Rehabilitation training feedback adjusting system based on brain-computer interface and myoelectricity sensing

Abstract

The invention relates to the technical field of medical data processing and human-computer interaction, and particularly discloses a rehabilitation training feedback adjusting system based on a brain-computer interface and myoelectricity sensing. The system comprises a multi-modal physiological signal acquisition module, a signal fusion and intention decoding module, a rehabilitation state dynamic evaluation module, a self-adaptive feedback strategy generation module and a multi-channel feedback execution module. By synchronously collecting and fusing the electroencephalogram signals and the electromyographic signals to decode the motion intention, dynamically evaluate the rehabilitation state and predict the trend, a multi-sensory collaborative feedback strategy is adaptively generated, personalized and dynamic precise adjustment of rehabilitation training is achieved, and the training effect and the patient participation degree are improved. The system can autonomously switch between a reinforcement learning strategy and an auxiliary guiding strategy according to the real-time performance and the long-term trend of a patient, and dynamically optimize various feedback parameters.

Description

Technical Field

[0001] This invention belongs to the field of medical data processing and human-computer interaction technology, specifically relating to a rehabilitation training feedback regulation system based on brain-computer interface and electromyography. Background Technology

[0002] In the field of neurorehabilitation and motor function reconstruction, effectively promoting the remodeling of damaged neural pathways and the recovery of motor function is a core research topic in clinical medicine and rehabilitation engineering. Active rehabilitation training systems based on biosignal feedback, which collect and analyze patients' physiological signals and transform them into visual and audible feedback information to guide patients in conscious, goal-oriented movement attempts, have become a key technological direction for improving rehabilitation outcomes.

[0003] Such systems often rely on a single signal modality, such as using only surface electromyography (EMG) or electroencephalography (EEG) signals for feedback control. However, EMG signals are susceptible to interference from factors such as muscle fatigue, electrode displacement, and changes in skin impedance, making it difficult to guarantee signal quality and stability of motor intention decoding during long-term training. While EEG signals based on brain-computer interfaces can directly reflect motor cortex activity, their signal-to-noise ratio is low, limiting the accuracy of decoding complex and subtle motor intentions, and they are easily affected by the user's psychological state and external environment. More importantly, the recovery of a patient's neurological function is a dynamic and nonlinear process. Single-modal feedback systems cannot comprehensively and adaptively characterize this complex process, making it difficult to accurately predict the individualized recovery trajectory of the patient. Consequently, they cannot intelligently and proactively adjust intervention strategies and feedback parameters during training. Summary of the Invention

[0004] The purpose of this invention is to provide a rehabilitation training feedback regulation system based on brain-computer interface and electromyography to solve the problems of unstable feedback, limited decoding accuracy, and inability to dynamically and adaptively adjust according to individualized rehabilitation progress caused by reliance on a single physiological signal mode in the prior art.

[0005] This invention provides a rehabilitation training feedback regulation system based on brain-computer interface and electromyography, which includes a multimodal physiological signal acquisition module, a signal fusion and intention decoding module, a rehabilitation status dynamic assessment module, an adaptive feedback strategy generation module, and a multi-channel feedback execution module.

[0006] The multimodal physiological signal acquisition module is used to simultaneously acquire the patient's electroencephalogram (EEG) and surface electromyography (EMG) signals. This module includes a high-density EEG electrode array and a multi-channel EMG sensor array. The high-density EEG electrode array is arranged according to the international 10-20 system extension standard, covering the primary motor cortex, supplementary motor area, and prefrontal cortex, with a sampling frequency of no less than 1000 Hz. The multi-channel EMG sensor array is specifically deployed according to the anatomical location of the target rehabilitation muscle groups, covering agonist, antagonist, and synergist muscles, with a sampling frequency of no less than 2000 Hz. This module incorporates a high-precision synchronization clock unit to ensure microsecond-level precise alignment of EEG and EMG signals in the time dimension.

[0007] The signal fusion and intent decoding module is connected to the multimodal physiological signal acquisition module for preprocessing, feature extraction, and fusion decoding of synchronously acquired EEG and EMG signals. The preprocessing includes bandpass filtering of the EEG signals from 0.5 Hz to 40 Hz to remove baseline drift and high-frequency noise, and using independent component analysis (ICA) to remove artifacts from electrooculograms (EOG) and electrocardiograms (ECG); bandpass filtering of the EMG signals from 20 Hz to 500 Hz to extract effective EMG components, and full-wave rectification and smoothing. The feature extraction process includes extracting time-frequency energy features of event-related desynchronization and event-related synchronization, and temporal amplitude features of motor-related cortical potentials from the preprocessed EEG signals; and extracting root mean square (RMS) values, integral EMG values, and temporal pattern features of muscle activation from the preprocessed EMG signals.

[0008] The fusion decoding process is implemented using a cascaded deep neural network model. The first level of this model is a feature-level fusion network, which concatenates the extracted EEG and EMG features along the feature dimension and inputs them into a fully connected network with three hidden layers for preliminary fusion and dimensionality reduction. The second level is a decision-level fusion network, which receives the output of the first-level network and captures the temporal dependencies between multimodal signals through a long short-term memory network layer. Finally, it outputs a multidimensional motion intention vector, which quantifies the direction, force, and pattern of the target motion that the patient is attempting to perform.

