Implantable spinal cord electrical stimulation neuromotor function regulation method and system based on causal spatiotemporal graph twin network

Abstract

The application provides an implantable spinal cord electrical stimulation neuromotor function regulation method and system based on a causal spatiotemporal graph twin network, relates to the technical field of spinal cord electrical stimulation neuromodulation, and comprises the following steps: acquiring bilateral limb movement trajectory data and stimulation channels and parameter characteristics of an implantable spinal cord electrical stimulation device; inputting the trajectory data of the healthy side and the affected side into a double-branch spatiotemporal graph convolution network sharing weights respectively, extracting spatiotemporal characteristics and calculating asymmetric difference characteristics; modulating the difference characteristics by using a causal gating weight vector output by a causal gating neuromodulation module, and fusing the difference characteristics with the spatiotemporal characteristics of the affected side to obtain input characteristics; inputting the input characteristics and the stimulation parameter characteristics into an evaluation network to output a limb motor function evaluation result. In this way, the causal regulation path of the stimulation channel to joint movement is determined, and the bilateral asymmetric difference characteristics are targetedly modulated and fused, so that precise, objective and interpretable quantitative evaluation of limb motor function is achieved.

Description

Technical Field

[0001] This application relates to the field of spinal cord electrical stimulation neuromodulation technology, and in particular to an implantable spinal cord electrical stimulation neuromotor function modulation method and system based on a causal spatiotemporal graph twin network. Background Technology

[0002] In the context of post-stroke motor function rehabilitation, implantable spinal cord stimulation (SCS) has become a key neuromodulation method for reconstructing motor control and improving gait and trunk balance. By activating spinal cord circuits and reconstructing descending cortical regulation function through electrical stimulation, it can significantly improve limb motor ability and gait performance in stroke patients. Currently, clinical practice routinely relies on physicians' clinical experience to select electrode channels and adjust stimulation parameters such as voltage, current, and frequency. The stimulation protocol is optimized through multiple postoperative follow-ups and repeated trial and error. This approach has prominent problems such as long adjustment cycles, strong subjectivity, lack of quantitative evidence, and poor individual adaptability. It not only increases the medical burden on patients but also makes it difficult to achieve precise neuromodulation.

[0003] Therefore, there is an urgent need to provide a method for analyzing the neuromodulation of spinal cord electrical stimulation, in order to analyze the causal regulation pathway of stimulation channels on joint movement, and based on this causal relationship, to perform targeted modulation and fusion of the asymmetric differences in bilateral limbs, thereby achieving a precise, objective and interpretable quantitative assessment of limb motor function. Summary of the Invention

[0004] The purpose of this application is to provide an implantable spinal cord electrical stimulation neuromotor function modulation method and system based on causal spatiotemporal graph twin network. By clarifying the causal regulation path of stimulation channels on joint movement, and thereby targeting and fusing bilateral asymmetric differences, a precise, objective and interpretable quantitative assessment of limb motor function is achieved.

[0005] This application provides a method for modulating neuromotor function through implanted spinal cord electrical stimulation based on a causal spatiotemporal graph twin network, including: Bilateral limb movement trajectory data and stimulation parameter features of an implanted spinal cord stimulation device are acquired. The bilateral limb movement trajectory data includes trajectory data of the healthy side and trajectory data of the affected side. The trajectory data of the healthy side and the trajectory data of the affected side are respectively input into a two-branch spatiotemporal graph convolutional network with the same structure and shared weights to obtain spatiotemporal features of the healthy side and spatiotemporal features of the affected side. The absolute value of the feature difference between the spatiotemporal features of the affected side and the spatiotemporal features of the healthy side is calculated to obtain asymmetric difference features. The asymmetric difference features are modulated using the causal gated weight vector output by the causal gated neural modulation module to obtain modulated features. The modulated features are then fused with the spatiotemporal features of the affected side to obtain input features. The causal gated weight vector is used to characterize the modulation intensity applied by each stimulation channel to the movement features of the affected joint. The input features are concatenated with the stimulation parameter features to obtain a fused feature vector. The fused feature vector is then input into an evaluation network to obtain the limb movement function evaluation result.

[0006] Optionally, the stimulation parameter features include: a channel gating vector and a continuous stimulation parameter feature vector; the channel gating vector is set as a 16-dimensional binary vector to characterize the activation state of the 16 implanted stimulation channels of the spinal cord electrical stimulation device; the continuous stimulation parameter feature vector is a 3-dimensional continuous parameter vector to characterize the voltage, current and frequency of the implanted stimulation channels.

[0007] Optionally, the step of inputting the trajectory data of the healthy limb and the trajectory data of the affected limb into a bi-branch spatiotemporal graph convolutional network with the same structure and shared weights to obtain spatiotemporal features of the healthy side and the affected side includes: performing graph convolution processing and temporal convolution processing on the trajectory data of the healthy limb and the trajectory data of the affected limb respectively based on a fixed anatomical adjacency matrix to extract spatial feature representation and temporal feature representation, thereby obtaining spatiotemporal features of the healthy side and the affected side; wherein, in the adjacency matrix, 1 represents an adjacent joint and 0 represents a non-adjacent joint.

[0008] Optionally, before modulating the asymmetric difference features using the causal gating weight vector output by the causal gating neural modulation module to obtain the modulated features, the method further includes: using a 16-dimensional channel gating vector as input to the causal gating neural modulation module; projecting the channel gating vector into a 12-dimensional joint weight vector based on a causal regulation path within the causal gating neural modulation module, and selecting the weight vectors corresponding to each joint on the affected side from the joint weight vectors to obtain the causal gating weight vector; wherein, the magnitude of each weight in the joint weight vector corresponds to the modulation intensity applied by the stimulus gating to the motion features of each joint; the causal regulation path is used to characterize the causal relationship between the stimulus channel and each joint.

