A central-peripheral nerve regulation method and system based on a hybrid brain-computer interface and AI, an electronic device, and a storage medium

Abstract

This invention provides a method, system, electronic device, and storage medium for central-peripheral neural modulation based on a hybrid brain-computer interface and AI, belonging to the field of brain-computer interface technology. This method integrates four types of signals: EEG, fNIRS, EMG, and tactile pressure. It achieves dynamic weighted fusion through a cross-modal attention mechanism, introduces neurovascular coupling and cortical-muscle coherence as functional connectivity features, comprehensively reflecting the functional state of the central-peripheral system. Through domain adversarial training, it enables cross-individual generalization of motor intention decoding. Deep reinforcement learning is used to optimize electrical stimulation parameters online, simultaneously outputting multiple electrical stimuli to achieve a synergistic intervention effect between the central and peripheral systems. Through response effect evaluation and counterfactual reasoning optimization, adjustment suggestions are generated when the response effect is poor, improving the target subject's active participation and compliance. This invention fills a gap in the field of central-peripheral closed-loop modulation methods and has excellent application prospects.

Description

Technical Field

[0001] This invention relates to the field of brain-computer interface technology, specifically to a central-peripheral neural modulation method, system, electronic device, and storage medium based on a hybrid brain-computer interface and AI. Background Technology

[0002] As a crucial component of the core muscle group, the pelvic floor muscles directly impact an individual's urination, defecation control, and core stability. Decreased pelvic floor function or control is not only common in specific physiological stages (e.g., postpartum, old age) but also prevalent among sedentary, sedentary working professionals, directly leading to a decline in daily life quality and limited athletic performance. Therefore, developing intervention techniques to improve pelvic floor neuromuscular coordination and control has broad value in promoting health and enhancing function. Currently, peripheral interventions for pelvic floor function mainly include electromyographic biofeedback training, functional electrical stimulation, and Kegel exercises. These methods improve pelvic floor function by strengthening pelvic floor muscles and improving muscle coordination. However, the targets of these peripheral interventions are limited to the effector level, failing to directly mobilize the central nervous system's active control over the pelvic floor muscles. This hinders the effective reshaping and functional strengthening of neural pathways between the motor cortex and effectors, thus limiting their effectiveness in improving an individual's fine voluntary control over the pelvic floor muscles.

[0003] Current advancements in brain-computer interface (BCI) technology have provided new insights into neurological rehabilitation. By decoding motor intentions and translating them into control commands, BCIs can achieve closed-loop intervention between the central and peripheral nervous systems. Research indicates that BCIs based on motor imagery can induce cortical plasticity, and when combined with functional electrical stimulation (fNIRS), they can enhance motor function. Hybrid BCIs, by fusing multimodal signals, can achieve multidimensional perception of neuromuscular functional states. The fusion of electroencephalogram (EEG) signals and functional near-infrared spectroscopy (fNIRS) can effectively improve the decoding accuracy of motor imagery tasks. Cortical-muscle coherence analysis provides a powerful tool for quantitatively assessing central-peripheral functional connectivity.

[0004] Meanwhile, existing multimodal neuromodulation techniques offer the possibility of precise intervention at different targets: transcranial electrical stimulation (tES) can modulate cortical excitability and promote neural plasticity; taVNS can activate the basal forebrain cholinergic system through the nucleus tractus solitarius-locus coeruleus pathway, regulating neural function; and peripheral functional electrical stimulation (FES) can induce muscle contraction and promote muscle motor activation. The synergistic regulation of these three techniques can form a multi-level intervention effect of "cortical modulation-neuromodulation-muscle activation".

[0005] Existing typical neuromodulation schemes are as follows: Single-modal EEG is used to acquire EEG signals during motor imagery, or pelvic floor surface EMG is used to acquire muscle activity signals; μ / β band power or event-related desynchronization features are extracted from the EEG signals; integral electromyography, root mean square, and other temporal features are extracted from the EMG signals; traditional machine learning methods such as linear discriminant analysis or support vector machines are used to classify the motor imagery state and identify whether the user has a contraction intention; when a contraction intention is detected, pre-set fixed-parameter electrical stimulation is triggered to assist the user in completing pelvic floor muscle contraction; the contraction effect is evaluated through EMG or pressure probes, and feedback is provided to the user in visual or auditory form. This scheme has initially achieved intention-driven electrical stimulation, but it has the following limitations: it uses only a single modality of signal, resulting in incomplete information; the stimulation parameters are fixed and cannot be dynamically optimized according to the user's state; it lacks coordinated regulation between the central and peripheral systems; the model has poor generalization ability, and new users require long-term calibration. Currently, only a few studies have explored hybrid brain-computer interfaces that integrate EEG and fNIRS, but most remain at the offline analysis stage and have not achieved real-time closed-loop control. Some studies have also combined tES and FES, but adopted an open-loop mode and did not dynamically adjust according to intent. None of these studies have formed a complete central-peripheral closed-loop modulation method. Summary of the Invention

[0006] Purpose of the invention: The first purpose of this invention is to provide a central-peripheral neural modulation method based on hybrid brain-computer interface and AI to achieve central-peripheral coordinated closed-loop regulation. The second purpose is to provide a central-peripheral neural modulation system, electronic device and storage medium based on hybrid brain-computer interface and AI.

[0007] Technical solution: A central-peripheral neural modulation method based on hybrid brain-computer interface and AI, comprising the following steps:

[0008] S1. Simultaneously acquire the target object's electroencephalogram (EEG) signals, functional near-infrared spectral signals, pelvic floor electromyography (EMG) signals, and pelvic floor tactile pressure signals;

[0009] S2. Construct a multi-branch deep neural network to extract features from EEG signals, functional near-infrared spectroscopy signals, pelvic floor electromyography signals, and pelvic floor tactile pressure signals, respectively. Obtain fusion features through a cross-modal attention mechanism, calculate the neurovascular coupling strength and cortical-muscle coherence coefficient, and concatenate them with the fusion features to obtain a central-peripheral joint feature vector.

[0010] S3. Construct a motion intent decoder based on a temporal convolutional network and a multi-head self-attention mechanism. Input the central-peripheral joint feature vector sequence into the motion intent decoder to obtain the motion intent of the target object.

[0011] S4. Construct a biophysical model based on the neural mass model to simulate the neurodynamic process of the cortex-spinal cord-muscle pathway. Use a dual-delay deep deterministic strategy gradient algorithm to optimize the electrical stimulation parameters applied to the biophysical model. The state space of the dual-delay deep deterministic strategy gradient algorithm is composed of the central-peripheral joint feature vector, the confidence of motor intention decoding and the fatigue index of the target object. The action space is composed of the adjustment amount of the electrical stimulation parameters. By accumulating the adjustment amount, the optimized electrical stimulation parameters are obtained.

[0012] S5. Based on the optimized electrical stimulation parameters, design a timing coordination strategy to synchronously output multiple electrical stimulation control signals and dynamically modulate the electrical stimulation parameters.

[0013] S6. Construct a response prediction model, train it using historical response characteristics of electrical stimulation output, output expected response index, and trigger counterfactual inference when the actual response index is lower than the lower limit of the confidence interval of the expected response index to generate an electrical stimulation parameter adjustment plan.

[0014] Specifically, the multi-branch deep neural network includes: using a multi-scale temporal convolutional network to extract EEG signal features, constructing three parallel branches, including a high-frequency branch, a mid-frequency branch, and a low-frequency branch. The high-frequency branch is used to capture the rapid oscillation features of the EEG signal's γ band, obtaining 64-dimensional features through global average pooling. The mid-frequency branch is used to capture the event-related desynchronization features of the EEG signal's β band, obtaining 64-dimensional features through global average pooling. The low-frequency branch is used to capture the rhythmic activity features of the EEG signal's μ and α bands, obtaining 64-dimensional features through global average pooling. The outputs of the three parallel branches are concatenated and input into a two-layer fully connected network to obtain the deep feature vector of the EEG signal. The deep feature vector of the EEG signal is concatenated with the time-frequency features of the EEG signal to obtain the EEG signal feature vector.