[0009] The dynamic rehabilitation status assessment module is connected to the signal fusion and intention decoding module. Based on the decoded motor intention vector and the acquired raw physiological signals, it performs real-time and phased quantitative assessments of the patient's neuromuscular function recovery status. This module includes a real-time status assessment unit and a long-term trend prediction unit. The real-time status assessment unit calculates the decoding confidence of the motor intention vector, the time delay between the motor intention and the actual induced electromyographic activity, and the antagonistic muscle synergistic contraction ratio within the current training cycle, and fuses these three indicators into a real-time rehabilitation efficacy index. The long-term trend prediction unit is based on a time-series prediction model. This model uses the real-time rehabilitation efficacy index from multiple consecutive historical training cycles, the entropy change of the motor intention vector, and the electromyographic signal fatigue index as inputs to predict the trend of the rehabilitation efficacy index over the next 5 to 10 training cycles and outputs a rehabilitation stage prediction label, which includes a plateau phase, a progression phase, or a fluctuation phase.

[0010] The adaptive feedback strategy generation module connects to the rehabilitation status dynamic assessment module and the signal fusion and intent decoding module. It dynamically generates and adjusts feedback parameters and training task difficulty based on the real-time rehabilitation efficacy index, rehabilitation stage prediction labels, and the current movement intent vector. This module has a built-in strategy mapping rule base, which defines the feedback patterns, gain coefficients, and task parameter sets corresponding to different rehabilitation stages and efficacy index ranges. When the label output by the long-term trend prediction unit is "progression stage" and the real-time rehabilitation efficacy index is above the threshold of 0.7, this module executes a reinforcement learning strategy. Within the range allowed by the strategy mapping rule base, it uses increasing task success rate as the reward function, automatically exploring and updating the feedback gain coefficient and task difficulty parameters to match the patient's improved abilities. When the label is "plateau stage" or "fluctuation stage" and the real-time rehabilitation efficacy index is below the threshold of 0.5, this module switches to an assisted guidance strategy, reducing task difficulty and increasing the strength of the guiding components in the feedback signal, such as increasing the weight of virtual assistive forces or providing more explicit target trajectory prompts.

[0011] The multi-channel feedback execution module connects to the adaptive feedback strategy generation module and the signal fusion and intent decoding module. It transforms the generated feedback strategy into multi-sensory synergistic feedback stimuli and presents them to the patient. This module includes a visual feedback unit, an auditory feedback unit, and a tactile force feedback unit. The visual feedback unit renders a virtual rehabilitation training scene on a head-mounted display or tablet screen, mapping the decoded movement intent vectors to virtual limb movements in real time. Based on instructions from the adaptive feedback strategy generation module, it overlays and displays the target trajectory, a bar chart of force magnitude, and a curve showing the change in the real-time rehabilitation efficacy index. The auditory feedback unit generates audio feedback with different tones and rhythms based on the execution accuracy and force of the movement intent. Higher execution accuracy results in more harmonious tones, and achieving the target force produces a specific confirmation sound effect. The tactile force feedback unit applies programmable tactile force or resistance to the patient's limbs through a wearable exoskeleton or robotic end effector. The magnitude and direction of this force are calculated and controlled in real time by the adaptive feedback strategy generation module according to an assisted guidance strategy or resistance training strategy. This provides proprioceptive input and assists or challenges the patient's movement execution.

[0012] Furthermore, the cascaded deep neural network model used in the signal fusion and intent decoding modules is trained based on an offline-constructed multimodal physiological signal dataset. This dataset contains EEG and EMG signals synchronously recorded by healthy subjects and patients at different stages of rehabilitation while performing standardized rehabilitation tasks, along with action labels. The training process is optimized using backpropagation and adaptive moment estimation algorithms, with the loss function being the weighted sum of the mean squared error between the decoded motion intent vector and the true action label and the classification cross-entropy.

[0013] Furthermore, the long-term trend prediction unit in the dynamic assessment module of rehabilitation status employs a gated recurrent unit network (GRU) as its time series prediction model. The number of input layer nodes corresponds to the number of historical indicators used, and the output layer has three nodes, each representing the probability of one of the three categories of the rehabilitation stage prediction label. After normalization using the Softmax function, the category corresponding to the highest probability is taken as the final prediction label. During training, the network uses continuous training data from the past 30 days as samples and the rehabilitation efficacy index trend for the next 5 days as the supervision label.

[0014] Furthermore, the reinforcement learning strategy in the adaptive feedback strategy generation module is implemented using a proximal policy optimization algorithm. This algorithm defines the system state as a combination of the current real-time rehabilitation efficacy index, the predicted label of the rehabilitation stage, and the motor intention vector, and defines the action as an adjustment to the visual feedback gain, the slope of the auditory feedback mapping curve, and the magnitude of the tactile force feedback. The reward function design includes three parts: a base reward is the percentage of task completion in the current training cycle; an exploration reward encourages trying infrequently accessed combinations of feedback parameters; and a stability penalty is used to suppress drastic fluctuations in feedback parameters, ensuring the consistency of the patient's training experience.

