Neurodynamics parameter reasoning model training method and reasoning method based on scalp electroencephalogram

Abstract

The invention relates to the technical field of electroencephalogram signal analysis, and particularly provides a neurodynamic parameter reasoning model training method and reasoning method based on scalp electroencephalograph, comprising the following steps: inputting electroencephalogram characteristics to be reasoned into a reasoning model, and outputting neurodynamic parameters; constructing a target loss function according to the simulated electroencephalogram features, the reconstructed electroencephalogram features and the electroencephalogram features to be reasoned; iteratively training the first coding module according to the target loss function and the first decoding module until the target loss function converges, and obtaining an inference model; the inference model comprises a first coding module, and the electroencephalogram features to be inferred are mapped to a target neurodynamic parameter space through the first coding module to obtain neurodynamic parameters; the simulated electroencephalogram characteristics are obtained by processing an intracranial simulation signal, and the intracranial simulation signal is obtained by simulating according to a target neurodynamic equation; the reconstructed electroencephalogram features are obtained by reconstructing the neurodynamic parameters through the first decoding module. The neurodynamic parameters inferred by the inference model are high in clinical interpretability.

Description

Technical Field

[0001] The present invention relates to the technical field of electroencephalogram signal analysis, and in particular to a training method and an inference method for a neural dynamics parameter inference model based on scalp electroencephalogram. Background Art

[0002] Scalp electroencephalogram (EEG) is a non-invasive method for recording the activities of brain neurons. By placing electrodes on the scalp, scalp electroencephalogram can detect the potential changes generated by the activities of brain neurons and record them as scalp electroencephalogram signals. Based on the scalp electroencephalogram signals, real-time monitoring of the activities of brain neurons can be achieved. Therefore, scalp electroencephalogram is widely used in the fields of medicine, neuroscience, etc.

[0003] In the prior art, although the neuron activity information contained in scalp electroencephalogram signals can be quantified into data analysis from multiple perspectives, the overall clinical interpretability is relatively low. Summary of the Invention

[0004] The training method and the inference method for a neural dynamics parameter inference model based on scalp electroencephalogram provided by the embodiments of the present invention at least solve the problem of low clinical interpretability of the neuron activity information inferred from scalp electroencephalogram signals in the prior art, can establish the connection between scalp electroencephalogram signals and real neuron activities, and have high clinical interpretability.

[0005] In a first aspect, the present invention provides a training method for a neural dynamics parameter inference model based on scalp electroencephalogram, including inputting the electroencephalogram features to be inferred of the brain region to be inferred into an inference model, and outputting corresponding neural dynamics parameters by the inference model; wherein, the inference model includes a first encoding module, which is used to map the electroencephalogram features to be inferred into a target neural dynamics parameter space through the first encoding module to obtain the neural dynamics parameters; constructing a target loss function according to the simulated electroencephalogram features, the reconstructed electroencephalogram features and the electroencephalogram features to be inferred; wherein, the simulated electroencephalogram features are obtained by processing the intracranial simulation signals of the brain region to be inferred, and the intracranial simulation signals are simulated according to a target neural dynamics equation; the reconstructed electroencephalogram features are obtained by reconstructing the neural dynamics parameters through a first decoding module; and iteratively training the first encoding module according to the target loss function and the first decoding module until the target loss function converges, so as to obtain the trained inference model.

[0006] In one embodiment of the present invention, the target loss function includes a first loss function, a second loss function, and a third loss function. Constructing the target loss function based on the simulated EEG features, the reconstructed EEG features, and the EEG features to be inferred includes: constructing the first loss function based on the reconstructed EEG features and the EEG features to be inferred; constructing the second loss function based on the reconstructed EEG features and the simulated EEG features; and constructing the third loss function by performing weighted summation of the first loss function and the second loss function.

[0007] In one embodiment of the present invention, based on the target loss function and the first decoding module, iteratively training the first encoding module until the target loss function converges to obtain the trained inference model includes: iteratively training the first encoding module and the first decoding module according to the first loss function until the first loss function converges to obtain a second encoding module and a second decoding module; iteratively training the second decoding module according to the second loss function until the second loss function converges to obtain a third decoding module; and iteratively training the second encoding module and the third decoding module according to the third loss function until the third loss function converges to obtain the trained third encoding module and the inference model.

[0008] In one embodiment of the present invention, before constructing the target loss function based on the simulated EEG features, the reconstructed EEG features, and the EEG features to be inferred, it further includes: establishing an intracranial signal simulation sub-model corresponding to each brain region to be inferred according to the target neurodynamics equation; coupling the intracranial signal simulation sub-models according to the structural connection coefficients to obtain the intracranial simulation signal; inputting the intracranial simulation signal into the forward inference sub-model, and mapping the intracranial simulation signal by the forward inference sub-model to obtain the simulated EEG signal; and processing the simulated EEG signal to obtain the simulated EEG features.

[0009] In one embodiment of the present invention, before coupling the intracranial signal simulation sub-models according to the structural connection coefficients, it further includes: when the target imaging acquisition condition is satisfied, acquiring the target imaging data corresponding to the brain region to be inferred, and processing the target imaging data to obtain the structural connection coefficients; wherein the processing of the target imaging data includes connectivity functional analysis, structural connectivity analysis, and network analysis; when the target imaging acquisition condition is not satisfied, calculating the mean of the automated anatomical labeling structural connectivity matrices in the public dataset to obtain the structural connection coefficients.

[0010] In one embodiment of the present invention, before inputting the intracranial simulation signal into the forward inference sub-model, it further includes: obtaining a target head model; defining a target source position; setting a target source type; and performing forward solution according to the target head model, the target source position, and the target source type to establish a mapping from the target source to the target scalp potential, thereby obtaining the forward inference sub-model.

[0011] In one embodiment of the present invention, the target neural dynamics equation is set as the Jensen and Ritter equation. According to the Jensen and Ritter equation, an intracranial signal simulation sub-model corresponding to each brain region to be inferred is established, expressed as: , , , , , , , Wherein, and are the change rates of the pyramidal cell membrane potential, is the rate of change of the change rate of the membrane potential of the pyramidal neuron population, is the excitatory synaptic gain constant, is the reciprocal of the excitatory synaptic time constant, is the neuron activation function, is the deviation of the membrane from the resting potential, is the membrane potential of the excitatory neuron population, is the membrane potential of the inhibitory neuron population, is the membrane potential of the pyramidal neuron population, and are the change rates of the membrane potential of the excitatory neuron population, is the rate of change of the change rate of the membrane potential of the excitatory neuron group, is the external input, is the average number of connections from excitatory neurons to pyramidal neurons, is the average number of connections from pyramidal neurons to excitatory neurons, and are the change rates of the membrane potential of the inhibitory neuron population, is the rate of change of the change rate of the membrane potential of the inhibitory neuron group, is the inhibitory synaptic gain constant, is the average number of connections from inhibitory neurons to pyramidal neurons, is the reciprocal of the inhibitory synaptic time constant, is the average number of connections from pyramidal neurons to inhibitory neurons; is the maximum firing rate, is the slope of the neuron activation function, is the average firing threshold; According to the structural connection coefficient, each of the intracranial signal simulation sub-models is coupled to obtain the intracranial simulation signal, expressed as: , , , , , , , , where, and is an integer, is the total number of brain regions to be inferred, and are at the th brain region to be inferred, the rate of change of the pyramidal cell membrane potential, is at the th brain region to be inferred, how fast the rate of change of the pyramidal neuron population membrane potential changes, is at the th brain region to be inferred, the excitatory synaptic gain constant, is at the th brain region to be inferred, the reciprocal of the excitatory time constant, is at the th brain region to be inferred, the membrane potential of the excitatory neuron population, is at the th brain region to be inferred, the membrane potential of the inhibitory neuron population, is at the th brain region to be inferred, the membrane potential of the pyramidal cells, and are at the th brain region to be inferred, the membrane potential of the excitatory neuron population, is at the th brain region to be inferred, how fast the rate of change of the excitatory neuron population membrane potential changes, is at the th brain region to be inferred, the average number of connections from excitatory neurons to pyramidal neurons, is at the The average number of connections from pyramidal neurons to excitatory neurons in the brain region to be inferred, is the external input of the th brain region to be inferred, is the variance of the external input of the th brain region to be inferred, is the th standard deviation of the external input of the th brain region to be inferred, is the th variance of the neuronal firing rate of the and is the rate of change of the membrane potential of the inhibitory neuron population in the th brain region to be inferred, is the inhibitory synaptic gain constant of the th brain region to be inferred, is the reciprocal of the inhibitory time constant of the th brain region to be inferred; Input the intracranial simulation signal into the forward inference sub-model, and the forward inference sub-model maps the intracranial simulation signal to obtain a simulated electroencephalogram signal, expressed as: , where is the simulated electroencephalogram signal, is the intracranial simulation signal, is the mapping of the forward inference sub-model.

[0012] In one embodiment of the present invention, the electroencephalogram (EEG) features to be inferred of the brain region to be inferred are input into an inference model, and the corresponding neurodynamics parameters are output by the inference model, including: inputting the EEG features to be inferred into the convolutional sub-module of the first encoding module, extracting spatial features through the convolutional sub-module to obtain spatial EEG features; inputting the spatial EEG features into the temporal feature extraction sub-module of the first encoding module, capturing temporal dynamic features through the temporal feature extraction sub-module to obtain spatio-temporal EEG features; mapping the spatio-temporal EEG features to the target neurodynamics parameter space to obtain the neurodynamics parameters; or inputting the EEG features to be inferred into the temporal feature extraction sub-module of the first encoding module, capturing temporal dynamic features through the temporal feature extraction sub-module to obtain temporal EEG features; inputting the temporal EEG features into the convolutional sub-module of the first encoding module, extracting spatial features through the convolutional sub-module to obtain spatio-temporal EEG features; mapping the spatio-temporal EEG features to the target neurodynamics parameter space to obtain the neurodynamics parameters.

