Closed-loop nerve regulation and control system and method based on electroencephalogram detection

Abstract

The invention provides a closed-loop nerve regulation and control system and method based on electroencephalogram detection. The defects that a traditional open-loop method lags in stimulation, equipment is heavy and the like are overcome. The system integrates an electroencephalogram acquisition module, an electroencephalogram analysis module, a remote program control module and a stimulation output module to form a closed-loop link. Through 32-channel 1kHz signal acquisition and 1-100Hz filtering pretreatment, pathological features are extracted in real time by using a deep convolutional neural network, stimulation parameters are modified through a remote program control module, and multi-channel collaborative stimulation and safe electrical stimulation are realized in combination with a time division multiplexing architecture and bipolar optocoupler isolation output. The method has the advantages that the pathological feature recognition accuracy of the deep convolutional neural network is greater than or equal to 95%, and the dynamic response delay is less than 50ms; a time division multiplexing hardware architecture is designed, multi-channel collaborative stimulation is achieved through a single DAC, and the size is reduced by 30%; + / -5V bipolar output is realized by using an optocoupler isolated H bridge driver, and the IEC 60601 medical safety standard is met; wireless regulation and control of stimulation parameters are supported, the long-term implantation requirement is met, and the risk of a secondary operation is reduced.

Description

Technical Field

[0001] This invention provides a closed-loop neural modulation system and method based on electroencephalogram (EEG) detection. Background Technology

[0002] Neuromodulation technology involves targeted intervention of the nervous system through electrical or chemical stimulation, enabling precise intervention of pathological neural circuits and providing novel treatment pathways for intractable diseases such as epilepsy and Parkinson's disease. Building upon this foundation, closed-loop neuromodulation technology dynamically adjusts stimulation parameters by providing real-time feedback of the patient's electroencephalogram (EEG) signals (such as local field potentials, LFP), overcoming the static limitations of traditional open-loop stimulation and significantly improving treatment precision.

[0003] Traditional implantable neuromodulation devices primarily rely on open-loop stimulation modes and multi-channel parallel DAC hardware architectures, which suffer from three major drawbacks: First, the stimulation parameter adjustment is lagging, failing to respond in real time to changes in pathological states (such as pre-seizure aura), and easily causing side effects such as dyskinesia; second, miniaturization is hindered, with multi-channel independent signal source design resulting in circuit volume redundancy (>30%) and excessive power consumption (standby current >100μA), making it difficult to meet long-term implantation requirements; third, clinical control is a bottleneck, as parameter adjustments require a second surgery, leading to high infection risks and increased medical costs. These methods, limited by rigid hardware architecture and insufficient real-time algorithm performance, struggle to construct a closed-loop system that combines pathological prediction capabilities (accuracy >90%), ultra-low latency response (<50ms), and safe remote control.

[0004] Therefore, how to achieve closed-loop modulation of the nervous system to overcome the response lag bottleneck of open-loop modulation, how to solve the volume redundancy problem of multi-channel implantation through miniaturized hardware architecture, and how to design a closed-loop neuromodulation system that supports safe remote dynamic programming while reducing tissue damage are the urgent problems to be solved in the field of implantable EEG neuromodulation. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this invention proposes a closed-loop neuromodulation system and method that integrates real-time EEG analysis, remote programming technology, time-division multiplexing hardware architecture, and bipolar optocoupler isolated output. Through four modules—EEG signal acquisition, pathological feature identification, remote adjustment of dynamic parameters, and safe stimulation output—closed-loop modulation of the nervous system is achieved.

[0006] According to an embodiment of the present invention, a closed-loop neuromodulation system and method based on electroencephalogram (EEG) detection is provided, characterized by comprising: an EEG signal acquisition module, an EEG signal analysis and detection module, a remote programming module, and a neural stimulation output module. Wherein:

[0007] EEG signal acquisition module: Used to acquire EEG signals and transmit the acquired signals to the analysis and detection module.

[0008] EEG signal analysis and detection module: Used to receive the collected EEG signals, analyze and detect the EEG signals using machine learning methods, predict information on the onset of nervous system diseases, and send the prediction results to the remote control module.

[0009] Remote control module: Used to receive prediction results and user operation commands, programmatically modify stimulation parameters on the controller, and send modification instructions to the neural stimulation output module through remote programming.

[0010] Stimulation output module: Used to output electrical stimulation signals to achieve neural modulation, and the stimulation parameters of the output electrical stimulation can be modified according to the instructions of the remote programmable module.

[0011] in:

[0012] The EEG signal acquisition module supports 32-channel simultaneous acquisition with a sampling rate of 1kHz.