[0015] A graph convolutional network is used to extract functional near-infrared spectral signal features. The graph structure is defined as G=(V,E), where V is the channel node of the functional near-infrared spectrum and E represents the adjacency matrix constructed based on the physical distance between the functional near-infrared spectrum channels. The initial features of each channel node are the statistics of the time series of oxyhemoglobin and deoxyhemoglobin. A node feature matrix is ​​constructed, and the node features are mapped to 64 dimensions through the first layer of graph convolution. The second layer of graph convolution obtains the functional near-infrared spectral signal graph feature vector through global average pooling.

[0016] A one-dimensional convolutional network is used to extract the features of pelvic floor electromyography (EMG) signals. The one-dimensional convolutional network contains four convolutional layers, each followed by max pooling. Finally, global average pooling is used to obtain the depth feature vector of pelvic floor EMG signals. The depth feature vector of pelvic floor EMG signals is concatenated with the basic features of pelvic floor EMG signals to obtain the feature vector of pelvic floor EMG signals.

[0017] A spatiotemporal convolutional network is used to extract the pelvic floor tactile pressure signal features. The information of all pressure sensing units is fused through spatial convolution, and the dynamic features of pressure are extracted through temporal convolution. After two spatial and temporal convolutions, the pelvic floor tactile pressure signal feature vector is obtained through global average pooling and a fully connected layer.

[0018] The formula for calculating the fusion feature is:

[0019]

[0020]

[0021]

[0022] In the formula: As a feature of fusion, Representing time t, The weighting coefficients are m=1,2,3,4, which correspond to the four modalities of electroencephalogram (EEG) signal, functional near-infrared spectroscopy (FIR) signal, pelvic floor electromyography (EMG) signal, and pelvic floor tactile pressure signal, respectively. It is a linear projection matrix. For feature vectors, To score attention, For the attention score of the k-th modality, , , , For learnable parameters, The task context vector is obtained by concatenating the intensity of the motion intent and the fatigue index of the previous moment.

[0023] Specifically, the neurovascular coupling strength is the correlation between the β-band power of the electroencephalogram (EEG) signal and the change in oxyhemoglobin concentration in the functional near-infrared spectral signal, calculated using the following formula:

[0024]

[0025] In the formula: The correlation coefficient is... Represents mutual information analysis, Represents the maximum value. This refers to the beta band power of the EEG signal. This refers to the concentration of oxyhemoglobin. τ represents time t, τ ranges from 0 to 5s, and the mutual information analysis step size is 0.1s;

[0026] The cortical-muscular coherence coefficient is the coherence coefficient between EEG signals and pelvic floor muscle electromyography (EMG) signals, and its calculation formula is as follows:

[0027]

[0028] In the formula: The coherence coefficient, Representing time t, Represents frequency, This represents the EEG signal after wavelet transform. This represents the electromyographic signal of the pelvic floor muscles after wavelet transform.

[0029] Specifically, the motion intent decoder includes a temporal convolutional block, a multi-head self-attention layer, a global average pooling layer, and a multi-task output layer. The temporal convolutional block contains three temporal convolutional layers with residual connections. The multi-head self-attention layer captures long-range dependencies in the central-peripheral joint feature vector sequence through residual connections and layer normalization. The global average pooling layer maps the sequence into a fixed-length feature vector. The multi-task output layer contains four output branches: an intent strength regression branch, a contraction pattern classification branch, an expected time classification branch, and a contraction strength grading branch. The intent strength regression branch outputs continuous values ​​in the range of 0-1, representing the intensity of the motion intent. The contraction pattern classification branch outputs the probabilities of three different contraction patterns, including rapid contraction, continuous contraction, and relaxation. The expected time classification branch outputs a binary probability of whether contraction will occur within a set future time scale. The contraction strength grading branch outputs the probability of contraction at different intensities. The joint loss function formula for the motion intent decoder is:

[0030]

[0031] In the formula: For the joint loss function, , , , For balance coefficient, The mean squared error loss corresponds to the intention intensity regression branch. The cross-entropy loss is the corresponding to the shrinking pattern classification branch. The cross-entropy loss is the binary classification loss corresponding to the classification branch at the expected time. This represents the cross-entropy loss corresponding to the contraction intensity grade branch.

[0032] Specifically, step S3 also includes: introducing a domain adversarial neural network based on Wasserstein distance to achieve cross-individual adaptation of the motion intent decoder. The domain adversarial neural network includes a feature extractor, an intent decoder, and a domain discriminator. The feature extractor consists of temporal convolutional blocks and multi-head attention layers, used to map the central-peripheral joint feature vector sequence to features f. The intent decoder consists of a multi-task output layer, used to predict motion intent based on the input features f. The domain discriminator consists of a three-layer fully connected network, which outputs domain label probabilities after inputting features f. It is trained using an adversarial strategy. The loss function of the domain discriminator is binary cross-entropy, and the training objective is to minimize the intent decoding loss and maximize the domain discriminant loss.

[0033] Specifically, the biophysical model includes cortical excitatory neuron groups, cortical inhibitory neuron groups, spinal alpha motor neuron groups, and a muscle unit model. The cortical excitatory neuron groups receive external input and thalamic input, projecting to the cortical inhibitory neuron groups and the spinal alpha motor neuron groups. The cortical excitatory neuron groups perform local feedback inhibition. The spinal alpha motor neuron groups receive input from the cortical excitatory neuron groups and output to the muscle unit model. The muscle unit model uses second-order differential equations to describe muscle contraction dynamics. The input to the muscle unit model is the motor neuron firing sequence, and the output is muscle contractile force and surface electromyography (EMG) signals. The neuron groups are connected through excitatory and inhibitory synaptic models. The postsynaptic potential is described using an alpha function. The output of the biophysical model includes local cortical field potentials, motor neuron firing sequences, muscle contractile force, and surface EMG signals.

[0034] Specifically, in the state space, the central-peripheral joint feature vector is compressed to 32 dimensions through an autoencoder; the confidence level of motion intention decoding is the inverse of the variance of the 3-dimensional motion intention intensity prediction; the fatigue index of the target object is the estimated rate of decline of the median frequency of the 1-dimensional surface electromyography signal; the electrical stimulation parameter adjustments include the intensity and frequency adjustments of transcranial electrical stimulation, vagal nerve electrical stimulation, and functional electrical stimulation; the reward function formula of the dual-delay deep deterministic policy gradient algorithm is:

[0035]

[0036] In the formula: For the reward function, , , , , , These are the weighting coefficients. This represents the change in the cortical-muscle coherence coefficient after stimulation compared to before stimulation. This represents the change in neurovascular coupling strength after stimulation compared to before stimulation. Actual contraction range With the target contraction range The ratio represents the contractile efficacy. The fatigue index. To stimulate changes in energy consumption, Changes in stimulus parameters The L2 norm.

[0037] Specifically, the temporal synergistic strategy includes four stimulation phases: a prediction phase, a pre-activation phase, a synergistic stimulation phase, and a recovery phase. In the prediction phase, steps S1 to S4 are executed to obtain optimized electrical stimulation parameters, but no stimulation is output. In the pre-activation phase, based on the output results of the classification branch at the expected time, vagal nerve electrical stimulation and transcranial electrical stimulation are output in advance. Vagal nerve electrical stimulation uses high-frequency pulses with a set duration, while transcranial electrical stimulation uses anodic stimulation. In the synergistic stimulation phase, functional electrical stimulation and transcranial electrical stimulation are output simultaneously. The current intensity of functional electrical stimulation is adjusted synchronously according to the intensity of the motor intention, while transcranial electrical stimulation is switched to a simulated theta short-array rapid pulse stimulation mode. In the recovery phase, stimulation is stopped.