[0013] In one embodiment of the present invention, mapping the spatio-temporal EEG features to the target neurodynamics parameter space to obtain the neurodynamics parameters includes: through an adjustment function Mapping the spatio-temporal EEG features to the target neurodynamics parameter space is expressed as: , where represents the minimum range of the neurodynamics parameters, represents the maximum range of the neurodynamics parameters, is an activation function, is the neurodynamics parameter.

[0014] In one embodiment of the present invention, before inputting the EEG features to be inferred of the brain region to be inferred into the inference model, it further includes: acquiring the scalp electroencephalogram (EEG) signal to be inferred; preprocessing the scalp EEG signal to be inferred to obtain a preprocessed scalp EEG signal; where the preprocessing includes at least one of resampling, rereferencing, removing power frequency, removing baseline drift, removing invalid high frequencies, removing electrooculogram, electrocardiogram, and electromyogram; performing time slicing and artifact removal on the preprocessed scalp EEG signal to obtain an artifact-removed slice set; extracting features from the artifact-removed slices in the artifact-removed slice set to obtain the EEG features to be inferred.

[0015] In a second aspect, the present invention also provides a method for inferring neurodynamics parameters based on scalp electroencephalogram, including acquiring EEG features to be inferred; inputting the EEG features to be inferred into an inference model, and outputting corresponding neurodynamics parameters by the inference model; where the inference model is trained according to the method for training a neurodynamics parameter inference model based on scalp electroencephalogram described in any one of the above.

[0016] Thirdly, the present invention further provides a training device for a neural dynamics parameter inference model based on scalp electroencephalogram, which is applied to the training method for a neural dynamics parameter inference model based on scalp electroencephalogram described in any one of the above. The device includes an inputter for inputting the electroencephalogram features to be inferred of the brain region to be inferred into the inference model, and the inference model outputs the corresponding neural dynamics parameters. Among them, the inference model includes a first encoding module for mapping the electroencephalogram features to be inferred into the target neural dynamics parameter space through the first encoding module to obtain the neural dynamics parameters; a builder for constructing a target loss function according to the simulated electroencephalogram features, the reconstructed electroencephalogram features and the electroencephalogram features to be inferred. Among them, the simulated electroencephalogram features are obtained by processing the intracranial simulation signals of the brain region to be inferred, and the intracranial simulation signals are simulated according to the target neural dynamics equation; the reconstructed electroencephalogram features are obtained by reconstructing the neural dynamics parameters through a first decoding module; a trainer for iteratively training the first encoding module according to the target loss function and the first decoding module until the target loss function converges to obtain the trained inference model.

[0017] Fourthly, the present invention further provides a neural dynamics parameter inference device based on scalp electroencephalogram, including an acquirer for acquiring the electroencephalogram features to be inferred; a speculator for inputting the electroencephalogram features to be inferred into the inference model, and the inference model outputs the corresponding neural dynamics parameters. Among them, the inference model is trained according to the training method for a neural dynamics parameter inference model based on scalp electroencephalogram described in any one of the above, or the inference model is trained according to the training device for a neural dynamics parameter inference model based on scalp electroencephalogram described in any one of the above.

[0018] Fifthly, the present invention further provides an electronic device, including a processor and a memory storing a program. The program includes instructions that, when executed by the processor, cause the processor to execute the training method for a neural dynamics parameter inference model based on scalp electroencephalogram described in any one of the above, or the instructions, when executed by the processor, cause the processor to execute the neural dynamics parameter inference method based on scalp electroencephalogram described in any one of the above.

[0019] The above technical solutions of the present invention have the following beneficial effects compared with the prior art:

[0020] The training method of the neural dynamics parameter inference model based on scalp electroencephalogram in the present invention first infers neural dynamics parameters according to real electroencephalogram features to be inferred, making the distribution of neural dynamics parameters more realistic. The electroencephalogram features to be inferred can be obtained from scalp electroencephalogram signals, and scalp electroencephalogram signals are easy to measure and obtain. Secondly, the simulated electroencephalogram features in the target loss function are obtained by simulation and processing according to the target neural dynamics equation, and the target neural dynamics equation can provide physical constraints to ensure the physical rationality of the final neural dynamics parameters. Finally, through the first decoding module and the target loss function constructed according to the simulated electroencephalogram features, reconstructed electroencephalogram features, and electroencephalogram features to be inferred, the iterative optimization of the first encoding module is realized, ensuring that the trained inference model can infer and output corresponding neural dynamics parameters according to the electroencephalogram features to be inferred, establishing the connection between scalp electroencephalogram signals and real neuron activities, realizing the interpretation of real brain activities, having high clinical interpretability, and being better applied to clinical practice.

[0021] In addition, through the training method of the neural dynamics parameter inference model based on scalp electroencephalogram in the present invention, the trained inference model can directly infer neural dynamics parameters through real electroencephalogram features to be inferred, and there will be no problems in existing neural dynamics methods, such as the need for linearization processing of nonlinear systems, resulting in errors, requiring accurate system noise and observation noise models, and being unable to process complex feature distributions that may be contained in scalp electroencephalograms. Brief Description of the Drawings

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other embodiments can be obtained according to these drawings. In the drawings:

[0023] Figure 1 It is a schematic flowchart of the training method of the neural dynamics parameter inference model based on scalp electroencephalogram in the preferred embodiment of the present invention.

[0024] Figure 2 It is a schematic structural diagram of the training device of the neural dynamics parameter inference model based on scalp electroencephalogram in the preferred embodiment of the present invention.

[0025] Figure 3 It is a schematic structural diagram of the neural dynamics parameter inference device based on scalp electroencephalogram in the preferred embodiment of the present invention.

[0026] Figure 4 It is a schematic structural diagram of the electronic device in the preferred embodiment of the present invention.

[0027] Among them, the above-mentioned accompanying drawings include the following reference numerals:

[0028] 10. Input device; 20. Builder; 30. Trainer; 40. Fetcher; 50. Speculator; 601. Computing unit; 602. ROM; 603. RAM; 604. Bus; 605. I / O interface; 606. Input unit; 607. Output unit; 608. Storage unit; 609. Communication unit. Detailed implementation manners

[0029] Embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. On the contrary, these embodiments are provided to more thoroughly and completely understand the present invention. It should be understood that the accompanying drawings and embodiments of the present invention are only for exemplary purposes and are not used to limit the protection scope of the present invention.

[0030] It should be noted that scalp electroencephalogram signals, as a multi-channel time series signal, contain rich brain activity information. However, these signals are complex and have a lot of noise, making it difficult to directly interpret. With the development of technology, in recent years, a large number of studies have emerged in the understanding, processing and application of electroencephalogram, including traditional feature extraction methods and deep learning methods.

[0031] Taking the traditional feature extraction method as an example, in the prior art, researchers transformed scalp electroencephalogram signals into features based on time-frequency or spatial relationships through signal processing, and used statistical methods to compare the features to obtain clinical hypothesis results. However, these feature extraction methods are limited by the randomness, complexity and noise components of the signals themselves. To ensure the credibility of the results, the processing process is relatively complex and cumbersome. At the same time, traditional feature extraction methods can only explain brain activities at the macroscopic level and cannot map the explanations to real neuron changes.

[0032] Taking the deep learning method as an example, in the prior art, researchers used the neural network of a deep model to learn the complex components in scalp electroencephalogram to assist in achieving clinical goals and perform scalp electroencephalogram feature extraction. However, as a data-driven model, issues such as a large demand for samples, the risk of overfitting, and cumbersome parameter adjustment all need to be considered. More importantly, deep learning models are usually regarded as "black box models", and there is no direct connection between the feature level and the substances actually existing in clinical practice, making them non-explainable.

[0033] In summary, how to improve clinical interpretability has become an urgent problem to be solved. Among them, clinical interpretability refers to the characteristics that the decision-making logic and output results of artificial intelligence models or algorithms in the medical field can be understood, verified, and trusted by clinicians, patients, or regulatory agencies.

[0034] To solve the above problems, referring to Figure 1 as shown, the present invention provides a training method for a neurodynamics parameter inference model based on scalp electroencephalogram. The training method for the neurodynamics parameter inference model based on scalp electroencephalogram includes:

[0035] First, input the electroencephalogram features to be inferred of the brain region to be inferred into the inference model, and the inference model outputs the corresponding neurodynamics parameters.

[0036] Among them, the inference model includes a first encoding module, which is used to map the electroencephalogram features to be inferred to the target neurodynamics parameter space through the first encoding module to obtain neurodynamics parameters.

[0037] Secondly, construct a target loss function according to the simulated electroencephalogram features, the reconstructed electroencephalogram features, and the electroencephalogram features to be inferred.

[0038] Among them, the simulated electroencephalogram features are obtained by processing the intracranial simulation signals of the brain region to be inferred, and the intracranial simulation signals are simulated according to the target neurodynamics equation. The reconstructed electroencephalogram features are obtained by reconstructing the neurodynamics parameters through the first decoding module.

[0039] Finally, according to the target loss function and the first decoding module, perform iterative training on the first encoding module until the target loss function converges to obtain a trained inference model.

[0040] Specifically, the electroencephalogram features to be inferred are scalp electroencephalogram features. The scalp electroencephalogram features are obtained by processing real scalp electroencephalogram signals. The electroencephalogram features to be inferred are usually a multi-dimensional array, representing time series data or feature vectors.

