Brain excitation and inhibition state depth estimation method and monitoring device

Abstract

The invention discloses a brain excitation and inhibition state depth estimation method and a monitoring device, and the method provides a time-frequency domain fusion deep neural network model, carries out the multi-stage decomposition of electroencephalogram data through discrete wavelet transform, and extracts different frequency band information; meanwhile, a convolutional neural network and a long-short-term memory neural network are integrated, and the brain excitation and inhibition state is accurately estimated from electroencephalogram data and multi-band information; an attention mechanism is introduced to adaptively fuse time-frequency features, and attention of the model to key features is enhanced; a time-frequency domain fusion deep neural network model is trained by adopting synthetic electroencephalogram data generated by a neural cluster model, so that the time-frequency domain fusion deep neural network model learns a mapping relation between neural activities and brain excitation and inhibition states, and the physiological challenge that real brain excitation and inhibition state tags are difficult to directly obtain is solved; through deep fusion of a deep learning method and a neural modeling technology, deep estimation and effective monitoring of brain excitation and inhibition states are realized.

Description

Technical Field

[0001] This patent belongs to the field of biomedical engineering technology, and specifically relates to a method and monitoring device for depth estimation of brain excitation and inhibition states. Background Art

[0002] Epilepsy is characterized by abnormal brain activity leading to motor, sensory, or behavioral disorders, and even coma. Epileptic seizures originate from an imbalance between excitatory and inhibitory neural activity in the brain, resulting in abnormally synchronized discharges in neural networks and temporary disruptions of brain function. Therefore, a deeper understanding of the changes in excitatory and inhibitory states during epileptic seizures is crucial for uncovering the neurophysiological mechanisms of epilepsy and could also open new avenues for precise monitoring. However, due to technical and ethical limitations, direct measurement of brain excitatory and inhibitory states through electrophysiological methods is difficult. In recent years, computational modeling methods that combine neural models with electroencephalographic (EEG) data have emerged as important tools for indirectly estimating brain excitatory and inhibitory states. These methods, combined with neural cluster models, can infer the neural dynamics parameters representing brain states from EEG data, demonstrating great potential for modeling and predicting epileptic activity. Among these, the Kalman filter and its extended models have been widely used due to their powerful time series modeling and state estimation capabilities.

[0003] However, these methods still have some limitations. Their performance depends on accurately setting the system's initial state, which is difficult to obtain in practical applications. They are essentially local search algorithms and are prone to falling into local optima. In addition, most of these methods analyze EEG data from the time domain, rarely considering frequency domain or multi-scale features, which limits the model's ability to depict complex brain dynamics. Therefore, a more robust and generalizable deep estimation method is needed that can mine the time-frequency information in EEG data from multiple scales and perspectives to accurately model and estimate the parameters of the brain's excitatory and inhibitory states. Summary of the Invention

[0004] The purpose of this patent is to overcome the shortcomings of the existing technology and provide a deep estimation method and monitoring device for the brain's excitation and inhibition states. It uses discrete wavelet transform to extract multi-level frequency domain information of electroencephalogram signals, and uses deep learning methods to capture the complex characteristics of the brain state from time domain and frequency domain signals, solving the problem of insufficient utilization of frequency domain information in the state estimation process of existing methods. The method of the present invention can be deployed in an end-to-end form to directly estimate the brain's excitation and inhibition states from electroencephalogram data without the need for complex initialization process or additional manual adjustment.

[0005] To achieve the above purpose, this patent adopts the following technical solutions:

[0006] In a first aspect, a method for estimating the depth of brain excitation and inhibition states comprises the following steps:

[0007] Step 1: Take EEG data X as input to obtain the excitatory synaptic gain state parameter that reflects the brain functional state. and inhibitory synaptic gain state parameters As output, a time-frequency domain fusion deep neural network model is constructed;

[0008] Step 2: Based on the neural cluster model, systematically adjust the excitatory synaptic gain state parameters in the neural cluster model and inhibitory synaptic gain state parameters , simulate the brain electrical activity process under at least one brain excitation and inhibition state, synthesize EEG data with different physiological mechanism characteristics; compare the synthesized EEG data with the corresponding excitatory synaptic gain state parameters and inhibitory synaptic gain state parameters Construct a complete training dataset for training time-frequency domain fusion deep neural network models;

[0009] Step 3: Use the training dataset constructed in step 2 to supervise the time-frequency domain fusion deep neural network model. Use the error back propagation algorithm to iteratively optimize the network parameters, minimize the loss function between the network output and the brain state parameters, and achieve a deep estimation of the brain's excitation and inhibition states.

