a multichannel EEG signal. Represents excitatory characteristic sequences The Middle Features corresponding to each time step This represents the weight matrix in the excitatory pathway. This indicates a bias in the excitatory pathway. Represents the normalization function. express Activation function This represents the weighting parameter in the excitatory pathway. This represents the state vector of an excitatory neuron. This represents the normalized result of the excitatory state. This indicates an excitatory pre-activated state.

[0089] Simultaneously, the inhibitory pathway receives the temporal feature sequence from the output of the LSTM network. For each time step, firstly, for each time step... Corresponding features Perform polynomial expansion to obtain the corresponding features Then follow the formula below:

[0090] ;

[0091] ;

[0092] ;

[0093] ;

[0094] ;

[0095] ;

[0096] Obtain the inhibitory characteristic sequence output by the inhibitory pathway The output is sent to the feature correlation analysis layer, where... This indicates the number of time steps in a multichannel EEG signal. Indicating inhibitory characteristic sequences The Middle Features corresponding to each time step , These represent the first and second weight matrices in the inhibitory pathway, respectively. , These represent the first and second biases in the inhibitory pathway, respectively. Represents the normalization function. express Activation function This represents the weighting parameter in the inhibitory pathway. Indicates the first The activation coefficient of each neuron in the first layer. Indicating the inhibitory pathway The output of the first layer of neurons, Indicates the first The layer normalization result output by the layer. Indicating the inhibitory pathway The output of the first layer of neurons, This indicates an excitatory pre-activated state.

[0097] EEG signals have strong temporal dependence, and the E / I imbalance needs to capture dynamic changes. Therefore, LSTM networks are needed to encode temporal features. LSTM uses standard gating mechanisms, including forget gate, input gate and output gate, which can selectively retain or forget historical information. 30% Dropout regularization is added between two LSTM layers to prevent overfitting.

[0098] Excitatory characteristic sequences output by excitatory pathways and the inhibitory signature sequences output by inhibitory pathways. The data will be fed to the feature correlation analysis layer. In practical applications, the feature correlation analysis layer embeds the Wilson-Cowan neural population dynamics equation and receives the excitatory feature sequence output from the excitatory pathway. Inhibitory signature sequences from inhibitory pathway outputs The following formula is first executed by the feature correlation analysis layer;

[0099] ;

[0100] ;

[0101] in, This represents the sigmoid activation function. , , , These are the excitatory-excitatory connection weights, excitatory-inhibitory connection weights, inhibitory-excitatory connection weights, and inhibitory-inhibitory connection weights, respectively. , These represent the time constants of the corresponding excitatory pathway and the time constants of the corresponding inhibitory pathway, respectively.

[0102] Next, the feature correlation analysis layer uses the pre-trained state transition neural network corresponding to the excitatory pathway. and pre-trained state transition neural networks corresponding to the inhibitory pathways. Execute the following formula:

[0103] ;

[0104] ;

[0105] Obtain the predicted value of the rate of change of excitability state Predicted value of rate of change of inhibitory state ,in, Represents excitatory characteristic sequences and inhibitory characteristic sequences The splicing representation, and These represent the training parameters for the corresponding excitatory pathway and the training parameters for the corresponding inhibitory pathway, respectively.

[0106] In practical applications, the excitatory pathway uses the ReLU activation function and residual link to ensure that the output is non-negative, which is consistent with the positive propagation characteristic of excitatory neurons. Conversely, since inhibitory neurons may produce negative modulation, the Tanh activation function and residual link are used to make the output range between [-1.1], allowing the existence of negative values.

[0107] Then, the feature correlation analysis layer is based on the step size in the dynamic evolution process. and the first step characteristics in the excitatory dynamics evolution process. Preset initial values, characteristics of the first step in the inhibitory dynamic evolution process The preset initial value is used to update the state using the Euler method according to the following formula;

[0108] ;

[0109] ;

[0110] Obtain the temporal characteristic sequence of excitatory dynamics evolution Inhibitory dynamics evolution time sequence ,in, This represents the number of steps in the dynamic evolution process. This represents the predicted rate of change of excitability state corresponding to the step size in the dynamic evolution process. This represents the predicted rate of change of the state of inhibition corresponding to the step size of the dynamic evolution process.