[0041] Exemplarily, after collecting real scalp electroencephalogram signals, perform corresponding preprocessing on the scalp electroencephalogram signals and perform feature extraction to obtain scalp electroencephalogram features. The acquisition technology of real scalp electroencephalogram signals has been relatively mature and can be easily measured through existing acquisition devices.

[0042] After obtaining the electroencephalogram features to be inferred, input the electroencephalogram features to be inferred into the inference model. Specifically, input the electroencephalogram features to be inferred into the first encoding module, and map the electroencephalogram features to be inferred to the target neurodynamics parameter space through the first encoding module to obtain neurodynamics parameters.

[0043] The first encoding module is a deep encoding network built based on a deep learning framework. Those skilled in the art can screen the parameters to be trained from the target neural dynamics equation according to the scalp electroencephalogram signals that need to be analyzed actually, use the corresponding electroencephalogram features to be inferred as inputs, and use the latent space (parameter space) with the same scale as the number of parameters to be trained as outputs, so that the electroencephalogram features to be inferred are mapped into the prior value range of neural dynamics parameters, and then the corresponding neural dynamics parameters are obtained.

[0044] Neural dynamics parameters have higher interpretability, and their changes can be described as changes in the neural activities of brain regions.

[0045] Among them, those skilled in the art can design the architecture of the first encoding module according to the types of parameters that need to be encoded actually, including but not limited to, input layer, convolutional layer, pooling layer, recurrent layer or Transformer layer, fully connected layer, output layer, etc.

[0046] Those skilled in the art can set the corresponding target neural dynamics equation according to actual needs. Exemplarily, in the prior art, neural dynamics equations include Jansen & Rit equation, Hodgkin-Huxley equation, Fitzhugh-Nagumo equation, etc. Each equation has its own characteristics and scope of application, and can be used to study the dynamic behaviors of the nervous system at different levels, such as the dynamics of single units, neuron populations, and cerebral cortex regions.

[0047] It should be noted that at the beginning of training, the neural dynamics parameters output by the first encoding module may not be accurate. Therefore, it is necessary to construct a target loss function based on the simulated electroencephalogram features, reconstructed electroencephalogram features, and electroencephalogram features to be inferred, so as to iteratively train the first encoding module through the target loss function.

[0048] Among them, the simulated electroencephalogram features are obtained by processing the intracranial simulation signals of the brain region to be inferred. Exemplarily, the intracranial simulation signals of the brain region to be inferred are processed to obtain simulated electroencephalogram signals. The simulated electroencephalogram signals are simulated scalp electroencephalogram signals. By processing the simulated electroencephalogram signals, the corresponding simulated electroencephalogram features can be obtained.

[0049] The intracranial simulation signals are simulated according to the target neural dynamics equation. During iterative training, physical constraints are provided by the target neural dynamics equation, which can ensure the physical rationality of the final neural dynamics parameters.

[0050] The brain region to be inferred is the brain region corresponding to the EEG features to be inferred. Preferably, using the Automated Anatomical Labeling (AAL) standard, select or directly define the brain regions for which simulated EEG signals need to be generated, providing a location basis for subsequent forward inference of simulated EEG signals based on intracranial simulation signals.

[0051] Automated Anatomical Labeling is a labeling and system for standardizing the naming of brain structures. It provides a unified naming scheme for each region of the brain, facilitating communication and comparison among researchers during the research process. This system assigns a unique label to each region of the brain and is commonly used in neuroimaging research. It usually includes structures in the left and right cerebral hemispheres, such as the frontal lobe, temporal lobe, parietal lobe, occipital lobe, etc. These labels enable the identification and comparison of the same regions of the brain between different studies and even between different imaging techniques.

[0052] Exemplarily, using the Automated Anatomical Labeling standard, directly define ninety brain regions. In the case of only wanting to obtain the parameters of frontal lobe neurons, only select the regions of interest (ROI) in the frontal lobe and correspondingly generate only the scalp EEG signals corresponding to that brain region.

[0053] The reconstructed EEG features are obtained by reconstructing the neural dynamics parameters through the first decoding module. Among them, the first decoding module is also a deep network. The structures of the first decoding module and the first encoding module are symmetric, and the two form a deep model with an auto - encoding architecture, capable of handling complex non - linear relationships.

[0054] After mapping and outputting the EEG features to be inferred through the first encoding module to obtain neural dynamics parameters, the first decoding module extracts information from the corresponding parameter space to generate a reconstructed EEG signal that is highly similar to the true EEG features to be inferred. The reconstructed EEG signal is the scalp EEG signal reconstructed by the first decoding module. Processing based on the reconstructed EEG signal can obtain the corresponding reconstructed EEG features, realizing the reconstruction from neural dynamics parameters to EEG features.

[0055] The first decoding module acts as an agent module, optimizing the problem that some neural dynamics behaviors are non - differentiable. It can still propagate gradients through a differentiable neural network structure, thus allowing the use of gradient - based optimization methods, effectively avoiding the problem that gradients cannot be backpropagated when only using neural dynamics equations for encoding.

[0056] In the case where the input is scalp electroencephalogram signals, the autoencoder architecture deep model composed of the first decoding module and the first encoding module realizes an end-to-end training process. Among them, end-to-end learning means directly obtaining the final target result (such as neurodynamics parameters) from the original data (such as scalp electroencephalogram signals), without the need for complex and cumbersome processing like traditional feature extraction methods in the middle.

[0057] Those skilled in the art can construct corresponding objective loss functions according to actual needs, including but not limited to, constructing based on reconstructed EEG features and EEG features to be inferred, constructing based on reconstructed EEG features and simulated EEG features, and so on. Specific loss functions include but are not limited to mean squared error loss (MSE, Mean Squared Error), mean absolute error loss (MAE, Mean Absolute Error), etc. Of course, it can also be a difference function quantified based on the statistical characteristics of features, such as waveform similarity or the quantification of differences in other related features.

[0058] After obtaining the objective loss function, the first encoding module is iteratively trained according to the objective loss function and the first decoding module until the objective loss function converges, ensuring that the inference from the real EEG features to be inferred to the neurodynamics parameters can be realized, and an inference model that has completed training is obtained.

[0059] In summary, for the method for training a neurodynamics parameter inference model based on scalp EEG of the present invention, first, the neurodynamics parameters are inferred based on the real EEG features to be inferred, making the distribution of the neurodynamics parameters more realistic. The EEG features to be inferred can be obtained from scalp electroencephalogram signals, and scalp electroencephalogram signals are easy to measure and obtain. Secondly, the simulated EEG features in the objective loss function are obtained by simulating and processing according to the target neurodynamics equation, and the target neurodynamics equation can provide physical constraints to ensure the physical rationality of the final neurodynamics parameters. Finally, through the first decoding module and the objective loss function constructed based on the simulated EEG features, reconstructed EEG features, and EEG features to be inferred, the iterative optimization of the first encoding module is realized, ensuring that the inference model that has completed training can infer and output the corresponding neurodynamics parameters according to the EEG features to be inferred, establishing a connection between scalp electroencephalogram signals and real neuron activities, realizing real brain activity interpretation, having high clinical interpretability, and being better applicable to clinical practice.

[0060] It should be noted that in existing neurodynamics methods, there are also methods for inferring kinetic model parameters from scalp electroencephalogram data, such as Kalman filtering and variational Bayesian inference.

[0061] Among them, although Kalman filtering can provide parameter estimation, it requires linearization processing for nonlinear systems, which may introduce errors, and accurate system noise and observation noise models are needed. Although variational Bayesian inference can provide an estimate of the complete posterior distribution and can incorporate the prior knowledge provided by real electroencephalograms, it is sensitive to the choice of prior distribution and cannot handle the complex feature distributions that may be contained in scalp electroencephalograms.

[0062] However, through the method for training a neural dynamics parameter inference model based on scalp electroencephalogram described in the present invention, the trained inference model can directly infer neural dynamics parameters through real electroencephalogram features to be inferred, and the above problems will not occur.

[0063] In some embodiments, the method for training a neural dynamics parameter inference model based on scalp electroencephalogram described in the present invention further includes, before inputting the electroencephalogram features to be inferred of the brain region to be inferred into the inference model:

[0064] First, obtain the scalp electroencephalogram signal to be inferred.

[0065] Exemplarily, collect the scalp electroencephalogram signal according to the clinical standard electroencephalogram amplifier to obtain the scalp electroencephalogram signal to be inferred. Among them, the sampling rate range is set to 200 to 1000 Hz, and the number of channels range is set to 16 to 32.

[0066] Secondly, preprocess the scalp electroencephalogram signal to be inferred to obtain a preprocessed scalp electroencephalogram signal. Among them, the preprocessing includes at least one of resampling, rereferencing, removing power frequency, removing baseline drift, removing invalid high frequencies, and removing electrooculogram, electrocardiogram, and electromyogram.

[0067] Preferably, resample, rereference, remove power frequency, remove baseline drift, remove invalid high frequencies, and remove electrooculogram, electrocardiogram, and electromyogram from the scalp electroencephalogram signal to be inferred.

[0068] Among them, through resampling processing, the data volume can be effectively reduced and the calculation speed can be improved. Preferably, the sampling rate is set to 128 Hz to satisfy the sampling theorem while reducing the data volume.

[0069] Through rereferencing processing, the influence of the reference electrode position can be removed, the volume conduction effect can be reduced, and it is convenient for standardization and comparison. Those skilled in the art can set different rereferencing methods according to actual needs, including but not limited to average reference and ear-level reference.

[0070] Preferably, design a notch filter (such as 50 Hz) to eliminate power frequency noise interference and remove power frequency noise.

[0071] Through the de - baseline drift processing, the influence brought by data drift can be effectively eliminated. Preferably, the de - baseline drift processing is achieved by means of high - pass filtering. Exemplarily, the high - pass filtering frequency can be set to 1 Hz.