[0015] Furthermore, the tactile force feedback unit in the multi-channel feedback execution module applies tactile force calculated by an impedance control algorithm. This algorithm takes the patient's limb position and velocity as input, and calculates the required auxiliary force or resistance based on the target impedance parameters set by the adaptive feedback strategy generation module, including virtual stiffness, virtual damping, and virtual mass. Under the assisted guidance strategy, the target impedance parameters are set to a lower value, allowing the system to gently guide the limb towards the target movement; under the resistance training strategy, the target impedance parameters, especially the virtual damping value, are significantly increased to increase the resistance the patient needs to overcome during movement.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0017] 1. This invention fundamentally overcomes the limitations of a single signal source by constructing a multimodal physiological signal acquisition and fusion decoding architecture. The system utilizes high-precision synchronization technology to spatiotemporally align the central motor intention reflected by EEG signals with the peripheral muscle activation state embodied by EMG signals, and then performs deep fusion through a cascaded deep neural network. This fusion mechanism ensures that the decoding of motor intentions relies not only on easily disturbed cortical electrical activity or easily fatigued muscle electrical activity, but also on the mutual verification and complementarity of both, thereby significantly improving the robustness, stability, and accuracy of motor intention recognition, laying a reliable data foundation for subsequent precise feedback control.

[0018] 2. This invention innovatively introduces a dynamic assessment and prediction mechanism for rehabilitation status, transforming rehabilitation training from a static, experience-driven process into a dynamic, data-driven, adaptive process. By calculating the rehabilitation efficacy index in real time and predicting trends in the rehabilitation stages, the system can quantitatively and proactively grasp the individualized trajectory of a patient's neurological function recovery. This allows the system to perceive different states of a patient's abilities, such as plateaus and progression phases, thus providing an objective basis for personalized intervention.

[0019] 3. Based on dynamic evaluation results, this invention constructs an adaptive feedback strategy generation engine, realizing intelligent real-time adjustment of training parameters and feedback modes. The system can autonomously switch between reinforcement learning strategies and assisted guidance strategies according to the patient's real-time performance and long-term trends, and dynamically optimize various feedback parameters. This adaptive capability ensures that the difficulty of the training task always matches the patient's current ability level, avoiding frustration and ineffective training due to overly difficult tasks, and preventing neural over-adaptation and recovery stagnation due to overly easy tasks. Thus, it continuously challenges the patient's neural plasticity boundaries within a safe range, maximizing the promotion of neural pathway remodeling and functional reorganization.

[0020] 4. This invention provides an immersive, multi-sensory collaborative feedback experience through a multi-channel feedback execution module. The organic combination of visual, auditory, and tactile feedback not only strengthens the patient's understanding of the relationship between motor intention and execution result from multiple dimensions, but tactile feedback also directly provides proprioceptive input, which is crucial for reconstructing damaged motor sensory integration circuits. The multi-channel feedback works collaboratively under the scheduling of adaptive strategies, forming a closed-loop, reinforced learning environment that greatly enhances the patient's motivation to actively participate in training and the efficiency of neural remodeling. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0022] Figure 2 This is a schematic diagram of the core principle framework of the signal fusion and intent decoding module based on cascaded deep neural networks in this invention;

[0023] Figure 3 This is a logical flowchart of the dynamic assessment and prediction of rehabilitation status in this invention;

[0024] Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow of the adaptive feedback strategy generation module in this invention, which switches strategies and adjusts parameters based on evaluation results.

[0025] Figure 5 This is a schematic diagram illustrating the principle framework of the collaborative operation of the multi-channel feedback execution module in this invention. Detailed Implementation

[0026] Example 1: The overall technical architecture of the rehabilitation training feedback regulation system based on brain-computer interface and electromyography proposed in this invention is shown in the attached figure. Figure 1As shown in the attached diagram, the system consists of five core functional units: a multimodal physiological signal acquisition module, a signal fusion and intent decoding module, a dynamic assessment module for rehabilitation status, an adaptive feedback strategy generation module, and a multi-channel feedback execution module. These modules communicate with each other via a high-speed data bus and control command channel, achieving low-latency, high-reliability information exchange and closed-loop control. The following will be discussed in conjunction with the attached diagram. Figure 1 To be continued Figure 5 The specific implementation methods of this system are described in detail.

[0027] First, the multimodal physiological signal acquisition module, as the sensing front end of the entire system, undertakes the crucial task of simultaneously acquiring the patient's central nervous system activity and peripheral muscle activation state. This module includes a high-density EEG electrode array and a multi-channel surface electromyography (EMG) sensor array. The high-density EEG electrode array is spatially arranged strictly according to the international 10-20 system extension standard, covering brain regions closely related to motor planning, initiation, and execution, such as the primary motor cortex (M1 area), supplementary motor area (SMA), and prefrontal cortex (PFC). The number of electrodes is no less than 64 to ensure that the spatial resolution of motor-related cortical potentials meets the requirements of subsequent feature extraction. The sampling frequency is set to 1000 Hz or higher to fully capture key time-frequency dynamics such as event-related desynchronization (ERD) and event-related synchronization (ERS).

