A closed-loop multimodal neural stimulation system and method based on time interference

Abstract

The present invention discloses a closed-loop multimodal neural stimulation system and method based on time interference, which belongs to the field of neural stimulation technology, including a time interference stimulation system, an EEG-functional near-infrared sampling system and a host control system; the time interference stimulation system uses the beat electric field generated by two sets of electrodes to accurately stimulate the designated area of ​​the brain; the EEG-functional near-infrared sampling system is a dual-modal collector that couples EEG and functional near-infrared, including signal extraction and correlation analysis, and analyzes the stimulation effect and adjusts the stimulation scheme by integrating the integrated brain signal data of EEG time accuracy and fNIRS spatial accuracy; the host control system includes dual-modal fusion model calculation, graph convolutional neural network prediction and stimulation scheme formulation. The present invention provides a closed-loop multimodal neural stimulation system and method based on time interference, which performs deep nerve stimulation without electrode implantation surgery, reducing time costs while improving safety.

Description

Technical Field

[0001] The present invention belongs to the technical field of neural stimulation, and in particular, relates to a closed-loop multimodal neural stimulation system and method based on time interference. Background Art

[0002] The activity of the brain is mainly manifested in neuronal discharges, so short-term or long-term changes in brain function can be induced by regulating the potential activity of neurons. Direct artificial regulation of neuronal activity is highly relevant to the treatment of neurological and psychiatric diseases. It can also be used clinically to treat cranial and neurological diseases, has great potential, and can also contribute to the scientific study of the potential mechanisms of human brain nerves in sensation, movement and cognition.

[0003] Existing neural stimulation methods mainly include electrical stimulation, magnetic stimulation, optogenetic stimulation and ultrasonic stimulation. Electrical stimulation activates neurons by transmitting electric current to the nervous system. There are two major routes: deep brain stimulation with invasive electrodes and non-invasive transcranial electrical stimulation. Magnetic stimulation affects brain activity by placing a coil on the scalp to generate a magnetic field. Optogenetic stimulation transfects cells with photosensitive proteins through genetic engineering, and stimulates these proteins with light to cause neuronal excitement. Ultrasonic stimulation stimulates neurons through focused ultrasound. Electrical stimulation and optogenetic stimulation are widely used in the study of brain function and the treatment of neurological diseases because of their obvious effects and high reliability. From the perspective of stimulation methods, the existing mainstream stimulation methods have obvious defects: deep brain stimulation requires electrodes to be implanted in the brain. This invasive operation may lead to potential complications such as infection and bleeding, which increases the complexity and risk of the implantation surgery, and patients also need a longer recovery time. In addition, the electrode implantation method has great disadvantages in stimulation selectivity and spatial resolution due to the limitations of implantation depth and electrode positioning. Although transcranial electrical stimulation is non-invasive, transcranial current can only provide subthreshold stimulation. This method can increase the activity of neurons and make them easier to activate, but it cannot directly activate nerves. Moreover, transcranial current is non-directional and any area covered by the current will be affected. These two characteristics greatly limit the application of transcranial electrical stimulation. Optogenetics has high stimulation accuracy and fast response speed, but because it requires surgical implantation of the stimulator and involves transgenic technology, it is also very limited in practical applications.

[0004] From the perspective of system structure, traditional stimulation methods do not have a closed-loop control system and lack correction and optimization methods. Due to the large individual differences among users, the stimulation scheme and the user have a long running-in period, which is not conducive to its application. Summary of the invention

[0005] The purpose of the present invention is to provide a closed-loop multimodal neural stimulation system and method based on time interference, so as to solve the problems existing in the above-mentioned technology that the traditional stimulation method has no closed-loop control system, lacks correction and optimization means, and the stimulation scheme has a long running-in period with the user, which is not conducive to application.

[0006] To achieve the above-mentioned purpose, the present invention provides a closed-loop multimodal neural stimulation system based on time interference, including a time interference stimulation system, an EEG-functional near-infrared sampling system and a host control system; the time interference stimulation system utilizes the beat electric field generated by two sets of electrodes to accurately stimulate the designated area of ​​the brain; the EEG-functional near-infrared sampling system is a dual-modal collector that couples EEG and functional near-infrared, including signal extraction and correlation analysis, and analyzes the stimulation effect and adjusts the stimulation scheme by integrating brain signal data with EEG time accuracy and fNIRS spatial accuracy; the host control system includes dual-modal fusion model calculation, graph convolutional neural network prediction and stimulation scheme formulation.