[0013] The EEG signal analysis and detection module includes an adaptive filtering unit and a machine learning unit. The adaptive filtering unit is used to dynamically suppress noise in the acquired EEG signals. The machine learning unit uses a deep convolutional neural network algorithm to extract and recognize features from the preprocessed EEG signals. The machine learning unit includes ten network layers: an initial feature mapping layer, a channel normalization layer, a feature enhancement layer, a deep convolutional layer, an activation function layer, a Dropout layer, a channel normalization layer, a global average pooling layer, a fully connected layer, and a Softmax activation layer.

[0014] The remote control module includes an external programmer for the closed-loop neural modulation system and an internal BLE receiver module. The external programmer is equipped with a touchscreen serial port, supporting parameter viewing and modification. The internal BLE receiver module uses an ESP32-C3 single-chip solution; its integrated Bluetooth 5 module supports dual-mode communication in the 2.4GHz ISM band, with power consumption as low as 1.3μA in deep sleep mode and a communication distance of ≥10 meters. The design utilizes the heterogeneous multi-core architecture of the ESP32-C3, ensuring that the BLE Bluetooth program and the main program are executed on separate cores, thus guaranteeing the operation of the main control program.

[0015] The stimulus output module includes a main controller circuit, a time-division multiplexing circuit, and a bipolar output circuit. The main controller circuit functions based on a time-division multiplexing scheduling algorithm (τ). i =T / N) controls the DAC to generate a time-domain superimposed signal, realizing single-chip driving of eight-channel output; the time-division multiplexing circuit is responsible for dynamically selecting a single channel within the Δt time slot through the CD4051 analog switch, satisfying t pulseTiming constraint ≤1 / (8f); the bipolar output circuit converts the unipolar signal into ±5V bidirectional pulses via an H-bridge with PC817C optocoupler isolation (5000Vrms), eliminating the electrode polarization effect.

[0016] The main advantages of the closed-loop neuromodulation system and method based on electroencephalogram (EEG) detection of this invention include:

[0017] 1. This invention integrates an EEG signal analysis and detection algorithm to achieve real-time identification and prediction of pathological features of neurological diseases such as epilepsy. Based on this, a dynamic parameter control mechanism is constructed, which solves the problem of delayed stimulation parameter adjustment caused by the inability of traditional open-loop DBS systems to respond to EEG changes in real time, and significantly improves the treatment effect.

[0018] 2. This invention designs a unique time-division multiplexing hardware architecture, using a single DAC chip in conjunction with a CD4051 multiplexer to achieve eight independent outputs, through a precise timing control algorithm (τ). i =T / N) ensures that the signals of each channel do not interfere with each other. Compared with the traditional multi-DAC parallel solution, the circuit size is reduced by 30% and the power consumption is reduced by 50%, providing key technical support for the miniaturization of implantable devices.

[0019] 3. This invention constructs a highly safe bipolar stimulation output system, which achieves ±5V bidirectional current output through an H-bridge circuit, effectively avoiding tissue electrode polarization problems caused by long-term unipolar stimulation; at the same time, it adopts PC817C optocoupler isolation technology to provide electrical isolation protection of up to 5000Vrms, ensuring patient safety and meeting the safety standards for medical devices.

[0020] 4. This invention proposes a remote control scheme based on BLE. Utilizing the heterogeneous multi-core architecture of the ESP32-C3 chip, it achieves a wireless communication distance of ≥10 meters while ensuring system real-time performance. Doctors can remotely adjust stimulation parameters (frequency 50-250Hz, pulse width 60-450μs) through an external programmer, avoiding the drawbacks of traditional equipment that requires secondary surgery for parameter adjustment, and significantly reducing medical risks and patient burden. Attached Figure Description

[0021] Figure 1 This is an overall block diagram of a closed-loop neural modulation system according to an embodiment of the present invention.

[0022] Figure 2 This is a flowchart of a closed-loop neural modulation system according to an embodiment of the present invention.

[0023] Figure 3 This is a functional block diagram of an H-bridge circuit according to an embodiment of the present invention. Detailed Implementation

[0024] According to the present invention, a closed-loop neuromodulation system and method based on electroencephalogram (EEG) detection is provided, characterized by comprising: an EEG signal acquisition module, an EEG signal analysis and detection module, a remote programming module, and a stimulation output module (see...). Figure 1 ).in:

[0025] The aforementioned EEG signal acquisition module is used to acquire EEG signals and transmit the acquired signals to the analysis and detection module.

[0026] The aforementioned EEG signal analysis and detection module is used to receive the collected EEG signals, analyze and detect the EEG signals using machine learning methods, predict information on the onset of nervous system diseases, and send the prediction results to the remote control module.

[0027] The remote control module is used to receive prediction results and user operation commands, programmatically modify stimulation parameters on the controller, and send modification instructions to the neural stimulation output module through remote programming.

[0028] The stimulation output module is used to output electrical stimulation signals to achieve neural modulation, and the stimulation parameters of the output electrical stimulation can be specifically modified according to the instructions of the remote programmable module.