[0038] Specifically, dynamic modulation of electrical stimulation parameters includes: using a proportional-integral controller to modulate the intensity of functional electrical stimulation, as shown in the following formula:

[0039]

[0040] In the formula: The intensity of the functional electrical stimulation current. For baseline intensity, For the intended decoding strength, This represents the actual contraction rate in the previous cycle. and This is a preset constant;

[0041] The vagal nerve electrical stimulation is dynamically adjusted based on the fatigue index, as shown in the following formula:

[0042]

[0043] In the formula: The intensity of the vagus nerve electrical stimulation current. And α are preset constants;

[0044] The electrode polarity configuration of transcranial electrical stimulation is dynamically optimized based on the activation mode of functional near-infrared spectroscopy. When ipsilateral cortical overactivation is detected, cathode stimulation is used to inhibit ipsilateral activation; when bilateral activation imbalance is detected, bilateral hemispheric balanced stimulation is used.

[0045] Specifically, step S6 includes:

[0046] A response prediction model was constructed, which is a time-series prediction model based on LSTM, including two LSTM layers, a fully connected layer, and an output layer. The response prediction model was trained using historical response features extracted during electrical stimulation output, and the expected response index was output, which is the pelvic floor muscle strength level. A structural causal model describing the electrical stimulation output process was constructed, including observable variables, latent variables, and outcome variables. The observable variables include electrical stimulation intensity, electrical stimulation frequency, electrode position, target subject compliance, and baseline function. The latent variables include neural plasticity response ability, electrical stimulation fit, and fatigue recovery ability. The outcome variable is the pelvic floor muscle strength level.

[0047] When the actual response index is lower than the lower limit of the confidence interval of the expected response index, counterfactual inference is triggered. Counterfactual inference includes: inferring the posterior distribution of latent variables based on observable and outcome variables; using variational inference methods, approximating the posterior distribution through an encoder network composed of three fully connected layers, and outputting the mean and variance of a Gaussian distribution; performing intervention operations on each observable variable, setting the observable variable as a candidate value in the structural causal model to cut off the influence of other variables on the observable variable; under the intervention conditions, predicting the posterior predicted distribution of the outcome variable based on the inferred posterior distribution of latent variables; obtaining a sample of the outcome variable by sampling latent variables and running the structural causal model; calculating the expected value of the outcome variable; calculating the counterfactual effect; selecting the observable variable with the largest counterfactual effect as the key constraint; generating an electrical stimulation parameter adjustment plan based on the key constraint; and adjusting the electrical stimulation parameters according to the electrical stimulation parameter adjustment plan.

[0048] This invention also provides a central-peripheral neural modulation system based on a hybrid brain-computer interface and AI, comprising:

[0049] Signal acquisition module: used to simultaneously acquire EEG signals, functional near-infrared spectral signals, pelvic floor electromyography signals, and pelvic floor tactile pressure signals of the target object;

[0050] Feature fusion module: used to construct a multi-branch deep neural network, extract features from EEG signals, functional near-infrared spectroscopy signals, pelvic floor electromyography signals and pelvic floor tactile pressure signals respectively, obtain fused features through cross-modal attention mechanism, calculate neurovascular coupling strength and cortical-muscle coherence coefficient, and concatenate with fused features to obtain central-peripheral joint feature vector;

[0051] Motion Intent Decoding Module: This module is used to construct a motion intent decoder based on a temporal convolutional network and a multi-head self-attention mechanism. It inputs the central-peripheral joint feature vector sequence into the motion intent decoder to obtain the motion intent of the target object.

[0052] Electrical stimulation parameter generation module: used to construct a biophysical model based on the neural mass model to simulate the neurodynamic process of the cortex-spinal cord-muscle pathway. It uses a dual-delay deep deterministic policy gradient algorithm to optimize the electrical stimulation parameters applied to the biophysical model. The state space of the dual-delay deep deterministic policy gradient algorithm is composed of the central-peripheral joint feature vector, the confidence of motor intention decoding, and the fatigue index of the target object. The action space is composed of the adjustment amount of the electrical stimulation parameters. By accumulating the adjustment amount, the optimized electrical stimulation parameters are obtained.

[0053] Electrical stimulation output modulation module: used to design a timing coordination strategy based on the optimized electrical stimulation parameters, synchronously output multiple electrical stimulation control signals, and dynamically modulate the electrical stimulation parameters;

[0054] Response assessment and adjustment module: This module is used to build a response prediction model. It is trained using historical response characteristics of electrical stimulation output and outputs expected response indicators. When the actual response indicator is lower than the lower limit of the confidence interval of the expected response indicator, counterfactual inference is triggered to generate an electrical stimulation parameter adjustment plan.

[0055] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the central-peripheral neural modulation method based on the above-described hybrid brain-computer interface and AI.

[0056] The present invention also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described central-peripheral neural modulation method based on hybrid brain-computer interface and AI.

[0057] Beneficial effects: Compared with the prior art, the significant effects of the present invention are:

[0058] 1. This invention simultaneously integrates four types of signals: EEG (high temporal resolution), fNIRS (high spatial resolution), EMG (muscle electrophysiology), and tactile pressure (muscle geometry). Through a cross-modal attention mechanism, it achieves dynamic weighted fusion, comprehensively reflecting the entire chain of information from "cortical activity to blood oxygen metabolism, muscle electrical activity, and muscle geometric deformation." Furthermore, it introduces neurovascular coupling and cortical-muscle coherence as functional connectivity features, enabling a more comprehensive assessment of the functional state of the central-peripheral nervous system.

[0059] 2. This invention enables the feature extractor to learn domain-invariant features through domain adversarial training. When facing new target objects, it can achieve higher decoding accuracy using shorter calibration data, significantly improving the applicability of the solution.

[0060] 3. This invention simultaneously outputs transcranial electrical stimulation (to modulate the cortex), vagus nerve electrical stimulation (to activate the nervous system), and pelvic floor functional electrical stimulation (to activate muscles), forming a multi-level intervention effect of "cortical modulation - neuromodulation - muscle activation." It also employs deep reinforcement learning to optimize electrical stimulation parameters online, using the strength of central-peripheral functional connectivity as the optimization target, achieving closed-loop adaptive regulation. Compared to traditional open-loop schemes that only stimulate the peripheral nervous system (FES), this invention simultaneously activates the cortex (tES), the neuromodulation system (taVNS), and muscles (FES), forming a bidirectional regulatory loop of "top-down" and "bottom-up," producing a synergistic effect of 1+1+1>3. Traditional schemes cannot adjust stimulation according to the real-time state of the target subject, while this invention, through deep reinforcement learning, optimizes parameters online with the strength of central-peripheral connectivity as the target, avoiding overstimulation and adaptation plateaus, thus improving intervention efficiency. Traditional schemes rely on human experience for adjustment, resulting in lag and subjectivity, while this invention can automatically attribute causes and adjust strategies when the effect is unsatisfactory.

[0061] 4. This invention decodes fine intent parameters and dynamically adjusts the intensity of functional electrical stimulation based on intent intensity errors to achieve a physiological synergistic stimulation mode, thereby enhancing the active participation and compliance of the target subject. Attached Figure Description

[0062] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention. Detailed Implementation

[0063] A preferred embodiment of the present invention will be further described below with reference to the accompanying drawings.