[0072] Through the removal of invalid high - frequency processing, the high - frequency component noise that does not contain physiological or pathological significance in the scalp electroencephalogram signal to be inferred can be effectively removed. Preferably, a low - pass filter (such as 40 Hz) is designed to remove high - frequency components.

[0073] Preferably, electro - oculogram, electrocardiogram, and electromyogram are removed through independent component analysis (ICA). Exemplarily, by performing a linear transformation on the original signal, the signal is decomposed into a set of mutually independent components using the non - Gaussianity of statistical characteristics. Each component represents a different source. The components with sources of electro - oculogram, electrocardiogram, and electromyogram are removed, and then the remaining components are inverse - transformed back to the original signal.

[0074] After completing the corresponding pre - processing operations, time slicing and artifact removal are performed on the pre - processed scalp electroencephalogram signal to obtain an artifact - removed slice set.

[0075] Among them, through time slicing of the pre - processed scalp electroencephalogram signal, the scalp electroencephalogram signal is sliced into multiple segments to obtain a segment set.

[0076] Exemplarily, first set the observation time unit T t , and slice with T t as the step size. After completing the slicing, for segments with a time greater than T t , slice again with T t as the step size. Those skilled in the art can set the specific step size according to actual needs. Through time slicing, the EEG acquisition time with an unfixed length can be determined to an estimated length, ensuring the consistency of subsequent feature extraction and facilitating input and improving the overall efficiency.

[0077] After completing the time slicing, artifact removal is performed to remove EEG segments containing artifacts, avoiding the adverse effects caused by large - amplitude artifacts due to human body jitter, poor equipment contact, etc.

[0078] Exemplarily, set an artifact threshold, record the time points where the amplitude exceeds the threshold, mark the remaining time points as non - artifact times, and retain the EEG segments with a continuous duration greater than T t as the non - artifact segment set. The non - artifact segment set with a time greater than T t is sliced with a step size of T t . After completing the relevant operations, an artifact - removed slice set is obtained.

[0079] Finally, feature extraction is performed on the artifact-removed slices in the artifact-removed slice set to obtain the EEG features to be inferred.

[0080] Exemplarily, select the channels to focus on. Within each artifact-removed slice and for each channel, use a dimensionality reduction method or a feature extraction method to transform the EEG signal into low-dimensional feature data, obtaining the EEG features to be inferred. Among them, relevant features such as spectrum and connectivity can be used. Spectrum features include absolute and relative band energies calculated after short-time Fourier transform, etc. Connectivity features include amplitude connectivity obtained by calculating the correlation coefficient of the EEG signal, or phase-locking values calculated from the phases of the signals, etc.

[0081] By performing feature extraction on the artifact-removed slices, it is possible to extract the effective information in the signal, remove more noise components, enabling the inference model to be associated with more effective EEG features.

[0082] It should be noted that in the case of processing real scalp EEG signals, obtaining the EEG features to be inferred, and constructing an objective loss function based on the EEG features to be inferred, simulated EEG features, and reconstructed EEG features to cooperate with the first decoding module to iteratively train the first encoding module, an end-to-end training process can be achieved.

[0083] In some embodiments of the method for training a neural dynamics parameter inference model based on scalp EEG of the present invention, the architecture of the first encoding module can be flexibly adjusted according to the actual training effect. Exemplarily, the first encoding module includes a convolutional sub-module and a temporal feature extraction sub-module.

[0084] Among them, the convolutional sub-module includes a multi-layer convolutional neural network and a residual structure. The convolutional sub-module is used to extract spatial features. The convolutional sub-module identifies local features, such as waveforms, sharp waves, slow waves, etc., through a series of convolutional layers. These convolutional layers may include different convolutional kernel sizes, strides, and padding strategies to capture features at different scales.

[0085] The temporal feature extraction sub-module includes multiple temporal neural networks, such as long short-term memory networks (LSTM), recurrent neural networks (RNN), etc. The temporal feature extraction sub-module can capture temporal dynamic features. Through the temporal feature extraction sub-module, it is possible to capture the changes of features over time and understand the temporal dynamics of the signal.

[0086] Exemplarily, input the EEG features to be inferred of the brain region to be inferred into the inference model, and the inference model outputs the corresponding neural dynamics parameters, including:

[0087] First, input the electroencephalogram (EEG) features to be inferred into the convolutional sub-module of the first encoding module. Extract spatial features through the convolutional sub-module to obtain spatial EEG features. The spatial EEG features represent the feature representation of the scalp electroencephalogram signal to be inferred in the spatial dimension.

[0088] Secondly, input the spatial EEG features into the temporal feature extraction sub-module of the first encoding module. Capture temporal dynamic features through the temporal feature extraction sub-module to obtain spatio-temporal EEG features. Based on the spatial EEG features, the spatio-temporal EEG features also encode the dynamic changes of the scalp electroencephalogram signal to be inferred in the temporal dimension.

[0089] Finally, map the spatio-temporal EEG features to the target neurodynamics parameter space to obtain neurodynamics parameters.

[0090] In some other embodiments, the processing order of the convolutional sub-module and the temporal feature extraction sub-module can also be adjusted, including:

[0091] First, input the EEG features to be inferred into the temporal feature extraction sub-module of the first encoding module. Capture temporal dynamic features through the temporal feature extraction sub-module to obtain temporal EEG features. The temporal EEG features encode the dynamic changes of the scalp electroencephalogram signal to be inferred in the temporal dimension.

[0092] Secondly, input the temporal EEG features into the convolutional sub-module of the first encoding module. Extract spatial features through the convolutional sub-module to obtain spatio-temporal EEG features. Based on the temporal EEG features, the spatio-temporal EEG features also represent the feature representation of the scalp electroencephalogram signal to be inferred in the spatial dimension.

[0093] Finally, map the spatio-temporal EEG features to the target neurodynamics parameter space to obtain neurodynamics parameters.

[0094] Preferably, in some embodiments of the method for training a neurodynamics parameter inference model based on scalp electroencephalogram of the present invention, the first encoding module includes an input layer, a convolutional layer, a pooling layer, a recurrent layer or a Transformer layer, a fully connected layer, and an output layer.

[0095] First, receive the EEG features to be inferred through the input layer of the first encoding module, and input the EEG features to be inferred into the convolutional layer of the first encoding module. Extract local features through the convolutional layer to obtain the first EEG features. Through the extraction of local features by the convolutional layer, the spatial and temporal patterns in the EEG features to be inferred can be recognized. Among them, those skilled in the art can select an appropriate convolutional kernel size according to actual needs and determine the stride of the convolutional operation to ensure the continuity of the time series.

[0096] Secondly, input the first EEG feature into the pooling layer of the first encoding module. The pooling layer reduces the data dimension while retaining important feature information to obtain the second EEG feature.

[0097] Next, input the second EEG feature into the recurrent layer or Transformer layer of the first encoding module. The recurrent layer or Transformer layer captures the long-term dependencies in the time series data to obtain the third EEG feature.

[0098] Then, input the third EEG feature into the fully connected layer of the first encoding module. The fully connected layer maps the third EEG feature to the target neurodynamics parameter space.

[0099] Finally, output through the output layer of the first encoding module to obtain the neurodynamics parameters.

[0100] To keep the value range of the neurodynamics parameters within a reasonable range, it is necessary to determine the range of the neurodynamics parameters in advance according to the selected neurodynamics equation, define the activation function, and implement the adjustment of the distribution of the target neurodynamics parameter space to ensure that the correct mapping logic can be achieved.

[0101] Among them, the mapping scheme includes but is not limited to using a scaled Sigmoid function.

[0102] Preferably, taking the Sigmoid function as an example of the activation function, by adjusting the function map the spatio-temporal EEG feature to the target neurodynamics parameter space, expressed as:

[0103] .

[0104] In the formula, represents the minimum range of the neurodynamics parameter, represents the maximum range of the neurodynamics parameter. The minimum range and maximum range of the neurodynamics parameter are both determined in advance according to the selected target neurodynamics equation. is the neurodynamics parameter output by the first encoding module.

[0105] By setting the adjustment function, the rationality of the finally trained inference model can be effectively guaranteed, so that the parameters are correctly mapped, facilitating the correct inference of the corresponding target neurodynamics equation.

[0106] In some embodiments of the method for training a neurodynamics parameter inference model based on scalp EEG of the present invention, before constructing the target loss function according to the simulated EEG feature, the reconstructed EEG feature, and the EEG feature to be inferred, it further includes:

[0107] First, according to the target neurodynamics equation, an intracranial signal simulation sub-model corresponding to each brain region to be inferred is established.

[0108] Exemplarily, equations such as the Jensen and Rit equations, the Hodgkin-Huxley equation, and the FitzHugh-Nagumo equation can be selected as the target neurodynamics equation.

[0109] Secondly, according to the structural connection coefficients, each intracranial signal simulation sub-model is coupled to obtain an intracranial simulation signal.

[0110] The structural connection coefficient is the basis for coupling each intracranial signal simulation sub-model. Through the structural connection coefficient, an additional input of the intracranial signal simulation sub-model at each individual position is obtained, and then a coupled model for multi-brain region coupling is realized.

[0111] The structural connection coefficient needs to be processed before coupling each intracranial signal simulation sub-model. Exemplarily, after selecting or directly defining the brain regions for which simulated electroencephalogram signals need to be generated through the automated anatomical labeling standard, it is necessary to determine whether the target object has the target imaging acquisition conditions. For example, whether it has magnetic resonance imaging acquisition conditions.

[0112] In the case of having the target imaging acquisition conditions, the target imaging data corresponding to the brain region to be inferred is obtained, and the target imaging data is processed to obtain the structural connection coefficient.