[0009] Optionally, the spatiotemporal features of the affected side include features of the six joints on the affected side, and the spatiotemporal features of the healthy side include features of the six joints on the healthy side; calculating the absolute value of the difference between the spatiotemporal features of the healthy side and the spatiotemporal features of the affected side to obtain asymmetric difference features includes: calculating the absolute value of the difference between the features of the six joints on the affected side and the features of the six corresponding joints on the healthy side to obtain six-dimensional asymmetric difference features.

[0010] Optionally, the step of inputting the fused feature vector into the evaluation network to obtain the limb motor function evaluation result includes: sequentially performing a first fully connected layer mapping, layer normalization processing, random deactivation processing, and nonlinear activation processing on the fused feature vector to obtain a first dimensionality reduction feature; inputting the first dimensionality reduction feature into a second fully connected layer for dimensionality reduction processing to obtain a second dimensionality reduction feature; and inputting the second dimensionality reduction feature into a final fully connected layer for linear mapping, outputting a linear result as the limb motor function evaluation result.

[0011] Optionally, the training steps of the shared-weight bi-branch spatiotemporal graph convolutional network, the evaluation network, and the channel-to-joint mapping layer include: acquiring a training sample set; the training sample set includes: historical bilateral limb movement trajectory data and historical stimulus parameter features; inputting the historical bilateral limb movement trajectory data and the historical stimulus parameter features into the evaluation network to be trained to obtain a predicted training result; calculating a loss value based on the smoothed absolute value error between the predicted training result and the real clinical evaluation label, and updating the parameters of the evaluation network to be trained based on the loss value until a preset convergence condition is met.

[0012] Optionally, the causal regulation path is obtained based on the following steps: extracting the channel joint weight matrix output by the causal gating neural regulation module and calculating the global mean of the channel joint weight matrix; using the global mean as a threshold to filter out significant weights with weights greater than the threshold from the channel joint weight matrix; and determining the causal regulation path of the stimulation channel on joint movement based on the significant weights and the correspondence between the stimulation channel and the joint represented by the significant weights.

[0013] Optionally, the dual-branch spatiotemporal graph convolutional network includes at least one cascaded spatiotemporal residual block; each spatiotemporal residual block includes: a spatial graph convolutional sub-layer for extracting spatial topological features based on the fixed anatomical adjacency matrix, and a temporal convolutional sub-layer for extracting temporal dynamic features along the time dimension; a normalization layer and a nonlinear activation layer are respectively provided between the spatial graph convolutional sub-layer and the temporal convolutional sub-layer, and after the temporal convolutional sub-layer; the input features and output features of each spatiotemporal residual block are added element-wise through residual connections.

[0014] This application also provides an implantable spinal cord electrical stimulation neuromotor function modulation system based on a causal spatiotemporal graph twin network, which is used to implement the steps of the implantable spinal cord electrical stimulation neuromotor function modulation method based on a causal spatiotemporal graph twin network as described above.

[0015] This application also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the implantable spinal cord electrical stimulation neuromotor function modulation method based on causal spatiotemporal graph twin networks as described above.

[0016] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the implantable spinal cord electrical stimulation neuromotor function modulation method based on causal spatiotemporal graph twin network as described above.

[0017] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the implantable spinal cord electrical stimulation neuromotor function modulation method based on causal spatiotemporal graph twin networks as described above.

[0018] This application provides a method and system for modulating neuromotor function through implantable spinal cord stimulation based on a causal spatiotemporal graph twin network. The method acquires bilateral limb movement trajectory data and stimulation parameter features of an implantable spinal cord stimulation device. The bilateral limb movement trajectory data includes trajectory data of the healthy side and trajectory data of the affected side. The healthy side trajectory data and the affected side trajectory data are respectively input into a two-branch spatiotemporal graph convolutional network with identical structure and shared weights to obtain spatiotemporal features of the healthy side and the affected side. The absolute value of the feature difference between the corresponding joints in the spatiotemporal features of the affected side and the healthy side is calculated to obtain asymmetric difference features. The asymmetric difference features are modulated using a causal gated weight vector output by a causal gated neural modulation module to obtain modulated features. The modulated features are then fused with the affected side spatiotemporal features to obtain input features. The causal gated weight vector is used to characterize the modulation intensity applied by each stimulation channel to the movement features of the affected joint. The input features are concatenated with the stimulation parameter features to obtain a fused feature vector, which is then input into an evaluation network to obtain limb motor function evaluation results. Thus, by clarifying the causal regulatory pathways of stimulation channels on joint movement, and using this to target and fuse bilateral asymmetric differences, a precise, objective, and interpretable quantitative assessment of limb motor function is achieved. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the architecture of the implantable spinal cord electrical stimulation neuromotor function modulation system based on causal spatiotemporal graph twin network provided in this application; Figure 2 This is a flowchart illustrating the implantable spinal cord electrical stimulation neuromotor function modulation method based on causal spatiotemporal graph twin network provided in this application. Figure 3 This is a schematic diagram of the structure of the twin causal spatiotemporal graph convolutional network provided in this application; Figure 4 This is a schematic diagram of the causal gating neural modulation module provided in this application; Figure 5 This is a schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of terms can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship. All actions involving the acquisition of signal information or data in this application are performed in accordance with the relevant data protection laws and policies of the country where the application is located and with authorization from the owner of the relevant device.

[0023] The technical solutions in related technologies are difficult to interpret and demonstrate the electrical stimulation channels and joint modulation mechanisms, making it impossible for doctors to clearly understand the mechanism of stimulation during parameter adjustment. At the same time, traditional deep learning models are difficult to operate stably and cannot support stable and reliable decision output. In addition, existing technologies lack precise causal analysis methods for the neuromodulation mechanism after SCS surgery in stroke patients, making it difficult to clarify the causal relationship between stimulation channels and joint movement, and failing to provide doctors with clear control basis. Furthermore, the hardware configuration of existing related systems is complex and has poor adaptability, making it difficult to meet the needs of accurate assessment in medical institutions and convenient use in home scenarios. The lack of standardized data transmission and interaction design further limits the clinical implementation and widespread application of the technology.