[0064] Example 1

[0065] Please see Figure 1 This embodiment provides a central-peripheral neural modulation method based on a hybrid brain-computer interface and AI, including the following steps:

[0066] S1. Simultaneously acquire the target object's electroencephalogram (EEG) signals, functional near-infrared spectral signals, pelvic floor electromyographic signals, and pelvic floor tactile pressure signals.

[0067] In this embodiment, four modalities of data—electroencephalography (EEG), functional near-infrared spectroscopy (fNIRS), pelvic floor electromyography (EMG), and pelvic floor tactile pressure signals—are first acquired simultaneously. Time alignment is achieved using a synchronization trigger to ensure that the timestamp error between different modalities is less than 1 ms. EEG signals are acquired according to the international 10-10 system layout, covering brain regions related to pelvic floor muscle control, such as the Cz region. fNIRS signal acquisition covers the primary motor cortex, supplementary motor area, and prefrontal cortex. EMG signals are acquired via vaginal or rectal electrodes. Pelvic floor tactile pressure signals are obtained in real-time by acquiring the pressure distribution generated by pelvic floor muscle contractions using a flexible tactile pressure sensor array placed on the surface of the pelvic floor rehabilitation probe.

[0068] S2. Construct a multi-branch deep neural network to extract features from EEG signals, functional near-infrared spectroscopy signals, pelvic floor electromyography signals, and pelvic floor tactile pressure signals. Obtain fusion features through a cross-modal attention mechanism, calculate the neurovascular coupling strength and cortical-muscle coherence coefficient, and concatenate them with the fusion features to obtain a central-peripheral joint feature vector.

[0069] Multi-branch deep neural networks specifically include:

[0070] A multi-scale temporal convolutional network was used to extract EEG signal features. Three parallel branches were constructed: a high-frequency branch, a mid-frequency branch, and a low-frequency branch. The high-frequency branch was used to capture the rapid oscillation features of the EEG signal in the γ band (30-100Hz), and a 64-dimensional feature was obtained through global average pooling. The mid-frequency branch was used to capture the event-related desynchronization (ERD) features of the EEG signal in the β band (13-30Hz), and a 64-dimensional feature was obtained through global average pooling. The low-frequency branch was used to capture the rhythmic activity features of the EEG signal in the μ band and α band (8-13Hz), and a 64-dimensional feature was obtained through global average pooling. The outputs of the three parallel branches were concatenated and input into a two-layer fully connected network to obtain the EEG signal depth feature vector. At the same time, traditional EEG signal time-frequency features were extracted as a supplement, including the power spectral density of each band, the area under the ERD / ERS curve, and the lateralization coefficient. The EEG signal depth feature vector was concatenated with the EEG signal time-frequency features to obtain the EEG signal feature vector. The final EEG signal feature vector has a dimension of 256.

[0071] Graph Convolutional Networks (GCNs) are used to extract functional near-infrared spectral signal features. The graph structure is defined as G=(V,E), where V represents the channel nodes of the functional near-infrared spectrum, and E represents the adjacency matrix constructed based on the physical distances between the functional near-infrared spectral channels. Specifically, if the distance between two channels is less than 4 cm, an edge connection exists, with the edge weight being the reciprocal of the distance. The initial features of each channel node are statistical measures of the time series of oxyhemoglobin and deoxyhemoglobin, including mean, slope, peak value, area under the curve, and average activation intensity within the task-related time period. A node feature matrix is ​​constructed. Taking 24 channels as an example, there are 24 nodes in total, with each node having a feature dimension of 10. The node feature matrix can then be represented as: X∈R 24×10 The GCN structure contains two layers of graph convolution. The first layer of graph convolution maps node features to 64 dimensions, and the second layer of graph convolution obtains the functional near-infrared spectral signal map feature vector through global average pooling. The feature vector of the functional near-infrared spectral signal map has a dimension of 128.

[0072] A one-dimensional convolutional network was used to extract pelvic floor electromyography (EMG) signal features. The one-dimensional convolutional network contained four convolutional layers with kernel sizes of 64, 32, 16, and 8, corresponding to the time lengths, respectively. The number of kernels were 16, 32, 64, and 128, respectively. Each convolutional layer was followed by max pooling, and finally, global average pooling was used to obtain the depth feature vector of the pelvic floor EMG signal. The depth feature vector of the pelvic floor EMG signal had a dimension of 128. The depth feature vector of the pelvic floor EMG signal was concatenated with the basic features of the pelvic floor EMG signal to obtain the feature vector of the pelvic floor EMG signal with a dimension of 256.

[0073] A spatiotemporal convolutional network is used to extract pelvic floor tactile pressure signal features. Spatial convolution fuses information from all pressure sensing units, and temporal convolution extracts dynamic pressure features. The spatiotemporal convolutional network consists of two spatiotemporal convolutional blocks. Each spatiotemporal convolutional block contains spatial convolution + temporal convolution + batch normalization + ReLU activation function. After two spatial and temporal convolutions, global average pooling and fully connected layers are used to obtain the pelvic floor tactile pressure signal feature vector. The pelvic floor tactile pressure signal feature vector has a dimension of 64.

[0074] The fusion weights are adaptively generated based on the relevance of each modal signal to the current task. The formula for calculating the fusion features is as follows:

[0075]

[0076]

[0077]

[0078] In the formula: As a feature of fusion, Representing time t, The weighting coefficients are m=1,2,3,4, which correspond to the four modalities of electroencephalogram (EEG) signal, functional near-infrared spectroscopy (FIR) signal, pelvic floor electromyography (EMG) signal, and pelvic floor tactile pressure signal, respectively. A linear projection matrix projects features onto the same dimension. For feature vectors, To score attention, For the attention score of the k-th modality, , , , For learnable parameters, The task context vector is obtained by concatenating the intensity of the motion intent and the fatigue index of the previous moment.

[0079] This invention also introduces functional connectivity feature calculation, specifically by calculating the neurovascular coupling strength and the cortical-muscle coherence coefficient to construct a central-peripheral joint feature vector.

[0080] The neurovascular coupling strength is the correlation between the β-band power of the electroencephalogram (EEG) signal and the change in oxyhemoglobin concentration in the functional near-infrared spectral signal. The calculation formula is as follows:

[0081]

[0082] In the formula: The correlation coefficient is... Represents mutual information analysis, Represents the maximum value. This refers to the beta band power of the EEG signal. This refers to the concentration of oxyhemoglobin. Let t represent time t, τ range from 0 to 5s, and the mutual information analysis step size is 0.1s.

[0083] The cortical-muscular coherence coefficient is the coherence coefficient between EEG signals and pelvic floor muscle electromyography (EMG) signals, and its calculation formula is as follows:

[0084]

[0085] In the formula: The coherence coefficient, Representing time t, Represents frequency, This represents the EEG signal after wavelet transform. This represents the electrophysiological signal of the pelvic floor muscles after wavelet transform. This embodiment focuses on calculating the coherence peak (CMC) of the β-band related to pelvic floor muscle contraction. β Coherence peak CMC of the γ band γ .

[0086] The result CMC β CMC γ and The features are spliced ​​together to form the final central-peripheral joint feature vector. ∈R 256+3 .

[0087] S3. Construct a motion intent decoder based on a temporal convolutional network and a multi-head self-attention mechanism. Input the central-peripheral joint feature vector sequence into the motion intent decoder to obtain the motion intent of the target object.