[0007] Preferably, the time interference stimulation system includes a communication control module, a power module and a time interference stimulation module; the communication module adopts a SoC chip, which integrates the functions of a processor and wireless communication, and completes the communication between the time interference stimulation system and the host control system and the control of the stimulation current; the power module adopts a low-noise, low-voltage dropout regulator with an additional shielding layer design; the time interference stimulation module is composed of two beat electrode groups, and two sine wave generators generate 2KHz AC signals with a phase difference of 10-20Hz. The two signals interfere in the skull via an electrode group attached to the surface of the scalp, and the low-frequency envelope current generated by the interference stimulates the nerve cells in the specified area.

[0008] Preferably, the EEG-functional near-infrared sampling system includes an EEG EEG acquisition module and an fNIRS functional near-infrared brain signal acquisition module; the EEG EEG acquisition module uses epidermal electrodes to record signals; the EEG signal generated by time interference stimulation in the EEG EEG acquisition module is filtered and denoised by a preprocessing circuit, and then the signal is converted into a digital form by an analog-to-digital converter and sent to the communication and control subsystem, and the EEG signal is uploaded to the upper control system after removing artifacts and baseline drift by the MCU; the fNIRS functional near-infrared brain signal acquisition module measures the activity level of a specific area of ​​the brain by monitoring the absorption changes of brain hemoglobin and oxygenated hemoglobin within the near-infrared spectrum, the fNIRS functional near-infrared brain signal acquisition module uses a light-emitting diode as a light source to generate 650nm near-infrared light, and a photodiode as a light signal detector to convert the light signal passing through the brain tissue into an electrical signal for analysis, the obtained electrical signal is denoised and filtered by preprocessing, and then the signal is converted into a digital form by an analog-to-digital converter and sent to the communication and control subsystem, the MCU performs brain tissue blood oxygen calculation to obtain a brain activity signal and uploads the activity signal to the upper control system.

[0009] Preferably, the dual-modality fusion model calculation extracts features from EEG and fNIRS signal data and combines the two into a fusion model and then performs correlation analysis.

[0010] Preferably, the graph convolutional neural network prediction predicts the stimulation position according to the desired stimulation effect and formulates a stimulation plan, which is specifically expressed as follows:

[0011] Let G = (V; E) represent a graph network, V represents the node set, E represents the edge set, if two nodes v i ,v j There are edges connected, then (v i ,v j )∈E, the entire network is considered to be composed of nodes composed of different brain functional areas connected by connecting nerves, and the whole is a structure composed of subgraphs of a finite number of nodes;

[0012] Among them, the four brain waves in a certain brain region node for a fixed period of time are recorded as different dimensions, the number of dimensions is D, and the number of samples is τ. The input brain activity feature matrix of N brain network nodes at time t is specifically expressed as follows:

[0013] χ=(X 1 ,X 2 ,...,X t ,...,X τ )∈R N×D×τ

[0014] Where X represents the brain activity feature matrix of N nodes in the figure at time t, which is specifically expressed as follows:

[0015]

[0016] For the output brain activity feature matrix with K sampling numbers, it is specifically expressed as follows:

[0017]

[0018] Where Y represents the brain activity matrix of N nodes at the Kth time point;

[0019] The prediction task uses the data of T adjacent time samples of N brain network nodes and expresses the brain activity feature prediction task as a function ψ. The specific expression is as follows:

[0020] γ=ψ θ (X t-T+1 ,X t-T ,...,X t ; G).

[0021] The present invention also provides a closed-loop multimodal neural stimulation method based on time interference, comprising the following steps:

[0022] Step 1: Determine the stimulation target link, determine the planned stimulation location and the expected four brain wave characteristic signal responses, including δ, θ, α, and β; the frequency of δ is 0.5-3.5Hz, the frequency of θ is 4-7Hz, the frequency of α is 8-13Hz, and the frequency of β is 14-30Hz;

[0023] Step 2: Setting stimulation parameters. If it is the first stimulation, there is no available graph convolutional neural network feedback result at this time. It is necessary to calculate the placement of the two sets of electrodes and the current frequency based on the generation characteristics of the envelope electric field. If it is not the first stimulation, the placement of the electrodes and the frequency parameter setting are determined based on the results predicted by the graph convolutional neural network.