[0024] To address the aforementioned technical problems in related technologies, this application provides an implantable spinal cord electrical stimulation neuromotor function modulation method based on a causal spatiotemporal graph twin network. This method is based on an interpretable model of a causal spatiotemporal graph twin network and combined with standardized hardware configuration to achieve integrated support for modulation analysis and hardware adaptation, breaking through the limitations of existing technologies and meeting the diverse needs of clinical individualized programming and family rehabilitation assessment.

[0025] like Figure 1 The diagram shown is an architecture diagram of an implantable spinal cord electrical stimulation neuromotor function modulation system based on a causal spatiotemporal graph twin network provided in this application embodiment. It mainly includes four core parts: data preprocessing, a lightweight causal spatiotemporal graph twin network core architecture, a causal-gated neuromodulation module, and an evaluation network. The flow and connection relationships of each part are as follows: I. Data preprocessing, which includes three steps: Trajectory data reshaping: The format of the original motion trajectory data is adjusted to meet the model input requirements; Healthy / affected limb joint splitting: The motion data is split into healthy (unaffected) and affected (affected) limbs; Data standardization: The split data is normalized to eliminate dimensional differences.

[0026] II. Lightweight Causal Spatiotemporal Graph Siamese Network Core Architecture: The core network architecture is structured in the middle, employing a dual-branch design (healthy side branch + affected side branch) with shared weights. These branches are connected by a fixed anatomical adjacency matrix (defining the physical connections between joints), and each branch has a consistent structure: a shared twin backbone network (weight sharing): each branch contains two sub-networks: a Graph Convolutional Network (GCN): based on the fixed anatomical adjacency matrix, it aggregates features of adjacent joints in the spatial dimension; a Temporal Convolutional Network (TCN): it captures local dynamic patterns of motion sequences in the temporal dimension. Output features: the healthy side branch outputs the spatiotemporal features of the healthy side, and the affected side branch outputs the spatiotemporal features of the affected side.

[0027] III. Calculation of Differential Features and Causal Gated Regulation Networks: The spatiotemporal features of the healthy and affected sides output by the core architecture enter the "Difference Feature Calculation (Absolute Difference)" module, which calculates the absolute difference in the bilateral limb movement features to obtain the "Asymmetric Difference Feature". This feature is simultaneously input into two modules: the causal gating neural modulation module (mapping and weighting): receiving the electrical stimulation gating vector (a 16-dimensional binary vector representing the activation state of the stimulation channel), projecting the gating vector into a joint weight vector through a mapping layer, and performing element-wise weighted modulation of the difference feature to simulate the targeted regulation of joint movement by electrical stimulation; and intensity modulation: adjusting the intensity of the difference feature to enhance or inhibit specific features.

[0028] IV. Evaluation Network: The outputs of the causal gating module and the intensity modulation are input into the evaluation network, and the final result is generated through the following steps: Feature fusion: The modulated features are fused with the original differential features, electrical stimulation parameters, etc., to integrate motor defects and regulatory information; Dimension transformation: The fused features are reduced in dimension and transformed to adapt to the requirements of the output layer; Final output: The limb motor function evaluation results are output (such as quantitative indicators such as support phase ratio and shoulder imbalance angle).

[0029] In this way, the overall architecture enables the correlation analysis between electrical stimulation parameters and limb motor function, providing an interpretable quantitative assessment for spinal cord electrical stimulation neuromodulation.

[0030] The following description, in conjunction with the accompanying drawings, details the implantable spinal cord electrical stimulation neuromotor function modulation method based on causal spatiotemporal graph twin networks provided in this application, through specific embodiments and application scenarios.

[0031] like Figure 2 As shown in the embodiment of this application, an implantable spinal cord electrical stimulation method for modulating neuromotor function based on a causal spatiotemporal graph twin network is provided. This method may include the following steps 201 to 204: Step 201: Obtain bilateral limb movement trajectory data and stimulation parameter characteristics of the implanted spinal cord electrical stimulation device.

[0032] The bilateral limb movement trajectory data includes: trajectory data of the healthy side limb and trajectory data of the affected side limb. The bilateral limb movement trajectory data can be collected using wearable inertial sensors or optical motion capture devices, and the stimulation parameter characteristics of implanted spinal cord stimulation devices can be directly read through the device's communication interface. This data serves as the basic input for subsequent motor function assessment.

[0033] For example, video footage of the subject's limb movements was acquired using a visual motion capture device, and three-dimensional motion trajectories were extracted. To accurately quantify changes in motor function, 16 core key points on the subject's body were selected, and 12 key joint indicators were further extracted from them to construct a kinematic tensor. For hemiplegic patients, where one side of the limb has relatively normal motor function while the other side has motor dysfunction, the above trajectory data was distinguished into trajectory data of the healthy side and trajectory data of the affected side, providing a comparative basis for subsequent calculation of bilateral motor asymmetry differences.

[0034] For example, the stimulation parameter features include: a channel gating vector and a continuous stimulation parameter feature vector. The stimulation parameter features contain information about the current operating state of the implanted spinal cord stimulation device, which is divided into discrete gating states and continuous physical quantity parameters. The channel gating vector is set as a 16-dimensional binary vector to characterize the activation state of the 16 implanted stimulation channels of the spinal cord stimulation device.

[0035] For example, implantable spinal cord stimulation devices are typically equipped with multi-channel electrode arrays. Each element in the 16-dimensional binary vector takes a value of 0 or 1, corresponding to the closing or opening of an implantable stimulation channel, thereby accurately characterizing which channels are currently receiving electrical stimulation. The continuous stimulation parameter feature vector is a 3-dimensional continuous parameter vector used to characterize the voltage, current, and frequency of the implantable stimulation channels.

[0036] For example, in addition to the on / off state of the electrical stimulation channel, the intensity and rhythm of the electrical stimulation are controlled by continuous parameters. The 3D continuous parameter vector records the voltage amplitude, current magnitude, and pulse frequency of the currently activated implanted stimulation channel, which directly affect the depth and range of neuromodulation.