[0028] The multi-channel surface electromyography (EMG) sensor array is individually deployed according to the anatomical distribution of the target rehabilitation muscle groups. Typical applications include monitoring the anterior deltoid, biceps brachii, triceps brachii, wrist flexors, and wrist extensors during upper limb rehabilitation, or monitoring the rectus femoris, vastus lateralis, tibialis anterior, and gastrocnemius muscles during lower limb rehabilitation. Each muscle group is equipped with at least one pair of differential electrodes to suppress common-mode interference. The EMG signal sampling frequency is no less than 2000 Hz to accurately reproduce the high-frequency components of muscle activation. Crucially, the module incorporates a high-precision synchronization clock unit. This unit employs a hardware-level timestamp mechanism, simultaneously sending a synchronization trigger signal to both the EEG and EMG acquisition subsystems at the start of each signal acquisition cycle, ensuring that the alignment error of the two signals on the time axis does not exceed 10 microseconds. This microsecond-level synchronization capability is a prerequisite for subsequent multimodal signal fusion decoding, avoiding feature distortion and misinterpretation of intent caused by time misalignment.

[0029] The acquired raw physiological signals are then transmitted to the signal fusion and intent decoding module. Please refer to the appendix. Figure 2The core processing flow of this module is divided into three stages: preprocessing, feature extraction, and fusion decoding. In the preprocessing stage, the EEG signal first passes through a digital bandpass filter with a passband range of 0.5 Hz to 40 Hz, effectively filtering out high-frequency noise such as DC drift, power frequency interference, and electromyography artifacts. Subsequently, the independent component analysis (ICA) algorithm is used to perform blind source separation on the filtered signal, identifying and eliminating artifact components generated by eye movements, blinking, or electrocardiographic activity, while retaining pure cortical source signals.

[0030] The electromyographic (EMG) signal is then passed through another bandpass filter with a passband of 20 Hz to 500 Hz to remove motion artifacts and low-frequency noise while retaining the effective frequency band generated by muscle contraction. The filtered EMG signal is then subjected to full-wave rectification, which flips all negative half-cycle signals to positive, and then passes through a sliding window smoothing filter (the window length is usually 200 milliseconds) for envelope extraction to obtain a continuous time-series signal reflecting the intensity of muscle activation.

[0031] In the feature extraction stage, the system calculates two core features from the preprocessed EEG signals: First, event-related desynchronization / synchronization time-frequency energy features, quantifying the dynamic patterns of cortical activation by performing wavelet transforms or short-time Fourier transforms on energy changes in specific frequency bands (e.g., μ rhythms from 8 Hz to 12 Hz, β rhythms from 18 Hz to 26 Hz) during motor imagery or attempted execution. Second, motor-related cortical potential (MRCP) temporal amplitude features, extracting the peak amplitude and slope of negative slow potentials within a time window from 500 ms before the action initiation to 300 ms after the action execution. For EMG signals, the system extracts root mean square (RMS), integrated electromyography (iEMG), and temporal pattern features of muscle activation. RMS reflects instantaneous muscle activation intensity, iEMG reflects the total activation over a period, and the temporal pattern features are constructed by calculating the relative onset time and duration of activation between different muscle groups to construct a muscle co-activation matrix, used to characterize the maturity of motor control strategies.

[0032] The extracted EEG and EMG feature vectors are then fed into the fusion decoding stage. (See attached image.) Figure 2As shown, this stage employs a cascaded deep neural network model. The first stage is a feature-level fusion network, whose input layer receives the concatenated multimodal feature vectors (typically 50 to 100 dimensions). These vectors undergo nonlinear mapping and feature dimensionality reduction via three fully connected hidden layers (each with 128, 64, and 32 neurons respectively), outputting a 32-dimensional fused feature representation. The second stage is a decision-level fusion network, whose core is a Long Short-Term Memory (LSTM) network layer containing 64 memory units. This LSTM layer receives the fused feature sequence output from the first-stage network (typically with a time step of 1 second, corresponding to 1000 sampling points). Through its gating mechanism, it captures the temporal dependencies between EEG intention signals and EMG execution signals, such as the time difference between intention initiation and EMG activation, and the matching degree between intention intensity and EMG amplitude. Finally, the output of the LSTM layer is fed into a fully connected output layer, generating a 5-dimensional motion intention vector. Each dimension of this vector quantifies the direction (e.g., X and Y axis components), force (scalar value), movement pattern (e.g., discrete category coding such as grasping, extension, and pronation), and execution confidence (a continuous value between 0 and 1) of the target movement the patient attempts to perform. This movement intention vector serves as the core intermediate variable of the system and is transmitted in real time to the downstream dynamic assessment module of rehabilitation status and the multi-channel feedback execution module.