[0029] According to a specific embodiment of the present invention, the sampling rate of the EEG signal acquisition module is 1kHz, supports up to 32 acquisition channels to be acquired simultaneously, and performs signal preprocessing with 1-100Hz bandpass filtering.

[0030] According to one embodiment of the present invention, the electroencephalogram (EEG) signal analysis and detection module includes an adaptive filtering unit and a machine learning unit, wherein:

[0031] A) The adaptive filtering unit is used to perform dynamic noise suppression on the acquired EEG signals;

[0032] The machine learning unit employs a deep convolutional neural network algorithm for feature extraction and recognition of preprocessed EEG signals. The machine learning unit comprises ten network layers: an initial feature mapping layer, a channel normalization layer, a feature enhancement layer, a deep convolutional layer, an activation function layer, a Dropout layer, a channel normalization layer, a global average pooling layer, a fully connected layer, and a Softmax activation layer.

[0033] B1) The initial feature mapping layer uses N1 one-dimensional convolutional kernels to perform cross-channel feature recombination on the input signal in the form of pointwise convolutional nodes, generating N1 equal-length feature subbands; in the input data matrix X (m,n) In this context, m represents the number of original signal channels, and n is the number of time-domain sampling points. Therefore, the i-th convolutional kernel passes through an m-dimensional weight vector. and scalar bias b i(1) The linear combination of the inputs is expressed as:

[0034]

[0035] Where i = 1, 2, 3, ..., N1, and m is the number of input signal channels. Let x represent the output feature value of the i-th feature channel in the n-th layer network at the j-th time-domain sampling point. p,j It is the j-th sampling point of the p-th channel of the input sample. It is the scalar bias of the i-th convolutional kernel. It is the weight vector of the p-th channel of the i-th convolutional kernel.

[0036] B2) The channel normalization layer achieves dynamic equalization of cross-channel responses through statistical feature calibration, and normalizes the output of the previous layer:

[0037]

[0038] Where: γ and β are the scaling and translation parameters of the channel normalization layer; μ β , These are the mean and variance of the mini-batch data, respectively; ∈ is a minimal constant to prevent division by zero errors.

[0039] The output of the feature enhancement layer (B3) is:

[0040]

[0041] Where K is the number of parallel branches in the multi-scale feature enhancement layer, and α i,k It is the attention weight of the k-th branch of the i-th output channel. These are the weight parameters of the p-th channel in the k-th branch.

[0042] B4) The deep convolutional layer is used to extract feature information from P-dimensional input data. Each new signal sample input to this layer is convolved with a certain deep convolutional node in the deep convolutional layer. The parameter set of the i-th deep convolutional node includes l weight coefficients and 1 bias parameter. The feature output is as follows:

[0043]

[0044] Where p is the temporal width of the convolution kernel, and R is the temporal offset of the convolution kernel. It is the weight of the temporal offset r of the i-th convolutional kernel's p-th channel. These are the bias parameters of the deep convolutional layer.

[0045] B5) The activation function layer with parameterized correction optimizes feature representation by dynamically adjusting the activation threshold, and its output is:

[0046]

[0047] Where α and β are the slope and threshold parameters of the parameterized activation function, respectively. This is the output of a deep convolutional layer. ReLU(·) represents a non-linear activation function, and exp(·) represents the natural exponential function.

[0048] B6) The Dropout layer suppresses overfitting through probabilistic feature masking, and its output is:

[0049]

[0050] Where p is the probability of neurons being dropped in the random deactivation layer.

[0051] B7) The channel normalization layer achieves cross-channel response equalization through statistical feature calibration; its output is:

[0052]

[0053] Where γ and β are the scaling and translation parameters of the channel normalization layer; μ β , , respectively, represent the mean and variance of the mini-batch data; ∈ is the numerical stability constant.

[0054] B8) The global average pooling layer generates a fixed-length feature vector by reducing the spatial dimension. Its output is:

[0055]

[0056] Where n is the length of the time-domain sequence. It is the output of the global average pooling layer.

[0057] B9) The fully connected layer achieves feature mapping through affine transformation, and its output is:

[0058]

[0059] in This is the weighted output of the fully connected layer, where C is the number of input features to the fully connected layer, and k is the output class index. These are the weight parameters of the fully connected layer. These are the bias parameters of the fully connected layer.

[0060] B10) Finally, the Softmax activation layer outputs the final discrimination result through probability normalization:

[0061]

[0062] in This represents the output probability of the Softmax layer, where c is the target class index (1 ≤ c ≤ K). It is the weighted input of class c in the fully connected layer.

[0063] According to one embodiment of the present invention, the remote control module receives the prediction results from the EEG signal analysis and detection module via a wired connection or WIFI wireless communication, and receives user commands through a 3.5-inch touchscreen user interface. The programmable controller can select appropriate stimulation parameters based on the prediction results, or allow the user to specify the stimulation parameters. The programmable controller remotely connects to the stimulation output module via BLE communication, and issues modification commands to achieve remote programming of the stimulation parameters.