[0037] Step 202: Input the trajectory data of the healthy side limb and the trajectory data of the affected side limb into a two-branch spatiotemporal graph convolutional network with the same structure and shared weights to obtain spatiotemporal features of the healthy side and spatiotemporal features of the affected side. Calculate the absolute value of the feature difference between the corresponding joints in the spatiotemporal features of the affected side and the spatiotemporal features of the healthy side to obtain asymmetric difference features.

[0038] For example, before feature extraction, the trajectory data of the healthy and affected sides are preprocessed, including trajectory reshaping, data normalization, and limb segmentation (Healthy / Affected). The bi-branch spatiotemporal graph convolutional network adopts a bi-branch structure with shared weights, which can simultaneously capture the spatial topological features and temporal dynamic changes of limb joint movements, ensuring alignment of the feature space. Calculating the absolute value of the difference between corresponding joint features can quantify the degree of movement deviation of the affected side relative to the healthy side, thereby objectively reflecting the degree of motor dysfunction.

[0039] Specifically, in step 202 above, the processing of the dual-branch spatiotemporal graph convolutional network may include the following step 202a: Step 202a: Based on a fixed anatomical adjacency matrix, perform graph convolution and temporal convolution on the trajectory data of the healthy limb and the trajectory data of the affected limb respectively, extract spatial feature representation and temporal feature representation, and obtain the spatiotemporal features of the healthy side and the spatiotemporal features of the affected side.

[0040] In the adjacency matrix, 1 represents an adjacent joint and 0 represents a non-adjacent joint.

[0041] For example, such as Figure 3 As shown, the overall structure of the Siamese causal spatiotemporal graph convolutional network is illustrated. This architecture mainly consists of three parts: data preprocessing, lightweight Siamese STGCN, and causal FiLM modulation.

[0042] 1. Data Preprocessing: Trajectory Reshaping: Reshaping the original motion trajectory data to conform to the model input format requirements; Data Normalization: Normalizing the motion data to eliminate the influence of dimensions; Limb Splitting (Healthy / Affected): Segmenting the motion data into two parts: the healthy limb and the affected limb.

[0043] 2. Lightweight Siamese STGCN: Fixed Anatomical Adjacency Matrix: Defines the physical connections between human joints; Shared Siamese Backbone Network (GCN+TCN): Contains two branches with shared weights, processing healthy and affected side data respectively; Healthy Limb Trajectory Data; Affected Limb Trajectory Data; Healthy Limb Spatiotemporal Features: Spatiotemporal features of the healthy side extracted through Graph Convolutional Network (GCN) and Temporal Convolutional Network (TCN); Affected Limb Spatiotemporal Features: Spatiotemporal features of the affected side extracted through Graph Convolutional Network (GCN) and Temporal Convolutional Network (TCN); 3. Causal FiLM Modulation Network: Difference Feature Calculation: Calculates the differences in spatiotemporal features between the healthy and affected sides; 16 Stimulation Gates: Characterize the activation status of the 16 implanted stimulation channels of the spinal cord electrical stimulation device; Electrical Stimulation Parameters: Includes 3-dimensional continuous parameters such as voltage, current, and frequency; CausalFiLM Neural Modulation Module: Performs causal-gated modulation on the difference features; Element-wise Weighting: Uses causal-gated weight vectors to perform element-wise weighting on the difference features; Feature Fusion: Fuses the modulated features with the spatiotemporal features of the affected side; Dimension Conversion: Performs dimensional transformation on the fused features; Rehabilitation Effect Prediction: Outputs the final assessment results of limb motor function.

[0044] For example, such as Figure 3As shown, the network combines a fixed anatomical adjacency matrix with a graph convolutional network (GCN) and a temporal convolutional network (TCN). The fixed anatomical adjacency matrix defines the physical connections between human joints. Graph convolution processing aggregates the features of adjacent joints in the spatial dimension based on this matrix; temporal convolution processing captures the local dynamic patterns of motion sequences in the temporal dimension. For example, the shoulder and elbow joints are anatomically adjacent, and their corresponding positions in the adjacency matrix are 1; while the shoulder and wrist joints are not adjacent, and their corresponding positions are 0. This binary matrix definition is clear and facilitates graph convolution operations.

[0045] In another embodiment, the spatiotemporal features of the affected side include features of six joints on the affected side, and the spatiotemporal features of the healthy side include features of six joints on the healthy side. For a single limb, features of six key joints (e.g., shoulder, elbow, wrist, hip, knee, and ankle) are selected for representation. These six joints can comprehensively depict the movement state of the upper and lower limbs.

[0046] Specifically, regarding the calculation of asymmetric difference characteristics, step 202 above may further include the following step 202b: Step 202b: Calculate the absolute value of the difference between the features of the 6 joints on the affected side and the features of the 6 corresponding joints on the healthy side to obtain 6-dimensional asymmetric difference features.

[0047] For example, a 6-dimensional vector is formed by calculating the absolute value of the difference between the features of the healthy side and the affected side joint by joint. The calculation of feature differences is defined as follows: ; in, The spatiotemporal feature vector representing the trajectory of the unaffected limb. This represents the spatiotemporal feature vector of the trajectory of the affected limb. To quantify the degree of motion deviation, the absolute value of the difference is calculated to obtain the asymmetric difference feature. Each dimension of this vector represents the degree of motion deviation at a specific joint; a larger value indicates a more severe asymmetry.

[0048] Specifically, regarding the internal structure of the aforementioned bi-branch spatiotemporal graph convolutional network: the bi-branch spatiotemporal graph convolutional network includes at least one cascaded spatiotemporal residual block; each spatiotemporal residual block includes: a spatial graph convolutional sub-layer for extracting spatial topological features based on the fixed anatomical adjacency matrix, and a temporal convolutional sub-layer for extracting temporal dynamic features along the time dimension; a normalization layer and a nonlinear activation layer are respectively provided between the spatial graph convolutional sub-layer and the temporal convolutional sub-layer, and after the temporal convolutional sub-layer; the input features and output features of each spatiotemporal residual block are added element-wise through residual connections.