[0033] The dynamic assessment module for rehabilitation status is attached. Figure 3 As shown, its function is to quantitatively assess the patient's neuromuscular function recovery status in a dual dimension: real-time status assessment and long-term trend prediction. The real-time status assessment unit calculates three key indicators per training cycle (typically 30 seconds to 2 minutes). The first indicator is the decoding confidence of the motor intention vector, which is the average of the confidence dimension in the output layer, reflecting the reliability of the system's judgment of the current intention. The second indicator is the time delay between the motor intention and the actual induced electromyographic activity, obtained by calculating the difference between the time when the directional component in the intention vector first exceeds the threshold and the time when the RMS value of the target muscle group's electromyographic signal first rises significantly. A typical healthy value should be less than 300 milliseconds; the longer the delay, the lower the neuromuscular transmission efficiency. The third indicator is the antagonist muscle co-contraction ratio, defined as the reciprocal of the ratio of the iEMG value of the target agonist muscle to the iEMG value of the corresponding antagonist muscle. The closer this ratio is to 1, the more immature the muscle co-control, which is common in the early rehabilitation stage. After normalization, these three indicators are weighted and summed (with weights of 0.4, 0.3, and 0.3 respectively) to form a Real-Time Rehabilitation Efficacy Index (REI) between 0 and 1. This index directly reflects the effectiveness of the current training.

[0034] The long-term trend prediction unit uses time-series analysis to prospectively assess the patient's rehabilitation progress. This unit employs a gated recurrent unit (GRU) network as its core model. The network's input layer receives historical data from the past 10 consecutive training cycles. Each cycle's data includes REI values, Shannon entropy of the motor intention vector (measuring the diversity and complexity of intention expression), and median frequency decline rate of electromyography signals (quantifying muscle fatigue). The input feature dimension is 30 (3 indicators × 10 cycles). The GRU network contains 128 hidden units, effectively learning long-term dependencies. The output layer has 3 nodes, corresponding to the probabilities of predicted labels for the three rehabilitation stages: plateau, progression, and fluctuation. After normalizing the output using the Softmax function, the category corresponding to the highest probability is taken as the final prediction result. Before system deployment, this model was trained offline using continuous training data from the past 30 days, with the REI trend (e.g., continuous increase, flattening, or significant fluctuation) over the next 5 days used as a supervision label to ensure the accuracy and clinical relevance of the predictions.

[0035] The adaptive feedback strategy generation module is attached. Figure 4 As shown, its core task is to dynamically adjust feedback parameters and training task difficulty based on the REI output by the rehabilitation status dynamic assessment module and the predicted label of the rehabilitation stage. This module has a built-in policy mapping rule base, which stores feedback policy templates for different state combinations. When the system detects that the predicted label of the rehabilitation stage is in the progress phase and the current REI is higher than 0.7, the module activates a reinforcement learning policy. This policy is implemented using the proximal policy optimization (PPO) algorithm.

[0036] System status Defined as The combination, in which This is the current REI value. One-hot encoding for stage labels. This is the current motion intent vector. (Action) Defined as the adjustment amount for three feedback parameters: visual feedback gain coefficient. Slope of auditory feedback mapping curve tactile feedback magnitude The reward function r consists of three parts: the basic reward... Equal to the percentage of task completion in the current training cycle (e.g., the percentage of times a virtual hand successfully grasps a target object); exploration reward In action A positive reward is given when the parameter falls into a low-access-frequency range; a stability penalty is applied. This is proportional to the second difference of the parameter adjustment amount, used to suppress drastic parameter jumps. The policy network updates its parameters by maximizing the cumulative discount reward, thereby autonomously exploring better combinations of feedback parameters while ensuring training stability.

[0037] Conversely, when the rehabilitation phase prediction label is a plateau or fluctuating phase, and the REI is below 0.5, the module switches to an assisted guidance strategy. Under this strategy, the system automatically reduces the geometric complexity of the training task (e.g., shortening the target trajectory length, increasing the target tolerance radius) or the dynamic difficulty (e.g., reducing the virtual resistance to be overcome). Simultaneously, the guiding component in the feedback signal is enhanced: the visual feedback unit overlays a thicker, brighter target trajectory line in the virtual scene; the auditory feedback unit plays encouraging sound effects when the patient approaches the correct movement direction; and the tactile force feedback unit applies an assistive force pointing towards the target location. The magnitude of this assistive force is calculated in real-time by an impedance control algorithm, and its target impedance parameter (virtual stiffness) is used to determine the target location. Virtual damping Virtual quality ) is set to a lower value (e.g. =5 Newtons / meter =2 Newtons per second (m), enabling exoskeletons or robots to gently guide limb movements and reconstruct correct motion sensation patterns.