[0111] Finally, the feature correlation analysis layer executes the following formula:

[0112] ;

[0113] ;

[0114] Obtain the first in the dynamic evolution process Theoretical rate of change of excitatory state With the first in the dynamic evolution process Theoretical rate of change of the step-inhibitory state And execute the following formula:

[0115] ;

[0116] Obtain the corresponding physical residual .

[0117] The excitatory dynamics evolution time series characteristic sequence obtained after processing by the feature correlation analysis layer Inhibitory dynamics evolution time sequence It will be further transported to the fusion layer, where the fusion layer will be selected. The final step of the feature composition , The final step of the feature composition Then press Perform the splicing, and follow the formula below:

[0118] ;

[0119] ;

[0120] Obtain parameters The output is sent to the classification layer, where, , These represent the first and second weights in the fusion layer, respectively. , These represent the first and second biases in the fusion layer, respectively. Represents the ReLU activation function. Represents the temporal characteristic sequence of excitatory dynamics evolution. middle The final feature, Represents the temporal characteristic sequence of inhibitory dynamics evolution. middle The final feature, Represents the concatenated E / I feature vector. This represents the first layer of fused feature vectors.

[0121] The classification layer receives parameters from the output of the fusion layer. For each preset category, the following formula is used:

[0122] ;

[0123] ;

[0124] Obtaining multi-channel EEG signals regarding the first The probability of each category ,in, Indicates the number of categories. Indicates the first Weight matrices corresponding to each category Indicates the first The bias corresponding to each category This indicates that the multi-channel EEG signal corresponds to the first Data values ​​for each category, Indicates the original category The output of the final classifier.

[0125] The network to be trained designed above first uses a multinomial feature extraction layer to extract nonlinear features, then uses an LSTM network to capture the long-term temporal dependencies of EEG signals, then implements a neural network architecture that separates E / I pathways, connecting excitatory and inhibitory pathways respectively, then uses a feature correlation analysis layer with embedded Wilson-Cowan neural population dynamics equations for physical constraints, and finally connects a fusion layer and a classification layer in sequence, with the classification layer completing the construction of the network to be trained.

[0126] Step C. Based on each sample, using the multi-channel EEG signals in the sample as input and the predicted category corresponding to the multi-channel EEG signals in the sample as output, the network to be trained is trained in combination with the target loss function to obtain the EEG signal classification model.

[0127] In practical applications, target loss functions include cross-entropy loss function, physical residual loss function, E / I balance loss function, and L2 regularization loss function. The cross-entropy loss function is as follows:

[0128] ;

[0129] Obtain the cross-entropy loss result ,in, Indicates the number of samples. Indicates the number of categories. For the first The sample belongs to the first The true category of each category, This represents the predicted probability of the model. , They represent the first The multi-channel EEG signals in each sample and the predicted categories corresponding to the multi-channel EEG signals.

[0130] The physical residual loss function is as follows:

[0131] ;

[0132] Obtain physical residual loss results ,in, This represents the number of steps in the dynamic evolution process. This represents the rate of change during the dynamic evolution of excitability. This represents the rate of change during the inhibitory dynamics evolution process. Represents the first step in the dynamic evolution process. The theoretical rate of change of the excitatory state. Represents the first step in the dynamic evolution process. The theoretical rate of change of the step-inhibitory state.

[0133] The E / I balance loss function is as follows:

[0134] ;

[0135] Obtain E / I balance loss results ,in, Represents the temporal characteristic sequence of excitatory dynamics evolution. middle The final feature, Represents the temporal characteristic sequence of inhibitory dynamics evolution. middle The final feature, This represents the function for calculating the average.

[0136] The L2 regularization loss function is as follows:

[0137] ;

[0138] Obtain L2 regularization loss results ,in, This represents the parameters to be trained in the network. This represents the set of all parameters to be trained in the network to be trained. This represents the squared L2 norm of the parameter.