[0049] For example, by stacking multiple spatiotemporal residual blocks, the network can abstract higher-level spatiotemporal motion features layer by layer, while the residual structure helps alleviate the gradient vanishing problem in deep network training. Spatial graph convolutional sublayers handle the spatial dependencies between different joints at the same time step, while temporal convolutional sublayers handle the temporal variations of the same joint at different times. Normalization layers accelerate network convergence and improve stability, while nonlinear activation layers introduce nonlinear mapping capabilities. Residual connections directly add the block's input to its output, allowing the network to focus on learning residual mappings during training, thus making it easier to train deep models.

[0050] Step 203: Modulate the asymmetric difference features using the causal gating weight vector output by the causal gating neural modulation module to obtain modulated features, and fuse the modulated features with the spatiotemporal features of the affected side to obtain input features.

[0051] The causal gating weight vector is used to characterize the modulation intensity applied by each stimulation channel to the motion characteristics of the affected joint.

[0052] For example, such as Figure 4 The diagram shows the detailed structure of the CausalFiLM neural modulation module, which is mainly used to simulate the neuromodulation effects of spinal cord electrical stimulation to achieve targeted regulation of motor characteristics. Specifically, it includes: 1. Channel-to-Joint Mapping Layer: Linear layer (Linear(16→12, bias=False)): Projects the 16-dimensional gating vector into a 12-dimensional joint weight vector; Initialization ([Init: uniform_(0, 0.5)]): The weights are initialized in a uniform distribution; Select Affected Weights (Indices 1,3,5,7,9,11)): Selects the weights corresponding to the 6 joints on the affected side from the 12-dimensional vector.

[0053] 2. Causal FiLM Physical Modulation: Element-wise Multiplication (⊙): Uses joint weight vectors to perform element-wise weighted modulation of differential features; Activation (activation: torch.clamp(x, 0, 1)): Uses the clamp activation function to limit the output range to [0,1]; Targeted modulation simulating neurotransmitter scaling: Simulates the scaling effect of neurotransmitters on motion features.

[0054] 3. Feature Fusion: Concat operation: Concatenates the modulated features (48-dimensional), 16-dimensional stimulus gating vector, and 3-dimensional continuous parameter vector; Input dimension ([B,8,1,6] → 48 dim): Feature dimension transformation process.

[0055] For example, Figure 4 The overall process shown is as follows: Sixteen stimulation gates are used as input and mapped to a 12-dimensional joint weight vector through a linear layer. The weights corresponding to the six joints on the affected side are selected and element-wise multiplied with the differential features to obtain the modulation features. The modulation features are then fused with the spatiotemporal features of the affected side and finally input into the subsequent network for rehabilitation effect prediction.

[0056] For example, combined Figure 3 ,like Figure 4 The diagram showing the causal-gated CausalFiLM neural modulation module illustrates how causal-gated weight vectors are used to perform element-wise weighted modulation of asymmetric difference features. This amplifies joint differences with high causal correlation and suppresses noise with low correlation. The calculation formula for the modulation process is as follows: ; in, W is the causal gating weight vector, and ⊙ represents the Hadamard product (element-by-element multiplication). The modulated features are then fused with the original spatiotemporal features of the affected side, so that the input features contain both motor defect information and prior knowledge of stimulus modulation.

[0057] Optionally, before performing the modulation operation in step 203 above, the method may further include steps 205 and 206 to obtain the desired causal gating weight vector: Step 205: Use the 16-dimensional channel gating vector as the input to the causal gating neural modulation module.

[0058] For example, such as Figure 4 As shown, the causal-gated neural regulation module obtains the gating vectors representing the activation states of 16 stimulation channels, and uses these as the basis to deduce the regulatory intensity of each joint.

[0059] Step 206: In the causal-gated neural modulation module, the channel gating vector is projected into a 12-dimensional joint weight vector based on the causal regulation path, and the weight vectors corresponding to each joint on the affected side are selected from the joint weight vectors to obtain the causal-gated weight vector.

[0060] The magnitude of each weight in the joint weight vector corresponds to the modulation intensity applied by the stimulus gating to the motion characteristics of each joint; the causal regulation path is used to characterize the causal relationship between the stimulus channel and each joint.

[0061] For example, the system reverse-engineers the causal regulatory path after training is complete. Figure 4 As shown, the module contains a channel-to-joint mapping layer, which projects a 16-dimensional gated vector into a 12-dimensional joint weight vector through an unbiased linear layer (Linear 16→12). W ∈[0,1]. Since hemiplegic patients focus on the affected side, weights corresponding to the six joints on the affected side are selected from the 12-dimensional vector (for example, weights with indices 1, 3, 5, 7, 9, 11 are extracted) to form a 6-dimensional causal gating weight vector for modulation. The larger the weight value, the greater the modulation intensity applied to the motion characteristics of the corresponding joint by the stimulus gating.

[0062] Step 204: Concatenate the input features with the stimulus parameter features to obtain a fused feature vector, and input the fused feature vector into the evaluation network to obtain the limb motor function evaluation result.

[0063] For example, combined Figure 3 ,like Figure 4 As shown, the input features containing rich motion and modulation information are concatenated with the original stimulus parameter features. Specifically, the modulated and flattened features (48 dimensions) are fused and concatenated with a 16-dimensional stimulus gating vector and a 3-dimensional continuous parameter vector to form a complete 67-dimensional fused feature vector. The evaluation network performs a nonlinear mapping based on this vector, ultimately outputting the limb motor function assessment results, providing clinicians with a quantitative assessment basis.

[0064] Specifically, in step 204 above, the fused feature vector is input into the evaluation network, which may include the following steps 204a1 and 204a2: Step 204a1: The fused feature vector is sequentially mapped by the first fully connected layer, processed by layer normalization, random deactivation, and nonlinear activation to obtain the first dimensionality reduction feature. The first dimensionality reduction feature is then input into the second fully connected layer for dimensionality reduction processing to obtain the second dimensionality reduction feature.