[0088] In this embodiment, the motion intent decoder includes a temporal convolutional block, a multi-head self-attention layer, a global average pooling layer, and a multi-task output layer. The input is a feature sequence of the central-peripheral joint feature vector within a sliding window. The temporal convolutional block contains three residually connected temporal convolutional layers with dilation rates of 1, 2, and 4, and a kernel size of 3. Each layer includes batch normalization and ReLU activation, with a dropout rate of 0.2. The multi-head self-attention layer captures long-range dependencies in the central-peripheral joint feature vector sequence through residual connections and layer normalization. The multi-head self-attention layer contains 8 heads, each with a dimension of 64. A global average pooling layer maps the variable-length sequence to a fixed-length feature vector. The multi-task output layer contains four output branches: an intent intensity regression branch, a contraction pattern classification branch, an expected time classification branch, and a contraction intensity grading branch. The intent intensity regression branch outputs continuous values ​​in the range of 0-1, representing the intensity of motion intent. The contraction pattern classification branch outputs the probabilities of three different contraction patterns, including rapid contraction, continuous contraction, and relaxation. The expected time classification branch outputs a binary classification probability of whether contraction will occur within a set future time scale (e.g., 0.5s). The contraction intensity grading branch outputs the probability of contraction at different intensities (e.g., light / medium / heavy). The joint loss function formula for the motion intent decoder is:

[0089]

[0090] In the formula: For the joint loss function, , , , As a balance coefficient, in this embodiment, it is taken as 1, 0.5, 0.5, and 0.5 respectively. The mean squared error loss corresponds to the intention intensity regression branch. The cross-entropy loss is the corresponding to the shrinking pattern classification branch. The cross-entropy loss is the binary classification loss corresponding to the classification branch at the expected time. This represents the cross-entropy loss corresponding to the contraction intensity grade branch.

[0091] To achieve cross-individual adaptation of the motion intent decoder, this embodiment introduces a domain adversarial neural network based on Wasserstein distance. The domain adversarial neural network includes a feature extractor G. f Intent decoder G y Domain Discriminator G d Feature extractor G f It consists of temporal convolutional blocks and multi-head attention layers, used to process the central-peripheral joint feature vector sequence. Mapped to feature f=G f (X), Intent decoder G y It consists of a multi-task output layer, used to predict motion intention y=G based on the input features f. y (f) Domain discriminator G d It consists of a three-layer fully connected network, which takes the feature f as input and outputs the domain label probability d=G. d (f), where d=0 represents the source domain, i.e., the existing dataset, and d=1 represents the target domain, i.e., the current target object. An adversarial strategy is used for training, specifically the feature extractor G. f Attempting to extract information that can deceive the domain discriminator G d The characteristics of the domain discriminator G d The goal is to accurately distinguish the source of feature f. During forward propagation, the gradient inversion layer (GRL) directly passes feature f to the domain discriminator G. d During backpropagation, GRL will use the domain discriminator G d The gradient is multiplied by -λ (λ is a tradeoff parameter) and then passed to the feature extractor. The loss function of the domain discriminator is the binary cross-entropy. The formula is:

[0092]

[0093] in For the domain discriminator G d The output of represents the probability that the discriminator determines that the input feature f belongs to the target domain.

[0094] The training objective of a domain adversarial neural network is to minimize the intent decoding loss. Maximize the domain discrimination loss, therefore the overall training objective loss function Represented as:

[0095]

[0096] S4. Construct a biophysical model based on the neural mass model to simulate the neurodynamic process of the cortex-spinal cord-muscle pathway. Use a dual-delay deep deterministic policy gradient algorithm to optimize various electrical stimulation parameters applied to the biophysical model. The state space of the dual-delay deep deterministic policy gradient algorithm is composed of the central-peripheral joint feature vector, the confidence of motor intention decoding, and the fatigue index of the target object. The action space is composed of the adjustment amount of the electrical stimulation parameters. By accumulating the adjustment amount, the optimized electrical stimulation parameters are obtained.

[0097] The biophysical model comprises cortical excitatory neuronal groups, cortical inhibitory neuronal groups, spinal alpha motor neuron groups, and a myounit model. The cortical excitatory neuronal groups receive external and thalamic inputs and project to the cortical inhibitory neuronal groups and the spinal alpha motor neuron groups. The cortical excitatory neuronal groups perform local feedback inhibition. The spinal alpha motor neuron groups receive inputs from the cortical excitatory neuronal groups and output to the myounit model. The myounit model uses second-order differential equations to describe muscle contraction dynamics. The input to the myounit model is the firing sequence of motor neurons, and the output is muscle contractile force and surface electromyography (EMG) signals. The neuronal groups are connected through excitatory and inhibitory synaptic models. Postsynaptic potentials are described using alpha functions. The outputs of the biophysical model include local cortical field potentials (corresponding to EEG), motor neuron firing sequences, muscle contractile force (corresponding to tactile pressure), and surface EMG signals. Using resting-state and task-state data of the target object, such as resting-state EEG power spectrum, task-state ERD, and EMG contraction curves, the parameters of the biophysical model are calibrated. Individualized parameters, such as synaptic connection strength, time constant, and threshold, are estimated using particle swarm optimization algorithm. The optimization objective is to minimize the mean square error between the model output and the real data.

[0098] For the dual-delay deep deterministic policy gradient algorithm (TD3), in its state space, the central-peripheral joint feature vector is compressed to 32 dimensions by an autoencoder. The confidence of motion intention decoding is the inverse of the variance of the 3-dimensional motion intention intensity prediction. The fatigue index of the target object is the estimated rate of decline of the median frequency of the 1-dimensional surface electromyography signal. The electrical stimulation parameter adjustment includes the intensity and frequency adjustment of transcranial electrical stimulation, vagal nerve electrical stimulation, and functional electrical stimulation. Therefore, the action space consists of a 6-dimensional continuous vector. The actual electrical stimulation parameters are obtained by accumulating the adjustment and are limited to a safe range.

[0099] The reward function formula for the dual-delay deep deterministic policy gradient algorithm is:

[0100]

[0101] In the formula: For the reward function, , , , , , These are weighting coefficients, which have been normalized to ensure that the contributions of each item are equivalent. This represents the change in the cortical-muscle coherence coefficient after stimulation compared to before stimulation. This represents the change in neurovascular coupling strength after stimulation compared to before stimulation. Actual contraction range With the target contraction range The ratio represents the contractile efficacy. The fatigue index. To stimulate changes in energy consumption, Changes in stimulus parameters The L2 norm.

[0102] S5. Based on the optimized electrical stimulation parameters, design a timing coordination strategy to synchronously output multiple electrical stimulation control signals and dynamically modulate the electrical stimulation parameters.

[0103] In this embodiment, the temporal synergistic strategy includes four stimulation phases: a prediction phase, a pre-activation phase, a synergistic stimulation phase, and a recovery phase. In the prediction phase, steps S1 to S4 are executed, acquiring and processing four modal signals, decoding the motor intent, obtaining optimized electrical stimulation parameters, and no stimulation is output. In the pre-activation phase, based on the output results of the classification branches at the expected time, vagal nerve electrical stimulation and transcranial electrical stimulation are output in advance. Vagal nerve electrical stimulation uses high-frequency pulses (greater than 20Hz) for a set duration to activate the ascending cholinergic system and enhance cortical excitability. Transcranial electrical stimulation uses anodic stimulation to enhance the excitability of the pelvic floor representation area. In the synergistic stimulation phase, functional electrical stimulation and transcranial electrical stimulation are output simultaneously. The functional electrical stimulation adjusts the current intensity synchronously according to the intensity of the motor intent, achieving physiological synergy where stronger intent leads to stronger stimulation. Transcranial electrical stimulation is switched to a simulated theta short-array rapid pulse stimulation mode to simulate the physiological theta rhythm. In the recovery phase, stimulation is stopped, feedback signals are acquired, and the effect of electrical stimulation is further evaluated.