[0010] Step 4: obtaining the subject's EEG data through an EEG acquisition device and preprocessing the EEG data;

[0011] Step 5: Input the EEG data pre-processed in step 4 into the time-frequency domain fusion deep neural network model trained in step 3, and predict the excitatory synaptic gain state parameters of the subject's brain through real-time regression. and inhibitory synaptic gain state parameters .

[0012] Furthermore, step 1 of this technical solution to construct a time-frequency domain fusion deep neural network model specifically includes the following steps:

[0013] (1) Input EEG data Use discrete wavelet transform to perform multi-level decomposition to obtain frequency domain components of different frequency bands 、 、 and ,in, is the third-order approximation coefficient, 、 and They are the 1st to 3rd level detail coefficients respectively, and the discrete wavelet transform calculation process is as follows:

[0014]

[0015] in, is the j-th level approximation coefficient, representing the low-frequency signal component; is the j-th level detail coefficient, representing the high-frequency signal component; is the scaling function, is the wavelet function, t=2 j nk; j is the decomposition level, k is the translation factor, and n is the sampling time;

[0016] (2) EEG signal and frequency domain components of different frequency bands 、 、 and Construct a multi-branch one-dimensional convolutional neural network to extract time domain and frequency domain features respectively. Each branch convolutional neural network consists of multiple convolution layers, pooling layers and normalization layers. The number of convolution layers in different branches is designed according to the input signal length of the channel. The output signal length of all channels is unified as ;

[0017] (3) Use the SE-Net attention mechanism to perform channel attention weighting for all channels to enhance the model's ability to select channels of different frequency bands;

[0018] (4) All weighted frequency band features are spliced ​​in the channel dimension, and the fused features are compressed and reconstructed using a one-dimensional convolutional network. The fused time-frequency features are input into a long short-term memory neural network to extract the long-term dependencies and dynamic change patterns in the EEG data;

[0019] (5) The output features of the long short-term memory neural network are mapped to the excitatory synaptic gain state parameters through a fully connected neural network. and inhibitory synaptic gain state parameters , achieving deep estimation of the brain's excitation and inhibition states.

[0020] Furthermore, step 2 of constructing a training data set in this technical solution specifically includes the following steps:

[0021] (1) Use the JR neural cluster model to generate EEG signals. The dynamic equation of the JR neural cluster model is as follows:

[0022]

[0023] in, is the output of the JR neural cluster model, representing the synthetic EEG signal; and represent the excitatory synaptic gain state parameters and inhibitory synaptic gain state parameters respectively; and represents the time constant, with values ​​of 100 and 50 respectively; represents the average number of synaptic connections between pyramidal neuron subclusters and interneuron subclusters, with a value of [135, 108, 33.75, 33.75]; It is an external input, using a mean of 90 , with a variance of 30 Gaussian white noise representation; is a sigmoid function, where is the maximum discharge rate, which is 5. The discharge rate is The corresponding average postsynaptic membrane potential is 6. is the discharge rate, yes Function steepness, the value is 0.56;

[0024] (2) Systematically changing the excitatory synaptic gain state parameters in the JR neural cluster model and inhibitory synaptic gain state parameters , simulate the brain electrical activity process under at least one brain excitation and inhibition state, and synthesize EEG data with different physiological characteristics;

[0025] (3) Construct a large-scale synthetic EEG-excitation-inhibition state parameter dataset that matches the pathological mechanism for training a time-frequency domain fusion deep neural network model.

[0026] Furthermore, step 4 of this technical solution collects the subject's EEG data and performs preprocessing, which specifically includes the following steps:

[0027] (1) Obtain the EEG data of the subjects using EEG acquisition equipment;

[0028] (2) Preprocessing the collected EEG data, including at least one of bandpass filtering, artifact removal, segmentation, and removal of electrooculogram and electromyography interference;

[0029] (3) The preprocessed EEG data is linearly normalized to make its amplitude range consistent with the synthetic EEG data used for training, ensuring consistency of model input.

