Functional electrical stimulation method and system based on central intention and peripheral collaborative mode collaborative decoding

Abstract

The application relates to the technical field of biomedical engineering and neural rehabilitation, and discloses a functional electrical stimulation method and system based on central intention and peripheral collaborative decoding of a central-peripheral collaborative mode. In a training stage, central brain signals of a diseased side and peripheral multi-channel electromyographic signals of a healthy side are synchronously collected, the central brain signals of the diseased side are input into a central intention feature extraction module to output low-dimensional central intention latent space feature sequences, the peripheral multi-channel electromyographic signals of the healthy side are decomposed to obtain a collaborative primitive matrix and a collaborative activation coefficient sequence, a mapping model is trained with the central intention latent space feature sequences as input and the collaborative activation coefficient sequence as target; in an application stage, the central brain signals of the diseased side are collected in real time to output the central intention latent space feature sequences, the mapping model is input to predict the collaborative activation coefficient sequence, and the multi-channel functional electrical stimulation instructions are reconstructed by using the collaborative primitive matrix and the collaborative activation coefficient sequence and applied to the limbs of the diseased side. The method realizes coordinated natural rehabilitation movements, and improves system robustness and individualization.

Description

Technical Field

[0001] This application relates to the fields of biomedical engineering and neurorehabilitation technology, and in particular to a neural signal decoding and coordinated stimulation control technology that combines brain-computer interface with functional electrical stimulation. Background Technology

[0002] Functional electrical stimulation (FES) is an important neurorehabilitation technique that applies low-intensity electrical currents to stimulate denervated muscles or peripheral nerves to assist patients with stroke, spinal cord injury, or cerebral palsy in performing functionally meaningful movements such as grasping, standing, or walking. In recent years, to enhance patient involvement and promote neuroplasticity, brain-computer interfaces (BCIs) have been introduced into FES systems, forming BCI-FES rehabilitation systems. In this system, the patient's motor intentions (usually detected via electroencephalography, EEG) are used as the source of commands to trigger or control FES.

[0003] However, existing BCI-FES systems still face significant technical bottlenecks in practical applications, especially in reproducing fine motor coordination movements of the upper limbs (such as grasping). In a typical rehabilitation scenario, for example, a hemiplegic patient wants to reach for a water glass. Physiologically, this requires extremely precise dynamic coordination of multiple forearm muscles (such as finger flexors, finger extensors, and the radial and ulnar wrist stabilizing muscles) in terms of time, space, and intensity. Most existing BCI-FES systems employ a "threshold trigger-fixed program" control logic, whereby the system decodes the patient's brain signals into a discrete "switch" command (e.g., when the detected ERD / ERS phenomenon related to motor intention exceeds a certain threshold, it is determined as "intention generation"). Once this "switch" is triggered, the system immediately executes a pre-set FES stimulation program with fixed timing and intensity (e.g., fixed stimulation of flexors at 70% intensity and wrist stabilizing muscles at 30% intensity for 2 seconds).

[0004] This "switch-on" control logic leads to specific technical problems in practical applications: 1. Stiff and uncoordinated movements (lack of spatial coordination): A natural grasp requires dynamic balance between flexor and extensor muscles (as antagonistic muscles) and coordinated activation of wrist stabilizing muscles. "Fixed programs," however, are often rigid and cannot simulate this complex spatial coupling between multiple muscles, resulting in mechanical and uncoordinated movements. For example, excessive wrist flexion or disordered finger flexion sequence may occur during grasping, preventing the formation of an effective "grasping" action. 2. Lack of fine-grained force and speed control (lack of temporal coordination): Patients' motor intentions are inherently continuous and dynamic. For example, they may want to grasp "lightly" (hold an egg) or "forcefully" (lift a kettle), or they may want to grasp "fast" or "slowly." Existing "switch-on" systems cannot recognize these continuous changes in intention. As long as the intention exceeds a threshold, the output FES is the same fixed program, preventing patients from exercising fine-grained and voluntary control over the intensity and speed of stimuli. 3. Low degree of personalization and poor adaptability: The preset stimulation programs are usually generic or manually adjusted by the therapist, which makes it difficult to adapt to the unique physiological characteristics (such as muscle strength, degree of spasticity, residual motor function) or inherent movement patterns of different patients.

[0005] Therefore, existing technologies urgently need to solve a specific technical problem: how to decode the patient's continuous and dynamic central motor intentions in real time into multi-channel, time-varying FES stimulation commands that conform to the patient's own physiological coordination rules, so as to get rid of "on-off" control and achieve coordinated and natural rehabilitation movements. Summary of the Invention

[0006] The purpose of this application is to provide a functional electrical stimulation method and system based on the collaborative decoding of central intention and peripheral synergistic patterns, so as to solve the problems mentioned in the background art.

[0007] This application discloses a functional electrical stimulation method based on the collaborative decoding of central intention and peripheral coordination patterns, including a training phase and an application phase. The output of each step is used as the input for subsequent steps. The method includes:

[0008] a) Training phase:

[0009] i. Simultaneously acquire the patient's central brain signals on the affected side and multi-channel electromyographic signals on the healthy side, wherein the patient generates motor intentions on the affected side and performs corresponding or mirrored movements on the healthy side;

[0010] ii. Central Intent Latent Space Feature Extraction: The affected side central brain signal from step i) is input into the central intent feature extraction module, which outputs a continuous low-dimensional central intent latent space feature sequence. ;

[0011] iii. Peripheral synergistic pattern decomposition: The contralateral multichannel electromyographic signal from step i) is decomposed to obtain the synergistic primitive matrix. With co-activation coefficient sequence ,in The dimension is lower than the number of channels in the healthy side multi-channel electromyography signal. As a teacher signal, the Stored;

[0012] iv. Intent-Co-mapping Model Training: Following step ii) For input, step iii) To achieve the goal, a mapping model is trained. ;

[0013] b) Application phase:

[0014] v. Real-time acquisition of central brain signals from the affected side of the patient, and outputting a real-time central intention latent space feature sequence using the central intention feature extraction module described in step ii). ;

[0015] vi. (The following is a continuation of step v) Input the mapping model of step iv). Output the predicted sequence of co-activation coefficients ;

[0016] vii. Utilize the cooperative primitive matrix stored in step iii). And step vi) prediction The reconstruction yields multi-channel functional electrical stimulation commands;

[0017] viii. Apply the multichannel functional electrical stimulation command from step vii) to the affected limb of the patient.

[0018] In a preferred embodiment, the decomposition in step iii) employs a nonnegative matrix factorization algorithm, satisfying an approximation relationship. ,in, This is a preprocessed multi-channel electromyography (EMG) signal matrix from the healthy side, with dimensions of [dimension number missing]. ; The co-factor matrix has dimensions of . ; The co-activation coefficient matrix has a dimension of . ; For the number of channels, For the number of collaborative modes, For time points.

[0019] In a preferred embodiment, the co-activation coefficient sequence Dimensions The number of channels is less than the number of channels in the multi-channel electromyography signal of the healthy side. And the dimension The variance explanation rate criterion is used to determine the ability to reconstruct... For the original The variance explained is not lower than the minimum preset threshold. value.

[0020] In a preferred embodiment, the central intent feature extraction module in step ii) is a deep neural network containing convolutional layers and / or recurrent layers, used to extract the spatiotemporal features of the affected side's central brain signals to form a continuous low-dimensional sequence. .

[0021] In a preferred embodiment, the mapping model in step iv) is a recurrent neural network, a long short-term memory network, or a gated recurrent unit, used to learn from... arrive The sequence-to-sequence nonlinear mapping.

[0022] In a preferred embodiment, the reconstruction in step vii) is achieved by calculating the cooperating primitive matrix. With the predicted co-activation coefficient sequence To achieve this, use the product of:

[0023]

[0024] in, The reconstructed multichannel functional electrical stimulation command vector has a dimension of ; The cooperative primitive matrix stored in step iii); The predicted co-activation coefficients in step vi) at time [time value missing] The value of , dimension is .

[0025] In a preferred embodiment, the central brain signals acquired in steps i) and v) on the affected side are electroencephalogram (EEG) signals.

[0026] In a preferred embodiment, step viii) includes converting the multichannel functional electrical stimulation command into stimulation pulse width, stimulation frequency, or stimulation amplitude parameters of the functional electrical stimulation device.

[0027] In a preferred embodiment, the method further includes: simultaneously acquiring actual motion feedback signals of the affected limb during the application phase, and adaptively adjusting the multichannel functional electrical stimulation commands by utilizing the error between the actual motion feedback signals and the expected motion pattern.

[0028] In a preferred embodiment, the decomposition in step iii) employs sparse nonnegative matrix decomposition or factor analysis.

[0029] In a preferred embodiment, the central brain signals acquired in steps i) and v) on the affected side are electrocorticography (ECG) signals or functional near-infrared spectroscopy signals.

[0030] In a preferred embodiment, both the training and application phases employ sliding time windows. Sequence modeling is performed to continuously represent the intensity and temporal changes of motion intention.

[0031] In a preferred embodiment, the for A continuous time series with each dimension corresponding to the change in the intensity of a cooperative activation mode over time, serving as the instructional input for cooperative reconstruction.

[0032] In a preferred embodiment, the healthy side multichannel electromyography signal undergoes preprocessing including filtering, rectification, and smoothing before decomposition to obtain the input matrix for nonnegative matrix factorization. .