[0113] Among them, structural magnetic resonance imaging data (sMRI), functional magnetic resonance imaging data (fMRI), diffusion tensor imaging data (DTI), etc. can be selected as the target imaging data, and the processing of the target imaging data includes connectivity functional analysis, structural connectivity analysis, and network analysis.

[0114] Taking functional magnetic resonance imaging data and diffusion tensor imaging data as examples, after collecting the two types of data, first use the functional magnetic resonance imaging data to calculate the time series correlation between different automated anatomical labeling brain regions to evaluate functional connectivity. Secondly, use the diffusion tensor imaging data to reconstruct the protein fiber bundles between brain regions through fiber tracking technology. Then, analyze the connection pattern of the fiber bundles to determine the structural connection between the automated anatomical labeling brain regions. Finally, use methods such as graph theory to regard the representative positions of the automated anatomical labeling brain regions as the nodes of the network center and the connection relationship as the edges, and perform data analysis to obtain the structural connection coefficient.

[0115] In the case where the target imaging acquisition conditions are not met, calculate the mean of the automated anatomical landmark structural connection matrix in the public dataset to obtain the structural connection coefficient.

[0116] Subsequently, after obtaining the intracranial simulation signal, input the intracranial simulation signal into the forward inference sub-model, and the forward inference sub-model maps the intracranial simulation signal to obtain the simulated electroencephalogram (EEG) signal.

[0117] The forward inference sub-model is pre-constructed, and it can provide a model basis for inferring the simulated EEG signal from the intracranial simulation signal.

[0118] The steps for constructing the forward inference sub-model include: obtaining the target head model, defining the target source location, and setting the target source type. Based on the target head model, target source location, and target source type, perform forward solution to establish the mapping from the target source to the target scalp potential, thereby obtaining the forward inference sub-model.

[0119] The target head model can be set according to whether the target object has the target imaging acquisition conditions.

[0120] Exemplarily, in the case where the target object does not have the acquisition conditions for structural magnetic resonance imaging (MRI) or functional MRI, a standard head model in the public dataset can be used.

[0121] Common standard head models include, but are not limited to: the International 10 / 20 System model, the Colin 27 model, the MNI model (Montreal Neurological Institute), etc. These models usually come with detailed anatomical structure and conductivity information.

[0122] After determining the EEG signal source to be simulated, establish the mapping from brain activity to scalp potential, including defining the source location, determining the source time course, establishing a standard head model, using the standard head model to define the geometry and conductivity of the head, calculating the forward solution, establishing a linear system from source activity to scalp potential, and obtaining the forward inference sub-model.

[0123] In the case where the target object has the acquisition conditions for structural MRI or functional MRI, the structural MRI data of the target object can be used to create the target head model. The created target head model includes the geometry and conductivity of different tissue layers such as the scalp, skull, cerebrospinal fluid, cerebral gray matter, and white matter.

[0124] Subsequently, the target source position is defined. Exemplarily, based on functional magnetic resonance imaging data, the region with the strongest activity is determined to achieve the positioning of the target source position. Then, a suitable source model, such as a point source, a dipole source, or a distributed source model, is selected to simulate brain activity.

[0125] Finally, by calculating the diffusion of the electric field, the potential distribution of brain activity on the scalp surface is estimated, and a mapping from the target source to the target scalp potential is established to obtain the forward inference sub-model. The calculation methods include, but are not limited to, the boundary element method (BEM), the finite element method (FEM), etc.

[0126] Subsequently, after mapping the intracranial simulation signal through the forward inference sub-model to obtain the simulated electroencephalogram signal, the simulated electroencephalogram signal is processed to obtain simulated electroencephalogram features. Among them, the same steps as those for processing the real scalp electroencephalogram signal to be inferred can be selected to process the simulated electroencephalogram signal to obtain the simulated electroencephalogram features, which will not be elaborated here.

[0127] Similarly, when decoding and reconstructing through the first decoding module, the reconstructed electroencephalogram signal is also obtained, and the reconstructed electroencephalogram signal needs to be processed accordingly to obtain the reconstructed electroencephalogram features. Among them, the same steps as those for processing the real scalp electroencephalogram signal to be inferred can be selected to process the reconstructed electroencephalogram signal to obtain the reconstructed electroencephalogram features, which will not be elaborated here.

[0128] It should be noted that the consistency of each feature should be ensured.

[0129] Furthermore, in some embodiments of the method for training the neural dynamics parameter inference model based on scalp electroencephalogram of the present invention, the target neural dynamics equation is set as the Jensen and Rit equations.

[0130] The Jensen and Rit equations are classical models in neuroscience for simulating the activity of population neurons in the cerebral cortex. They describe the interactions between neuron populations through differential equations, especially focusing on the dynamic characteristics of the visual cortex, and can simulate the rhythmic activities in scalp electroencephalograms.

[0131] The Jensen and Rit equations mainly include three neuron populations, namely pyramidal neurons, excitatory neurons, and inhibitory neurons. Pyramidal neurons are used to receive external inputs (such as sensory stimuli) and local feedback, excitatory neurons are used to provide positive feedback, and inhibitory neurons are used to provide negative feedback.

[0132] According to the Jensen and Rit equations, an intracranial signal simulation sub-model corresponding to each brain region to be inferred is established, expressed as:

[0133] ,

[0134] ,

[0135] ,

[0136] ,

[0137] ,

[0138] ,

[0139] .

[0140] Among them, the established intracranial signal simulation sub-model is a sixth-order state space model.

[0141] In the formula, and are the change rates of the pyramidal cell membrane potential, is the rate of change of the change rate of the pyramidal neuron population membrane potential, is the excitatory synaptic gain constant, is the reciprocal of the excitatory synaptic time constant, is the neuron activation function, is the deviation of the membrane from the resting potential, is the membrane potential of the excitatory neuron population, is the membrane potential of the inhibitory neuron population, is the membrane potential of the pyramidal neuron population, and are the change rates of the membrane potential of the excitatory neuron population, is the rate of change of the change rate of the excitatory neuron population membrane potential, is the external input, is the average number of connections from excitatory neurons to pyramidal neurons, is the average number of connections from pyramidal neurons to excitatory neurons, and are the change rates of the membrane potential of the inhibitory neuron population, is the rate of change of the change rate of the inhibitory neuron population membrane potential, is the inhibitory synaptic gain constant, is the average number of connections from inhibitory neurons to pyramidal neurons, is the reciprocal of the inhibitory synaptic time constant, is the average number of connections from pyramidal neurons to inhibitory neurons; is the maximum firing rate, is the slope of the neuron activation function, is the average firing threshold.

[0142] Under the assumption that the co-activation of excitatory neurons and inhibitory cells is proportional:

[0143] ,

[0144] .

[0145] To represent the activities of neuronal activities in the whole brain, according to the structural connection coefficients, each intracranial signal simulation sub-model is coupled so that each intracranial signal simulation sub-model is coupled through excitatory connections to obtain an intracranial simulation signal, expressed as:

[0146] ,

[0147] ,

[0148] ,

[0149] ,

[0150] ,

[0151] ,

[0152] ,

[0153] .

[0154] In the formula, and is an integer, is the total number of brain regions to be inferred, and are at the th brain region to be inferred, the change rate of the pyramidal cell membrane potential, is at the th brain region to be inferred, how fast the change rate of the pyramidal neuron population membrane potential changes, is at the th brain region to be inferred, the excitatory synaptic gain constant, is at the th brain region to be inferred, the reciprocal of the excitatory time constant, is at the th brain region to be inferred, the membrane potential of the excitatory neuron population, is at the th brain region to be inferred, the membrane potential of the inhibitory neuron population, is at the th brain region to be inferred, the membrane potential of the pyramidal cells, and are at the th brain region to be inferred, the membrane potential of the excitatory neuron population, The rate of change of the membrane potential of the excitatory neuron population in the th brain region to be inferred, The average number of connections from excitatory neurons to pyramidal neurons in the th brain region to be inferred, The average number of connections from pyramidal neurons to excitatory neurons in the th brain region to be inferred, The external input to the th brain region to be inferred, The sum of the structural connection coefficients between the th brain region to be inferred and all other brain regions to be inferred except its own location, The variance of the external input to the th brain region to be inferred, The adjustment coefficient, The standard deviation of the external input to the th brain region to be inferred, The directional connection coefficient between the th brain region to be inferred and the th brain region to be inferred, The standard deviation of the neuronal firing rate in the th brain region to be inferred, The variance of the neuronal firing rate in the th brain region to be inferred, and The rate of change of the membrane potential of the inhibitory neuron population in the th brain region to be inferred, The rate of change of the rate of change of the membrane potential of the inhibitory neuron population in the th brain region to be inferred, The inhibitory synaptic gain constant in the th brain region to be inferred, The average number of connections from inhibitory neurons to pyramidal neurons in the th brain region to be inferred, The reciprocal of the inhibitory time constant in the th brain region to be inferred, The average number of connections from pyramidal neurons to inhibitory neurons in the th brain region to be inferred.

[0155] Among them, in order to isolate the influence of coupling on the oscillatory dynamics and ensure the maintenance of the balanced activity of each population, by scaling the influence of other regions relative to this neuronal region, the mean and variance of the presynaptic input of each population are ensured not to change with coupling.

[0156] Meanwhile, the adjustment coefficient The conservation constraint on the input of other inputs to the neuron population at this position is also realized.

[0157] The variance of the neuron firing rate of the th brain region to be inferred In, the symbol " " indicates that the equation can change the output representation at other positions according to the task requirements. Exemplarily, a time delay parameter can be added. By confirming the variance of the neuron firing rate of the th brain region to be inferred, the transient influence can be removed when determining the input of the neural population at other positions. Multiply the variance of the neuron firing rate of the th brain region to be inferred by the adjustment coefficient, so as to ensure that the mean and variance of the input of each population do not change with the change of coupling.