[0139] Then, according to the following formula:

[0140] ;

[0141] Weighted total loss ,in, , , These represent the weights of the physical residual loss, the E / I balance loss, and the L2 regularization loss, respectively.

[0142] The four loss function components each focus on a different optimization objective: cross-entropy loss ensures classification accuracy, physical residual loss ensures physical consistency, excitation-inhibition balance loss ensures physiological rationality, and L2 regularization loss ensures model simplicity. Through multi-objective optimization, the model can find an equilibrium point under multiple constraints, maintaining good classification performance while conforming to neuroscience principles.

[0143] In practical applications, the above technical solution can be used to acquire EEG signals from the CHB-MIT epilepsy EEG dataset, which includes 123 multichannel EEG recording files. Preprocessed multichannel EEG signals are extracted from each multichannel EEG recording file. The multichannel EEG signal field is a three-dimensional array with the following data shape: ,in This indicates the number of EEG data segments in the multichannel EEG recording file, where 512 represents the number of time sampling points for each EEG data segment, and 16 represents the number of EEG signal channels. The data is arranged chronologically, with the first half of the EEG data segments representing the pre-ictal state, the second half representing the ictal state, and the intermediate segments representing the transitional state. The data is labeled and the indices of the transitional EEG data segments are calculated. The number of valid EEG data segments is obtained after removing the EEG data segment at that location. The first half of the EEG data segments was labeled as label 0 (pre-ictal phase) and the second half as label 1 (ictal phase), ensuring that the number of labels for each class in each multi-channel EEG recording file was exactly equal, in order to eliminate the impact of class imbalance on model training.

[0144] The standardization process uses the StandardScaler method per channel. First, the training set is reshaped into a two-dimensional array with dimensions (...). (×512, 16), so that each column represents the observations of one channel at all time points. Then calculate the mean for each channel. and standard deviation Standardize the training set:

[0145] ;

[0146] After standardization, the two-dimensional array is reshaped back into its original three-dimensional shape. Through standardization, the mean of each channel's data is close to 0, and the standard deviation is close to 1, eliminating dimensional differences between channels and enabling the model to learn the characteristics of each channel equally.

[0147] After training the network to be trained, an EEG signal classification model is obtained. In application, while ensuring the accuracy of EEG signal classification prediction, the biological interpretability is significantly enhanced.

[0148] In this experiment, the processed CHB-MIT dataset was divided according to multi-channel EEG recording files, with the test set accounting for 10% and the remaining 90% of the training set being used as the validation set. Figure 2 Compared to traditional random forest methods and deep learning methods such as pure LSTM and Transformer, the accuracy is improved to a certain extent when the method based on physical information residual multinomial network proposed in this invention is adopted. Furthermore, all four metrics (area under the ROC curve, accuracy, sensitivity, and specificity) are well-balanced and stable. From the perspective of model interpretability, the model explicitly outputs the E / I ratio, which can quantify the brain's excitability / inhibition balance (preic E / I ≈ 0.76, ictal E / I ≈ 1.45), consistent with clinical observations of E / I imbalance during epileptic seizures. Simultaneously, the weight parameters (…) It also has a clear physical meaning (excitatory self-feedback, inhibitory regulation, etc.).

[0149] like Figure 2 As shown, this diagram presents a comprehensive comparison of the algorithms across key performance metrics after model training, reflecting the overall performance of the models during the stable phase; for example... Figure 3 As shown, the convergence trend of the model during training is demonstrated. As the number of training rounds increases, the accuracy of the training set and the validation set gradually increases and tends to stabilize, indicating that the model can converge effectively and has good generalization performance.