[0104] Dynamic modulation of electrical stimulation parameters includes: using a proportional-integral controller to modulate the intensity of functional electrical stimulation, as shown in the following formula:

[0105]

[0106] In the formula: The intensity of the functional electrical stimulation current. Baseline intensity, For the intended decoding strength, This represents the actual contraction rate in the previous cycle. and This is a preset constant;

[0107] The vagal nerve electrical stimulation is dynamically adjusted based on the fatigue index, as shown in the following formula:

[0108]

[0109] In the formula: The intensity of the vagus nerve electrical stimulation current. And α are preset constants;

[0110] The electrode polarity configuration of transcranial electrical stimulation is dynamically optimized based on the activation mode of functional near-infrared spectroscopy. When ipsilateral cortical overactivation is detected, cathode stimulation is used to inhibit ipsilateral activation. When bilateral activation imbalance is detected, bihemispheric balanced stimulation is used. The polarity configuration is updated every 30 seconds.

[0111] During electrical stimulation, safety indicators such as stimulation current intensity, skin impedance, heart rate variability, and pain feedback are monitored in real time. Stimulation is automatically terminated when safety indicators exceed limits.

[0112] S6. Construct a response prediction model, train it using historical response characteristics of electrical stimulation output, output expected response index, and trigger counterfactual inference when the actual response index is lower than the lower limit of the confidence interval of the expected response index to generate an electrical stimulation parameter adjustment plan.

[0113] In this embodiment, a response prediction model is first constructed. This model is an LSTM-based temporal prediction model, consisting of two LSTM layers, a fully connected layer, and an output layer. The response prediction model is trained using historical features extracted during electrical stimulation output. These historical response features include pelvic floor muscle strength levels, ICI-Q-SF scores, and CMC scores. β Peak value, NVC peak value, and compliance score were calculated, and expected response indicators were output. The rehabilitation indicator was pelvic floor muscle strength grade. A structural causal model describing the electrical stimulation output process was constructed, including observable variables, latent variables, and outcome variables. The observable variables included electrical stimulation intensity. Electrical stimulation frequency Electrode position Target audience compliance and baseline function Latent variables include neural plasticity response capacity. Electrical stimulation fit and fatigue recovery ability The outcome variable is the pelvic floor muscle strength grade. .

[0114] The equation model of the structural causal model is as follows:

[0115]

[0116]

[0117]

[0118] In the formula: As a latent variable, , For observable variables, , For exogenous noise, Represents the set of parent nodes, function , and The parameters are obtained by using a neural network and learning from historical data.

[0119] When the actual response indicators When the response falls below the lower limit of the confidence interval of the expected response indicator, counterfactual reasoning is triggered. Counterfactual reasoning includes:

[0120] Based on observable and outcome variables, the posterior distribution of latent variables is inferred. A variational inference method is employed, using an encoder network composed of three fully connected layers to approximate the posterior distribution, outputting the mean and variance of a Gaussian distribution. Intervention operations are performed on each observable variable. In the structural causal model, the observable variables... Set as an alternative value (For example, increasing the stimulus intensity by 10%), cutting off the influence of other variables on the observable variable, i.e., implementing... Under intervention conditions, the posterior distribution of the outcome variable is predicted by combining the inferred posterior distribution of the latent variables. A sample of the outcome variable is obtained by sampling the latent variables and running a structural causal model. The expected value of the outcome variable is calculated, and the counterfactual effect is also calculated. The formula is:

[0121]

[0122] The observable variable with the largest counterfactual effect is selected as the key constraint. If the counterfactual effect values ​​of multiple observable variables are similar, the key constraint with the lowest intervention cost is selected, such as adjusting the stimulus intensity, which is easier to implement than adjusting the electrode position.

[0123] Counterfactual reasoning first identifies key constraints, then converts these constraints into corresponding text templates based on a predefined mapping knowledge base, ultimately outputting structured suggestion text. For example, if the key constraint is electrical stimulation intensity, the output would be: "Stimulation intensity may be insufficient; it is recommended to increase the stimulation current." If the key constraint is electrode position, the output would be: "The electrodes may be misaligned." If the key constraint is compliance, the output would be: "The target's compliance is low; it is recommended to increase incentives."

[0124] Strategy adjustment suggestions are presented to operators through an interactive interface. After operator confirmation, the system automatically adjusts the corresponding parameters. The adjusted parameters serve as the new base values ​​for subsequent reinforcement learning optimization. Simultaneously, the results of counterfactual reasoning, such as key factors, intervention plans, and actual effects, are stored in a knowledge base for continuous optimization of the causal model, completing the closed-loop operation of the strategy and method.

[0125] In summary, the central-peripheral nerve modulation method provided by this invention, through motor intention decoding and multi-target synergistic neural modulation, can achieve active closed-loop enhancement of pelvic floor neuromuscular function. This method can also serve as an efficient training tool for individuals with normal pelvic floor function, improving the fine control of pelvic floor muscles, enhancing core stability, and preventing the natural decline of pelvic floor function due to aging or unhealthy lifestyle habits, thus possessing universal application value.

[0126] Example 2

[0127] This embodiment first provides a central-peripheral neural modulation system based on a hybrid brain-computer interface and AI, corresponding to the central-peripheral neural modulation method based on a hybrid brain-computer interface and AI described in Embodiment 1, including the following modules:

[0128] Signal acquisition module: used to simultaneously acquire EEG signals, functional near-infrared spectral signals, pelvic floor electromyography signals, and pelvic floor tactile pressure signals of the target object;

[0129] Feature fusion module: used to construct a multi-branch deep neural network, extract features from EEG signals, functional near-infrared spectroscopy signals, pelvic floor electromyography signals and pelvic floor tactile pressure signals respectively, obtain fused features through cross-modal attention mechanism, calculate neurovascular coupling strength and cortical-muscle coherence coefficient, and concatenate with fused features to obtain central-peripheral joint feature vector;

[0130] Motion Intent Decoding Module: This module is used to construct a motion intent decoder based on a temporal convolutional network and a multi-head self-attention mechanism. It inputs the central-peripheral joint feature vector sequence into the motion intent decoder to obtain the motion intent of the target object.

[0131] Electrical stimulation parameter generation module: used to construct a biophysical model based on the neural mass model to simulate the neurodynamic process of the cortex-spinal cord-muscle pathway. It uses a dual-delay deep deterministic policy gradient algorithm to optimize various electrical stimulation parameters applied to the biophysical model. The state space of the dual-delay deep deterministic policy gradient algorithm is composed of the central-peripheral joint feature vector, the confidence of motor intention decoding, and the fatigue index of the target object. The action space is composed of the adjustment amount of the electrical stimulation parameters. By accumulating the adjustment amount, the optimized electrical stimulation parameters are obtained.

[0132] Electrical stimulation output modulation module: used to design a timing coordination strategy based on the optimized electrical stimulation parameters, synchronously output multiple electrical stimuli, dynamically modulate the electrical stimulation parameters, and complete electrical stimulation rehabilitation training;

[0133] The response assessment and adjustment module is used to build a response prediction model. It is trained using historical characteristics of electrical stimulation rehabilitation training and outputs expected response indicators. When the actual response indicator is lower than the lower limit of the confidence interval of the expected response indicator, counterfactual inference is triggered to generate electrical stimulation rehabilitation training adjustment suggestions. The electrical stimulation parameters are adjusted according to the electrical stimulation rehabilitation training adjustment suggestions.

[0134] This embodiment also provides an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the central-peripheral neural modulation method based on hybrid brain-computer interface and AI as described in Embodiment 1.

[0135] This embodiment also provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the central-peripheral neural modulation method based on hybrid brain-computer interface and AI as described in Embodiment 1.