[0158] Of course, in some other embodiments, other neural dynamics equations can also be selected, which will not be elaborated here.

[0159] Taking the th brain region to be inferred as an example, after coupling the intracranial signal simulation submodels to obtain the coupled model, in the actual simulation task, when the parameters in the formula are determined, the state variables are initialized to obtain the intracranial simulation signals of the th brain region to be inferred. The determination of the parameters and the initialization of the relevant variables both belong to the prior art and will not be elaborated here.

[0160] Subsequently, the intracranial simulation signals are input into the forward inference submodel, and the forward inference submodel maps the intracranial simulation signals to obtain the simulated electroencephalogram signals, which are expressed as:

[0161] .

[0162] Among them, is the simulated electroencephalogram signal, is the intracranial simulation signal, is the mapping of the forward inference submodel.

[0163] In some embodiments of the neural dynamics parameter inference model training method based on scalp electroencephalogram according to the present invention, the target loss function includes a first loss function, a second loss function, and a third loss function.

[0164] Constructing the target loss function according to the simulated electroencephalogram features, the reconstructed electroencephalogram features, and the electroencephalogram features to be inferred includes:

[0165] Construct a first loss function based on the reconstructed EEG features and the EEG features to be inferred. By constructing a loss function with the reconstructed EEG features and the EEG features to be inferred, during subsequent training, the first encoding module can achieve reasonable feature compression from the real EEG features to be inferred to the latent space, ensuring successful encoding.

[0166] Construct a second loss function based on the reconstructed EEG features and the simulated EEG features. By constructing a loss function with the reconstructed EEG features and the simulated EEG features, during subsequent training, the first decoding module can be trained to simulate the simulation process based on neurodynamics.

[0167] After completing the construction of the above two loss functions, perform a weighted sum of the first loss function and the second loss function to construct a third loss function. Among them, those skilled in the art can set corresponding weights according to actual needs. Through the third loss function, the corresponding modules can be adapted to each other and infer from EEG features to neurodynamics parameters in the best way, thereby establishing the connection between scalp electroencephalogram signals and real neuron activities, realizing real brain activity interpretation, improving clinical interpretability, and better applying it to clinical practice.

[0168] Specific loss functions include but are not limited to mean squared error loss, mean absolute error loss, etc. Of course, it can also be a difference function quantified based on the statistical characteristics of features, such as waveform similarity or the quantification of differences in other relevant features.

[0169] Furthermore, according to the target loss function and the first decoding module, iteratively train the first encoding module until the target loss function converges to obtain a trained inference model, including:

[0170] First, according to the first loss function, iteratively train the first encoding module and the first decoding module until the first loss function converges to obtain a second encoding module and a second decoding module.

[0171] It can be understood that the second encoding module is the first encoding module that has been initially trained, and the second decoding module is the first decoding module that has been initially trained. In the above steps, the first encoding module and the first decoding module are iteratively trained according to the first loss function, enabling the first encoding module to learn how to accurately encode the real EEG features to be inferred into the target neurodynamics parameter space, and enabling the first decoding module to learn how to decode and reconstruct the reconstructed EEG features. After completion of training, the real EEG features to be inferred can be successfully encoded into the target neurodynamics parameter space.

[0172] Second, according to the second loss function, iteratively train the second decoding module until the second loss function converges to obtain a third decoding module.

[0173] It can be understood that the third decoding module is the second decoding module that has completed training. In the above steps, by guiding the second decoding module to achieve optimization based on the simulated EEG features obtained according to the target neurodynamics equation, physical constraints on the second decoding module can be realized, ensuring the physical rationality of the parameters, and ensuring that the output of the third decoding module that has completed training is not only similar to the real EEG features but also conforms to the simulation results of neurodynamics.

[0174] Finally, according to the third loss function, the second encoding module and the third decoding module are iteratively trained until the third loss function converges, and the third encoding module and the inference model that have completed training are obtained.

[0175] It can be understood that the third encoding module is the second encoding module that has completed training. In the above steps, the second encoding module and the third decoding module can be made to adapt to each other, ensuring that the third encoding module that has completed training can, in the best way, realize the inference from real EEG features to neurodynamics parameters, thereby establishing the connection between scalp electroencephalogram signals and real neuron activities, realizing the interpretation of real brain activities, improving clinical interpretability, and being better applied to clinical practice. In addition, the problem that only using the neurodynamics equation as the encoding module and the gradient cannot be backpropagated is also avoided.

[0176] It should be noted that corresponding convergence conditions need to be set during training, including but not limited to that the loss function no longer significantly decreases within a certain number of training epochs, etc. Once the stop condition is met, the training can be stopped.

[0177] The present invention also provides a method for inferring neurodynamics parameters based on scalp electroencephalogram. The method for inferring neurodynamics parameters based on scalp electroencephalogram includes:

[0178] Obtain the EEG features to be inferred.

[0179] Among them, the EEG features to be inferred are scalp electroencephalogram features. The scalp electroencephalogram features are obtained by processing real scalp electroencephalogram signals. The EEG features to be inferred are usually a multi-dimensional array representing time series data or feature vectors. Exemplarily, after collecting real scalp electroencephalogram signals, corresponding preprocessing and feature extraction are performed on the scalp electroencephalogram signals to obtain scalp electroencephalogram features.

[0180] Input the EEG features to be inferred into the inference model, and the inference model outputs the corresponding neurodynamics parameters.

[0181] Among them, the inference model is trained according to the method for training a neurodynamics parameter inference model based on scalp electroencephalogram described in any of the above embodiments. On this basis, the obtained neurodynamics parameters can realize the interpretation of real brain activities, have high clinical interpretability, and can be well applied to clinical practice.

[0182] Preferably, for the method for inferring neurodynamics parameters based on scalp electroencephalogram according to the present invention, obtaining the electroencephalogram features to be inferred includes:

[0183] First, obtain the scalp electroencephalogram signal to be inferred.

[0184] Exemplarily, collect the scalp electroencephalogram signal according to the clinical standard electroencephalogram amplifier to obtain the scalp electroencephalogram signal to be inferred. Among them, the sampling rate range is set to 200 to 1000 Hz, and the number of channels range is set to 16 to 32.

[0185] Second, preprocess the scalp electroencephalogram signal to be inferred to obtain the preprocessed scalp electroencephalogram signal. Among them, the preprocessing includes at least one of resampling, rereferencing, removing power frequency, removing baseline drift, removing invalid high frequencies, and removing electrooculogram, electrocardiogram, and electromyogram.

[0186] Preferably, perform resampling, rereferencing, removing power frequency, removing baseline drift, removing invalid high frequencies, and removing electrooculogram, electrocardiogram, and electromyogram on the scalp electroencephalogram signal to be inferred in sequence.

[0187] Among them, through resampling processing, the data volume can be effectively reduced and the calculation speed can be improved. Preferably, set the sampling rate to 128 Hz so as to reduce the data volume while satisfying the sampling theorem.

[0188] Through rereferencing processing, the influence of the reference electrode position can be removed, the volume conduction effect can be reduced, and it is convenient for standardization and comparison. Those skilled in the art can set different rereferencing methods according to actual needs, including but not limited to average reference and ear-level reference.

[0189] Preferably, design a notch filter (such as 50 Hz) to eliminate power frequency noise interference and remove power frequency noise.

[0190] Through removing baseline drift processing, the influence brought by data drift can be effectively eliminated. Preferably, the removing baseline drift processing is realized by means of high-pass filtering. Exemplarily, the high-pass filtering frequency can be set to 1 Hz.

[0191] Through removing invalid high frequencies processing, the high-frequency component noise that does not contain physiological or pathological significance in the scalp electroencephalogram signal to be inferred can be effectively removed. Preferably, design a low-pass filter (such as 40 Hz) to remove high-frequency components.

[0192] Preferably, electrooculogram (EOG), electrocardiogram (ECG), and electromyogram (EMG) are removed through independent component analysis (ICA). Exemplarily, a linear transformation is performed on the original signal, and the signal is decomposed into a set of independent components by using the non-Gaussianity of statistical characteristics. Each component represents a different source. The components with sources of EOG, ECG, and EMG are removed, and then the remaining components are inversely transformed back to the original signal.

[0193] After completing the corresponding preprocessing operations, time slicing and artifact removal are performed on the preprocessed scalp electroencephalogram (EEG) signals to obtain an artifact-removed slice set.

[0194] Among them, through time slicing of the preprocessed scalp EEG signals, the scalp EEG signals are sliced into multiple segments to obtain a segment set.

[0195] Exemplarily, first set the observation time unit T t , and slice with T t as the step size. After completing the slicing, for the segments with a time greater than T t , slice again with T t as the step size. Those skilled in the art can set the specific step size according to actual needs. Through time slicing, the EEG acquisition time with an unfixed length can be determined to an estimated length, ensuring the consistency of subsequent feature extraction, facilitating input, and improving the overall efficiency.

[0196] After completing the time slicing, artifact removal is performed to remove the EEG segments containing artifacts, avoiding the adverse effects caused by large-amplitude artifacts due to human body jitter, poor device contact, etc.

[0197] Exemplarily, set an artifact threshold, record the time points where the amplitude exceeds the threshold, mark the remaining time points as non-artifact times, and retain the EEG segments with a continuous duration greater than T t in the non-artifact times as the non-artifact segment set. The non-artifact segment set with a duration greater than T t is sliced with a step size of T t . After completing the relevant operations, an artifact-removed slice set is obtained.

[0198] Finally, feature extraction is performed on the artifact-removed slices in the artifact-removed slice set to obtain the EEG features to be inferred.