[0030] In the second aspect, this patent provides a device for monitoring the excitation and inhibition state of the brain, including an EEG acquisition device, a storage medium, a processor, and a brain excitation and inhibition state depth estimation program stored on the storage medium and capable of running on the processor. The EEG acquisition device can acquire EEG signals, and the storage medium stores a brain excitation and inhibition state depth estimation program. When the processor runs the program, it implements the steps of the brain excitation and inhibition state depth estimation method as described in any one of the first aspects.

[0031] This patent provides a method and device for estimating the depth of brain excitation and inhibition states, which has the following beneficial effects:

[0032] (1) In response to the problems of insufficient utilization of frequency domain information and difficulty in selecting the initial state when estimating brain states in traditional Kalman filter and other methods, the present invention constructs a time-frequency domain fusion deep neural network model that integrates the multi-scale features of the EEG time domain and frequency domain to enhance the model's perception ability; it uses a deep learning algorithm to directly estimate the brain's excitation and inhibition states from the EEG signal without the need for complex initialization or additional manual adjustments;

[0033] (2) Using the neural cluster model to generate a large-scale synthetic EEG training dataset covering multiple brain states, it solves the problem of being unable to directly measure the brain's excitatory and inhibitory state labels in clinical practice, and improves the excitatory synaptic gain state parameters. and inhibitory synaptic gain state parameters Estimation accuracy;

[0034] (3) EEG data is acquired through EEG acquisition equipment, and the excitation and inhibition states in the brain are estimated with the help of a trained time-frequency domain fusion deep neural network model, achieving accurate depth estimation and monitoring without the need for invasive detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of this patent, the following is a brief introduction to the drawings required for the description of the embodiments. Obviously, the drawings described below are only some embodiments of this patent. For those skilled in the art, other drawings can be obtained based on these drawings without creative work, among which:

[0036] Figure 1 A flowchart of a method for estimating the depth of brain excitation and inhibition states provided by the present invention;

[0037] Figure 2 This is a structural diagram of a time-frequency domain fusion deep neural network model in a method for deep estimation of brain excitation and inhibition states provided by the present invention;

[0038] Figure 3A schematic diagram of the architecture of a device for monitoring brain excitation and inhibition states provided by the present invention. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of this patent to clearly and completely describe the technical solutions in the embodiments of this patent. Obviously, the embodiments described are only part of the embodiments of this patent, not all of them. Based on the embodiments of this patent, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this patent.

[0040] The embodiments of this patent provide a method and monitoring device for deep estimation of brain excitation and inhibition states, which realizes accurate estimation of brain excitation and inhibition states by constructing a time-frequency domain fusion deep neural network model. The method uses discrete wavelet transform to perform multi-level decomposition of electroencephalogram (EEG) signals to extract information of different frequency bands; integrates convolutional neural networks and long short-term memory networks to extract complex features of neural activity from EEG signals and information of different frequency bands; and introduces an attention mechanism to adaptively fuse time-frequency features to enhance the model's attention to key features; at the same time, synthetic EEG data generated by a neural quality model is used to train a time-frequency domain fusion deep neural network, so that it learns the mapping relationship between neural activity and brain excitation and inhibition states, solving the physiological challenge of directly obtaining true brain excitation and inhibition state labels.

[0041] First, see Figure 1 , the embodiment of this patent discloses a method for estimating the depth of brain excitation and inhibition states, comprising the following steps:

[0042] Step 1: Take EEG data X as input to obtain the excitatory synaptic gain state parameter that reflects the brain functional state. and inhibitory synaptic gain state parameters As output, a time-frequency domain fusion deep neural network model is constructed;

[0043] Step 2: Based on the neural cluster model, systematically adjust the excitatory synaptic gain state parameters in the neural cluster model and inhibitory synaptic gain state parameters , simulate the brain electrical activity process under at least one brain excitation and inhibition state, synthesize EEG data with different physiological mechanism characteristics; compare the synthesized EEG data with the corresponding excitatory synaptic gain state parameters and inhibitory synaptic gain state parameters Construct a complete training dataset for training time-frequency domain fusion deep neural network models;

[0044] Step 3: Use the training dataset constructed in step 2 to supervise the time-frequency domain fusion deep neural network model. Use the error back propagation algorithm to iteratively optimize the network parameters, minimize the loss function between the network output and the brain state parameters, and achieve a deep estimation of the brain's excitation and inhibition states.