[0033] In a preferred embodiment, the affected side central brain signal is subjected to spatial and temporal filtering before being input into the convolutional and / or recurrent layers. The spatial filtering includes cospatial pattern or discriminative spatial projection, and the temporal filtering is used to extract frequency band components related to motor intent.

[0034] In a preferred embodiment, the mapping model is trained using the co-activation coefficient sequence obtained in step iii). Supervised training is conducted using teacher signals, enabling the model to learn from... arrive The correspondence.

[0035] In a preferred embodiment, the number of collaborative modes It is an integer between 3 and 6.

[0036] In a preferred embodiment, the multichannel functional electrical stimulation command satisfies the following channel-by-channel expansion:

[0037]

[0038] in, Indicates the first Each stimulation channel at time The number of instructions; For the cooperative primitive matrix The Line 1 Column elements; For the first A collaborative activation mode at any time The intensity.

[0039] In a preferred embodiment, the cooperative primitive matrix stored in step iii) For the patient's personalized muscle synergy pattern spatial structure, step vii) calls the synergy primitive matrix. The same matrix is ​​stored in step iii).

[0040] In a preferred embodiment, the number of channels of the unaffected side multi-channel electromyography signal The number of channels is the same as that of the multi-channel functional electrical stimulation command, and each channel corresponds to the same or mirror-symmetrical muscle location.

[0041] In a preferred embodiment, the preset threshold is 90% or higher.

[0042] In a preferred embodiment, the processing flow of the deep neural network is as follows: the affected side central brain signal, after spatial and temporal filtering, is input into a convolutional neural network to extract local temporal features; the local temporal features are then input into a long short-term memory network to extract time series dependencies; and finally, the low-dimensional central intention latent spatial feature sequence is output. .

[0043] This application also discloses a functional electrical stimulation system based on the collaborative decoding of central intention and peripheral coordination patterns, comprising:

[0044] The data acquisition module is used to simultaneously acquire the patient's central brain signals on the affected side and multi-channel electromyographic signals on the healthy side of the peripheral brain;

[0045] The central intent feature extraction module is used to receive the central brain signals from the affected side and output a continuous low-dimensional central intent latent space feature sequence. Or real-time central intent latent space feature sequence ;

[0046] The peripheral synergistic mode decomposition module is used to decompose the multi-channel electromyographic signals of the healthy side peripheral region to obtain a synergistic primitive matrix. With co-activation coefficient sequence ,in The dimension is lower than the number of channels in the healthy side multi-channel electromyography signal. As a signal from the teacher;

[0047] Storage module, used to store the collaborative primitive matrix and the trained mapping model ;

[0048] The mapping model training module is used to train the model using the aforementioned model. For input, the The mapping model is obtained by training the target. ;

[0049] The collaborative activation prediction module is used to generate the real-time central intent latent space feature sequence. Input the mapping model Output the predicted sequence of co-activation coefficients ;

[0050] The functional electrical stimulation instruction reconstruction module is used to utilize the cooperative primitive matrix stored in the storage module. and the output of the co-activation prediction module The reconstruction yields multi-channel functional electrical stimulation commands; and

[0051] A functional electrical stimulation execution module is used to apply the multi-channel functional electrical stimulation commands to the patient's affected limb.

[0052] The method described in this application effectively solves the technical problems commonly found in existing brain-computer interface-controlled functional electrical stimulation systems, such as stiff movements, lack of coordination, poor robustness, and low personalization, through a systematic technical arrangement of the training and application phases and close coordination among various technical means, achieving outstanding technical effects in many aspects.

[0053] In addressing the problem of motor coordination, the core breakthrough of this application lies in employing peripheral signal decomposition technology based on muscle synergy theory. By performing nonnegative matrix decomposition on the multi-channel electromyography (EMG) signals of the healthy side, the superficial, high-dimensional, and redundant EMG phenomena are decomposed into a few low-dimensional, essential synergistic primitive matrices W and synergistic activation coefficient sequences H(t). This decomposition process is not a simple mathematical dimensionality reduction, but rather extracts the inherent muscle synergistic working patterns in the human motor control system. The synergistic primitive matrix W encodes the physiological knowledge of the spatial proportion at which each muscle should be activated synergistically, while the synergistic activation coefficient sequence H(t) describes the dynamic evolution of these synergistic patterns over time. Using H(t) as a teacher signal for subsequent mapping model training elevates the model's learning objective from "replicating surface EMG signals" to "learning underlying synergistic control strategies." This shift in objective fundamentally ensures that the system can understand and reproduce the inherent coordination of human movement. In the application phase, when reconstructing multi-channel functional electrical stimulation commands using matrix multiplication FES_Command(t) = W × Ĥ(t), the W matrix itself encodes the physiological coordination structure. Therefore, the reconstructed commands automatically follow the patient's own physiological coordination pattern in terms of temporal relationships, intensity ratios, and activation sequences among the channels, thus ensuring the endogenous coordination of stimulation both mathematically and physiologically. This coordination is not obtained through manual experience adjustments but is automatically inherited from the patient's healthy side physiological system through mathematical operations. Consequently, the resulting rehabilitation movements are significantly improved in terms of fluency, naturalness, and similarity to movements on the healthy side.

[0054] To address the system robustness issue, this application achieves effective suppression of noise and interference by simultaneously performing feature extraction and latent space mapping at both the input and output ends. At the input end, the central intent feature extraction module extracts the high-dimensional, noisy EEG signal into a low-dimensional central intent latent space feature sequence X_intent(t) through spatial filtering, temporal filtering, and multi-level processing using a deep neural network. This feature sequence has a higher signal-to-noise ratio and stronger discriminative power, and can continuously and stably represent the intensity and temporal changes of motor intent without being significantly affected by artifacts such as electrooculograms (EOG) and electromyograms (EMG). At the output end, the coactivation coefficient sequence H(t), being the main component obtained through NMF decomposition, has already filtered out a large amount of noise and redundant information from the original EMG signal, representing the main mode of motor control. More importantly, the mapping relationship learned by the intent-co-mapping model is established between two low-dimensional latent spaces that have undergone dimensionality reduction and denoising. This "low-dimensional to low-dimensional" mapping is statistically more stable than a direct "high-dimensional to high-dimensional" mapping, and it is more likely to converge to the global optimum during the optimization process. It is also less sensitive to random noise and outliers in the training data. Therefore, the trained model can still maintain high decoding accuracy and small prediction variance when facing adverse conditions such as signal perturbations, electrode displacement, and muscle fatigue during the testing phase, and the overall robustness of the system is substantially improved.

[0055] The low-dimensional latent space mapping architecture of this application offers significant advantages in improving training efficiency and reducing data requirements. By reducing high-dimensional EEG signals (e.g., 64 leads) and EMG signals (e.g., 8 channels) to low-dimensional intent features (e.g., 8 dimensions) and covariance coefficients (e.g., 4 dimensions), respectively, the parameter space that the mapping model needs to learn is significantly reduced. Taking the LSTM model as an example, the number of network parameters required for mapping from 8-dimensional input to 4-dimensional output is far less than that for a direct mapping from 64 dimensions to 8 dimensions. This not only significantly shortens the model training time but, more importantly, reduces the demand for training data. In the low-dimensional space, fewer training samples are sufficient to cover the main regions of the feature space, avoiding the "curse of dimensionality" problem prevalent in high-dimensional spaces. Furthermore, since both the input and output are high-quality signals after feature extraction, with high signal-to-noise ratios and strong correlations, the model can more easily discover and learn the inherent patterns between them, resulting in faster convergence and fewer iterations required during training. The combined effect of these factors enables this application to achieve satisfactory mapping accuracy in a relatively short training time and with relatively little training data, greatly reducing the burden on patients during the training phase and improving the system's practicality and clinical acceptability.

[0056] In terms of achieving a high degree of personalization, the technical solution of this application has unique advantages. The collaborative primitive matrix W is extracted entirely from the actual movement data of the patient's own healthy limb using the NMF algorithm. It reflects the patient's unique muscle distribution, muscle strength level, nerve innervation pattern, and long-established movement habits. Due to differences in age, gender, degree of injury, rehabilitation stage, and other factors, the W matrix of different patients will show significant individual differences in spatial structure. The system of this application extracts and stores the W matrix independently for each patient and uses the patient-specific W matrix for collaborative reconstruction during the application stage. This ensures that the generated functional electrical stimulation commands are naturally adapted to the patient's physiological characteristics. This personalization is not achieved through manual parameter adjustment, but through automatic learning from the patient's own data, thus possessing higher physiological rationality and effectiveness. At the same time, the central intention feature extraction module and mapping model can also be trained individually for each patient, so that the entire system, from signal acquisition, feature extraction, mapping prediction to stimulation reconstruction, fully considers the individual differences of the patient at every stage, realizing a truly customized rehabilitation program.

[0057] In achieving continuous dynamic control, this application overcomes the limitations of traditional "on / off" control. By employing a sliding time window mechanism to continuously extract and update the central intention latent space feature sequence X_intent(t), the system can track the intensity changes and temporal evolution of the patient's motor intention in real time with high temporal resolution. The co-activation coefficient sequence Ĥ(t) output by the mapping model is also a continuous time-varying signal, rather than a discrete state label. This continuity continues until the final functional electrical stimulation command FES_Command(t), enabling the stimulation intensity of each channel to change smoothly and gradually, rather than abruptly jumping. The technical effects of continuous control are multifaceted: First, the generated movements are smoother and more natural, avoiding the shock of sudden starts or stops; second, the system can achieve fine force adjustment, allowing patients to flexibly control the grip strength or movement speed according to task requirements; third, the continuous intention-action association helps strengthen the patient's subjective sense of control and motor-sensory feedback loop, which is of great significance for inducing neural plasticity and promoting the recovery of motor function. This continuous control capability is something that "switching" systems cannot achieve, demonstrating the significant advancements in control precision and user experience of this application.