[0024] Step 3: Perform the stimulation phase, using two sets of electrodes placed on the scalp to generate transcranial alternating current with a slight frequency difference. The two sets of currents will stimulate a low-frequency envelope electric field in a specific area of ​​the brain to activate neurons;

[0025] Step 4: Dual-modal data acquisition and preprocessing: EEG and fNIRS acquisition systems jointly acquire brain signals, pre-process the acquired signals, and package them according to the corresponding sampling time;

[0026] Step 5: Data processing and correlation analysis: The upper control system extracts features from the two modal data, performs correlation analysis and calculations, and obtains time component features and space component features.

[0027] Step 6: Determine whether the stimulation has achieved the expected goal by analyzing the collected four brain wave signals and physiological reactions of the stimulated person;

[0028] Step 7: Graph convolutional neural network prediction phase: Use the previous stimulation data set and the stimulation data set collected during the stimulation to train the graph convolutional neural network prediction model, and use the four brain waves of the expected brain activity characteristics as the starting point to reversely predict the original stimulated points and activation conditions;

[0029] Step 8: Stimulation scheme optimization: Use the graph convolutional neural network to predict the stimulated position and activation degree to adjust the electrode placement and the generated envelope electric field frequency.

[0030] Preferably, in step 2, the specific expression for calculating the placement positions of the two groups of electrodes according to the generation characteristics of the envelope electric field is as follows:

[0031]

[0032]

[0033]

[0034]

[0035] Where E is the electric field intensity, J is the current density, ρ is the spatial current density, and B is the magnetic field intensity. Finite element simulation can be used to determine the distribution of the envelope electric field in the skull and the placement of the electrode. represents the Hamiltonian operator, ε 0 represents the vacuum dielectric constant, j represents an imaginary number, μ 0 represents the magnetic permeability of vacuum, and → represents a vector.

[0036] Preferably, the specific process of dual-modal data collection and preprocessing in step 4 is as follows:

[0037] S41, turning on the near-infrared light source to capture electrical signals and optical signals from the sensor at the same time, obtaining EEG signals and fNIRS signals;

[0038] S42, performing analog-to-digital conversion on the acquired signal after removing noise and filtering clutter;

[0039] S43, collecting the data to the data acquisition system MCU for artifact removal and baseline drift removal processing;

[0040] S44, the processed EEG and fNIRS data are packaged according to a unified clock, and the data packets are sent to the upper control system in time sequence.

[0041] Preferably, the specific process of data processing and correlation analysis in step 5 is as follows:

[0042] S51, extract data features from EEG and fNIRS data;

[0043] S52, using the CCA algorithm to perform correlation analysis on the extracted features to obtain time component features and space component features;

[0044] S53. Use the time component features and space component features obtained by analysis to perform cross-validation and summarize the bimodal fusion data.

[0045] Therefore, the present invention adopts the above-mentioned closed-loop multimodal neural stimulation system and method based on time interference, which can accurately and customizedly determine the actual stimulation plan under different individual conditions; in addition, deep nerve stimulation can be performed without electrode implantation surgery, which greatly reduces costs and greatly improves safety; finally, the complete closed-loop system set up by the present invention enables the operator to efficiently optimize and adjust the stimulation plan, greatly reducing the time cost in actual use.

[0046] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A system schematic diagram of a closed-loop multimodal neural stimulation method based on time interference according to the present invention;

[0048] Figure 2 A structural diagram of a closed-loop multimodal neural stimulation system based on time interference according to the present invention;

[0049] Figure 3 This is a flow chart of correlation analysis of the present invention;

[0050] Figure 4 This is a schematic diagram of the brain network of the present invention;

[0051] Figure 5 This is a structural diagram of the graph convolutional neural network of the present invention. DETAILED DESCRIPTION

[0052] Example

[0053] The following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] See also Figure 1-5, a closed-loop multimodal neural stimulation system based on time interference, including a time interference stimulation system, an EEG-functional near-infrared sampling system and a host control system; the time interference stimulation system uses the beat electric field generated by two sets of electrodes to accurately stimulate the designated area of ​​the brain; the EEG-functional near-infrared sampling system is a dual-modal collector that couples EEG and functional near-infrared, including signal extraction and correlation analysis, and analyzes the stimulation effect and adjusts the stimulation plan by integrating the integrated brain signal data with EEG time accuracy and fNIRS spatial accuracy; the host control system includes dual-modal fusion model calculation, graph convolutional neural network prediction and stimulation plan formulation;