[0199] Exemplarily, select the channels to focus on. In each artifact-removed slice and for each channel, use a dimensionality reduction method or a feature extraction method to convert the EEG signals into low-dimensional feature data, obtaining the EEG features to be inferred. Among them, relevant features such as spectrum and connectivity can be used. Spectrum features include absolute and relative band energies calculated after short-time Fourier transform, etc. Connectivity features include amplitude connectivity obtained by calculating the correlation coefficient of EEG signals, or phase-locking values calculated from the phases of the signals, etc.

[0200] By performing feature extraction on the artifact-removed slices, it is possible to extract the effective information in the signals, remove more noise components, enabling the inference model to be associated with more effective EEG features.

[0201] Referring to Figure 2 As shown, an embodiment of the present invention further provides a training device for a neural dynamics parameter inference model based on scalp EEG, which is applied to the training method for a neural dynamics parameter inference model based on scalp EEG described in any of the above embodiments. The training device for a neural dynamics parameter inference model based on scalp EEG includes an input unit 10, a construction unit 20, and a training unit 30.

[0202] The input unit 10 is configured to input the EEG features to be inferred of the brain region to be inferred into the inference model, and the inference model outputs the corresponding neural dynamics parameters. Among them, the inference model includes a first encoding module, which is used to map the EEG features to be inferred to the target neural dynamics parameter space through the first encoding module to obtain the neural dynamics parameters.

[0203] The construction unit 20 is configured to construct a target loss function according to the simulated EEG features, the reconstructed EEG features, and the EEG features to be inferred. Among them, the simulated EEG features are obtained by processing the intracranial simulation signals of the brain region to be inferred, and the intracranial simulation signals are simulated according to the target neural dynamics equation; the reconstructed EEG features are obtained by reconstructing the neural dynamics parameters through the first decoding module.

[0204] The training unit 30 is configured to iteratively train the first encoding module according to the target loss function and the first decoding module until the target loss function converges, obtaining the trained inference model.

[0205] Referring to Figure 3 As shown, the present invention further provides a neural dynamics parameter inference device based on scalp EEG, including an acquisition unit 40 and a speculation unit 50.

[0206] The acquisition unit 40 is configured to acquire the EEG features to be inferred.

[0207] The predictor 50 is used to input the electroencephalogram features to be inferred into the inference model, and the corresponding neurodynamics parameters are output by the inference model. Among them, the inference model is trained according to the method for training the neurodynamics parameter inference model based on scalp electroencephalogram described in any of the above embodiments, or the inference model is trained according to the device for training the neurodynamics parameter inference model based on scalp electroencephalogram described in any of the above embodiments.

[0208] The present invention also provides a non-transitory machine-readable medium storing a computer program. Wherein, the computer program, when executed by a processor of a computer, is used to cause the computer to execute the method for training the neurodynamics parameter inference model based on scalp electroencephalogram described in any of the above embodiments, or execute the method for inferring neurodynamics parameters based on scalp electroencephalogram described in any of the above embodiments.

[0209] The present invention also provides a computer program product, including a computer program. Wherein, the computer program, when executed by a processor of a computer, is used to cause the computer to execute the method for training the neurodynamics parameter inference model based on scalp electroencephalogram described in any of the above embodiments, or execute the method for inferring neurodynamics parameters based on scalp electroencephalogram described in any of the above embodiments.

[0210] The present invention also provides an electronic device, including at least one processor and a memory communicatively connected to the at least one processor. The memory stores a computer program that can be executed by the at least one processor. The computer program, when executed by the at least one processor, is used to cause the electronic device to execute the method for training the neurodynamics parameter inference model based on scalp electroencephalogram described in any of the above embodiments, or execute the method for inferring neurodynamics parameters based on scalp electroencephalogram described in any of the above embodiments.

[0211] Refer to Figure 4 As shown, the structural block diagram of an electronic device that can be a server or a client and can be an embodiment of the present invention will now be described. It is an example of a hardware device that can be applied to various aspects of the present invention. The electronic device is intended to represent various forms of digital electronic computer devices, such as, laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as, personal digital processors, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present invention described and / or claimed herein.

[0212] The electronic device includes a computing unit 601, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 602 or a computer program loaded from a storage unit 608 into a random access memory (RAM) 603. In the RAM 603, various programs and data required for the operation of the electronic device can also be stored. The computing unit 601, the ROM 602, and the RAM 603 are connected to each other via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0213] Multiple components in the electronic device are connected to the I / O interface 605, including: an input unit 606, an output unit 607, a storage unit 608, and a communication unit 609. The input unit 606 can be any type of device capable of inputting information into the electronic device. The input unit 606 can receive input digital or character information, and generate key signal inputs related to the user settings and / or function controls of the electronic device. The output unit 607 can be any type of device capable of presenting information, and can include but is not limited to a display, a speaker, a video / audio output terminal, a vibrator, and / or a printer. The storage unit 608 can include but is not limited to a magnetic disk, an optical disk. The communication unit 609 allows the electronic device to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks, and can include but is not limited to a modem, a network card, an infrared communication device, and / or a wireless communication transceiver, such as a Bluetooth device, a WiFi device, a WiMax device, a cellular communication device, and / or the like.

[0214] The computing unit 601 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 601 include but are not limited to a CPU, a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing units, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 601 executes the various methods and processes described above. For example, in some embodiments, the method embodiments of the present invention can be implemented as a computer program, which is tangibly contained in a machine-readable medium, such as the storage unit 608. In some embodiments, part or all of the computer program can be loaded and / or installed onto the electronic device via the ROM 602 and / or the communication unit 609. In some embodiments, the computing unit 601 can be configured to execute the scalp electroencephalogram-based neural dynamics parameter inference model training method described in any one of the above embodiments, or execute the neural dynamics parameter inference method described in any one of the above embodiments in any other appropriate manner (for example, by means of firmware).

[0215] The computer program for implementing the method of the embodiments of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing devices, such that when executed by the processor or controller, the functions / operations specified in the flowchart and / or block diagram are implemented. The computer program can be executed entirely on the machine, partially on the machine, executed partially on the machine as an independent software package and partially on a remote machine, or executed entirely on a remote machine or server.

[0216] In the context of the embodiments of the present invention, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable signal medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, or infrared system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0217] It should be noted that the term "including" and its variations used in the embodiments of the present invention are open-ended, that is, "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The modifications of "one" and "a plurality" mentioned in the embodiments of the present invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless clearly specified otherwise in the context, it should be understood as "one or more".

[0218] In the method embodiments provided by the embodiments of the present invention, the various steps recorded can be executed in different orders and / or executed in parallel. In addition, the method embodiments can include additional steps and / or omit the steps shown. The protection scope of the present invention is not limited in this regard.

[0219] The term "embodiment" in this specification means that the specific features, structures, or characteristics described in connection with the embodiments may be included in at least one embodiment of the present invention. The phrase appears at various positions in the specification and does not necessarily mean the same embodiment, nor does it mean being independent or alternative to other embodiments and mutually exclusive. The various embodiments in this specification are described in a related manner, and the same or similar parts among the embodiments are cross-referred to. In particular, for the embodiments of devices, equipment, and systems, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts refer to the partial description of the method embodiments.

[0220] The above-described embodiments merely represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the protection scope. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several deformations and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the appended claims.

Claims

1. A training method for a neural dynamics parameter inference model based on scalp electroencephalogram, characterized in that, Including: Input the electroencephalogram (EEG) features to be inferred of the brain region to be inferred into an inference model, and the inference model outputs corresponding neurodynamics parameters; wherein, the inference model includes a first encoding module, which is used to map the EEG features to be inferred into a target neurodynamics parameter space through the first encoding module to obtain the neurodynamics parameters; Construct a target loss function according to the simulated EEG features, the reconstructed EEG features, and the EEG features to be inferred; wherein, the simulated EEG features are obtained by processing the intracranial simulation signals of the brain region to be inferred, and the intracranial simulation signals are simulated according to the target neurodynamics equation; the reconstructed EEG features are obtained by reconstructing the neurodynamics parameters through a first decoding module; According to the target loss function and the first decoding module, iteratively train the first encoding module until the target loss function converges to obtain the trained inference model.

2. The training method of the neurodynamics parameter inference model based on scalp electroencephalogram according to claim 1, wherein The target loss function includes a first loss function, a second loss function, and a third loss function. Constructing the target loss function according to the simulated EEG features, the reconstructed EEG features, and the EEG features to be inferred includes: Construct the first loss function according to the reconstructed EEG features and the EEG features to be inferred; Construct the second loss function according to the reconstructed EEG features and the simulated EEG features; Perform weighted summation on the first loss function and the second loss function to construct the third loss function.

3. The training method of the neurodynamics parameter inference model based on scalp electroencephalogram according to claim 2, wherein According to the target loss function and the first decoding module, iteratively train the first encoding module until the target loss function converges to obtain the trained inference model, including: According to the first loss function, iteratively train the first encoding module and the first decoding module until the first loss function converges to obtain a second encoding module and a second decoding module; According to the second loss function, iteratively train the second decoding module until the second loss function converges to obtain a third decoding module; According to the third loss function, iteratively train the second encoding module and the third decoding module until the third loss function converges to obtain the trained third encoding module and the inference model.

4. The training method of the neural dynamics parameter inference model based on scalp electroencephalogram according to claim 1, characterized in that, Before constructing the target loss function according to the simulated EEG features, the reconstructed EEG features, and the EEG features to be inferred, it further includes: According to the target neurodynamics equation, establish an intracranial signal simulation sub-model corresponding to each brain region to be inferred; Couple each intracranial signal simulation sub-model according to the structural connection coefficient to obtain the intracranial simulation signal; Input the intracranial simulation signal into a forward inference sub-model, and the forward inference sub-model maps the intracranial simulation signal to obtain a simulated EEG signal; Process the simulated EEG signal to obtain the simulated EEG features.