[0058] The various technical methods in this application are not isolated but form an organically integrated technical system. Peripheral coordination pattern decomposition provides high-quality teacher signals for mapping learning, while central intent feature extraction provides high-quality input signals for mapping learning. Together, they ensure that the mapping model can learn stable and accurate intent-coordination relationships. The coordination primitive matrix W stored during the training phase is reused for coordination reconstruction during the application phase, realizing the effective transfer of training knowledge to application execution. The mapping model adopts a sequence modeling architecture (such as LSTM), enabling the system to capture not only instantaneous intent-coordination correspondences but also the patterns of temporal evolution. This, combined with the continuous feature extraction mechanism of the sliding window, achieves continuous dynamic control. It is the synergistic cooperation of these technical methods under a specific architecture that enables this application to achieve significantly better results than existing technologies in multiple dimensions such as coordination, robustness, training efficiency, personalization, and control fineness, providing a more effective, natural, and personalized solution for the rehabilitation of patients with neurological injuries.

[0059] The specification of this application contains numerous technical features distributed across various technical solutions. Listing all possible combinations of these technical features (i.e., technical solutions) would make the specification excessively lengthy. To avoid this problem, the various technical features disclosed in the above-described invention, the various technical features disclosed in the following embodiments and examples, and the various technical features disclosed in the accompanying drawings can be freely combined to form various new technical solutions (all of which are considered to have been described in this specification), unless such a combination of technical features is technically infeasible. For example, one example discloses feature A+B+C, and another example discloses feature A+B+D+E. Features C and D are equivalent technical means that serve the same function, and technically only one needs to be used; they cannot be used simultaneously. Feature E can technically be combined with feature C. Therefore, the solution A+B+C+D should not be considered as described because it is technically infeasible, while the solution A+B+C+E should be considered as described. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating a functional electrical stimulation method based on the collaborative decoding of central intent and peripheral synergistic patterns according to the first embodiment of this application.

[0061] Figure 2 This is a schematic diagram of the structure of a functional electrical stimulation system based on the collaborative decoding of central intent and peripheral synergistic modes according to the second embodiment of this application. Detailed Implementation

[0062] In the following description, many technical details are presented to help the reader better understand this application. However, those skilled in the art will understand that the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments.

[0063] Explanation of some concepts:

[0064] Functional electrical stimulation (FES) is a technique that induces functionally meaningful muscle contractions by applying external electrical pulses to the muscles or nerves of paralyzed or impaired limbs. FES stimulation parameters include pulse amplitude, pulse width, and stimulation frequency; precise control of these parameters allows for adjustment of the intensity and duration of muscle contractions.

[0065] Brain-Computer Interface (BCI): A technology system that extracts a user's intentions or commands directly from brain signals without relying on peripheral nerves and muscles, and converts them into control signals for external devices. BCI typically includes signal acquisition, signal processing, feature extraction, intention decoding, and feedback.

[0066] Electroencephalography (EEG) is a non-invasive neural signal acquisition technique that records the electrical activity of neuronal populations in the cerebral cortex using electrodes placed on the scalp. EEG signals reflect neural activity patterns in different brain regions under various cognitive or motor tasks and are one of the most commonly used input signals for brain-computer interface (BCI) systems.

[0067] Surface electromyography (sEMG) records the electrophysiological signals generated during muscle contraction using electrodes placed on the skin surface. The amplitude and frequency characteristics of sEMG signals reflect the activation level and contractile state of muscles, making it an important tool for studying motor control and assessing muscle function.

[0068] Muscle Synergy: The muscle synergy theory posits that when the central nervous system controls complex movements, it does not independently control each muscle, but rather achieves movement by activating a few pre-coded "synergistic modules." Each synergistic module represents a group of muscles activated in a specific spatial proportion. This synergistic control strategy greatly simplifies the complexity of motor control while ensuring the coordination of movements.

[0069] Non-negative matrix factorization (NMF) is a matrix factorization algorithm that decomposes a non-negative matrix into the product of two non-negative matrices. In muscle synergy analysis, NMF is used to decompose multi-channel sEMG signal matrices into synergistic primitive matrices (describing the spatial combination patterns of muscles) and synergistic activation coefficient matrices (describing the temporal activation patterns of synergistic patterns). The non-negativity constraint of NMF allows for a clear physiological interpretation of the decomposition results.

[0070] Common Spatial Pattern (CSP): A spatial filtering algorithm for EEG signal processing. Its goal is to find a set of spatial filters that maximizes the variance difference of EEG signals from different motion states (such as left-hand and right-hand movements) under the action of these filters. CSP is one of the most commonly used feature extraction methods in motion imagery BCI.

[0071] Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network (RNN) that selectively memorizes and forgets information in a sequence by introducing gating mechanisms (input gate, forget gate, output gate), thus effectively solving the gradient vanishing or gradient exploding problems faced by traditional RNNs when learning long-term dependencies. LSTMs are particularly suitable for processing and learning time-series data.

[0072] Latent space features are low-dimensional abstract representations extracted from high-dimensional raw data through feature extraction or dimensionality reduction algorithms. Latent space features typically have a higher signal-to-noise ratio, stronger discriminative power, and more explicit semantic meaning than the original data, making them more suitable as inputs or outputs of machine learning models.

[0073] Synergy Matrix: Within the framework of muscle synergy theory, this matrix is ​​extracted from multi-channel electromyographic signals using matrix factorization algorithms (such as NMF). Each column represents the spatial structure of a synergistic pattern, i.e., the relative activation ratio of each muscle within that pattern. The Synergy Matrix encodes the spatial organization rules of motor control.

[0074] Co-activation coefficient: In muscle synergistic decomposition, it describes the time series of changes in activation intensity of various synergistic modes over time. The dimensionality of the co-activation coefficient is usually much smaller than the number of channels in the original electromyographic signal; it represents the temporal evolution of motor control.

[0075] Variance Accounted For (VAF): A metric used to evaluate the quality of matrix factorization or model fitting, defined as the percentage of variance explained by the reconstructed signal relative to the original signal. A VAF value closer to 100% indicates a closer reconstructed signal to the original signal, signifying higher quality decomposition or fitting. In muscle synergy analysis, VAF is often used to determine the optimal number of synergistic patterns.

[0076] The following is a brief summary of some of the innovative aspects of this application:

[0077] In summary, the technical problem addressed in this application is highly complex and unique. It is not a simple signal transmission or mapping issue, but rather involves establishing a closed-loop control link between the motor intention of the affected brain (represented by signals from the affected central brain) and the coordinated motor output of the affected limb, ensuring both real-time performance and physiological coordination. The fundamental reason why traditional BCI-FES systems produce stiff and uncoordinated movements is that these systems attempt to directly establish a mapping between two high-dimensional, high-noise, and high-redundancy raw signal spaces (64-lead EEG and 8-channel sEMG), or more simply, use a threshold-triggered fixed program approach. Both of these methods ignore the essential laws of human motor control—that is, the central nervous system does not achieve movement by independently controlling each muscle, but by activating a few pre-coded coordinating modules to achieve the coordinated execution of complex movements. The technical concept of this application precisely grasps this physiological essence and creatively transforms it into a complete engineering implementation scheme.

[0078] Specifically, in the training phase, this application performs non-negative matrix decomposition on the multi-channel electromyographic signal of the healthy side through step iii (i.e., step 300). The inventive significance of this step lies in the fact that it is not simply a dimensionality reduction or filtering of the signal, but rather a profound physiological insight based on the muscle synergy theory, which decomposes the high-dimensional, superficial electromyographic phenomenon ( The matrix is ​​stripped down into two low-dimensional, essential physiological control elements—the synergistic primitive matrix. (214) and co-activation coefficient sequence (215). Here (214) is not a feature transformation matrix in the general sense, but rather encodes the patient's personalized spatial structure knowledge of "how muscles work together," with each column defining a physiologically self-consistent muscle combination pattern; and (215) represents the temporal activation trajectory of these collaborative patterns, and its dimensions are... Much smaller than the original number of channels This dimensionality reduction is not a loss of information, but rather the extraction of the "main theme" of motion control from redundant noise. The key innovation lies in... (215) In this application, the special status of “teacher signal” means that the goal of the subsequent mapping learning (step iv, i.e. step 400) is no longer “replicating surface electromyographic signals”, but “learning underlying collaborative control strategies”. This essential change in goal elevates the entire system from a “signal replicator” to a “motor strategy transferor”.