[0055] The temporal interference stimulation system includes a communication control module, a power module and a temporal interference stimulation module; the communication module adopts the Nordic nRF52832 series SoC chip, which integrates the STM32M4F-level processor and the 2.4GHz wireless communication function, and is responsible for the communication between the stimulation system and the host computer and the control of the stimulation current; the power module adopts a low-noise, low-voltage dropout regulator with an external shielding layer design to ensure that the precise high-frequency signal of the electrode part will not be interfered by noise; the temporal interference stimulation module is composed of two-way differential beat electrode groups, and two sine wave generators generate 2KHz AC signals with a phase difference of 10-20Hz. The two signals will interfere in the skull through the electrode group attached to the surface of the scalp. The low-frequency envelope current generated by the interference can stimulate the nerve cells in the designated area. This stimulation method can directly stimulate the nerve cells in the designated area to produce action potentials, and at the same time will not affect the brain areas adjacent to the designated stimulation area.

[0056] The EEG-functional near-infrared sampling system includes an EEG acquisition module and an fNIRS functional near-infrared brain signal acquisition module; the EEG acquisition module uses epidermal electrodes to record signals, and arranges electrodes according to the international 10-20 system or a customized layout; in this module, the EEG signals generated by time interference stimulation are filtered and denoised by a preprocessing circuit, and then converted into digital form by an analog-to-digital converter (ADC) and sent to the communication and control subsystem. The EEG signals will be uploaded to the host computer after the MCU removes artifacts and baseline drift. The fNIRS functional near-infrared brain signal acquisition module measures the activity level of a specific area of ​​the brain by monitoring the absorption changes of brain hemoglobin and oxygenated hemoglobin within the near-infrared spectrum. The module uses a light-emitting diode as a light source to generate 650nm near-infrared light. A photodiode is used as a light signal detector to convert the light signal passing through the brain tissue into an electrical signal for analysis. The obtained electrical signal will be denoised and filtered by pre-processing and then digitized by an analog-to-digital converter and sent to the communication and control subsystem. The MCU will perform brain tissue blood oxygen calculation to obtain brain activity signals and upload the activity signals to the host computer. In the present invention, the signal acquisition system couples EEG and fNIRS to perform multimodal acquisition of brain signals. By combining the high temporal resolution of EEG (which can capture millisecond-level brain activity) and the high spatial resolution of fNIRS (which can identify brain activity in areas of different depths), more detailed and comprehensive brain activity information is obtained in the time domain and spatial domain at the same time. At the same time, this method complementarily combines the sensitivity of EEG to neural electrical activity and the sensitivity of fNIRS to changes in cerebral blood flow and cerebral blood oxygen, providing more comprehensive brain activity information in terms of types. EEG and fNIRS will be integrated into one hardware system using the same standard clock to solve the timing accuracy problem caused by different sampling rates of the two acquisition methods. This solution will make the time error of brain activity data collected by the two methods less than 1.5ms, so that the data collected from the two modules can be accurately located at a time node. Provide effective and comprehensive brain activity information for subsequent analysis of stimulation effects.

[0057] The dual-modal fusion model extracts features from EEG and fNIRS signal data and combines the two into a fusion model and then performs correlation analysis, such as Figure 3 , as follows:

[0058] The feature extraction of EEG mainly involves extracting the response related to the target stimulus, combining the ERP (event-related potential) of the target stimulus into a data matrix, and then convolving the data matrix with the hemodynamic response function to obtain the EEG data features that can be used for subsequent correlation analysis to form the feature matrix X EEG; The feature extraction of fNIRS is to construct a data matrix by sampling the blood oxygen data within 40 seconds after nerve stimulation in time-division and sorting them by time to obtain the fNIRS data features corresponding to the EEG data features, forming the feature matrix X fNIRS ;

[0059] The construction of the fusion model is to calculate the maximum correlation matrix A of the two feature data sets through the CCA (canonical correlation analysis) algorithm. EEG and A fNIRS Then, the temporal component represented by the EEG data and the spatial component C represented by the fNIRS data are fitted by the least squares method. EEG and C fNIRS In this algorithm, the following model formula can be constructed:

[0060] X i =A i C i ,i=1,2,...,2N

[0061] In this formula A i ∈K T×D , i is the number of modes, T is the number of sampling time points of the feature, V i is the characteristic variable of the characteristic matrix, and D is the minimum rank of the characteristic matrix.