[0065] For example, the first fully connected layer maps the high-dimensional fused features to a dimension suitable for network processing, layer normalization stabilizes the training process, random deactivation prevents model overfitting, and nonlinear activation introduces nonlinear expressive power. Subsequently, the second fully connected layer further reduces the dimensionality and extracts the core evaluation features.

[0066] Step 204a2: Input the second dimensionality reduction feature into the final fully connected layer for linear mapping, and output a linear result as the limb motor function evaluation result.

[0067] For example, the final fully connected layer maps the features to specific evaluation index values, and the network directly outputs a linear result as the final limb motor function evaluation result.

[0068] Optionally, in the application embodiment, the training steps of the shared-weights dual-branch spatiotemporal graph convolutional network, the evaluation network, and the channel-to-joint mapping layer include the following steps 207 to 209: Step 207: Obtain the training sample set.

[0069] The training sample set includes historical bilateral limb movement trajectory data and historical stimulus parameter features. Before applying the model, a large amount of historical clinical data is needed for training to optimize the network parameters. The training process first requires obtaining a sufficiently large training sample set.

[0070] For example, the data format in the training sample set is consistent with that in actual applications, including historically collected bilateral limb movement trajectory data and the historical stimulation parameter characteristics of the corresponding implanted spinal cord stimulation device, ensuring the consistency of the distribution of training data with the real scene.

[0071] Step 208: Input the historical bilateral limb movement trajectory data and the historical stimulus parameter features into the evaluation network to be trained to obtain the predicted training results.

[0072] For example, historical data is input into the network to be trained and evaluated. The network forward propagation outputs the predicted training result under the current parameters. This result represents the model's predicted value for the limb motor function evaluation of historical samples.

[0073] Step 209: Calculate the loss value based on the smoothed absolute value error between the predicted training results and the real clinical assessment labels, and update the parameters of the assessment network to be trained based on the loss value until the preset convergence condition is met.

[0074] For example, smoothed absolute error is used as the loss function, which is more robust to outliers than mean squared error. The loss value is calculated using the backpropagation algorithm, and the parameters of the network to be trained and evaluated are updated using an optimizer. Training ends when the loss value no longer decreases or meets the preset convergence condition, and the optimal model parameters are obtained.

[0075] In another implementation, since the causal regulation path is obtained by reverse engineering after network training is completed, the causal regulation path is obtained based on the following steps 210 to 212: Step 210: Extract the channel joint weight matrix output by the causal gating neural regulation module, and calculate the global mean of the channel joint weight matrix.

[0076] Step 211: Use the global mean as a threshold to filter out significant weights with weights greater than the threshold from the channel joint weight matrix.

[0077] For example, retaining only weights above the global mean can filter out weak or irrelevant connections and highlight the core relationships that truly have a significant regulatory effect.

[0078] Step 212: Based on the significant weights and the correspondence between the stimulation channels and joints represented by the significant weights, determine the causal regulation path of the stimulation channels on joint movement.

[0079] For example, through the above screening, the causal-gated neural modulation module can construct a clear causal regulation path, clarifying which stimulation channels have a dominant influence on the movement of which specific joints, thereby tracing the causal path between electrical stimulation intervention and motor response. This path will be used for gating vector projection operations in subsequent inference processes.

[0080] The implantable spinal cord stimulation neuromotor function modulation analysis method based on causal spatiotemporal graph twin network provided in this embodiment extracts bilateral spatiotemporal features and calculates asymmetric difference features through a weighted bi-branch spatiotemporal graph convolutional network. It is then modulated by combining causal gating weight vectors and back-inferring the causal regulation path of the stimulation channel on joint movement after training, highlighting the movement deviation of key joints. Finally, the evaluation network outputs multi-dimensional limb motor function assessment results, realizing a precise, objective and interpretable quantitative assessment of the recovery of motor function under spinal cord stimulation intervention.

[0081] The method for modulating neuromotor function of implantable spinal cord stimulation based on causal spatiotemporal graph twin network provided in this application acquires bilateral limb movement trajectory data and stimulation parameter features of the implantable spinal cord stimulation device. The bilateral limb movement trajectory data includes: trajectory data of the healthy side limb and trajectory data of the affected side limb. The trajectory data of the healthy side limb and the trajectory data of the affected side limb are respectively input into a two-branch spatiotemporal graph convolutional network with the same structure and shared weights to obtain spatiotemporal features of the healthy side and spatiotemporal features of the affected side. The absolute value of the feature difference between the spatiotemporal features of the affected side and the spatiotemporal features of the healthy side limb is calculated to obtain asymmetric difference features. The asymmetric difference features are modulated using the causal gated weight vector output by the causal gated neural modulation module to obtain modulated features. The modulated features are then fused with the spatiotemporal features of the affected side to obtain input features. The causal gated weight vector is used to characterize the modulation intensity applied by each stimulation channel to the movement features of the affected joint. The input features are concatenated with the stimulation parameter features to obtain a fused feature vector. The fused feature vector is then input into an evaluation network to obtain the limb motor function evaluation result. Thus, by clarifying the causal regulatory pathways of stimulation channels on joint movement, and using this to target and fuse bilateral asymmetric differences, a precise, objective, and interpretable quantitative assessment of limb motor function is achieved.

[0082] It should be noted that the implantable spinal cord electrical stimulation neuromotor function modulation method based on causal spatiotemporal graph twin network provided in this application embodiment can be executed by an implantable spinal cord electrical stimulation neuromotor function modulation system based on causal spatiotemporal graph twin network, or by a control module in the implantable spinal cord electrical stimulation neuromotor function modulation system based on causal spatiotemporal graph twin network for executing the implantable spinal cord electrical stimulation neuromotor function modulation method based on causal spatiotemporal graph twin network.