[0045] Step 4: obtaining the subject's EEG data through an EEG acquisition device and preprocessing the EEG data;

[0046] Step 5: Input the EEG data pre-processed in step 4 into the time-frequency domain fusion deep neural network model trained in step 3, and predict the excitatory synaptic gain state parameters of the subject's brain through real-time regression. and inhibitory synaptic gain state parameters .

[0047] For further information, see Figure 2 This is a structural diagram of a time-frequency domain fusion deep neural network model. Step 1 of constructing a time-frequency domain fusion deep neural network model specifically includes the following steps:

[0048] (1) Input EEG data Use discrete wavelet transform to perform multi-level decomposition to obtain frequency domain components of different frequency bands 、 、 and ,in, is the third-order approximation coefficient, 、 and They are the 1st to 3rd level detail coefficients respectively, and the discrete wavelet transform calculation process is as follows:

[0049]

[0050] in, is the j-th level approximation coefficient, representing the low-frequency signal component; is the j-th level detail coefficient, representing the high-frequency signal component; is the scaling function, is the wavelet function, t=2 j nk; j is the decomposition level, k is the translation factor, and n is the sampling time. In this embodiment, since the original EEG signal is filtered by a 0.5-70 Hz bandpass filter, the three-level wavelet transform results in , , and The frequency bands represented are 0.5-8.75 Hz, 8.75-17.5 Hz, 17.5-35 Hz, and 35-70 Hz;

[0051] (2) EEG signal and frequency domain components of different frequency bands 、 、 and Construct a multi-branch one-dimensional convolutional neural network to extract time domain and frequency domain features respectively. Each branch convolutional neural network consists of multiple convolution layers, pooling layers and normalization layers. The number of convolution layers in different branches is designed according to the input signal length of the channel. The output signal length of all channels is unified as ;

[0052] (3) Use the SE-Net attention mechanism to perform channel attention weighting for all channels to enhance the model's ability to select channels of different frequency bands;

[0053] (4) All weighted frequency band features are spliced ​​in the channel dimension, and the fused features are compressed and reconstructed using a one-dimensional convolutional network. The fused time-frequency features are input into a long short-term memory neural network to extract the long-term dependencies and dynamic change patterns in the EEG data;

[0054] (5) The output features of the long short-term memory neural network are mapped to the excitatory synaptic gain state parameters through a fully connected neural network. and inhibitory synaptic gain state parameters , achieving deep estimation of the brain's excitation and inhibition states.

[0055] Furthermore, step 2 of constructing a training data set specifically includes the following steps:

[0056] (1) Use the JR neural cluster model to generate EEG signals. The dynamic equation of the JR neural cluster model is as follows:

[0057]

[0058] in, is the output of the JR neural cluster model, representing the synthetic EEG signal; and represent the excitatory synaptic gain state parameters and inhibitory synaptic gain state parameters respectively; and represents the time constant, with values ​​of 100 and 50 respectively; represents the average number of synaptic connections between pyramidal neuron subclusters and interneuron subclusters, with a value of [135, 108, 33.75, 33.75]; It is an external input, using a mean of 90 , with a variance of 30 Gaussian white noise representation; is a sigmoid function, where is the maximum discharge rate, which is 5. The discharge rate is The corresponding average postsynaptic membrane potential is 6. is the discharge rate, yes Function steepness, the value is 0.56;

[0059] (2) Systematically changing the excitatory synaptic gain state parameters in the JR neural cluster model and inhibitory synaptic gain state parameters , simulate at least one brain electrical activity process under the state of brain excitation and inhibition, and synthesize EEG data with different physiological characteristics; in this embodiment, the excitatory synaptic gain parameter The value range is [2,5], the inhibitory synaptic gain parameter The value range is [10,40];

[0060] (3) Construct a large-scale synthetic EEG-excitation-inhibition state parameter dataset that matches the pathological mechanism for training a time-frequency domain fusion deep neural network model.