[0064] According to one embodiment of the present invention, the adjustment of the stimulation parameters includes pulse amplitude, pulse width, pulse frequency, and waveform. The pulse amplitude ranges from 0.25 mA to 3.5 mA, with an adjustment step of 0.05 mA; the pulse width ranges from 150 to 1000 μs, with an adjustment step of 50 μs; the pulse frequency ranges from 1 to 30 Hz, with an adjustment step of 1 Hz. Square wave, triangular wave, and sine wave are all selectable pulse waveforms. The main method for selecting the stimulation parameters is the calculation of charge density. Current is the flow of charge, and the total charge Q of the electrical stimulation can be calculated by integrating the stimulation current over the stimulation time or by calculating the product of the current amplitude and the pulse width. The calculation formula is as follows:

[0065]

[0066] or

[0067] The relationship between stimulus duration and the amount of charge required for excitation can be directly derived from the intensity-duration relationship, since Q = PW * I, and the charge-duration relationship is:

[0068] Q th =I rh [PW+T ch (13)

[0069] Among them, Q th For the threshold charge, I rh Where P is the base current, PW is the pulse width, and T is the pulse width. ch Time-dependent.

[0070] According to one embodiment of the present invention, the stimulation output module receives programming instructions from the remote programming module using BLE communication, supports multi-channel coordinated stimulation, and supports bipolar output.

[0071] According to one embodiment of the present invention, the BLE communication method adopts the ESP32-C3 single-chip solution, whose integrated Bluetooth 5 module supports dual-mode communication (broadcast / connection) in the 2.4GHz ISM band. The design utilizes the heterogeneous multi-core architecture of the ESP32-C3, ensuring that the execution of the BLE Bluetooth part of the program and the main program are not on the same core, thus guaranteeing the operation of the main control program.

[0072] According to one embodiment of the present invention, the multi-channel co-stimulation uses a time-division multiplexing method to achieve multi-channel output from a single DAC signal source, supporting up to eight independent stimulation output channels. The implementation principle of the time-division multiplexing method is as follows: based on time-division multiplexing technology, by dynamically allocating fixed time windows, the output signal of a single digital-to-analog converter (DAC) is periodically divided into multiple independent time slots. Each time slot corresponds to the signal loading and output of one stimulation channel, thereby achieving parallel processing of multi-channel signals in the time domain. The specific implementation process is as follows:

[0073] (1) Setting parameters for the time-sharing multiplexing model

[0074] Let the total stimulation period of the stimulation system be TTT, and the maximum number of channels supported by the system be Nmax = 8. Divide the total period T into N equal durations (i = 1, 2, ..., N) satisfying:

[0075]

[0076] Each time slot τ i A stimulation signal loading window corresponds to an independent channel. Within this window, the voltage pulse signal U(t) output by the DAC is routed to the analog switch of the i-th channel and output to the target neural target through the electrode.

[0077] (2) To ensure the independence of stimulation parameters (such as pulse amplitude and pulse width) for each channel, channel switching is achieved through a multi-channel analog switch (such as the CD4051 chip). The channel selection signal S2S1S0 (binary code) is dynamically generated by the microcontroller (MCU) according to the time slot index iii, satisfying:

[0078] S2S1S0=Binary(i-1),i=1,2,...,8 (15)

[0079] By combining timer interrupt control with the on / off timing of the analog switch, it is ensured that in time slot τ i Only the i-th channel is active.

[0080] (3) Pulse signal timing constraints: To ensure the effectiveness of stimulation, the single-channel pulse time t must be satisfied. pulse With time slot width τ i Matching relationships:

[0081]

[0082] Where f is the stimulation frequency (f = 1 ~ 30 Hz), it can be deduced that the maximum allowable pulse width can reach 4167 μs, which meets the requirement of 1000 μs.

[0083] (4) Multi-channel synchronous control, the specific strategy of which includes generating a high-precision clock signal through the timer module built into the MCU, triggering the interrupt service routine (ISR) in a loop with a period of T, and performing the following operations in the interrupt service routine in sequence:

[0084] S1: Calculate the current time slot index

[0085] S2: Update DAC output voltage (a i (This refers to the 8-bit DAC input value of the i-th channel);

[0086] S3: Pull the analog switch enable signal of the corresponding channel high for a duration τ. i ;

[0087] S4: In τ i When the time slot ends, turn off the switch and proceed to the next time slot.

[0088] This strategy achieves precise synchronization and isolation of multi-channel signals through hardware and software collaboration, ensuring that the requirements are met. Under these conditions, the stimulation parameters of each channel can be adjusted independently without interfering with each other.