5. The training method of the neural dynamics parameter inference model based on scalp electroencephalogram according to claim 4, characterized in that, Before coupling each intracranial signal simulation sub-model according to the structural connection coefficient, it further includes: Under the condition of having the target imaging acquisition condition, obtain the target imaging data corresponding to the brain region to be inferred, and process the target imaging data to obtain the structural connection coefficient; wherein, the processing of the target imaging data includes connection function analysis, structural connection analysis, and network analysis; Under the condition of not having the target imaging acquisition condition, calculate the mean value of the automated anatomical labeling structural connection matrix in the public dataset to obtain the structural connection coefficient.

6. The training method of the neural dynamics parameter inference model based on scalp electroencephalogram according to claim 4, wherein, Before inputting the intracranial simulation signal into the forward inference sub-model, it further includes: Obtain a target head model; Define a target source position; Set a target source type; According to the target head model, the target source position, and the target source type, perform forward solution to establish a mapping from the target source to the target scalp potential, and obtain the forward inference sub-model.

7. The method for training a neural dynamics parameter inference model based on scalp electroencephalogram according to claim 4, wherein: The target neural dynamics equation is set as the Jensen and Rit equation. According to the Jensen and Rit equation, establish an intracranial signal simulation sub-model corresponding to each brain region to be inferred, expressed as: ， ， ， ， ， ， ， Among them, and are the change rates of the cone cell membrane potential, is the speed of change of the change rate of the population membrane potential of cone neurons, is the excitatory synaptic gain constant, is the reciprocal of the excitatory synaptic time constant, is the neuron activation function, is the deviation of the membrane from the resting potential, is the membrane potential of the excitatory neuron population, is the membrane potential of the inhibitory neuron population, is the membrane potential of the cone neuron population, and are the change rates of the membrane potential of the excitatory neuron population, is the speed of change of the change rate of the membrane potential of the excitatory neuron population, is the external input, is the average number of connections from excitatory neurons to cone neurons, is the average number of connections from cone neurons to excitatory neurons, and are the change rates of the membrane potential of the inhibitory neuron population, is the speed of change of the change rate of the membrane potential of the inhibitory neuron population, is the inhibitory synaptic gain constant, is the average number of connections from inhibitory neurons to cone neurons, is the reciprocal of the inhibitory synaptic time constant, is the average number of connections from cone neurons to inhibitory neurons; is the maximum firing rate, is the slope of the neuron activation function, is the average firing threshold; According to the structural connection coefficient, couple each intracranial signal simulation sub-model to obtain the intracranial simulation signal, expressed as: ， ， ， ， ， ， ， ， Wherein, and is an integer, is the total number of brain regions to be inferred, and are the change rates of the pyramidal cell membrane potential in the th brain region to be inferred, is the speed of change of the change rate of the population membrane potential of pyramidal neurons in the th brain region to be inferred, is the excitatory synaptic gain constant in the th brain region to be inferred, is the reciprocal of the excitatory time constant in the th brain region to be inferred, is the membrane potential of the excitatory neuron population in the th brain region to be inferred, is the membrane potential of the inhibitory neuron population in the th brain region to be inferred, is the membrane potential of the pyramidal cells in the th brain region to be inferred, and are the membrane potential of the excitatory neuron population in the th brain region to be inferred, is the speed of change of the change rate of the membrane potential of the excitatory neuron population in the th brain region to be inferred, is the average number of connections from excitatory neurons to pyramidal neurons in the th brain region to be inferred, is the average number of connections from pyramidal neurons to excitatory neurons in the th brain region to be inferred, is the external input in the th brain region to be inferred, is the sum of the structural connection coefficients between the th brain region to be inferred and all other brain regions to be inferred except its own position, is the variance of the external input in the th brain region to be inferred, is the adjustment coefficient, is the standard deviation of the external input in the th brain region to be inferred, is the directional connection coefficient between the th brain region to be inferred and the th brain region to be inferred, is the standard deviation of the neuron firing rate in the th brain region to be inferred, is the variance of the neuron firing rate in the th brain region to be inferred, and is the membrane potential change rate of the inhibitory neuron population in the th brain region to be inferred, is the speed of change of the membrane potential change rate of the inhibitory neuron population in the th brain region to be inferred, is the inhibitory synaptic gain constant in the th brain region to be inferred, is the average number of connections from inhibitory neurons to pyramidal neurons in the th brain region to be inferred, is the reciprocal of the inhibitory time constant in the th brain region to be inferred, is the average number of connections from pyramidal neurons to inhibitory neurons in the th brain region to be inferred; Input the intracranial simulation signal into the forward inference sub-model, and the forward inference sub-model maps the intracranial simulation signal to obtain a simulated electroencephalogram signal, expressed as: ， Among them, is the simulated electroencephalogram signal, is the intracranial simulation signal, is the mapping of the forward inference sub-model.

8. The method for training a neural dynamics parameter inference model based on scalp electroencephalogram according to claim 1, wherein Input the electroencephalogram feature to be inferred of the brain region to be inferred into the inference model, and the inference model outputs the corresponding neural dynamics parameters, including: Input the electroencephalogram feature to be inferred into the convolutional sub-module of the first encoding module, and extract spatial features through the convolutional sub-module to obtain spatial electroencephalogram features; Input the spatial electroencephalogram features into the temporal feature extraction sub-module of the first encoding module, and capture temporal dynamic features through the temporal feature extraction sub-module to obtain spatio-temporal electroencephalogram features; Map the spatio-temporal electroencephalogram features to the target neural dynamics parameter space to obtain the neural dynamics parameters; Or, Input the electroencephalogram feature to be inferred into the temporal feature extraction sub-module of the first encoding module, and capture temporal dynamic features through the temporal feature extraction sub-module to obtain temporal electroencephalogram features; Input the temporal electroencephalogram features into the convolutional sub-module of the first encoding module, and extract spatial features through the convolutional sub-module to obtain spatio-temporal electroencephalogram features; Map the spatio-temporal electroencephalogram features to the target neural dynamics parameter space to obtain the neural dynamics parameters.

9. The method for training a neural dynamics parameter inference model based on scalp electroencephalogram according to claim 8, characterized in that Mapping the spatio-temporal electroencephalogram features to the target neural dynamics parameter space to obtain the neural dynamics parameters includes: By adjusting the function Map the spatio-temporal electroencephalogram features to the target neurodynamics parameter space, expressed as: ， Among them, represents the minimum range of the neural dynamics parameters, represents the maximum range of the neural dynamics parameters, is the activation function, is the neural dynamics parameter.

10. The training method of the neurodynamics parameter inference model based on scalp electroencephalogram according to claim 1, wherein Before inputting the electroencephalogram feature to be inferred of the brain region to be inferred into the inference model, it further includes: Obtain the electroencephalogram signal of the scalp to be inferred; Preprocess the electroencephalogram signal of the scalp to be inferred to obtain a preprocessed electroencephalogram signal of the scalp; wherein, the preprocessing includes at least one of resampling, rereferencing, removing power frequency, removing baseline drift, removing invalid high frequencies, and removing electrooculogram, electrocardiogram, and electromyogram. Perform time slicing and artifact removal on the preprocessed scalp electroencephalogram (EEG) signals to obtain an artifact-removed slice set; Extract features from the artifact-removed slices in the artifact-removed slice set to obtain the EEG features to be inferred.

11. A method for inferring neurodynamics parameters based on scalp electroencephalogram, characterized in that, Including: Obtain the EEG features to be inferred; Input the EEG features to be inferred into an inference model, and the inference model outputs corresponding neurodynamics parameters; wherein, the inference model is trained according to the method for training a neurodynamics parameter inference model based on scalp EEG as described in any one of claims 1 to 10.

12. A training device for a neural dynamics parameter inference model based on scalp electroencephalogram, which is applied to the method for training a neural dynamics parameter inference model based on scalp electroencephalogram according to any one of claims 1 to 10, characterized in that Including: An inputter for inputting the EEG features to be inferred of the brain region to be inferred into an inference model, and the inference model outputs corresponding neurodynamics parameters; wherein, the inference model includes a first encoding module for mapping the EEG features to be inferred to a target neurodynamics parameter space through the first encoding module to obtain the neurodynamics parameters; A builder for constructing a target loss function according to the simulated EEG features, the reconstructed EEG features, and the EEG features to be inferred; wherein, the simulated EEG features are obtained by processing the intracranial simulated signals of the brain region to be inferred, and the intracranial simulated signals are simulated according to a target neurodynamics equation; the reconstructed EEG features are obtained by reconstructing the neurodynamics parameters through a first decoding module; A trainer for iteratively training the first encoding module according to the target loss function and the first decoding module until the target loss function converges to obtain the trained inference model.

13. A neural dynamics parameter inference device based on scalp electroencephalogram, characterized in that, Including: An acquirer for acquiring the EEG features to be inferred; A speculator for inputting the EEG features to be inferred into an inference model, and the inference model outputs corresponding neurodynamics parameters; wherein, the inference model is trained according to the method for training a neurodynamics parameter inference model based on scalp EEG as described in any one of claims 1 to 10, or the inference model is trained through the device for training a neurodynamics parameter inference model based on scalp EEG as described in claim 12.

14. An electronic device, comprising: A processor and a memory storing a program, characterized in that the program includes instructions that, when executed by the processor, cause the processor to execute the method for training a neurodynamics parameter inference model based on scalp EEG as described in any one of claims 1 to 10, or the instructions, when executed by the processor, cause the processor to execute the method for inferring neurodynamics parameters based on scalp EEG as described in claim 11.

Images