[0083] This application provides an implantable spinal cord stimulation neuromotor function modulation system based on a causal spatiotemporal graph twin network. It acquires bilateral limb movement trajectory data and stimulation parameter features of the implantable spinal cord stimulation device. The bilateral limb movement trajectory data includes trajectory data of the healthy side and trajectory data of the affected side. The healthy side trajectory data and the affected side trajectory data are respectively input into a two-branch spatiotemporal graph convolutional network with identical structure and shared weights to obtain spatiotemporal features of the healthy side and the affected side. The absolute value of the feature difference between the corresponding joints in the spatiotemporal features of the affected side and the healthy side is calculated to obtain asymmetric difference features. The asymmetric difference features are modulated using a causal gated weight vector output by a causal gated neural modulation module to obtain modulated features. The modulated features are then fused with the affected side spatiotemporal features to obtain input features. The causal gated weight vector is used to characterize the modulation intensity applied by each stimulation channel to the movement features of the affected joint. The input features are concatenated with the stimulation parameter features to obtain a fused feature vector, which is then input into an evaluation network to obtain limb motor function evaluation results. Thus, by clarifying the causal regulatory pathways of stimulation channels on joint movement, and using this to target and fuse bilateral asymmetric differences, a precise, objective, and interpretable quantitative assessment of limb motor function is achieved.

[0084] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5As shown, the electronic device may include: a processor 510, a communication interface 520, a memory 530, and a communication bus 540. The processor 510, communication interface 520, and memory 530 communicate with each other via the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a method for modulating neuromotor function through implanted spinal cord stimulation based on a causal spatiotemporal graph twin network. This method includes: acquiring bilateral limb movement trajectory data and stimulation parameter characteristics of the implanted spinal cord stimulation device; the bilateral limb movement trajectory data includes: trajectory data of the healthy side limb and trajectory data of the affected side limb; inputting the trajectory data of the healthy side limb and the trajectory data of the affected side limb into a two-branch spatiotemporal graph convolutional network with identical structure and shared weights to obtain spatiotemporal features of the healthy side and the affected side, and calculating the spatiotemporal features of the affected side. The absolute value of the difference between the features and the corresponding joint features in the spatiotemporal features of the healthy side is used to obtain the asymmetric difference features. The asymmetric difference features are modulated using the causal gating weight vector output by the causal gating neural modulation module to obtain modulated features. These modulated features are then fused with the spatiotemporal features of the affected side to obtain input features. The causal gating weight vector characterizes the modulation intensity applied by each stimulation channel to the joint movement features of the affected side. The input features are concatenated with the stimulation parameter features to obtain a fused feature vector, which is then input into the evaluation network to obtain the limb motor function evaluation result. Thus, by clarifying the causal regulation path of stimulation channels on joint movement and using this to target and fuse bilateral asymmetric difference features, a precise, objective, and interpretable quantitative evaluation of limb motor function is achieved.

[0085] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0086] On the other hand, this application also provides a computer program product, which includes a computer program stored on a computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by a computer, the computer is able to execute the implantable spinal cord electrical stimulation neuromotor function modulation method based on causal spatiotemporal graph twin network provided by the above methods. The method includes: acquiring bilateral limb movement trajectory data and stimulation parameter characteristics of an implantable spinal cord electrical stimulation device; the bilateral limb movement trajectory data includes: trajectory data of the healthy side limb and trajectory data of the affected side limb; inputting the trajectory data of the healthy side limb and the trajectory data of the affected side limb into a twin network with the same structure and shared weights. A branched spatiotemporal graph convolutional network is used to obtain spatiotemporal features of the healthy side and the affected side. The absolute value of the feature difference between the corresponding joints in the spatiotemporal features of the affected side and the healthy side is calculated to obtain asymmetric difference features. The asymmetric difference features are modulated using a causal gating weight vector output from a causal gating neural modulation module to obtain modulated features. These modulated features are then fused with the spatiotemporal features of the affected side to obtain input features. The causal gating weight vector characterizes the modulation intensity applied by each stimulation channel to the joint movement features of the affected side. The input features are concatenated with the stimulation parameter features to obtain a fused feature vector, which is then input into an evaluation network to obtain the limb motor function evaluation result. Thus, by clarifying the causal regulation path of stimulation channels on joint movement and using this to target and fuse bilateral asymmetric difference features, a precise, objective, and interpretable quantitative evaluation of limb motor function is achieved.

[0087] Furthermore, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned implantable spinal cord electrical stimulation neuromotor function modulation methods based on causal spatiotemporal graph twin networks. The method includes: acquiring bilateral limb movement trajectory data and stimulation parameter characteristics of an implantable spinal cord electrical stimulation device; the bilateral limb movement trajectory data includes: trajectory data of the unaffected limb and trajectory data of the affected limb; inputting the trajectory data of the unaffected limb and the trajectory data of the affected limb into a two-branch spatiotemporal graph convolutional network with identical structure and shared weights, respectively, to obtain spatiotemporal features of the unaffected limb and... The spatiotemporal features of the affected side are analyzed, and the absolute value of the feature difference between the corresponding joints in the spatiotemporal features of the affected side and the healthy side is calculated to obtain asymmetric difference features. The asymmetric difference features are modulated using the causal gating weight vector output by the causal gating neural modulation module to obtain modulated features. These modulated features are then fused with the spatiotemporal features of the affected side to obtain input features. The causal gating weight vector characterizes the modulation intensity applied by each stimulation channel to the joint movement features of the affected side. The input features are concatenated with the stimulation parameter features to obtain a fused feature vector, which is then input into the evaluation network to obtain the limb motor function evaluation result. Thus, by clarifying the causal regulation path of stimulation channels on joint movement and using this to target and fuse bilateral asymmetric difference features, a precise, objective, and interpretable quantitative evaluation of limb motor function is achieved.