[0079] Meanwhile, the processing of the affected side's central brain signals in step ii (i.e., step 200) also demonstrates a high degree of ingenuity. This application does not simply use traditional frequency band energy features (such as ERD / ERS) as intent representation, but instead designs a multi-level feature extraction module that includes spatial filtering (such as CSP), temporal filtering, and a deep neural network (CNN-LSTM hybrid architecture) to automatically learn and extract a low-dimensional, continuous, and highly discriminative central intent latent spatial feature sequence from high-dimensional noisy EEG signals. (203). The non-obviousness of this feature extraction process lies in the fact that it does not use predefined fixed features (such as the power of a certain frequency band), but rather adaptively learns abstract representations from the data through deep learning that are most strongly correlated with the motion intention. These representations continuously and dynamically reflect the entire temporal evolution of the intention from nothing to something, from weak to strong, and from preparation to execution, rather than simply a binary state of "on" or "off." More importantly, (203) dimensions and The dimensions of (215) are all significantly lower than the dimensions of their respective original signals, but the dimensional relationship between the two (e.g.) It is 8-dimensional. The 4-dimensionality is not arbitrarily set, but rather emerges naturally through their respective independent optimization processes (maximizing the discriminative power of CSP and maximizing the variance explanatory power of NMF). This architecture of "independent dimensionality reduction on both sides and intermediate mapping connection" itself constitutes a non-obvious technical solution.

[0080] Mapping model in step iv (i.e., step 400) The training process of (222) involves aligning and associating the two latent spaces in the time dimension. The innovation here lies in the fact that this mapping is not a complex high-dimensional to high-dimensional mapping (64-dimensional EEG to 8-dimensional sEMG) in the original signal space, but rather a mapping between two low-dimensional latent spaces that have undergone physiologically heuristic dimensionality reduction (8-dimensional...). To 4D This "low-dimensional to low-dimensional" mapping is mathematically more tractable (easier to optimize), statistically more robust (insensitive to noise), and physiologically more meaningful (learning the essential association between intention and collaborative strategy, rather than the accidental correlation between intention and surface electromyography). The mapping model uses sequence models such as LSTM, which can capture... (203) to (215) The long-term dependency between them enables the system to not only understand "the current intention corresponds to the current collaboration", but also "the evolution trajectory of the intention corresponds to the evolution trajectory of the collaboration", thereby achieving truly continuous and smooth motion decoding.

[0081] In step vii (i.e., step 700) of the application phase, the formula is used. The matrix multiplication operation achieves collaborative reconstruction, a seemingly simple linear operation that actually contains profound technical implications. The key lies in: (214) is a "physiological blueprint" that encodes the patient's personalized synergistic structure, extracted and stored from the patient's healthy side during the training phase; and It is a "control command" representing the intensity of the current intention, predicted in real time from the patient's affected side brain during the application phase. The product of these two commands is mathematically a reconstruction from a low-dimensional synergistic space to a high-dimensional muscle space, but physiologically it is the process of expanding the "affected side intention" into a "multi-channel stimulation command for the affected side" according to the "healthy side synergistic rule." The non-obviousness of this reconstruction lies in its guarantee of the reconstructed multi-channel FES command. The timing relationships, intensity ratios, and activation sequences between the various channels strictly follow the physiological coordination pattern of the patient's own healthy side, thereby fundamentally ensuring the coordination of stimulation—a coordination that is not achieved through manual adjustment or experience-based design, but rather an intrinsic coordination that is automatically inherited from the patient's own physiological system through mathematical calculations.

[0082] The various technical features of this application form a tightly coupled, progressively layered technical chain: step iii (300) transforms the "appearance" of multi-channel electromyographic signals into their "essence" through NMF decomposition. (214) and (215) provides high-quality teacher signals for subsequent learning; step ii (200) purifies the "noise" of the affected side's central brain signals into "intention" through deep feature extraction. (203) provides high-quality input signals for subsequent learning; step iv (400) establishes a nonlinear relationship between the two low-dimensional latent spaces through a sequence mapping model. (222) realizes the knowledge transfer from intent to collaborative strategy; step vii (700) utilizes the stored information through collaborative reconstruction. (214) Predicted co-activation Convert to multi-channel coordination instructions This approach achieves a mapping from abstract strategies to concrete execution. These four steps are indispensable and interdependent: without the collaborative decomposition in step iii, the teacher signal remains high-dimensional noise, making mapping difficult to learn; without the intent extraction in step ii, the input signal remains low-discriminative features, making accurate mapping difficult; without the low-dimensional mapping in step iv, the association between intent and collaboration cannot be efficiently established; and without the collaborative reconstruction in step vii, the predicted abstract collaboration cannot be transformed into actual multi-channel coordinated stimuli. It is precisely the organic combination and synergistic cooperation of these four stages that enables this application to fundamentally solve the core problem of "stiff and uncoordinated movements" in existing technologies.

[0083] From a technical perspective, this application achieves several prominent and interrelated beneficial effects. Firstly, because the FES instructions are based on the patient's own physiological coordination pattern (… (214) is reconstructed, and its mathematical structure naturally inherits the synergistic characteristics of the healthy side's natural movements. Therefore, the resulting rehabilitation movements far surpass existing fixed-procedure stimulation or simple mapping methods in terms of coordination, fluency, and naturalness. Secondly, since mapping learning is carried out between two low-dimensional latent spaces, the number of model parameters is greatly reduced, the training data requirement is significantly reduced, the training time is significantly shortened, and the model's robustness to signal noise and diurnal variability is significantly improved. Thirdly, because (214) Originating entirely from the patient's own healthy side, the system is highly personalized for each patient, automatically adapting to their unique muscle distribution, strength levels, and movement habits. Finally, because the system achieves continuous intent decoding and continuous collaborative reconstruction, the generated stimulus commands are smooth and gradual, rather than abrupt and jumpy. This continuous control is more in line with the characteristics of natural human movement and is more conducive to inducing neural plasticity. These technical effects are not simply a linear superposition of technical means, but rather the inevitable result of the synergistic cooperation of various technical features under the unique technical concept of this application, reflecting the substantial characteristics and significant progress of this application compared to existing technologies.

[0084] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0085] This embodiment provides a functional electrical stimulation method based on the collaborative decoding of central intention and peripheral synergistic patterns. The core idea of ​​this method is to learn the physiological synergistic movement patterns of the patient's unaffected limbs during the training phase and establish a mapping relationship from the brain's motor intention on the affected side to this synergistic pattern. This allows for the generation of multi-channel functional electrical stimulation commands with endogenous coordination in real time based on the patient's motor intention during the application phase. Unlike the "on / off" control method in existing technologies that uses preset fixed stimulation programs, this application achieves continuous and coordinated control from central intention to peripheral synergistic movements through dual latent space mapping, resulting in more natural and fluid rehabilitation movements.

[0086] In this embodiment, the patient is a stroke rehabilitation patient with right-sided hemiplegia. The patient's left (unaffected) upper limb function is normal, while the right (affected) upper limb motor function is impaired but still retains some motor intention ability. The rehabilitation goal is to enable the patient, through the method of this application, to use the motor intention of their right brain to control the functional electrical stimulation system to stimulate the muscles of the right forearm, thereby completing daily living activities such as grasping.

[0087] The training and calibration phase is described below.

[0088] The purpose of the training and calibration phase is to collect bilateral neural signal data from the patient, extract key features, and train a model capable of establishing a mapping relationship between central intentions and peripheral coordination patterns. This phase is conducted offline, requiring the patient to simultaneously use the affected side to generate motor intentions and the unaffected side to execute the actual movements.

[0089] like Figure 1 As shown, step 100: Synchronous data acquisition

[0090] Furthermore, in step 100 of the training phase, the system simultaneously acquires central brain signals from the affected side and multi-channel electromyographic signals from the unaffected side. Specifically, the patient is instructed to attempt to generate a grasping motion using the affected (right) limb, guided by visual or auditory cues provided by the system, while simultaneously performing the same grasping action using the unaffected (left) limb. This synchronous training paradigm is based on the mirror neuron theory, which states that the bilateral motor cortex produces similar activation patterns when performing or imagining the same action. In this embodiment, each training trial lasts approximately 5 seconds, including a preparation phase, an execution phase, and a rest phase, with multiple trials repeated.

[0091] Step 110: Central brain signal acquisition

[0092] For the acquisition of central brain signals, this embodiment uses a multi-lead electroencephalography (EEG) system, with electrodes arranged according to the international 10-20 system and a sampling rate set to 1000Hz. The key brain regions monitored include the motor cortex of the cerebral hemisphere corresponding to the affected limb, particularly electrode locations related to motor control. EEG signals reflect central neural activity when the patient attempts to generate movement in the right limb.

[0093] Step 120: Peripheral electromyography signal acquisition

[0094] For the acquisition of peripheral electromyographic signals, this embodiment places multiple surface electromyography (sEMG) electrodes on the patient's unaffected (left) forearm, corresponding to the main muscles involved in grasping movements, including: flexor digitorum superficialis, flexor digitorum profundus, flexor pollicis longus, flexor carpi radialis, flexor carpi ulnaris, extensor digitorum commonis, extensor carpi radialis, and extensor carpi ulnaris. The sampling rate is also set to 1000Hz, and the signals are acquired synchronously with EEG. These sEMG signals realistically record the muscle activation patterns of the patient's unaffected limb when performing natural grasping movements, reflecting the body's inherent coordinated control strategies.

[0095] It should be noted that in this embodiment, the patient performs a "corresponding action," meaning that both hands attempt to grasp. In other embodiments, if the patient's motor injury is unilateral, a "mirror image" mode can be used; for example, when the patient attempts to grasp with their right hand, the left hand performs a corresponding mirror image grasping action. Regardless of the mode used, the key is to ensure that the actual movement performed on the healthy side matches the motor intention on the affected side in both time and function.