[0061] Furthermore, the step 4 collects the subject's EEG data and performs preprocessing, which specifically includes the following steps:

[0062] (1) Obtain the EEG data of the subjects using EEG acquisition equipment;

[0063] (2) Preprocessing the collected EEG data, including at least one of bandpass filtering, artifact removal, segmentation processing, and elimination of electrooculogram and electromyography interference; in this embodiment, the EEG is subjected to 0.5-70 Hz bandpass filtering;

[0064] (3) The preprocessed EEG data is linearly normalized to make its amplitude range consistent with the synthetic EEG data used for training, ensuring consistency of model input.

[0065] Second, see Figure 3 , an embodiment of this patent provides a brain state monitoring device 1, including an EEG acquisition device 2, a processor 3, a communication interface 4, a storage medium 5, a bus 6, and a brain excitation and inhibition state depth estimation program stored on the storage medium and capable of running on the processor. The EEG acquisition device 2, processor 3, communication interface 4 and storage medium 5 are connected through a bus 6. The EEG acquisition device can acquire EEG signals. The storage medium stores the computer program, EEG data and data involved in the calculation process involved in this embodiment. When the processor runs the program, the steps of the brain state depth estimation method as described in any one of the first aspect embodiments are implemented.

[0066] The above embodiments of this patent provide a method and monitoring device for deep estimation of brain excitation and inhibition states. (1) In view of the problems of insufficient utilization of frequency domain information and difficulty in selecting initial states when estimating brain states in traditional Kalman filter and other methods, the present invention constructs a time-frequency domain fusion deep neural network model, which integrates the multi-scale features of EEG time domain and frequency domain to enhance the model's perception ability; uses a deep learning algorithm to directly estimate the brain's excitation and inhibition states from EEG signals without the need for a complex initialization process or additional manual adjustments; (2) uses a neural cluster model to generate a large-scale synthetic EEG training data set covering multiple brain states, solving the problem of being unable to directly measure brain excitation and inhibition state labels in clinical practice, and improving the excitatory synaptic gain state parameters. and inhibitory synaptic gain state parameters Estimation accuracy; (3) EEG data is acquired through EEG acquisition equipment, and the excitation and inhibition states in the brain are estimated with the help of a trained time-frequency domain fusion deep neural network model, which can achieve accurate and in-depth estimation and monitoring of the excitation and inhibition states of the brain without invasive detection.

[0067] The above description is only an implementation method of this patent and does not limit the patent scope of this patent. Any equivalent structure or equivalent process transformation made using the contents of this patent specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of this patent.

Claims

1. A method for estimating the depth of brain excitation and inhibition states, characterized in that: The depth estimation method comprises the following steps: Step 1: Take EEG data X as input to obtain the excitatory synaptic gain state parameter that reflects the brain functional state. and inhibitory synaptic gain state parameters As output, a time-frequency domain fusion deep neural network model is constructed; Step 2: Based on the neural cluster model, systematically adjust the excitatory synaptic gain state parameters in the neural cluster model and inhibitory synaptic gain state parameters , simulate the brain electrical activity process under at least one brain excitation and inhibition state, synthesize EEG data with different physiological mechanism characteristics; compare the synthesized EEG data with the corresponding excitatory synaptic gain state parameters and inhibitory synaptic gain state parameters Construct a complete training dataset for training time-frequency domain fusion deep neural network models; Step 3: Use the training dataset constructed in step 2 to supervise the time-frequency domain fusion deep neural network model. Use the error back propagation algorithm to iteratively optimize the network parameters, minimize the loss function between the network output and the brain state parameters, and achieve a deep estimation of the brain's excitation and inhibition states. Step 4: obtaining the subject's EEG data through an EEG acquisition device and preprocessing the EEG data; Step 5: Input the EEG data pre-processed in step 4 into the time-frequency domain fusion deep neural network model trained in step 3, and predict the excitatory synaptic gain state parameters of the subject's brain through real-time regression. and inhibitory synaptic gain state parameters .