[0089] According to one embodiment of the present invention, the bipolar output adopts an optocoupler-isolated H-bridge driving scheme (e.g., Figure 3 As shown in the figure, the specific implementation is as follows:

[0090] To eliminate the electrode polarization effect caused by the DC component in traditional unipolar stimulation, the stimulation output module adopts a full-bridge H-bridge drive circuit. By dynamically switching the conduction state of the power switching transistors, the current direction is controlled to periodically reverse between positive and negative polarities, thereby generating a bidirectional pulse current. Figure 2 As shown, the H-bridge consists of four switches (Q1-Q4), and its conduction logic satisfies:

[0091] Positive current stage: Q1 and Q4 are on, Q2 and Q3 are off, and the current path is positive power supply → Q1 → electrode → tissue → Q4 → negative power supply.

[0092] Negative current stage: Q2 and Q3 are on, Q1 and Q4 are off, and the current path is positive terminal of power supply → Q3 → electrode → tissue → Q2 → negative terminal of power supply;

[0093] Zero current stage: Q1 and Q3 or Q2 and Q4 are turned off simultaneously, cutting off the current path.

[0094] By switching the above states at high frequency (controlled by PWM signal), a bidirectional pulse current I(t) is synthesized on the load electrode, and its waveform expression is as follows:

[0095]

[0096] Among them I + and I - ω represents the amplitude of the positive / negative polarity current, ω is the pulse frequency (angular frequency), and sgn is the sign function.

[0097] According to one embodiment of the present invention, the optocoupler used for the optocoupler isolation is a PC817c, which has an isolation voltage rating of 5kVrms, meeting the requirements for isolation withstand voltage (≥4kVrms) in the medical electrical equipment safety standard (IEC 60601-1).

Claims

1. A closed-loop neuromodulation system based on electroencephalographic detection, characterized in that The system comprises: a brain electrical signal acquisition module for acquiring brain electrical signals, a brain electrical signal analysis and detection module for receiving the acquired brain electrical signals, analyzing and detecting the brain electrical signals through a machine learning method, predicting nervous system disease information, and sending the prediction results to a remote program control module, a remote program control module for receiving the prediction results and receiving user operation commands, programmably modifying the stimulation parameters on a program control instrument, and issuing modification instructions to a neural stimulation output module through remote program control, a neural stimulation output module for outputting electrical stimulation signals to achieve neural regulation, wherein: the brain electrical signal analysis and detection module comprises an adaptive filtering unit and a machine learning unit, wherein: the adaptive filtering unit is used for dynamic noise suppression of the acquired brain electrical signals; the machine learning unit adopts a deep convolutional neural network algorithm for feature extraction and recognition of the preprocessed brain electrical signals; the machine learning unit comprises ten network layers, namely: an initial feature mapping layer, a channel normalization layer, a feature enhancement layer, a deep convolutional layer, an activation function layer, a Dropout layer, a channel normalization layer, a global average pooling layer, a full connection layer, and a Softmax activation layer, wherein: B1) The initial feature mapping layer adopts N1 one-dimensional convolution kernels to perform cross-channel feature reorganization on the input signal in the form of point-by-point convolution nodes, generating N1 equal-length feature subbands; in the input data matrix X (m,n) where m represents the original signal channel number, n is the time domain sampling point number, the i-th convolution kernel passes through an m-dimensional weight vector and a scalar bias b i (1) linear combination is performed on the input, and the expression is: where i = 1, 2, 3, …, N1, m is the number of input signal channels, represents the output feature value of the i-th feature channel in the n-th layer network at the j-th time domain sampling point, x p,j is the j-th sampling point of the p-th channel of the input sample, is the scalar bias of the i-th convolution kernel, is the weight vector of the p-th channel of the i-th convolution kernel, B2) The channel normalization layer realizes dynamic balance of cross-channel response through statistical feature calibration, and normalizes the output of the previous layer: where: γ, β are scaling and translation parameters of the channel normalization layer; μ β , are the mean and variance of the mini-batch data, respectively; ∈ is a small constant to prevent division by zero errors. B3) The output result of the feature enhancement layer is: where K is the number of parallel branches of the multi-scale feature enhancement layer, a i,k is the attention weight of the kth branch of the ith output channel, is the weight parameter of the pth channel of the kth branch. B4) The deep convolutional layer is used to extract feature information from P-dimensional input data, and each input signal sample to the layer is convolved with a certain deep convolutional node in the deep convolutional layer, wherein the parameter group of the i-th deep convolutional node includes l weight coefficients and 1 bias parameter, and the feature output result is: where p is the convolution kernel time domain width, R is the convolution kernel time domain offset, is the weight of the time domain offset r of the pth channel of the i th convolution kernel, is the depth convolution layer bias parameter B5) The activation function layer with parameterized correction optimizes feature expression by dynamically adjusting the activation threshold, and its output is: where a, b are the slope and threshold parameters of the parametric activation function, respectively, is the output of the deep convolutional layer, ReLU(·) represents a nonlinear activation function, and exp(·) represents the natural exponential function, B6) The Dropout layer suppresses overfitting through probabilistic feature shielding, and its output is: where p is the neuron dropout probability of the random inactivation layer, B7) The channel normalization layer realizes cross-channel response balance through statistical feature calibration; its output is: where γ, β are the scaling and translation parameters of the channel normalization layer; μ β , are the mean and variance of the mini-batch data, respectively; ∈ is a numerical stability constant, B8) The global average pooling layer generates a fixed-length feature vector through spatial dimension reduction, and its output is: where n is the time domain sequence length, is the global average pooling layer output, B9) The full connection layer realizes feature mapping through affine transformation, and its output is: wherein is the fully connected layer weighted output, C is the number of input features of the fully connected layer, k is the output class serial number, is the fully connected layer weight parameter, is the fully connected layer bias parameter, B10) Finally, the Softmax activation layer outputs the final discriminant result through probability normalization: wherein is the Softmax layer output probability, c is the target class index (1≤c≤K), is the weighted input of the fully connected layer for class c.