[0088] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0089] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for modulating neuromotor function through implanted spinal cord electrical stimulation based on causal spatiotemporal graph twin networks, characterized in that, include: Acquire bilateral limb movement trajectory data, as well as stimulation parameter characteristics of the implanted spinal cord electrical stimulation device; The bilateral limb movement trajectory data includes: trajectory data of the healthy limb and trajectory data of the affected limb; The trajectory data of the healthy side limb and the trajectory data of the affected side limb are respectively input into a two-branch spatiotemporal graph convolutional network with the same structure and shared weights to obtain spatiotemporal features of the healthy side and spatiotemporal features of the affected side. The absolute value of the feature difference between the spatiotemporal features of the affected side and the spatiotemporal features of the healthy side is calculated to obtain asymmetric difference features. The asymmetric difference features are modulated using the causal gating weight vector output by the causal gating neural modulation module to obtain modulated features, and the modulated features are fused with the spatiotemporal features of the affected side to obtain input features; the causal gating weight vector is used to characterize the modulation intensity applied by each stimulation channel to the joint motion features of the affected side; The input features are concatenated with the stimulus parameter features to obtain a fused feature vector, which is then input into the evaluation network to obtain the limb motor function evaluation result.

2. The method according to claim 1, characterized in that, The stimulation parameter features include: a channel gating vector and a continuous stimulation parameter feature vector; the channel gating vector is set as a 16-dimensional binary vector to characterize the activation state of the 16 implanted stimulation channels of the spinal cord electrical stimulation device; the continuous stimulation parameter feature vector is a 3-dimensional continuous parameter vector to characterize the voltage, current and frequency of the implanted stimulation channels.

3. The method according to claim 1 or 2, characterized in that, The step of inputting the trajectory data of the healthy limb and the trajectory data of the affected limb into a two-branch spatiotemporal graph convolutional network with identical structure and shared weights to obtain spatiotemporal features of the healthy side and the affected side includes: Based on a fixed anatomical adjacency matrix, graph convolution and temporal convolution are performed on the trajectory data of the healthy limb and the trajectory data of the affected limb respectively to extract spatial feature representation and temporal feature representation, thereby obtaining the spatiotemporal features of the healthy side and the spatiotemporal features of the affected side. In the adjacency matrix, 1 represents an adjacent joint and 0 represents a non-adjacent joint.

4. The method according to claim 1, characterized in that, Before modulating the asymmetric difference features using the causal gating weight vector output by the causal gating neural modulation module to obtain the modulated features, the method further includes: The 16-dimensional channel-gated vector is used as the input to the causal-gated neural modulation module; Within the causal-gated neural modulation module, the channel gating vector is projected into a 12-dimensional joint weight vector based on the causal regulation path, and the weight vectors corresponding to each joint on the affected side are selected from the joint weight vector to obtain the causal-gated weight vector. The magnitude of each weight in the joint weight vector corresponds to the modulation intensity applied by the stimulus gating to the motion characteristics of each joint; the causal regulation path is used to characterize the causal relationship between the stimulus channel and each joint.

5. The method according to claim 1, characterized in that, The spatiotemporal features of the affected side include features of the six joints on the affected side, and the spatiotemporal features of the healthy side include features of the six joints on the healthy side. The calculation of the absolute value of the difference between the spatiotemporal characteristics of the healthy side and the spatiotemporal characteristics of the affected side yields asymmetric difference features, including: The absolute values ​​of the differences between the features of the 6 joints on the affected side and the features of the corresponding 6 joints on the healthy side are calculated to obtain 6-dimensional asymmetric difference features.

6. The method according to claim 1, characterized in that, The step of inputting the fused feature vector into the evaluation network to obtain the limb motor function evaluation result includes: The fused feature vector is sequentially mapped by a first fully connected layer, processed by layer normalization, random deactivation, and nonlinear activation to obtain a first dimensionality-reduced feature. The first dimensionality-reduced feature is then input into a second fully connected layer for dimensionality reduction to obtain a second dimensionality-reduced feature. The second dimensionality reduction feature is input into the final fully connected layer for linear mapping, and a linear result is output as the limb motor function evaluation result.

7. The method according to claim 1, characterized in that, The training steps for the shared-weights bi-branch spatiotemporal graph convolutional network, the evaluation network, and the channel-to-joint mapping layer include: Obtain a training sample set; the training sample set includes: historical bilateral limb movement trajectory data, and historical stimulus parameter features; The historical bilateral limb movement trajectory data and the historical stimulus parameter features are input into the evaluation network to be trained to obtain the predicted training results. Based on the smoothed absolute error between the predicted training results and the actual clinical assessment labels, a loss value is calculated, and the parameters of the network to be trained are updated based on the loss value until a preset convergence condition is met.

8. The method according to claim 7, characterized in that, The causal regulation pathway is obtained based on the following steps: Extract the channel joint weight matrix output by the causal gating neural modulation module, and calculate the global mean of the channel joint weight matrix; The global mean is used as a threshold to filter out significant weights with weights greater than the threshold from the channel joint weight matrix; Based on the significant weights and the correspondence between the stimulation channels and joints represented by the significant weights, the causal regulation path of the stimulation channels on joint movement is determined.

9. The method according to claim 1, characterized in that, The dual-branch spatiotemporal graph convolutional network includes at least one cascaded spatiotemporal residual block; each spatiotemporal residual block includes: a spatial graph convolutional sub-layer for extracting spatial topological features based on the fixed anatomical adjacency matrix, and a temporal convolutional sub-layer for extracting temporal dynamic features along the time dimension; a normalization layer and a nonlinear activation layer are respectively provided between the spatial graph convolutional sub-layer and the temporal convolutional sub-layer, and after the temporal convolutional sub-layer; the input features and output features of each spatiotemporal residual block are added element-wise through residual connections.

10. An implantable spinal cord electrical stimulation neuromotor function modulation system based on a causal spatiotemporal graph twin network, characterized in that, The system is used to implement the implantable spinal cord electrical stimulation neuromotor function modulation method based on causal spatiotemporal graph twin network as described in any one of claims 1 to 9.

Images