[0096] Step 200: Extraction of latent spatial features of central intent

[0097] Step 200 of the training phase involves extracting a continuous low-dimensional central intent latent space feature sequence from the acquired central brain signals from the affected side. While the raw EEG signal contains rich neural activity information, its high dimensionality, low signal-to-noise ratio, and large amount of irrelevant information (such as eye movement artifacts and electromyographic interference) make direct decoding using the raw EEG inefficient and unstable. Therefore, this application introduces a central intent feature extraction module, which automatically extracts low-dimensional, robust feature representations strongly correlated with motor intent from high-dimensional, noisy EEG signals.

[0098] Step 210: Brain signal preprocessing

[0099] Specifically, the central intent feature extraction module in this embodiment employs a deep neural network structure containing convolutional and recurrent layers, more specifically, a hybrid architecture of convolutional neural networks (CNN) and long short-term memory networks (LSTM). This hybrid architecture can simultaneously capture the spatial patterns and temporal dynamic characteristics of EEG signals.

[0100] Furthermore, before inputting the raw EEG signals into the deep neural network, preprocessing and feature extraction steps are required. First, the raw EEG signals are bandpass filtered to retain frequency bands highly correlated with motor imagery and execution, including μ and β rhythms. Then, spatial filtering is performed using the Common Spatial Pattern (CSP) algorithm. CSP is a discriminative spatial projection method that aims to find a set of spatial filters that maximizes the variance difference between EEG signals in the motor intention state and the resting state. Through CSP transformation, multi-lead EEG signals are projected into a lower-dimensional space, resulting in projected signals with higher discriminative power and signal-to-noise ratio.

[0101] For example, the CSP transform can be expressed as:

[0102]

[0103] in, This is the original multi-lead EEG signal matrix, whose dimension is the original number of channels × the number of time points. This is the CSP spatial filter matrix, with superscript... Indicates matrix transpose. This is a dimensionality-reduced time series after spatial filtering. This transformation projects the high-dimensional EEG signal into a low-dimensional space that can best distinguish different motion states.

[0104] Step 220: Deep Neural Network Feature Extraction

[0105] After spatial and temporal filtering, the affected side's central brain signal is then input into a deep neural network. The network's processing flow is as follows: First, the preprocessed signal is divided into sliding time windows. The signal within each time window is flattened into a one-dimensional vector and then input into a one-dimensional convolutional neural network (1D-CNN). The CNN layer contains several convolutional kernels used to extract local temporal features, which reflect the pattern changes of the EEG signal on short time scales.

[0106] The output of the convolutional layer, after being processed by max pooling and activation functions, is fed into a Long Short-Term Memory (LSTM) network. The LSTM layer is used to extract time-series dependencies. Through its unique gating mechanism (input gate, forget gate, and output gate), the LSTM can selectively remember and forget information in the EEG feature sequence, thereby capturing the dynamic evolution of motion intent from the preparation phase to the execution phase.

[0107] Finally, the output of the LSTM is passed through a fully connected layer to reduce the dimensionality and obtain the central intent latent space feature sequence. This low-dimensional feature vector is updated at each time step, continuously representing the intensity and temporal changes of the patient's motor intention. Compared to the original multi-lead EEG signal, With lower dimensionality, higher signal-to-noise ratio, and stronger correlation with motion intent, it lays a good foundation for subsequent mapping learning.

[0108] Step 300: Decomposition of Peripheral Cooperative Patterns

[0109] Step 300 of the training phase involves decomposing the acquired multi-channel electromyographic signals from the healthy side to extract its inherent muscle synergy activation patterns. This step is one of the core innovations of this application, based on muscle synergy theory. This theory posits that during complex movements, the central nervous system does not independently control every muscle, but rather activates a few pre-coded "synergy modules," each representing a group of muscles with a fixed proportion of synergistic activation patterns. This synergistic control strategy greatly simplifies the complexity of motor control while ensuring the coordination and stability of movements.

[0110] Step 310: Electromyographic signal preprocessing

[0111] In this embodiment, the multi-channel sEMG signal acquired from the healthy side is first preprocessed. The preprocessing includes the following steps: bandpass filtering to retain effective frequency bands and remove motion artifacts and high-frequency noise; full-wave rectification to obtain the absolute value of the signal; low-pass filtering and downsampling to smooth the rectified signal, obtaining a signal reflecting the muscle activation envelope. After preprocessing, an electromyography signal matrix is ​​obtained. Its dimension is the number of channels × the number of time points.

[0112] Step 320: Nonnegative matrix factorization

[0113] Furthermore, this embodiment employs the nonnegative matrix factorization (NMF) algorithm to process the matrix. Decompose the non-negative matrix. NMF is an unsupervised learning algorithm whose goal is to decompose the non-negative matrix. It can be approximately decomposed into the product of two nonnegative matrices:

[0114]

[0115] in, This is a preprocessed multi-channel electromyography (EMG) signal matrix from the healthy side, with dimensions of [dimension number missing]. , For the number of channels, For time points; This is a synergy matrix with dimensions of . Each column represents the spatial structure of a synergistic pattern, that is, the relative activation ratio of each muscle in the pattern. This is the co-activation coefficient matrix, with dimensions of . Each row represents the change in the activation intensity of a collaborative mode over time. The number of collaborative patterns is usually much smaller than .

[0116] The NMF algorithm solves the problem by iteratively optimizing the following objective function. and :

[0117]

[0118] in, This represents the Frobenius norm, which is the square root of the sum of the squares of all elements of a matrix, and is used to measure reconstruction error. and This indicates that all elements in the matrix are non-negative. The algorithm typically uses a multiplicative update rule for iterative solving until the objective function converges to a local optimum.

[0119] Step 330: Determine the number of collaborative modes

[0120] It should be noted that the number of collaborative modes The choice of [aspect name] is crucial to the decomposition results. Values ​​that are too small will fail to adequately represent the complexity of the original electromyographic signal. Excessively large values ​​introduce redundancy and noise. This embodiment uses the Variance Accounted For (VAF) criterion to determine the optimal value. Value. VAF is defined as:

[0121]

[0122] in, For the sum of squares of the reconstruction error, The VAF is the sum of squares of the original signal, reflecting the reconstructed signal. For the original signal The goodness of fit is denoted by a value closer to 100%, indicating a better fit. In this embodiment, the preset threshold is set to 90%. The procedure in this embodiment is as follows: from... Start to gradually increase Value, for each The value is used to calculate the corresponding VAF. When the VAF first reaches or exceeds the preset threshold (90%), this option is selected. The value represents the final number of synergistic patterns. This means that when a patient's unaffected limb performs an action, the complex activation patterns of multiple muscles can be reproduced with high fidelity through a linear combination of a few basic synergistic patterns.

[0123] Step 340: Storage and Application of Collaborative Features

[0124] The co-factor matrix obtained by decomposition The matrix contains a spatial structure with several synergistic patterns. For example, the first synergistic pattern might primarily involve the synergistic activation of the finger flexor muscles, the second might involve the wrist stabilizing muscles, and other synergistic patterns correspond to different stages of extensor-flexor synergistic coordination. It is a patient's personalized spatial structure of muscle coordination patterns, which encodes the patient's unique motor control strategies. The matrix is ​​stored during the training phase and will be used to reconstruct functional electrical stimulation instructions during the application phase.

[0125] Meanwhile, the decomposition yields the cooperative activation coefficient matrix The temporal activation patterns of several cooperative modes throughout the entire action process are described. It is a low-dimensional time-varying signal that reflects the activation intensity of each cooperative mode at each time step. Compared with the original multidimensional sEMG signal, It has lower dimensionality, higher signal-to-noise ratio, and more explicit physiological meaning. In this application, It is used as a "teacher signal," that is, the target of the mapping model's learning. The key innovation of this design is that the goal of model learning is no longer to replicate surface electromyographic phenomena, but to learn the underlying collaborative control strategies.

[0126] Step 400: Intent-Co-mapping Model Training

[0127] Step 400 in the training phase involves training a mapping model that learns the mapping relationship from central intent latent space features to peripheral co-activation coefficients. This is the core of this application's implementation of continuous "intent-action" decoding.

[0128] Step 410: Mapping Model Construction

[0129] Specifically, the mapping model uses the central intent latent space feature sequence extracted in step 200. As input, the sequence of co-activation coefficients obtained in step 300. The objective is to perform supervised learning training. This is a sequence-to-sequence nonlinear mapping problem: given a low-dimensional time series representing the intention of a motion... The model needs to predict the lower-dimensional time series representing the co-activation of the corresponding representations. .

[0130] This embodiment uses a Long Short-Term Memory (LSTM) network as the main architecture of the mapping model. LSTM is an improved form of Recurrent Neural Network (RNN), particularly adept at processing and learning long-term dependencies in time-series data. The reason for choosing LSTM is that both motor intention and muscle activation are continuous temporal processes with obvious temporal dependencies; for example, actions include different phases such as preparation, execution, and maintenance. The gating mechanism of LSTM can selectively remember and forget information in the sequence, thereby effectively capturing the complex temporal correspondence between intention signals and coactivation.

[0131] More specifically, the LSTM model in this embodiment includes several LSTM layers and a fully connected output layer. The LSTM layer receives... The sequence is used as input, and high-level temporal features are extracted layer by layer. Finally, a fully connected layer maps the features to an output dimension equal to the number of cooperating modes, corresponding to the activation coefficients of each cooperating mode. The network outputs a prediction vector at each time step, constituting the predicted... sequence.