2. The closed-loop neural regulation system of claim 1, wherein: the sampling rate of the brain electrical signal acquisition module is 1 kHz, supports up to 32 acquisition channels for simultaneous acquisition, and performs 1-100 Hz band-pass filtering signal preprocessing, the remote program control module receives the prediction results of the brain electrical signal analysis and detection module through wired connection and / or WIFI wireless communication, and receives user instructions through a touch screen user interface, the program control instrument selects stimulation parameters according to the prediction results and / or receives user-specified stimulation parameters, the program control instrument is connected to the neural stimulation output module through BLE communication for remote issuance of modification instructions to realize remote program control of the stimulation parameters.

3. The closed-loop neural regulation system of claim 1, wherein: The adjustment of the stimulation parameters includes pulse amplitude, pulse width, pulse frequency, and waveform. in: The pulse amplitude range is 0.25-3.5mA, with an adjustment step of 0.05mA. The pulse width is 150-1000µs, and the adjustment step is 50µs. The pulse frequency is 1-30Hz, and the adjustment step is 1Hz. The pulse waveform is selected from at least one of square wave, triangle wave, and sine wave. The selection of stimulation parameters includes adjusting the charge density so that the total charge Q of the electrical stimulation can be calculated by integrating the stimulation current over the stimulation time and / or by calculating the product of the current amplitude and the pulse width.

4. The closed-loop neural modulation system according to claim 1, characterized in that: The neural stimulation output module uses BLE communication to receive programming commands from the remote programming module, supports multi-channel coordinated stimulation, and supports bipolar output. The multi-channel coordinated stimulation uses a time-division multiplexing method to achieve multi-channel output from a single DAC signal source, supporting up to eight independent stimulation output channels. Time-division multiplexing methods include: time-division multiplexing technology, which dynamically allocates fixed time windows to periodically divide the output signal of a single digital-to-analog converter (DAC) into multiple independent time slots. Each time slot corresponds to the signal loading and output of a stimulation channel, thereby achieving parallel processing of multi-channel signals in the time domain, including: 1) Set the parameters for the time-sharing multiplexing model, including: Let the total stimulation period of the stimulation system be TTT, and the maximum number of channels supported by the system be Nmax = 8. Divide the total period T into N equal durations, i = 1, 2, ..., N, satisfying: Each time slot τ i The stimulation signal loading window corresponding to an independent channel, in which the voltage pulse signal U(t) output by the DAC is routed to the analog switch of the i-th channel and output to the target nerve through the electrode, 2) To ensure the independence of the stimulation parameters of each channel, including pulse amplitude and pulse width, channel switching is realized through a multi-channel analog switch, and the channel selection signal is dynamically generated by the microcontroller MCU according to the time slot index iii, which satisfies: S2S1S0=Binary(i-1),i=1,2,...,8 (15) The on-off timing of the analog switch is controlled in combination with the timer interrupt to ensure that only the i-th channel is turned on within the time slot τ i where Binary(·) represents the binary representation, 3) Timing constraints of the voltage pulse signal: To ensure the effectiveness of stimulation, the single-channel pulse time t pulse must match the time slot width τ i . Where f is the stimulation frequency f = 1 to 30 Hz. 4) Multi-channel synchronous control, the strategy of which includes generating a high-precision clock signal through the MCU's built-in timer module, triggering the interrupt service routine (ISR) cyclically with a period of T, and performing the following operations in the interrupt service routine in sequence: S1 : Calculate current slot index S2: update the DAC output voltage where a i is the 8-bit DAC input value for the i-th channel; S3: pull high the analog switch enable signal of the corresponding channel, for a duration τ i ; S4: At τ i Switch is closed at the end of the time slot.