[0150] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for electroencephalogram signal classification based on physical information residual polynomial network, characterized in that: Perform steps A through C to obtain an EEG signal classification model, which is used to classify the detected multi-channel EEG signals into preset categories. Step A. Based on the preset categories, obtain a preset number of real categories corresponding to each preset category of multi-channel EEG signals to form each sample, and proceed to Step B; Step B. Construct the network to be trained, including a multinomial feature extraction layer, an LSTM network, an excitatory pathway, an inhibitory pathway, a feature correlation analysis layer, a fusion layer, and a classification layer. The input of the multinomial feature extraction layer constitutes the input of the network to be trained. The output of the multinomial feature extraction layer is connected to the input of the LSTM network. The output of the LSTM network is connected to the input of the excitatory pathway and the input of the inhibitory pathway, respectively. The outputs of the excitatory pathway and the inhibitory pathway are connected to the input of the feature correlation analysis layer. The output of the feature correlation analysis layer is sequentially connected in series with the fusion layer and the classification layer. The output of the classification layer constitutes the output of the network to be trained. Then proceed to step C. Wherein the feature correlation analysis layer is embedded with Wilson-Cowan neural population dynamics equation, the feature correlation analysis layer receives excitatory feature sequence from excitatory pathway output , inhibitory feature sequence from inhibitory pathway output ; Then the feature correlation analysis layer executes the following formula according to the pre-trained state transition neural network corresponding to the excitatory pathway and the pre-trained state transition neural network corresponding to the inhibitory pathway ​ ； ； obtaining an excitatory state rate of change prediction , an inhibitory state rate of change prediction wherein, denotes a concatenation of an excitatory feature sequence and an inhibitory feature sequence , and denote a training parameter corresponding to an excitatory pathway, a training parameter corresponding to an inhibitory pathway, respectively. Then the feature correlation analysis layer updates the state according to the step size of the dynamic evolution process and the preset initial value of the first step feature in the excitatory dynamic evolution process , the preset initial value of the first step feature in the inhibitory dynamic evolution process , using the Euler method to update the state; ； ； Obtaining an excitatory kinetics evolution time series , an inhibitory kinetics evolution time series wherein, denotes the number of steps of the kinetics evolution process, denotes the excitatory state change rate prediction value for the kinetics evolution process step length, denotes the inhibitory state change rate prediction value for the kinetics evolution process step length; The fusion layer receives the excitatory dynamics evolution time series feature sequence output from the feature correlation analysis layer , the inhibitory dynamics evolution time series feature sequence , the last step feature constitutes , the last step feature constitutes , then splicing is performed according to , and the following formula is used:​ ； ； obtaining parameters output to the classification layer, wherein , respectively represent the first weight, the second weight in the fusion layer, , respectively represent the first bias, the second bias in the fusion layer, represent the ReLU activation function, represent the excitatory dynamics evolution time sequence feature sequence in the last step feature , represent the inhibitory dynamics evolution time sequence feature sequence in the last step feature , represent the E / I feature splicing vector, represent the first layer fusion feature vector; The classification layer receives parameters from the fusion layer output respectively for each of the preset categories, according to the following formula: ； ； obtaining a probability of a multi-channel electroencephalogram signal about a first category wherein, denotes a number of categories, denotes a weight matrix corresponding to a first category, denotes a bias corresponding to a first category, denotes a data value of the multi-channel electroencephalogram signal corresponding to a first category, denotes an output of a final classifier of an original category ​ Step C. Based on each sample, take the multi-channel EEG signals in the sample as input, take the predicted category corresponding to the multi-channel EEG signals in the sample as output, and combine the target loss function to train the network to be trained to obtain the EEG signal classification model. The target loss function includes the cross-entropy loss function, the physical residual loss function, the E / I balance loss function, and the L2 regularization loss function. The cross-entropy loss function is as follows: ； Obtain the cross-entropy loss result ,in, Indicates the number of samples. Indicates the number of categories. For the first The sample belongs to the first The true category of each category, This represents the predicted probability of the model. , They represent the first Multichannel EEG signals in each sample, and the predicted categories corresponding to the multichannel EEG signals; The physical residual loss function is as follows: ； Obtain physical residual loss results ,in, This represents the number of steps in the dynamic evolution process. This represents the rate of change during the dynamic evolution of excitability. Represents the rate of change during the inhibitory dynamics evolution process. Represents the first step in the dynamic evolution process. The theoretical rate of change of the excitatory state. Represents the first step in the dynamic evolution process. The theoretical rate of change of the inhibited state; The E / I balance loss function is as follows: ； Obtain E / I balance loss results ,in, Represents the temporal characteristic sequence of excitatory dynamics evolution. middle The final feature, Represents the temporal characteristic sequence of inhibitory dynamics evolution. middle The final feature, This represents the function for finding the average. The L2 regularization loss function is as follows: ； Obtain L2 regularization loss results ,in, This represents the parameters to be trained in the network. This represents the set of all parameters to be trained in the network to be trained. This represents the squared L2 norm of the parameter; Then, according to the following formula: ； Weighted total loss ,in, , , These represent the weights of the physical residual loss, the E / I balance loss, and the L2 regularization loss, respectively.