[0132] Step 420: Supervised Training Process

[0133] The model is trained using supervised learning. The training data consists of paired samples from multiple trials, each sample being an aligned time series pair: the input is... (Extracted from EEG on the affected side), the target is (Originated from healthy side sEMG decomposition). The loss function uses mean squared error (MSE):

[0134]

[0135] in, The total number of training samples, These are the co-activation coefficients predicted by the model. The true cooperative activation coefficients obtained from NMF decomposition. Let L2 be the squared norm of a vector, which is the sum of the squares of all its elements. The model parameters are progressively adjusted using the backpropagation algorithm to minimize the prediction error.

[0136] For example, during training, when a patient attempts to make a grasping intention using the affected side (right side) and performs the grasping action using the unaffected side (left side), the system simultaneously records the EEG generated on the affected side. and the sEMG generated by the healthy side The goal of model learning is: when only the affected side is affected in the future... At the same time, it can accurately predict what will happen. This drives the affected side to produce movements that are as coordinated as the healthy side.

[0137] It is important to emphasize that this mapping is established between two "low-dimensional latent spaces": the input is low-dimensional intent features (rather than the original high-dimensional EEG), and the output is low-dimensional co-equalization coefficients (rather than the original multi-dimensional sEMG). This "low-dimensional to low-dimensional" mapping has multiple advantages: it significantly reduces model complexity, making training faster, more stable, and requiring less data; both input and output are high signal-to-noise ratio signals after feature extraction, allowing the model to learn a more fundamental and stable mapping relationship; and the model has stronger generalization ability and is more robust to signal perturbations during the testing phase.

[0138] After training, a trained mapping model is obtained. The model's parameters are saved for real-time decoding during the application phase.

[0139] The following describes the real-time application phase.

[0140] The goal of the real-time application phase is to enable patients to control the functional electrical stimulation system to provide coordinated stimulation to the affected limb in real time simply by generating the intention to move the affected limb, thereby producing natural rehabilitation movements. At this stage, patients no longer need to move their healthy limbs.

[0141] Step 500: Real-time central signal acquisition and intent decoding

[0142] Step 500 in the application phase involves real-time acquisition of central brain signals from the affected side of the patient and extraction of real-time latent spatial feature sequences of central intention. Specifically, the patient sits in a comfortable chair and only needs to concentrate on attempting to generate a grasping motion intention with the affected side (right side) limb, without performing any actual limb movements. The system acquires EEG signals from the motor cortex of the affected side (right side) of the brain in real time, with the acquisition settings identical to those in the training phase.

[0143] The real-time acquired EEG signals are continuously fed into the central intent feature extraction module, which has already been trained during the training phase. This module operates according to the same processing flow as the training phase: first, bandpass filtering and CSP spatial filtering are performed; then, the signal is divided into sliding time windows; the signal in each window is processed by a CNN-LSTM network to extract features; finally, the central intent latent space feature sequence is output in real time. The superscript "^" here indicates that this is a real-time predicted value.

[0144] Thanks to the sliding window mechanism, the system can continuously update intent features with high temporal resolution, thus enabling continuous representation of the intensity and temporal changes of motor intent. For example, as the patient gradually strengthens their grasping intent, Some dimensions show a gradual increasing trend; when the patient maintains a stable grasping intention, It will remain at a relatively stable level; when the patient intends to relax, It will gradually return to the baseline level. This continuous, dynamic decoding of intent is the basis for the natural coordinated actions achieved in this application.

[0145] Step 600: Co-activation prediction

[0146] Step 600 in the application phase involves using the trained mapping model to predict the corresponding co-activation coefficients based on the real-time decoded intent features. Specifically, this involves using the central intent latent space feature sequence output in real-time from step 500. The mapping model obtained by inputting into the training phase In the middle, the model outputs the predicted sequence of co-activation coefficients in real time. .

[0147] Due to the mapping model It is an LSTM network that considers the intent features from the current time step and several previous time steps, thereby capturing the temporal evolution of the intent signal. The model's output... It is a low-dimensional continuous time series, where each dimension corresponds to the change in the intensity of a co-activation mode over time. This sequence serves as the instructive input for co-reconstruction, carrying control instructions decoded from the central intent regarding "how to coordinate and activate multiple muscles".

[0148] It should be pointed out that, Although the signals are derived from electromyography (EMG) signals from the healthy side during the training phase, they are predicted from brain signals from the affected side during the application phase. This embodies the core idea of ​​"cross-modal transfer" in this application: learning the movement patterns from the healthy side as "teacher signals" during the training phase, and transferring these patterns to the affected side during the application phase, so that even if the affected side cannot generate coordinated movements voluntarily, it can still generate movements with the same level of coordination as the healthy side through the assistance of the FES system.

[0149] Step 700: Generation of multi-channel functional electrical stimulation instructions for collaborative reconstruction

[0150] Step 700 in the application phase involves reconstructing multichannel functional electrical stimulation commands using the predicted co-activation coefficients and the co-activation primitive matrix stored during the training phase. This is a key step in achieving coordinated stimulation in this application.

[0151] Specifically, the system first calls the patient's personalized collaborative primitive matrix stored in step 300 of the training phase. This matrix encodes the spatial structure of muscle synergy patterns in the patient's unaffected limb, defining the proportion at which each muscle should be activated within each synergy pattern. The system then uses this... The matrix and the co-activation coefficients predicted in step 600 Matrix multiplication is performed to reconstruct multi-channel functional electrical stimulation commands in real time.

[0152]

[0153] in, The reconstructed multichannel functional electrical stimulation command vector has a dimension of , Number of stimulation channels; The co-operative primitive matrix stored in step 300 of the training phase has a dimension of , This refers to the number of collaborative modes; The co-activation coefficients predicted in step 600 of the application phase at time [time value missing] The value of , dimension is Through this matrix multiplication, low-dimensional coactivation instructions are mapped back to a high-dimensional multichannel stimulation space.

[0154] Expanding formula (6), we can obtain the channel-wise expression for the instruction quantity of each stimulation channel:

[0155]

[0156] in, Indicates the first Each stimulation channel at time The amount of instructions, The value range is 1 to ; For the cooperative primitive matrix The Line 1 Column element, representing the first The first collaborative mode is for the first The contribution weight of bulk muscles; For the first A collaborative activation mode at any time The strength, The value range is 1 to ; Indicates to From 1 to Summation of .

[0157] For example, suppose that at a certain moment, the predicted co-activation coefficients show that the first co-activation mode is strongly activated, the second co-activation mode is moderately activated, and the other co-activation modes are weakly activated. Further assume the co-activation primitive matrix... The first column (the spatial structure of the first synergistic pattern) shows that this synergistic pattern primarily activates the first few muscles. Therefore, for the first stimulation channel, its instruction quantity will mainly be contributed by the strong activation of the first synergistic pattern, while also being superimposed with moderate contributions from other synergistic patterns. In this way, the system automatically calculates how much stimulation intensity should be applied to each stimulation channel at that moment, and this stimulation intensity distribution scheme strictly follows the synergistic pattern structure of the patient's own healthy side.

[0158] This reconstruction process mathematically guarantees the synergy of multi-channel stimulus commands. Because... The matrix, derived from the NMF decomposition of the patient's healthy side, itself encodes physiologically sound muscle synergy; and The relative relationships between the components are determined by the trained mapping model. Decoded from the central intent, it represents a co-activation strategy that matches the intensity of the intent. Therefore, the strategy reconstructed through formula (6) It fundamentally inherits this collaborative structure, and the temporal relationship and intensity ratio between its channels are in line with physiological laws.

[0159] Step 800: Application of electrical stimulation

[0160] Step 800 in the application phase involves applying the reconstructed multichannel functional electrical stimulation commands to the patient's affected limb, thereby driving muscle contraction to produce actual rehabilitation movements.

[0161] In this embodiment, the functional electrical stimulation system includes several independently controllable stimulation channels, each connected to multiple muscles on the affected (right) forearm corresponding to the unaffected side. Each channel consists of a pair of surface electrodes placed near the motor point of the target muscle. It should be noted that the number of channels for the multi-channel electromyography signal on the unaffected side... The number of channels is the same as that of multichannel functional electrical stimulation commands, and each channel corresponds to the same or mirror-symmetrical muscle location to ensure the coordination of the primitive matrix. The effectiveness.

[0162] The system will calculate the result in step 700. (A multidimensional, continuously changing time-varying signal) is converted into actual control parameters for the functional electrical stimulation device. Specifically, this embodiment employs a pulse width modulation (PWM) scheme: The amplitude is linearly mapped to the stimulation pulse width. The stimulation frequency and stimulation current amplitude are individually set according to the patient's tolerance. The system adjusts the pulse width parameters of each channel in real time at a certain update cycle, thereby achieving dynamic, continuous, and coordinated control of the stimulation intensity of multiple muscles.

[0163] In other implementations, frequency modulation or amplitude modulation schemes may also be used, that is... This is mapped to the stimulation frequency or stimulation current amplitude. Regardless of the modulation method used, the key is to ensure that the changes in stimulation parameters of each channel are consistent with... Maintain a consistent time sequence and proportional relationship.