[0038] Finally, the multi-channel feedback execution module is attached. Figure 5 As shown, the system is responsible for transforming the abstract parameters output by the adaptive feedback strategy generation module into multi-sensory stimuli that the patient can perceive. The visual feedback unit runs on a head-mounted display device or a fixed tablet screen, rendering a three-dimensional virtual rehabilitation scene. The decoded motion intention vector is mapped in real time to the joint angles and end-effector positions of virtual limbs (such as arms or legs), realizing intention-driven virtual movement. According to the strategy instructions, the system overlays and displays a semi-transparent target trajectory, a bar chart of real-time force magnitude (height proportional to the dimension of intention force), and a historical change curve of REI value (presented as a green broken line) in the scene.

[0039] The auditory feedback unit maintains a pitch-precision mapping table and a rhythm-velocity mapping table. When the movement direction error is less than 10 degrees, a C major tonic chord is played; as the error increases, the chord gradually changes to a dissonant interval. When the electromyographic activation intensity reaches more than 90% of the target value, a crisp "ding" sound is triggered as a confirmation effect. The tactile force feedback unit is implemented through a six-DOF wearable exoskeleton or desktop robot end effector. Its underlying controller receives the target impedance parameters set by the adaptive feedback strategy generation module and combines them with the patient's limb position read in real time from the exoskeleton joint encoder and force sensor. ,speed The required force is calculated using the following impedance control formula. : in, , , These represent the desired position, velocity, and acceleration trajectory, determined jointly by the current motion intention vector and the assist / resistance strategy. In assisted guidance mode, It is set as a smooth trajectory from the current position to the target position. and Choose a smaller value so that the system behaves as a low-impedance follower; in resistance training mode, Fixed or moving slowly, and (especially) ) was significantly improved (e.g. =15 Newton-seconds / meter), making the system behave as a high-damping environment, forcing the patient to actively exert force to overcome resistance, thereby enhancing muscle strength and neural drive efficiency.

[0040] The entire system forms a complete closed loop during operation: multimodal physiological signals are collected and decoded into motor intentions. These intentions, along with the original signals, are used to assess the rehabilitation status. The assessment results drive adaptive adjustments to the feedback strategy. The adjusted strategy then acts on the patient through multi-channel feedback, influencing their next round of motor attempts and changes in neural plasticity, thereby altering subsequent physiological signal characteristics. This closed-loop mechanism ensures that rehabilitation training is always on a dynamic optimization track of assessment-intervention-reassessment, achieving truly individualized, intelligent, and efficient neurorehabilitation.

[0041] Example 2: Based on Example 1, this example specifically adapts and optimizes the system for lower limb gait rehabilitation scenarios to verify its generalization ability and robustness in different rehabilitation tasks. In the multimodal physiological signal acquisition module, the coverage area of ​​the high-density EEG electrode array focuses on the lower part of the precentral gyrus (corresponding to the lower limb motor area) and the parietal cortex (involved in spatial navigation and balance control), and the number of electrodes can be reduced to 32 to improve wearing comfort. The surface electromyography sensor array is mainly deployed on the rectus femoris, biceps femoris, tibialis anterior, medial head of gastrocnemius, and gluteus maximus muscles, with a total of 5 pairs of electrodes, and the sampling frequency is still maintained at 2000 Hz. The accuracy requirement of the synchronization clock unit is further improved to within 5 microseconds to cope with the rapid alternation of muscle activation in the gait cycle.

[0042] In the signal fusion and intent decoding module, the feature extraction stage was adjusted to account for the periodicity and temporality specific to gait. In addition to retaining ERD / ERS energy, EEG features were supplemented with steady-state motion evoked potential amplitude features that are phase-locked with gait. EMG features incorporated the phase angle and duty cycle of muscle activation to characterize the activation timing and duration of each muscle group within the gait cycle. The time step of the LSTM layer in the fusion decoding network was adjusted to a complete gait cycle (approximately 1 second), and the output motion intent vector was expanded to 7 dimensions, with new dimensions including gait phase (continuous values ​​from 0 to 1), stride intent, and intention to switch between support and swing phases.

[0043] In the real-time status assessment unit of the dynamic assessment module for rehabilitation status, the time delay index has been replaced by gait phase synchronization error, which is the absolute difference between the gait phase predicted by EEG intent and the actual gait phase calculated based on electromyographic activation patterns. The antagonistic muscle synergistic contraction ratio focuses on the coordination between the quadriceps and hamstrings. The input features of the long-term trend prediction unit have been expanded to include the gait symmetry index and gait frequency variation coefficient to more comprehensively reflect the quality of lower limb motor function recovery.

[0044] The policy mapping rule base of the adaptive feedback policy generation module has been reconstructed for gait training. During the progress phase, the reward function of the reinforcement learning policy incorporates an incremental reward for improved gait symmetry. During the plateau phase, the assisted guidance strategy not only provides tactile assistance but also overlays virtual arrows of ground reaction force in the visual feedback, helping patients understand the mechanisms of weight transfer and propulsion. In the impedance control algorithm of the tactile force feedback unit, the target trajectory... The reference trajectory is set to conform to the spatiotemporal parameters of normal gait, and a virtual mass M is introduced to simulate the swinging inertia of the lower limbs, enhancing the realism of proprioception.