2. The EEG signal classification method based on physical information residual multinomial network according to claim 1, characterized in that: The polynomial feature extraction layer receives multi-channel EEG signals and performs polynomial expansion on the multi-channel feature data of each time step in the multi-channel EEG signals according to the preset maximum expansion order, so as to obtain the feature sequence corresponding to each time step and output it to the LSTM network. The LSTM network receives the feature sequences corresponding to each time step in the multi-channel EEG signal, and updates the hidden state of the LSTM network for each time step according to the following equation: ； ； ； ； ； ； Obtain the time-series feature sequence output by the LSTM network. These are respectively output to the excitatory pathway and the inhibitory pathway; among them, This indicates the number of time steps in a multichannel EEG signal. Indicates the first of the multichannel EEG signals Feature sequences at each time step Representing time-series feature sequences The Middle Features corresponding to each time step Indicates the output of the candidate memory unit. Indicates the output of the current memory unit. , , , These represent the weight matrices for the forget gate, input gate, output gate, and candidate memory units, respectively. , , , These represent the biases of the forget gate, input gate, output gate, and candidate memory unit, respectively. , , These represent the outputs of the forget gate, the input gate, and the output gate, respectively. express Activation function Indicates the preset initial value. This indicates element-wise multiplication. This indicates the sigmod activation function.

3. The EEG signal classification method based on physical information residual multinomial network according to claim 1, characterized in that: The excitatory pathway receives the temporal feature sequence output from the LSTM network. For each time step, firstly, for each time step... Corresponding features Perform polynomial expansion to obtain the corresponding features Then follow the formula below: ； ； ； ； Obtain the excitatory characteristic sequence output by the excitatory pathway The output is sent to the feature correlation analysis layer, where... This indicates the number of time steps in a multichannel EEG signal. Represents excitatory characteristic sequences The Middle Features corresponding to each time step This represents the weight matrix in the excitatory pathway. This indicates a bias in the excitatory pathway. Represents the normalization function. express Activation function This represents the weighting parameter in the excitatory pathway. This represents the state vector of an excitatory neuron. This represents the normalized result of the excitatory state. This indicates an excitatory pre-activated state.

4. The EEG signal classification method based on physical information residual multinomial network according to claim 1, characterized in that: The inhibitory pathway receives the temporal feature sequence from the output of the LSTM network. For each time step, firstly, for each time step... Corresponding features Perform polynomial expansion to obtain the corresponding features Then follow the formula below: ； ； ； ； ； ； Obtain the inhibitory characteristic sequence output by the inhibitory pathway The output is sent to the feature correlation analysis layer, where... This indicates the number of time steps in a multichannel EEG signal. Indicating inhibitory characteristic sequences The Middle Features corresponding to each time step , These represent the first and second weight matrices in the inhibitory pathway, respectively. , These represent the first and second biases in the inhibitory pathway, respectively. Represents the normalization function. express Activation function This represents the weighting parameter in the inhibitory pathway. Indicates the first The activation coefficient of each neuron in the first layer. Indicating the inhibitory pathway The output of the first layer of neurons, Indicates the first The layer normalization result output by the layer. Indicating the inhibitory pathway The output of the first layer of neurons This indicates an excitatory pre-activated state.

Images