[0164] When the patient makes a grasping intention, the EEG signal in the affected side of the brain is decoded as Then it is predicted by the mapping model as Finally, through the collaborative primitive matrix Reconstructed This command drives the electrodes in each channel to coordinately stimulate multiple muscles in the affected forearm. Because the timing and intensity of the stimulation are strictly organized according to the coordination pattern of the patient's own healthy side, the resulting movements are no longer stiff and mechanical, but rather smooth, natural, and coordinated. For example, in the initial stage of grasping, the system automatically increases stimulation of the flexor muscles and moderately activates the wrist stabilizing muscles; in the holding stage of grasping, the system maintains each muscle at a relatively stable activation level; in the release stage, the system gradually reduces flexor stimulation and increases extensor stimulation. Throughout the process, the synergistic relationship between the muscles is highly similar to that during natural movement on the healthy side.

[0165] The method described in this application has significant technical advantages and clinical efficacy compared to existing "switch-on" BCI-FES systems.

[0166] First, motor coordination is fundamentally improved. This is the core technical effect of this application. Traditional methods drive muscles through preset fixed stimulation programs, which cannot simulate the inherent fine coordination relationships of human movement, resulting in stiff and unnatural movements. This application, however, extracts the patient's own physiological coordination patterns through NMF (Natural Motion Modeling) and maps central intentions onto these patterns through neural network learning, thus automatically inheriting the physiological coordination structure when reconstructing stimulus instructions. This can be evaluated using indicators such as coordination indices (e.g., VAF) and kinematic similarity; in representative tests, it is expected to achieve higher coordination and trajectory similarity compared to fixed-program methods.

[0167] Secondly, the decoding robustness is significantly improved. By performing feature extraction and dimensionality reduction at both the input (EEG) and output (sEMG) ends, the system learns more essential and stable physiological signal features. In experiments, the decoding accuracy stability of the proposed method on the test set is superior to the control method that directly performs EEG-to-sEMG mapping, with a significantly reduced accuracy variance. This demonstrates that the proposed method substantially improves robustness against signal noise and diurnal variability.

[0168] Furthermore, training efficiency is significantly improved. By transforming the high-dimensional mapping problem into a low-dimensional mapping problem, the number of model parameters is greatly reduced, and the training time is significantly shortened. The amount of training data required is also reduced accordingly, alleviating the burden on patients. Compared with traditional methods that require a longer training time and a larger number of data samples, this application can complete model training in a relatively short time.

[0169] Furthermore, it offers a high degree of personalization. (Collaborative primitive matrix) The movement patterns originate entirely from the patient's own healthy side, therefore the reconstructed stimulus instructions are naturally adapted to the patient's physiological characteristics (such as muscle strength, muscle distribution, degree of spasticity, etc.). Different patients... The matrix exhibits significant differences, fully reflecting the personalized nature of the treatment. This allows this application to provide a customized rehabilitation plan for each patient, improving the targetedness and effectiveness of the rehabilitation outcome.

[0170] Finally, this application achieves continuous and dynamic motor control. Unlike traditional "on / off" control, this application continuously decodes motor intentions and reconstructs stimulus commands, producing smooth and gradual changes in movement that better reflect the characteristics of natural human movement. This continuous control not only improves the naturalness of the movement but also enhances the patient's sense of control over the FES system, which is beneficial for inducing neuroplasticity and restoring motor function.

[0171] The following examples further illustrate the alternative implementation methods.

[0172] Furthermore, this application may also include a feedback adjustment mechanism added during the application phase. Specifically, an inertial measurement unit (IMU) sensor or other motion feedback sensor is installed on the affected limb to monitor the actual motion effect produced by FES stimulation in real time, such as the hand's position, velocity, and acceleration. Simultaneously, the system compares the error between the actual motion trajectory and the desired motion trajectory (based on the teacher signal from the healthy side). If the error exceeds a preset threshold, the system activates an adaptive adjustment algorithm to fine-tune the process in real time. Amplitude or co-occurrence matrix The weights are adjusted to compensate for fatigue, spasm, or resistance changes in the affected muscles. This closed-loop feedback control further improves the robustness and long-term stability of the system.

[0173] In another alternative implementation, for patients with electrocorticography (ECoG) implantation, ECoG signals can be used instead of EEG signals as the source of central brain signals. ECoG has higher spatial resolution and signal-to-noise ratio, providing a more refined representation of motor intention. For patients unsuitable for EEG acquisition, functional near-infrared spectroscopy (fNIRS) signals can also be used as an alternative source of central brain signals. Regardless of the central signal used, the core mapping logic of this application (from intention latent space to co-intention latent space) remains unchanged.

[0174] In another alternative implementation, the cooperative pattern decomposition in step 300 can also employ other matrix factorization algorithms such as sparse nonnegative matrix factorization (Sparse NMF) or factor analysis. Sparse NMF introduces sparsity constraints, making the cooperative primitive matrix... The presence of more zero elements in the signal yields a more physiologically interpretable synergistic pattern. Factor analysis, on the other hand, decomposes the signal based on different statistical assumptions and may provide better decomposition results in some cases.

[0175] In step 200, the specific structure of the deep neural network can also be adjusted according to actual needs. For example, a gated recurrent unit (GRU) can be used instead of an LSTM, or other sequence modeling architectures such as a temporal convolutional network (TCN) can be used. The key is to ensure that the network can effectively extract the spatiotemporal features of the EEG signal and form a continuous low-dimensional latent space representation of intent.

[0176] In step 400, the mapping model can be bidirectional LSTM, GRU, or other recurrent neural network variants, in addition to LSTM. For certain applications, even attention-based architectures such as Transformer can be used. The choice of model should be based on a trade-off between specific data characteristics and real-time requirements.

[0177] To enable those skilled in the art to more clearly implement this application, the following is an exemplary example of a specific parameter configuration for the aforementioned central intent feature extraction module (deep neural network), but this example should not be considered as a limitation of this application:

[0178] The input layer receives data within a 500ms time window (i.e., 500 time points, with downsampling processing at a sampling rate of 1000Hz).

[0179] The first layer is a 1D convolutional layer with a kernel size of 3, a stride of 1, a filter count of 16, and ReLU activation.

[0180] The second layer is the max pooling layer, with a pooling window size of 2.

[0181] The third layer is an LSTM layer with 64 hidden units and a unidirectional structure.

[0182] The number of output nodes of the fully connected (Dense Layer) is equal to the number of feature sequences in the central intent latent space. The dimensions (e.g., 8 dimensions).

[0183] During training, the Adam optimizer was used with a learning rate of 0.001 and a batch size of 32.

[0184] The above description is merely an exemplary embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

[0185] The second embodiment of this application relates to a functional electrical stimulation system based on the collaborative decoding of central intention and peripheral coordination patterns, the structure of which is as follows: Figure 2 As shown, this functional electrical stimulation system based on the collaborative decoding of central intention and peripheral coordination patterns includes:

[0186] The data acquisition module is used to simultaneously acquire the patient's central brain signals on the affected side and multi-channel electromyographic signals on the healthy side of the peripheral brain;

[0187] The central intent feature extraction module is used to receive the central brain signals from the affected side and output a continuous low-dimensional central intent latent space feature sequence. Or real-time central intent latent space feature sequence ;

[0188] The peripheral synergistic mode decomposition module is used to decompose the multi-channel electromyographic signals of the healthy side peripheral region to obtain a synergistic primitive matrix. With co-activation coefficient sequence ,in The dimension is lower than the number of channels in the healthy side multi-channel electromyography signal. As a signal from the teacher;

[0189] Storage module, used to store the collaborative primitive matrix and the trained mapping model ;

[0190] The mapping model training module is used to train the model using the aforementioned model. For input, the The mapping model is obtained by training the target. ;

[0191] The collaborative activation prediction module is used to generate the real-time central intent latent space feature sequence. Input the mapping model Output the predicted sequence of co-activation coefficients ;

[0192] The functional electrical stimulation instruction reconstruction module is used to utilize the cooperative primitive matrix stored in the storage module. and the output of the co-activation prediction module The reconstruction yields multi-channel functional electrical stimulation commands; and

[0193] A functional electrical stimulation execution module is used to apply the multi-channel functional electrical stimulation commands to the patient's affected limb.

[0194] The first embodiment is a method embodiment corresponding to this embodiment. The technical details in the first embodiment can be applied to this embodiment, and the technical details in this embodiment can also be applied to the first embodiment.

[0195] The above embodiments have the following technical effects:

[0196] In addressing the problem of motor coordination, the core breakthrough of the above embodiments lies in employing peripheral signal decomposition technology based on muscle synergy theory. By performing nonnegative matrix decomposition on the multi-channel electromyography (EMG) signals of the healthy side, the superficial, high-dimensional, and redundant EMG phenomena are decomposed into a few low-dimensional, essential synergistic primitive matrices W and synergistic activation coefficient sequences H(t). This decomposition process is not a simple mathematical dimensionality reduction, but rather extracts the inherent muscle synergistic working patterns in the human motion control system. The synergistic primitive matrix W encodes the physiological knowledge of the spatial proportion at which each muscle should be activated synergistically, while the synergistic activation coefficient sequence H(t) describes the dynamic evolution of these synergistic patterns over time. Using H(t) as a teacher signal for subsequent mapping model training elevates the model's learning objective from "replicating surface EMG signals" to "learning underlying synergistic control strategies." This shift in objective fundamentally ensures that the system can understand and reproduce the inherent coordination of human movement. In the application phase, when reconstructing multi-channel functional electrical stimulation commands using matrix multiplication FES_Command(t) = W × Ĥ(t), the W matrix itself encodes the physiological coordination structure. Therefore, the reconstructed commands automatically follow the patient's own physiological coordination pattern in terms of temporal relationships, intensity ratios, and activation sequences among the channels, thus ensuring the endogenous coordination of stimulation both mathematically and physiologically. This coordination is not obtained through manual experience adjustments but is automatically inherited from the patient's healthy side physiological system through mathematical operations. Consequently, the resulting rehabilitation movements are significantly improved in terms of fluency, naturalness, and similarity to movements on the healthy side.