[0062] Therefore, in different correlation matrices A i The multiplication is equal to the correlation coefficient as shown in the following formula:

[0063]

[0064] Among them A i The vector of the jth column in is represented as Because the model is consistent with the principle of least squares, this method can be used to obtain the time component and spatial component C of the two data sets. i :

[0065]

[0066] Graph convolutional neural network prediction is to predict the stimulation position according to the desired stimulation effect and formulate a stimulation plan. The data of the graph structure usually includes nodes and edges, where nodes represent entities and edges represent the relationship between them. In the neural stimulation model, different brain regions can be constructed as nodes, and the different connection states between brain regions can be constructed as edges. Prediction at the node level is the prediction of the stimulation position. The specific expression is as follows:

[0067] Let G = (V; E) represent a graph network, V represents the node set, E represents the edge set, if two nodes vi ,v j There are edges connected, then (v i ,v j )∈E, the entire network is considered to be composed of nodes composed of different brain functional areas connected by connecting nerves, and the whole is a structure composed of subgraphs of a finite number of nodes;

[0068] Among them, the four brain waves in a certain brain region node for a fixed period of time are recorded as different dimensions, the number of dimensions is D, and the number of samples is τ. The input brain activity feature matrix of N brain network nodes at time t is specifically expressed as follows:

[0069] χ=(X 1 ,X 2 ,...,X t ,...,X τ )∈R N×D×τ

[0070] Where X represents the brain activity feature matrix of N nodes in the figure at time t, which is specifically expressed as follows:

[0071]

[0072] For the output brain activity feature matrix with K sampling numbers, it is specifically expressed as follows:

[0073]

[0074] Where Y represents the brain activity matrix of N nodes at the Kth time point;

[0075] The prediction task uses the data of T adjacent time samples of N brain network nodes and expresses the brain activity feature prediction task as a function ψ. The specific expression is as follows:

[0076] γ=ψ θ (X t-T+1 ,X t-T ,...,X t ; G).

[0077] A closed-loop multimodal neural stimulation method based on time interference comprises the following steps:

[0078] Step 1: Determine the stimulation target link, determine the planned stimulation location and the expected four brain wave characteristic signal responses, including δ, θ, α, and β; the frequency of δ is 0.5-3.5Hz, the frequency of θ is 4-7Hz, the frequency of α is 8-13Hz, and the frequency of β is 14-30Hz;

[0079] Step 2: Setting stimulation parameters. If it is the first stimulation, there is no available graph convolutional neural network feedback result at this time. It is necessary to calculate the placement position and current frequency of the two sets of electrodes according to the generation characteristics of the envelope electric field. If it is not the first stimulation, the placement position of the electrodes and the frequency parameter setting are determined according to the results predicted by the graph convolutional neural network. The specific expression for calculating the placement position of the two sets of electrodes according to the generation characteristics of the envelope electric field is as follows:

[0080]

[0081]

[0082]

[0083]

[0084] Where E is the electric field intensity, J is the current density, ρ is the spatial current density, and B is the magnetic field intensity. Finite element simulation can be used to determine the distribution of the envelope electric field in the skull and the placement of the electrode. represents the Hamiltonian operator, ε 0 represents the vacuum dielectric constant, j represents an imaginary number, μ 0 represents the magnetic permeability of vacuum, and → represents a vector.

[0085] Step 3: Perform the stimulation phase, using two sets of electrodes placed on the scalp to generate transcranial alternating current with a slight frequency difference. The two sets of currents will stimulate a low-frequency envelope electric field in a specific area of ​​the brain to activate neurons;

[0086] Step 4: Dual-modal data acquisition and preprocessing: EEG and fNIRS acquisition systems jointly acquire brain signals, preprocess the acquired signals, and package them according to the corresponding sampling time; the specific process is as follows:

[0087] S41, turning on the near-infrared light source to capture electrical signals and optical signals from the sensor at the same time, obtaining EEG signals and fNIRS signals;

[0088] S42, performing analog-to-digital conversion on the acquired signal after removing noise and filtering clutter;

[0089] S43, collecting the data to the data acquisition system MCU for artifact removal and baseline drift removal processing;

[0090] S44, the processed EEG and fNIRS data are packaged according to a unified clock, and the data packets are sent to the upper control system in time sequence.