2. A method for estimating the depth of brain excitation and inhibition states according to claim 1, characterized in that: The step 1 of constructing a time-frequency domain fusion deep neural network model specifically includes the following steps: (1) Input EEG data Use discrete wavelet transform to perform multi-level decomposition to obtain frequency domain components of different frequency bands 、 、 and ,in, is the third-order approximation coefficient, 、 and They are the 1st to 3rd level detail coefficients respectively, and the discrete wavelet transform calculation process is as follows: in, is the j-th level approximation coefficient, representing the low-frequency signal component; is the j-th level detail coefficient, representing the high-frequency signal component; is the scaling function, is the wavelet function, t=2 j nk; j is the decomposition level, k is the translation factor, and n is the sampling time; (2) EEG signal and frequency domain components of different frequency bands 、 、 and Construct a multi-branch one-dimensional convolutional neural network to extract time domain and frequency domain features respectively. Each branch convolutional neural network consists of multiple convolution layers, pooling layers and normalization layers. The number of convolution layers in different branches is designed according to the input signal length of the channel. The output signal length of all channels is unified as ; (3) Use the SE-Net attention mechanism to perform channel attention weighting for all channels to enhance the model's ability to select channels of different frequency bands; (4) All weighted frequency band features are spliced ​​in the channel dimension, and the fused features are compressed and reconstructed using a one-dimensional convolutional network. The fused time-frequency features are input into a long short-term memory neural network to extract the long-term dependencies and dynamic change patterns in the EEG data; (5) The output features of the long short-term memory neural network are mapped to the excitatory synaptic gain state parameters through a fully connected neural network. and inhibitory synaptic gain state parameters , achieving deep estimation of the brain's excitation and inhibition states.

3. A method for estimating the depth of brain excitation and inhibition states according to claim 1, characterized in that: The step 2 of constructing the training data set specifically includes the following steps: (1) Use the JR neural cluster model to generate EEG signals. The dynamic equation of the JR neural cluster model is as follows: in, is the output of the JR neural cluster model, representing the synthetic EEG signal; and represent the excitatory synaptic gain state parameters and inhibitory synaptic gain state parameters respectively; and represents the time constant, with values ​​of 100 and 50 respectively; represents the average number of synaptic connections between pyramidal neuron subclusters and interneuron subclusters, with a value of [135, 108, 33.75, 33.75]; It is an external input, using a mean of 90 , with a variance of 30 Gaussian white noise representation; is a sigmoid function, where is the maximum discharge rate, which is 5. The discharge rate is The corresponding average postsynaptic membrane potential is 6. is the discharge rate, yes Function steepness, the value is 0.56; (2) Systematically changing the excitatory synaptic gain state parameters in the JR neural cluster model and inhibitory synaptic gain state parameters , simulate the brain electrical activity process under at least one brain excitation and inhibition state, and synthesize EEG data with different physiological characteristics; (3) Construct a large-scale synthetic EEG-excitation-inhibition state parameter dataset that matches the pathological mechanism for training a time-frequency domain fusion deep neural network model.

4. A method for estimating the depth of brain excitation and inhibition states according to claim 1, characterized in that: The step 4 collects the subject's EEG data and performs preprocessing, specifically comprising the following steps: (1) Obtain the EEG data of the subjects using EEG acquisition equipment; (2) Preprocessing the collected EEG data, including at least one of bandpass filtering, artifact removal, segmentation, and removal of electrooculogram and electromyography interference; (3) The preprocessed EEG data is linearly normalized to make its amplitude range consistent with the synthetic EEG data used for training, ensuring consistency of model input.

5. A device for monitoring brain excitation and inhibition states, characterized in that: The monitoring device includes an EEG acquisition device, a storage medium, a processor, and a brain excitation and inhibition state depth estimation program stored in the memory and capable of running on the processor. The EEG acquisition device can acquire EEG signals. The storage medium stores the brain excitation and inhibition state depth estimation program. When the processor runs the program, it implements the steps of the brain excitation and inhibition state depth estimation method as described in any one of claims 1 to 4.

Images