[0197] To address the system robustness issue, the above embodiments effectively suppress noise and interference by simultaneously performing feature extraction and latent space mapping at both the input and output ends. At the input end, the central intent feature extraction module extracts the high-dimensional, noisy EEG signal into a low-dimensional central intent latent space feature sequence X_intent(t) through spatial filtering, temporal filtering, and multi-level processing using a deep neural network. This feature sequence has a higher signal-to-noise ratio and stronger discriminative power, and can continuously and stably represent the intensity and temporal changes of motor intent without being significantly affected by artifacts such as electrooculograms and electromyograms. At the output end, the coactivation coefficient sequence H(t), being the main component obtained through NMF decomposition, has already filtered out a large amount of noise and redundant information from the original electromyogram signal, representing the main mode of motor control. More importantly, the mapping relationship learned by the intent-co-mapping model is established between two low-dimensional latent spaces that have undergone dimensionality reduction and denoising. This "low-dimensional to low-dimensional" mapping is statistically more stable than a direct "high-dimensional to high-dimensional" mapping, and it is more likely to converge to the global optimum during the optimization process. It is also less sensitive to random noise and outliers in the training data. Therefore, the trained model can still maintain high decoding accuracy and small prediction variance when facing adverse conditions such as signal perturbations, electrode displacement, and muscle fatigue during the testing phase, and the overall robustness of the system is substantially improved.

[0198] The low-dimensional latent space mapping architecture described above offers significant advantages in improving training efficiency and reducing data requirements. By reducing high-dimensional EEG signals (e.g., 64 leads) and EMG signals (e.g., 8 channels) to low-dimensional intent features (e.g., 8 dimensions) and covariance coefficients (e.g., 4 dimensions), respectively, the parameter space that the mapping model needs to learn is significantly reduced. Taking the LSTM model as an example, the number of network parameters required for mapping from 8-dimensional input to 4-dimensional output is far less than that for a direct mapping from 64 dimensions to 8 dimensions. This not only significantly shortens the model training time but, more importantly, reduces the demand for training data. In the low-dimensional space, fewer training samples are sufficient to cover the main regions of the feature space, avoiding the "curse of dimensionality" problem prevalent in high-dimensional spaces. Furthermore, since both the input and output are high-quality signals after feature extraction, with high signal-to-noise ratios and strong correlations, the model can more easily discover and learn the inherent patterns between them, resulting in faster convergence and fewer iterations required during training. The combined effect of these factors enables the above embodiments to achieve satisfactory mapping accuracy in a relatively short training time and with relatively little training data, greatly reducing the burden on patients during the training phase and improving the system's practicality and clinical acceptability.

[0199] The technical solution of the above embodiments has unique advantages in achieving a high degree of personalization. The collaborative primitive matrix W is extracted entirely from the actual movement data of the patient's own healthy limb using the NMF algorithm. It reflects the patient's unique muscle distribution, muscle strength level, nerve innervation pattern, and long-established movement habits. Due to differences in age, gender, degree of injury, rehabilitation stage, and other factors, the W matrix of different patients will show significant individual differences in spatial structure. The system of the above embodiments extracts and stores the W matrix independently for each patient and uses the patient-specific W matrix for collaborative reconstruction during the application stage. This ensures that the generated functional electrical stimulation instructions are naturally adapted to the patient's physiological characteristics. This personalization is not achieved by manually adjusting parameters, but by the system automatically learning from the patient's own data, thus having higher physiological rationality and effectiveness. At the same time, the central intention feature extraction module and mapping model can also be trained individually for each patient, so that the entire system, from signal acquisition, feature extraction, mapping prediction to stimulation reconstruction, fully considers the individual differences of the patient at every stage, realizing a truly customized rehabilitation plan.

[0200] In achieving continuous dynamic control, the above embodiments overcome the limitations of traditional "on / off" control. By employing a sliding time window mechanism to continuously extract and update the central intention latent space feature sequence X_intent(t), the system can track the intensity changes and temporal evolution of the patient's motor intention in real time with high temporal resolution. The co-activation coefficient sequence Ĥ(t) output by the mapping model is also a continuous time-varying signal, rather than a discrete state label. This continuity continues until the final functional electrical stimulation command FES_Command(t), enabling the stimulation intensity of each channel to change smoothly and gradually, rather than abruptly jumping. The technical effects of continuous control are multifaceted: First, the generated movements are smoother and more natural, avoiding the shock of sudden starts or stops; second, the system can achieve fine force adjustment, allowing patients to flexibly control the grip strength or movement speed according to task requirements; third, continuous intention-action association helps strengthen the patient's subjective sense of control and motor-sensory feedback loop, which is of great significance for inducing neural plasticity and promoting the recovery of motor function. This continuous control capability is something that "switching" systems cannot achieve, demonstrating the significant progress made by the above embodiments in terms of control precision and user experience.

[0201] All documents mentioned in this application are considered to be incorporated in their entirety into the disclosure of this application so that they can serve as a basis for modifications if necessary. Furthermore, it should be understood that after reading the foregoing disclosure of this application, those skilled in the art can make various alterations or modifications to this application, and these equivalent forms also fall within the scope of protection claimed in this application.

Claims

1. A functional electrical stimulation system based on synergistic decoding of central intention and peripheral synergistic patterns, characterized by, Comprising: a data acquisition module for synchronously acquiring a central brain signal of a patient's affected side and a multi-channel peripheral electromyography signal of a patient's unaffected side; a central intention feature extraction module configured to receive the affected-side central brain signal and output a sequence of continuous low-dimensional central intention latent space features or a sequence of real-time central intention latent space features ​ a peripheral synergistic pattern decomposition module configured to decompose the contralateral peripheral multi-channel EMG signal to obtain a synergistic basis matrix with a synergistic activation coefficient sequence wherein a dimension of the synergistic basis matrix is lower than a number of channels of the contralateral peripheral multi-channel EMG signal, and the synergistic activation coefficient sequence serves as a teacher signal; a storage module configured to store the synergy basis matrix and the trained mapping model ; The mapping model training module is configured to train the mapping model with the input and the target. The mapping model training module is configured to train the mapping model with the input and the target. ​​ a synergistic activation prediction module configured to output a predicted synergistic activation coefficient sequence inputting the mapping model outputting the predicted synergistic activation coefficient sequence ​ The functional electrical stimulation instruction reconstruction module is used to utilize the cooperative primitive matrix stored in the storage module. and the output of the co-activation prediction module The reconstruction yields multi-channel functional electrical stimulation commands; and a functional electrical stimulation execution module for applying the multi-channel functional electrical stimulation instruction to a patient's affected side limb.

2. The system of claim 1, wherein, The peripheral cooperative mode decomposition module adopts a non-negative matrix decomposition algorithm, and satisfies an approximate relationship wherein, is a pretreated healthy side multi-channel electromyography signal matrix, with a dimension of ; is a cooperative base matrix, with a dimension of ; is a cooperative activation coefficient matrix, with a dimension of ; is a channel number, is a cooperative mode number, is a time point number.

3. The system of claim 2, wherein, The co-activation coefficient sequence Dimensions The number of channels is less than the number of channels in the multi-channel electromyography signal of the healthy side. And the dimension The variance explanation rate criterion is used to determine the ability to reconstruct... For the original The variance explained is not lower than the minimum preset threshold. value.

4. The system of claim 1, wherein, The central intention feature extraction module is a deep neural network comprising convolutional layers and / or recurrent layers configured to extract spatio-temporal features of the affected side central brain signal to form continuous low-dimensional .

5. The system of claim 1, wherein, The mapping model is a recurrent neural network, a long short-term memory network, or a gated recurrent unit configured to learn a sequence-to-sequence non-linear mapping from to ​ 6. The system of claim 1, wherein, The functional electrical stimulation instruction reconstruction module reconstructs the instruction by calculating the product of the synergy matrix and the predicted sequence of synergy activation coefficients ​ wherein, is a reconstructed multi-channel functional electrical stimulation instruction vector, with a dimension of ; is a cooperative basis matrix stored in the storage module; is a value of the cooperative activation coefficient predicted by the cooperative activation prediction module at time , with a dimension of .

7. The system of claim 1, wherein, The central brain signal of the affected side acquired by the data acquisition module is an electroencephalogram signal.

8. The system of claim 1, wherein, The functional electrical stimulation execution module is configured to convert the multi-channel functional electrical stimulation instruction into a stimulation pulse width, a stimulation frequency or a stimulation amplitude parameter of a functional electrical stimulation device.

9. The system of claim 1, wherein, The system further comprises a feedback adjustment module for simultaneously acquiring an actual movement feedback signal of the affected side limb during the application stage, and adaptively adjusting the multi-channel functional electrical stimulation instruction using an error between the actual movement feedback signal and an expected movement pattern.

Images