[0091] Step 5: Data processing and correlation analysis: The upper control system will extract features from the two modal data, perform correlation analysis and calculation, and obtain the time component features and space component features. The specific process is as follows:

[0092] S51, extract data features from EEG and fNIRS data;

[0093] S52, using the CCA algorithm to perform correlation analysis on the extracted features to obtain time component features and space component features;

[0094] S53. Use the time component features and space component features obtained by analysis to perform cross-validation and summarize the bimodal fusion data.

[0095] Step 6: Determine whether the stimulation has achieved the expected goal by analyzing the collected four brain wave signals and physiological reactions of the stimulated person;

[0096] Step 7: Graph convolutional neural network prediction phase: Use the previous stimulation data set and the stimulation data set collected during the stimulation to train the graph convolutional neural network prediction model, and use the four brain waves of the expected brain activity characteristics as the starting point to reversely predict the original stimulated points and activation conditions;

[0097] Step 8: Stimulation scheme optimization: Use the graph convolutional neural network to predict the stimulated position and activation degree to adjust the electrode placement and the generated envelope electric field frequency.

[0098] Therefore, the present invention adopts the above-mentioned closed-loop multimodal neural stimulation system and method based on time interference, uses time interference stimulation technology to perform precise non-invasive stimulation on cranial nerves; uses EEG coupled fNIRS dual-modal brain signal acquisition scheme to obtain comprehensive and accurate brain activity data; constructs an EEG and fNIRS data feature fusion model and uses a graph neural network prediction model to guide the formulation of the stimulation plan, which can continuously optimize its effect during use.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A closed-loop multimodal neural stimulation system based on time interference, characterized in that: It includes a time interference stimulation system, an EEG-functional near-infrared sampling system and a host control system; the time interference stimulation system uses the beat electric field generated by two sets of electrodes to accurately stimulate the designated area of ​​the brain; the EEG-functional near-infrared sampling system includes an EEG EEG acquisition module, an fNIRS functional near-infrared brain signal acquisition module and a communication and control subsystem, and the communication and control subsystem includes an MCU and wireless communication; the host control system includes dual-modal fusion model calculation, graph convolutional neural network prediction and stimulation program formulation; the dual-modal fusion model calculation extracts features from EEG and fNIRS signal data and combines the two into a fusion model and then performs correlation analysis; The system performs the following steps: Step 1: Determine the stimulation target link, determine the planned stimulation location and the expected four brain wave characteristic signal responses, including δ, θ, α, and β; the frequency of δ is 0.5-3.5Hz, the frequency of θ is 4-7Hz, the frequency of α is 8-13Hz, and the frequency of β is 14-30Hz; Step 2: Setting stimulation parameters. If it is the first stimulation, there is no available graph convolutional neural network feedback result at this time. It is necessary to calculate the placement of the two sets of electrodes and the current frequency based on the generation characteristics of the envelope electric field. If it is not the first stimulation, the placement of the electrodes and the frequency parameter setting are determined based on the results predicted by the graph convolutional neural network. Step 3: Perform the stimulation phase, using two sets of electrodes placed on the scalp to generate transcranial alternating current with a slight frequency difference. The two sets of currents will stimulate a low-frequency envelope electric field in a specific area of ​​the brain to activate neurons; Step 4: Dual-modal data acquisition and preprocessing: EEG and fNIRS acquisition systems jointly acquire brain signals, pre-process the acquired signals, and package them according to the corresponding sampling time; Step 5: Data processing and correlation analysis: The upper control system extracts features from the two modal data, performs correlation analysis and calculations, obtains time component features and space component features, and cross-validates the time component features and space component features to summarize the dual-modal fusion data. Step 6: Determine whether the stimulation has achieved the expected goal by analyzing the collected four brain wave signals and physiological reactions of the stimulated person; Step 7: Graph convolutional neural network prediction phase: Use the previous stimulation data set and the stimulation data set collected during the stimulation to train the graph convolutional neural network prediction model, and use the four brain waves of the expected brain activity characteristics as the starting point to reversely predict the original stimulated points and activation conditions; Step 8: Stimulation scheme optimization: Use the graph convolutional neural network to predict the stimulated position and activation degree to adjust the electrode placement and the generated envelope electric field frequency.