5. The closed-loop neural modulation system according to claim 4, characterized in that: The bipolar output employs an optocoupler-isolated H-bridge drive scheme, including: By dynamically switching the conduction state of the power switches, the current direction is controlled to periodically reverse between positive and negative polarities, thereby generating a bidirectional pulse current. The H-bridge consists of four switches Q1-Q4, and its conduction logic satisfies: Positive current stage: Q1 and Q4 are on, Q2 and Q3 are off, and the current path is positive power supply → Q1 → electrode → tissue → Q4 → negative power supply. Negative current stage: Q2 and Q3 are on, Q1 and Q4 are off, and the current path is positive terminal of power supply → Q3 → electrode → tissue → Q2 → negative terminal of power supply; Zero-current stage: Q1 and Q3 or Q2 and Q4 are turned off simultaneously, cutting off the current path. By switching the above conduction states at high frequency, a bidirectional pulse current I(t) is synthesized on the load electrode, and its waveform expression is as follows: where I + and I - are the positive / negative polarity current amplitudes, ω is the pulse frequency angular frequency, and sgn is the sign function.

6. A closed-loop neuromodulation method based on electroencephalogram detection, characterized in that include: The steps for acquiring electroencephalogram (EEG) signals are as follows: The electroencephalogram (EEG) signal analysis and detection step involves receiving the acquired EEG signals, analyzing and detecting them using machine learning methods, and predicting neurological disease information. The remote programming step is used to receive prediction results and user operation commands, programmatically modify stimulus parameters on the programmable controller, and issue modification instructions remotely. The neural stimulation output step is used to output electrical stimulation signals to achieve neural modulation. in: The steps for analyzing and detecting electroencephalogram (EEG) signals include: Dynamic noise suppression of the acquired EEG signals is performed using an adaptive filtering unit, including: feature extraction and recognition of the preprocessed EEG signals using a deep convolutional neural network algorithm; The machine learning unit comprises ten network layers: an initial feature mapping layer, a channel normalization layer, a feature enhancement layer, a deep convolutional layer, an activation function layer, a Dropout layer, a channel normalization layer, a global average pooling layer, a fully connected layer, and a Softmax activation layer, wherein: B1) The initial feature mapping layer adopts N1 one-dimensional convolution kernels to perform cross-channel feature reorganization on the input signal in the form of point-by-point convolution nodes, generating N1 equal-length feature subbands; in the input data matrix X (m,n) where m represents the original signal channel number, n is the time domain sampling point number, the i-th convolution kernel passes through an m-dimensional weight vector and a scalar bias b i (1) linear combination is performed on the input, and the expression is as follows: where i = 1, 2, 3, …, N1, m is the number of input signal channels, represents the output feature value of the i-th feature channel in the n-th layer network at the j-th time domain sampling point, x p,j is the j-th sampling point of the p-th channel of the input sample, is the scalar bias of the i-th convolution kernel, is the weight vector of the p-th channel of the i-th convolution kernel, B2) The channel normalization layer achieves dynamic equalization of cross-channel responses through statistical feature calibration, and normalizes the output of the previous layer: where: γ, β are scaling and translation parameters of the channel normalization layer; μ β , are the mean and variance of the mini-batch data, respectively; ∈ is a small constant to prevent division by zero errors, The output of the feature enhancement layer (B3) is: where K is the number of parallel branches of the multi-scale feature enhancement layer, a i,k is the attention weight of the kth branch of the ith output channel, is the weight parameter of the pth channel of the kth branch. B4) The deep convolutional layer is used to extract feature information from P-dimensional input data. Each new signal sample input to this layer is convolved with a certain deep convolutional node in the deep convolutional layer. The parameter set of the i-th deep convolutional node includes l weight coefficients and 1 bias parameter. The feature output is as follows: where p is the convolution kernel time domain width, R is the convolution kernel time domain offset, is the weight of the time domain offset r of the pth channel of the i th convolution kernel, is the depth convolution layer bias parameter B5) The activation function layer with parameterized correction optimizes feature representation by dynamically adjusting the activation threshold, and its output is: where a, b are the slope and threshold parameters of the parametric activation function, respectively, is the output of the deep convolutional layer, ReLU(·) represents a nonlinear activation function, and exp(·) represents the natural exponential function, B6) The Dropout layer suppresses overfitting through probabilistic feature masking, and its output is: Where p is the probability of neurons being dropped in the random deactivation layer. B7) The channel normalization layer achieves cross-channel response equalization through statistical feature calibration; its output is: where γ, β are the scaling and translation parameters of the channel normalization layer; μ β , are the mean and variance of the mini-batch data, respectively; ∈ is a numerical stability constant, B8) The global average pooling layer generates a fixed-length feature vector by reducing the spatial dimension. Its output is: where n is the time domain sequence length, is the global average pooling layer output, B9) The fully connected layer achieves feature mapping through affine transformation, and its output is: wherein is the fully connected layer weighted output, C is the number of input features of the fully connected layer, k is the output class serial number, is the fully connected layer weight parameter, is the fully connected layer bias parameter, B10) Finally, the Softmax activation layer outputs the final discrimination result through probability normalization: wherein is the Softmax layer output probability, c is the target class index (1≤c≤K), is the weighted input of the fully connected layer for class c.