[0045] The visual feedback unit of the multi-channel feedback execution module renders a virtual street scene, allowing the patient to control their virtual avatar's walking through intention. The auditory feedback unit plays rhythmic sounds synchronized with the footsteps, based on gait phase; the more accurate the phase, the more stable the rhythm. Tactile force feedback is achieved through a lower limb exoskeleton, providing forward assist during the swing phase and vertical upward support during the stance phase; the force magnitude is dynamically adjusted according to REI (Responsive Energy Injection). This embodiment fully demonstrates the high configurability and task adaptability of the system architecture of this invention, maintaining consistency in core logic and superior performance under different rehabilitation goals.

[0046] Example 3: This example further explores a lightweight deployment scheme for the system in a remote home rehabilitation scenario. The multimodal physiological signal acquisition module adopts the Bluetooth 5.0 wireless transmission protocol, and the EEG electrode array is simplified to a dry electrode flexible headband, covering five key points: Fz, Cz, C3, C4, and Pz, with the sampling frequency reduced to 500 Hz. The electromyography sensor adopts a miniaturized patch design, integrating signal conditioning circuitry, and transmits data via a Zigbee network. Despite the reduced hardware specifications, the signal fusion and intent decoding module introduces a transfer learning mechanism, using a large model trained in a hospital setting as the teacher network to perform knowledge distillation on a small student network (containing only one LSTM layer and 16 hidden units) in a home setting, thereby maintaining high decoding accuracy with limited computing power.

[0047] The long-term trend prediction unit of the rehabilitation status dynamic assessment module uses a lightweight linear dynamic system model instead of a GRU network. It only needs to store the mean and variance of the REI for the most recent 7 days to predict the trend, significantly reducing computational and storage overhead. The reinforcement learning strategy of the adaptive feedback strategy generation module is simplified to a rule-based heuristic adjustment: when the REI increases for 3 consecutive cycles, the task difficulty is automatically increased by one level; when the REI decreases for 2 consecutive cycles, the difficulty is decreased by one level. The multi-channel feedback execution module mainly relies on the screen and speakers of smartphones or tablets to provide audiovisual feedback, while tactile feedback is achieved through the intensity and frequency encoding of the phone's vibration motor. Although not as precise as an exoskeleton, it can still provide effective proprioceptive cues in a home setting.

Claims

1. A rehabilitation training feedback regulation system based on brain-computer interface and electromyography sensing, characterized in that, include: A multimodal physiological signal acquisition module is used to simultaneously acquire the patient's electroencephalogram (EEG) and surface electromyography (EMG) signals; The signal fusion and intent decoding module is connected to the multimodal physiological signal acquisition module and is used to preprocess, extract features, and fuse and decode the synchronously acquired electroencephalogram (EEG) and electromyogram (EMG) signals. The rehabilitation status dynamic assessment module is connected to the signal fusion and intention decoding module, and is used to perform real-time and phased quantitative assessment of the patient's neuromuscular function recovery status based on the decoded motion intention vector and the collected raw physiological signals. An adaptive feedback strategy generation module, connected to the rehabilitation status dynamic assessment module and the signal fusion and intent decoding module, is used to dynamically generate and adjust feedback parameters and training task difficulty based on the real-time rehabilitation efficacy index, rehabilitation stage prediction label and current movement intent vector. A multi-channel feedback execution module, connected to the adaptive feedback strategy generation module and the signal fusion and intent decoding module, is used to transform the generated feedback strategy into multi-sensory collaborative feedback stimuli.

2. The rehabilitation training feedback regulation system based on brain-computer interface and electromyography sensing according to claim 1, characterized in that, The multimodal physiological signal acquisition module includes a high-density EEG electrode array and a multi-channel surface electromyography sensor array. The high-density EEG electrode array covers the primary motor cortex, the supplementary motor area, and the prefrontal cortex. The multi-channel surface electromyography sensor array is arranged according to the anatomical location of the target rehabilitation muscle group, covering agonist muscles, antagonist muscles, and synergist muscles. The multimodal physiological signal acquisition module has a built-in high-precision synchronization clock unit to ensure that the electroencephalogram (EEG) signal and the electromyography (EMG) signal are precisely aligned at the microsecond level in the time dimension.

3. The rehabilitation training feedback regulation system based on brain-computer interface and electromyography sensing according to claim 1, characterized in that, The preprocessing includes bandpass filtering of the EEG signal and removal of artifacts from electrooculography and electrocardiography using independent component analysis algorithm, and bandpass filtering of the electromyography signal and full-wave rectification and smoothing. The feature extraction includes extracting time-frequency energy features of event-related desynchronization and event-related synchronization, as well as time-domain amplitude features of motor-related cortical potentials from the preprocessed EEG signal, and extracting root mean square value, integral electromyography value, and temporal pattern features of muscle activation from the preprocessed electromyography signal.