2. A closed-loop multimodal neural stimulation system based on time interference according to claim 1, characterized in that: The time interference stimulation system includes a communication control module, a power module and a time interference stimulation module; the communication module adopts a SoC chip, which integrates the functions of a processor and wireless communication, and completes the communication between the time interference stimulation system and the upper control system and the control of the stimulation current; the power module adopts a low-noise, low-voltage dropout regulator with an external shielding layer design; the time interference stimulation module is composed of two differential beat electrode groups, and two sine wave generators generate 2KHz AC signals with a phase difference of 10-20Hz. The two signals interfere with each other in the skull through an electrode group attached to the surface of the scalp, and the low-frequency envelope current generated by the interference stimulates the nerve cells in the designated area.

3. A closed-loop multimodal neural stimulation system based on time interference according to claim 1, characterized in that: The EEG acquisition module uses epidermal electrodes to record signals; the EEG signals generated by time interference stimulation in the EEG acquisition module are filtered and denoised by the preprocessing circuit, and then converted into digital form by the analog-to-digital converter and sent to the communication and control subsystem. The EEG signals will be removed by the MCU for artifact removal and baseline drift removal before being uploaded to the host control system; the fNIRS functional near-infrared brain signal acquisition module measures the activity level of specific areas of the brain by monitoring the absorption changes of brain hemoglobin and oxygenated hemoglobin in the near-infrared spectrum. The fNIRS functional near-infrared brain signal acquisition module uses a light-emitting diode as a light source to generate 650nm near-infrared light, and a photodiode as a light signal detector to convert the light signal passing through the brain tissue into an electrical signal for analysis. The obtained electrical signal will be denoised and filtered by preprocessing, and then converted into a digital form by the analog-to-digital converter and sent to the communication and control subsystem. The MCU calculates brain tissue blood oxygen to obtain brain activity signals and uploads the activity signals to the host control system.

4. A closed-loop multimodal neural stimulation system based on time interference according to claim 1, characterized in that: The graph convolutional neural network prediction is to predict the stimulation position according to the desired stimulation effect and formulate a stimulation plan, which is specifically expressed as follows: set up represents a graph network, represents a node set, Represents an edge set. If two nodes , If there are edges connected, , the entire network is considered to be composed of nodes composed of different brain functional areas connected by connecting nerves, and the whole is a structure composed of subgraphs of a finite number of nodes; Among them, the four brain waves in a certain brain region node for a fixed period of time are recorded as different dimensions, and the number of dimensions is , let the sampling number be ,but The brain network nodes Input brain activity feature matrix at each moment The specific expressions are as follows: in It represents the brain activity feature matrix of N nodes in the figure at time t, which is specifically expressed as follows: For The output brain activity feature matrix of the number of samples The specific expressions are as follows: in represents the brain activity matrix of N nodes at the Kth time point; Prediction task usage The brain network nodes are adjacent The data sampled at each time point is used to represent the brain activity feature prediction task as a function , the specific expression is as follows: 。 5. A closed-loop multimodal neural stimulation system based on time interference according to claim 4, characterized in that: In step 2, the specific expression for calculating the placement positions of the two sets of electrodes according to the generation characteristics of the envelope electric field is as follows: Where E is the electric field intensity, J is the current density, is the spatial current density, B is the magnetic field intensity, and finite element simulation is used to determine the distribution of the envelope electric field in the skull and the placement of the electrodes. represents the Hamiltonian operator, represents the vacuum dielectric constant, represents an imaginary number, represents the magnetic permeability of vacuum, and → represents a vector.

6. A closed-loop multimodal neural stimulation system based on time interference according to claim 5, characterized in that: The specific process of dual-modal data collection and preprocessing in step 4 is as follows: S41, turning on the near-infrared light source to capture electrical signals and optical signals from the sensor at the same time, obtaining EEG signals and fNIRS signals; S42, performing analog-to-digital conversion on the acquired signal after removing noise and filtering clutter; S43, the data is aggregated to the MCU for artifact removal and baseline drift removal processing; S44, the processed EEG and fNIRS data are packaged according to a unified clock, and the data packets are sent to the upper control system in time sequence.

7. A closed-loop multimodal neural stimulation system based on time interference according to claim 6, characterized in that: The specific process of data processing and correlation analysis in step 5 is as follows: S51, extract data features from EEG and fNIRS data; S52, using the CCA algorithm to perform correlation analysis on the extracted features to obtain time component features and space component features; S53. Use the time component features and space component features obtained by analysis to perform cross-validation and summarize the bimodal fusion data.

Images