Claims

1. A central-peripheral neural modulation method based on hybrid brain-computer interface and AI, characterized in that, Includes the following steps: S1. Simultaneously acquire the target object's electroencephalogram (EEG) signals, functional near-infrared spectral signals, pelvic floor electromyography (EMG) signals, and pelvic floor tactile pressure signals; S2. Construct a multi-branch deep neural network to extract features from EEG signals, functional near-infrared spectroscopy signals, pelvic floor electromyography signals, and pelvic floor tactile pressure signals, respectively. Obtain fusion features through a cross-modal attention mechanism, calculate the neurovascular coupling strength and cortical-muscle coherence coefficient, and concatenate them with the fusion features to obtain a central-peripheral joint feature vector. S3. Construct a motion intent decoder based on a temporal convolutional network and a multi-head self-attention mechanism. Input the central-peripheral joint feature vector sequence into the motion intent decoder to obtain the motion intent of the target object. S4. Construct a biophysical model based on the neural mass model to simulate the neurodynamic process of the cortex-spinal cord-muscle pathway. Use a dual-delay deep deterministic strategy gradient algorithm to optimize the electrical stimulation parameters applied to the biophysical model. The state space of the dual-delay deep deterministic strategy gradient algorithm is composed of the central-peripheral joint feature vector, the confidence of motor intention decoding and the fatigue index of the target object. The action space is composed of the adjustment amount of the electrical stimulation parameters. By accumulating the adjustment amount, the optimized electrical stimulation parameters are obtained. S5. Based on the optimized electrical stimulation parameters, design a timing coordination strategy to synchronously output multiple electrical stimulation control signals and dynamically modulate the electrical stimulation parameters. S6. Construct a response prediction model, train it using historical response characteristics of electrical stimulation output, output expected response index, and trigger counterfactual inference when the actual response index is lower than the lower limit of the confidence interval of the expected response index to generate an electrical stimulation parameter adjustment plan.

2. The central-peripheral neural modulation method based on hybrid brain-computer interface and AI according to claim 1, characterized in that, The multi-branch deep neural network includes: extracting EEG signal features using a multi-scale temporal convolutional network; constructing three parallel branches, including a high-frequency branch, a mid-frequency branch, and a low-frequency branch; the high-frequency branch is used to capture the rapid oscillation features of the γ band of the EEG signal, and obtains 64-dimensional features through global average pooling; the mid-frequency branch is used to capture the event-related desynchronization features of the β band of the EEG signal, and obtains 64-dimensional features through global average pooling; the low-frequency branch is used to capture the rhythmic activity features of the μ and α bands of the EEG signal, and obtains 64-dimensional features through global average pooling; the outputs of the three parallel branches are concatenated and input into a two-layer fully connected network to obtain a deep feature vector of the EEG signal; the deep feature vector of the EEG signal is concatenated with the time-frequency features of the EEG signal to obtain a feature vector of the EEG signal. A graph convolutional network is used to extract functional near-infrared spectral signal features. The graph structure is defined as G=(V,E), where V is the channel node of the functional near-infrared spectrum and E represents the adjacency matrix constructed based on the physical distance between the functional near-infrared spectrum channels. The initial features of each channel node are the statistics of the time series of oxyhemoglobin and deoxyhemoglobin. A node feature matrix is ​​constructed, and the node features are mapped to 64 dimensions through the first layer of graph convolution. The second layer of graph convolution obtains the functional near-infrared spectral signal graph feature vector through global average pooling. A one-dimensional convolutional network is used to extract the features of pelvic floor electromyography (EMG) signals. The one-dimensional convolutional network contains four convolutional layers, each followed by max pooling. Finally, global average pooling is used to obtain the depth feature vector of pelvic floor EMG signals. The depth feature vector of pelvic floor EMG signals is concatenated with the basic features of pelvic floor EMG signals to obtain the feature vector of pelvic floor EMG signals. A spatiotemporal convolutional network is used to extract the pelvic floor tactile pressure signal features. The information of all pressure sensing units is fused through spatial convolution, and the dynamic features of pressure are extracted through temporal convolution. After two spatial and temporal convolutions, the pelvic floor tactile pressure signal feature vector is obtained through global average pooling and a fully connected layer. The formula for calculating the fusion feature is as follows: In the formula: As a feature of fusion, Representing time t, The weighting coefficients are m=1,2,3,4, which correspond to the four modalities of electroencephalogram (EEG) signal, functional near-infrared spectroscopy signal, pelvic floor electromyography (EMG) signal, and pelvic floor tactile pressure signal, respectively. It is a linear projection matrix. For feature vectors, To score attention, For the attention score of the k-th modality, , , , For learnable parameters, The task context vector is obtained by concatenating the intensity of the motion intent and the fatigue index of the previous moment.

3. The central-peripheral neural modulation method based on hybrid brain-computer interface and AI according to claim 2, characterized in that, The neurovascular coupling strength is the correlation between the β-band power of the electroencephalogram (EEG) signal and the change in oxyhemoglobin concentration in the functional near-infrared spectral signal, and the calculation formula is as follows: In the formula: The correlation coefficient is... Represents mutual information analysis, Represents the maximum value. This refers to the beta band power of the EEG signal. This refers to the concentration of oxyhemoglobin. τ represents time t, τ ranges from 0 to 5s, and the mutual information analysis step size is 0.1s; The cortical-muscular coherence coefficient is the coherence coefficient between the electroencephalogram (EEG) signal and the pelvic floor muscle EEG signal, and its calculation formula is as follows: In the formula: The coherence coefficient, Representing time t, Represents frequency, This represents the EEG signal after wavelet transform. This represents the pelvic floor electromyographic signal after wavelet transform. represent . conjugate.

4. The central-peripheral neural modulation method based on hybrid brain-computer interface and AI according to claim 3, characterized in that, The motion intent decoder includes a temporal convolutional block, a multi-head self-attention layer, a global average pooling layer, and a multi-task output layer. The temporal convolutional block contains three residually connected temporal convolutional layers. The multi-head self-attention layer captures long-range dependencies in the central-peripheral joint feature vector sequence through residual connections and layer normalization. The global average pooling layer maps the sequence to a fixed-length feature vector. The multi-task output layer contains four output branches: an intent strength regression branch, a contraction pattern classification branch, an expected time classification branch, and a contraction strength grading branch. The intent strength regression branch outputs continuous values ​​in the range of 0-1, representing the intensity of the motion intent. The contraction pattern classification branch outputs probabilities of three different contraction patterns: rapid contraction, continuous contraction, and relaxation. The expected time classification branch outputs a binary probability of whether contraction will occur within a set future time scale. The contraction strength grading branch outputs contraction probabilities of different intensities. The joint loss function formula for the motion intent decoder is: In the formula: For the joint loss function, , , , For balance coefficient, The mean squared error loss corresponds to the intention intensity regression branch. The cross-entropy loss is the corresponding to the shrinking pattern classification branch. The cross-entropy loss is the binary classification loss corresponding to the classification branch at the expected time. This represents the cross-entropy loss corresponding to the contraction intensity grade branch.

5. The central-peripheral neural modulation method based on hybrid brain-computer interface and AI according to claim 4, characterized in that, Step S3 further includes: introducing a domain adversarial neural network based on Wasserstein distance to achieve cross-individual adaptation of the motion intent decoder. The domain adversarial neural network includes a feature extractor, an intent decoder, and a domain discriminator. The feature extractor consists of temporal convolutional blocks and multi-head attention layers, used to map the central-peripheral joint feature vector sequence to features f. The intent decoder consists of a multi-task output layer, used to predict motion intent based on the input features f. The domain discriminator consists of a three-layer fully connected network, which outputs domain label probabilities after inputting features f. It is trained using an adversarial strategy. The loss function of the domain discriminator is binary cross-entropy, and the training objective is to minimize the intent decoding loss and maximize the domain discriminant loss.