4. The rehabilitation training feedback regulation system based on brain-computer interface and electromyography sensing according to claim 3, characterized in that, The fusion decoding is implemented using a cascaded deep neural network model. The first stage of the cascaded deep neural network model is a feature-level fusion network, which concatenates the extracted EEG and EMG features along the feature dimension and inputs them into a fully connected network with three hidden layers for preliminary fusion and dimensionality reduction. The second stage of the cascaded deep neural network model is a decision-level fusion network, which receives the output of the first stage and captures the temporal dependencies between multimodal signals through a long short-term memory network layer, ultimately outputting a multidimensional motion intention vector. The motion intention vector quantifies the direction, force, and pattern of the target movement that the patient attempts to perform.

5. The rehabilitation training feedback regulation system based on brain-computer interface and electromyography as described in claim 4, characterized in that, The dynamic assessment module for rehabilitation status includes a real-time status assessment unit and a long-term trend prediction unit. The real-time status assessment unit is used to calculate the decoding confidence of the movement intention vector, the time delay between movement intention and actual induced electromyographic activity, and the antagonistic muscle synergistic contraction ratio within the current training cycle, and integrates these three indicators into a real-time rehabilitation efficacy index. The long-term trend prediction unit is based on a time series prediction model and uses the real-time rehabilitation efficacy index, the entropy change of the movement intention vector, and the electromyographic signal fatigue index from multiple consecutive historical training cycles as inputs to predict the trend of the rehabilitation efficacy index over the next 5 to 10 training cycles, and outputs a rehabilitation stage prediction label, which includes a plateau period, a progress period, or a fluctuation period.

6. The rehabilitation training feedback regulation system based on brain-computer interface and electromyography as described in claim 5, characterized in that, The adaptive feedback strategy generation module has a built-in strategy mapping rule library, which defines the feedback patterns, gain coefficients, and task parameter sets corresponding to different rehabilitation stages and efficacy index ranges. When the label output by the long-term trend prediction unit is the progress stage and the real-time rehabilitation efficacy index is higher than the threshold of 0.7, the adaptive feedback strategy generation module executes a reinforcement learning strategy. Within the range allowed by the strategy mapping rule library, it uses the improvement of task success rate as the reward function and automatically explores and updates the feedback gain coefficient and task difficulty parameters. When the label is the plateau stage or fluctuation stage and the real-time rehabilitation efficacy index is lower than the threshold of 0.5, the adaptive feedback strategy generation module switches to an auxiliary guidance strategy to reduce the task difficulty and increase the strength of the guiding component in the feedback signal.

7. The rehabilitation training feedback regulation system based on brain-computer interface and electromyography as described in claim 6, characterized in that, The multi-channel feedback execution module includes a visual feedback unit, an auditory feedback unit, and a tactile force feedback unit. The visual feedback unit renders a virtual rehabilitation training scene on a display device, maps the decoded movement intention vector into the movement of virtual limbs in real time, and overlays and displays the target trajectory, force magnitude bar chart, and real-time rehabilitation efficacy index change curve according to the instructions of the adaptive feedback strategy generation module. The auditory feedback unit generates audio feedback with different tones and rhythms based on the execution accuracy and force of the movement intention. The tactile force feedback unit applies programmable tactile force or resistance to the patient's limbs through a wearable exoskeleton or robotic end effector. The magnitude and direction of the tactile force or resistance are calculated and controlled in real time by the adaptive feedback strategy generation module according to the assisted guidance strategy or resistance training strategy.

8. The rehabilitation training feedback regulation system based on brain-computer interface and electromyography sensing according to claim 7, characterized in that, The training process of the cascaded deep neural network model used in the signal fusion and intent decoding module is based on an offline constructed multimodal physiological signal dataset. The multimodal physiological signal dataset includes EEG and EMG signals and action labels synchronously recorded by healthy subjects and patients at different rehabilitation stages when performing standardized rehabilitation action tasks. The training process is optimized using the backpropagation algorithm and the adaptive moment estimation algorithm. The loss function is the weighted sum of the mean square error between the decoded motion intent vector and the real action label and the classification cross-entropy.

9. The rehabilitation training feedback regulation system based on brain-computer interface and electromyography sensing according to claim 8, characterized in that, The long-term trend prediction unit in the dynamic assessment module of the rehabilitation status uses a gated recurrent unit network as its time series prediction model. The number of input layer nodes of the gated recurrent unit network corresponds to the number of historical indicators used, and the output layer has 3 nodes, which correspond to the probabilities of the 3 categories of the rehabilitation stage prediction label. After normalization by the Softmax function, the category corresponding to the maximum probability is taken as the final prediction label.

10. The rehabilitation training feedback regulation system based on brain-computer interface and electromyography sensing according to claim 9, characterized in that, The reinforcement learning strategy in the adaptive feedback strategy generation module is implemented using a proximal policy optimization algorithm. This algorithm defines the system state as a combination of the current real-time rehabilitation efficacy index, the rehabilitation stage prediction label, and the movement intention vector. It defines the action as an adjustment to the visual feedback gain, the slope of the auditory feedback mapping curve, and the magnitude of the tactile force feedback. The reward function design includes three parts: The base reward is the percentage of task completion in the current training cycle; The exploration reward encourages experimentation with combinations of feedback parameters that are not frequently accessed; Stability penalties are used to suppress drastic fluctuations in feedback parameters.

Images