7. The closed-loop neural modulation method according to claim 6, characterized in that: The EEG signal acquisition process uses a sampling rate of 1kHz, supports simultaneous acquisition from up to 32 channels, and performs signal preprocessing with 1-100Hz bandpass filtering. The remote-controlled process receives the prediction results of the EEG signal analysis and detection process via a wired connection and / or WIFI wireless communication, and receives user commands via a touchscreen user interface. The programmable controller selects stimulus parameters based on the prediction results and / or receives stimulus parameters specified by the user. The programmable controller connects remotely to the nerve stimulation output step via BLE communication, and issues modification commands to achieve remote programming of stimulation parameters.

8. The closed-loop neural modulation method according to claim 6, characterized in that: The adjustment of the stimulation parameters includes pulse amplitude, pulse width, pulse frequency, and waveform. in: The pulse amplitude range is 0.25-3.5mA, with an adjustment step of 0.05mA. The pulse width is 150-1000µs, and the adjustment step is 50µs. The pulse frequency is 1-30Hz, and the adjustment step is 1Hz. The pulse waveform is selected from at least one of square wave, triangle wave, and sine wave. The selection of stimulation parameters includes adjusting the charge density so that the total charge Q of the electrical stimulation can be calculated by integrating the stimulation current over the stimulation time and / or by calculating the product of the current amplitude and the pulse width.

9. The closed-loop neural modulation method according to claim 6, characterized in that: The neural stimulation output step uses BLE communication to receive programming instructions from the remotely programmed step, supports multi-channel coordinated stimulation, and supports bipolar output. The multi-channel coordinated stimulation uses a time-division multiplexing method to achieve multi-channel output from a single DAC signal source, supporting up to eight independent stimulation output channels. The time-division multiplexing method includes: time-division multiplexing technology, which dynamically allocates fixed time windows to periodically divide the output signal of a single digital-to-analog converter (DAC) into multiple independent time slots, each time slot corresponding to the signal loading and output of a stimulation channel, thereby achieving parallel processing of multi-channel signals in the time domain, including: 1) Set the parameters for the time-sharing multiplexing model, including: Let the total stimulation period of the stimulation system be TTT, and the maximum number of channels supported by the system be Nmax = 8. Divide the total period T into N equal durations, i = 1, 2, ..., N, satisfying: Each time slot τ i The stimulation signal loading window corresponding to an independent channel, in which the voltage pulse signal U(t) output by the DAC is routed to the analog switch of the i-th channel and output to the target nerve target point through the electrode, 2) To ensure the independence of stimulation parameters, including pulse amplitude and pulse width, for each channel, channel switching is achieved through a multiplexer analog switch. The channel selection signal is dynamically generated by the microcontroller (MCU) based on time slot index iii, satisfying the following: S2S1S0 = Binary(i - 1), i = 1, 2,..., 8 (15) Combining the on-off timing of the analog switch controlled by the timer interrupt, it is ensured that only the i-th channel is turned on in the time slot τ i and the i-th channel is turned off in the time slot τ 3) Timing constraints of the voltage pulse signal: To ensure the effectiveness of stimulation, the single-channel pulse time t pulse must match the time slot width τ i . Where f is the stimulation frequency f = 1 to 30 Hz. 4) Multi-channel synchronous control, the strategy of which includes generating a high-precision clock signal through the MCU's built-in timer module, triggering the interrupt service routine (ISR) cyclically with a period of T, and performing the following operations in the interrupt service routine in sequence: S1 : Calculate current slot index S2: update the DAC output voltage where a i is the 8-bit DAC input value for the i-th channel; S3: pull high the analog switch enable signal of the corresponding channel, for a duration τ i ; S4: At τ i Switch is closed at the end of the time slot.

10. The closed-loop neural modulation method according to claim 9, characterized in that: The bipolar output employs an optocoupler-isolated H-bridge drive scheme, including: By dynamically switching the conduction state of the power switches, the current direction is controlled to periodically reverse between positive and negative polarities, thereby generating a bidirectional pulse current. The H-bridge consists of four switches Q1-Q4, and its conduction logic satisfies: Positive current stage: Q1 and Q4 are on, Q2 and Q3 are off, and the current path is positive power supply → Q1 → electrode → tissue → Q4 → negative power supply. Negative current stage: Q2 and Q3 are on, Q1 and Q4 are off, and the current path is positive terminal of power supply → Q3 → electrode → tissue → Q2 → negative terminal of power supply; Zero-current stage: Q1 and Q3 or Q2 and Q4 are turned off simultaneously, cutting off the current path. By switching the above conduction states at high frequency, a bidirectional pulse current I(t) is synthesized on the load electrode, and its waveform expression is as follows: where I + and I - are the positive / negative polarity current amplitudes, ω is the pulse frequency angular frequency, and sgn is the sign function.

Images