6. The central-peripheral neural modulation method based on hybrid brain-computer interface and AI according to claim 5, characterized in that, The biophysical model includes a cortical excitatory neuron group, a cortical inhibitory neuron group, a spinal alpha motor neuron group, and a muscle unit model. The cortical excitatory neuron group receives external input and thalamic input, projecting to the cortical inhibitory neuron group and the spinal alpha motor neuron group. The cortical excitatory neuron group performs local feedback inhibition. The spinal alpha motor neuron group receives input from the cortical excitatory neuron group and outputs to the muscle unit model. The muscle unit model uses a second-order differential equation to describe muscle contraction dynamics. The input of the muscle unit model is the motor neuron firing sequence, and the output is muscle contractile force and surface electromyography (EMG) signal. The neuron groups are connected through excitatory and inhibitory synaptic models. The postsynaptic potential is described using an alpha function. The output of the biophysical model includes local cortical field potentials, motor neuron firing sequences, muscle contractile force, and surface EMG signal.

7. The central-peripheral neural modulation method based on hybrid brain-computer interface and AI according to claim 6, characterized in that, In the state space, the central-peripheral joint feature vector is compressed to 32 dimensions by an autoencoder; the confidence level of the motion intention decoding is the inverse of the variance of the 3-dimensional motion intention intensity prediction; the fatigue index of the target object is the estimated rate of decline of the median frequency of the 1-dimensional surface electromyography signal; the electrical stimulation parameter adjustment includes the intensity and frequency adjustment of transcranial electrical stimulation, vagus nerve electrical stimulation, and functional electrical stimulation; the reward function formula of the dual-delay deep deterministic policy gradient algorithm is: In the formula: For the reward function, , , , , , These are the weighting coefficients. This represents the change in the cortical-muscle coherence coefficient after stimulation compared to before stimulation. This represents the change in neurovascular coupling strength after stimulation compared to before stimulation. Actual contraction range With the target contraction range The ratio represents the contractile efficacy. The fatigue index. To stimulate changes in energy consumption, Changes in stimulus parameters The L2 norm.

8. The central-peripheral neural modulation method based on hybrid brain-computer interface and AI according to claim 7, characterized in that, The temporal coordination strategy includes four stimulation phases: a prediction phase, a pre-activation phase, a co-stimulation phase, and a recovery phase. In the prediction phase, steps S1 to S4 are executed to obtain optimized electrical stimulation parameters, but no stimulation is output. In the pre-activation phase, based on the output results of the classification branch at the expected time, vagal nerve electrical stimulation and transcranial electrical stimulation are output in advance. Vagal nerve electrical stimulation uses high-frequency pulses for a set duration, while transcranial electrical stimulation uses anodic stimulation. In the co-stimulation phase, functional electrical stimulation and transcranial electrical stimulation are output simultaneously. The functional electrical stimulation adjusts its current intensity synchronously according to the intensity of the motor intention, while the transcranial electrical stimulation switches to a simulated theta short-array rapid pulse stimulation mode. In the recovery phase, stimulation is stopped.

9. The central-peripheral neural modulation method based on hybrid brain-computer interface and AI according to claim 8, characterized in that, The dynamic modulation of the electrical stimulation parameters includes: using a proportional-integral controller to modulate the intensity of the functional electrical stimulation, as shown in the following formula: In the formula: The intensity of the functional electrical stimulation current. Baseline intensity, For the intended decoding strength, This represents the actual contraction rate in the previous cycle. and This is a preset constant; The vagal nerve electrical stimulation is dynamically adjusted based on the fatigue index, as shown in the following formula: In the formula: The intensity of the vagus nerve electrical stimulation current. And α are preset constants; The electrode polarity configuration of the transcranial electrical stimulation is dynamically optimized according to the activation mode of the functional near-infrared spectrum. When ipsilateral cortical overactivation is detected, cathode stimulation is used to inhibit ipsilateral activation; when bilateral activation imbalance is detected, bilateral hemispheric balanced stimulation is used.

10. The central-peripheral neural modulation method based on hybrid brain-computer interface and AI according to claim 9, characterized in that, Step S6 includes: A response prediction model is constructed, which is a time-series prediction model based on LSTM, including two LSTM layers, a fully connected layer, and an output layer. The response prediction model is trained using historical response features extracted during electrical stimulation output, and the expected response index is output, which is the pelvic floor muscle strength level. A structural causal model describing the electrical stimulation output process is constructed, including observable variables, latent variables, and outcome variables. The observable variables include electrical stimulation intensity, electrical stimulation frequency, electrode position, target subject compliance, and baseline function. The latent variables include neural plasticity response ability, electrical stimulation fit, and fatigue recovery ability. The outcome variable is the pelvic floor muscle strength level. When the actual response index is lower than the lower limit of the confidence interval of the expected response index, counterfactual inference is triggered. The counterfactual inference includes: inferring the posterior distribution of latent variables based on observable and outcome variables; using variational inference methods, approximating the posterior distribution through an encoder network composed of three fully connected layers, and outputting the mean and variance of a Gaussian distribution; performing intervention operations on each observable variable, setting the observable variable as a candidate value in the structural causal model to cut off the influence of other variables on the observable variable; under the intervention conditions, predicting the posterior predicted distribution of the outcome variable based on the inferred posterior distribution of latent variables; obtaining a sample of the outcome variable by sampling latent variables and running the structural causal model; calculating the expected value of the outcome variable; calculating the counterfactual effect; selecting the observable variable with the largest counterfactual effect as the key constraint; generating an electrical stimulation parameter adjustment plan based on the key constraint; and adjusting the electrical stimulation parameters according to the electrical stimulation parameter adjustment plan.

11. A central-peripheral neural modulation system based on a hybrid brain-computer interface and AI, characterized in that, include: Signal acquisition module: used to simultaneously acquire EEG signals, functional near-infrared spectral signals, pelvic floor electromyography signals, and pelvic floor tactile pressure signals of the target object; Feature fusion module: used to construct a multi-branch deep neural network, extract features from EEG signals, functional near-infrared spectroscopy signals, pelvic floor electromyography signals and pelvic floor tactile pressure signals respectively, obtain fused features through cross-modal attention mechanism, calculate neurovascular coupling strength and cortical-muscle coherence coefficient, and concatenate with fused features to obtain central-peripheral joint feature vector; Motion Intent Decoding Module: This module is used to construct a motion intent decoder based on a temporal convolutional network and a multi-head self-attention mechanism. It inputs the central-peripheral joint feature vector sequence into the motion intent decoder to obtain the motion intent of the target object. Electrical stimulation parameter generation module: used to construct a biophysical model based on the neural mass model to simulate the neurodynamic process of the cortex-spinal cord-muscle pathway. It uses a dual-delay deep deterministic policy gradient algorithm to optimize the electrical stimulation parameters applied to the biophysical model. The state space of the dual-delay deep deterministic policy gradient algorithm is composed of the central-peripheral joint feature vector, the confidence of motor intention decoding, and the fatigue index of the target object. The action space is composed of the adjustment amount of the electrical stimulation parameters. By accumulating the adjustment amount, the optimized electrical stimulation parameters are obtained. Electrical stimulation output modulation module: used to design a timing coordination strategy based on the optimized electrical stimulation parameters, synchronously output multiple electrical stimulation control signals, and dynamically modulate the electrical stimulation parameters; Response assessment and adjustment module: This module is used to build a response prediction model. It is trained using historical response characteristics of electrical stimulation output and outputs expected response indicators. When the actual response indicator is lower than the lower limit of the confidence interval of the expected response indicator, counterfactual inference is triggered to generate an electrical stimulation parameter adjustment plan.

12. An electronic device comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the central-peripheral neural modulation method based on hybrid brain-computer interface and AI as described in any one of claims 1 to 10.

13. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the central-peripheral neural modulation method based on hybrid brain-computer interface and AI as described in any one of claims 1 to 10.

Images