Method and system for upper limb grasping function reconstruction based on closed-loop brain-spinal cord interface with multi-modal biofeedback

Abstract

The application relates to the field of closed-loop neuromodulation and functional reconstruction, and discloses a closed-loop brain-spinal cord interface upper limb grasping function reconstruction method and system based on multi-modal biological feedback. The method synchronously collects at least three kinds of feedback information such as joint angle, grasping force and electromyographic signal, generates a comprehensive actual motion state through weighted fusion, creatively adopts a hierarchical adjustment strategy according to an error vector between a desired state and an actual state, and differentiates and adaptively adjusts parameters such as stimulation current intensity, frequency and pulse width, calculates a fatigue index based on electromyographic frequency domain characteristics (such as a median frequency and an average power frequency), and superimposes a fatigue compensation strategy to perform prospective intervention. The method effectively solves the problems of low control precision, insufficient single feedback dimension information and insufficient continuous motion capacity of existing systems, and significantly improves the precision, adaptability and long-term stability of motion control.

Description

Technical Field

[0001] This application relates to the fields of neuroengineering and rehabilitation medicine, and in particular to a closed-loop brain-spinal cord interface motor function reconstruction technology based on multimodal biofeedback. Background Technology

[0002] Spinal cord injury (SCI) is one of the leading causes of limb paralysis, especially when the injury occurs in the C5-C7 segment of the cervical spinal cord. Patients typically retain the ability to generate motor intentions in their brains, but due to the interruption of neural pathways, motor commands cannot be transmitted to the forearm muscles, resulting in severe or complete impairment of upper limb grasping function. Functional electrical stimulation (FES) has become a primary means of reconstructing motor function in these patients. It directly induces muscle contraction to achieve movement by applying electrical impulses to peripheral nerves or muscles.

[0003] However, traditional FES systems typically employ an open-loop control architecture, outputting fixed stimulation parameters according to a preset pattern, lacking a mechanism for sensing and correcting actual movement effects. When patients perform daily tasks such as grasping a cup or picking up tools, the preset stimulation parameters quickly become ineffective due to accumulated muscle fatigue, slight shifts in electrode-skin contact, or instantaneous changes in individual physiological state. This leads to unstable grip strength, inaccurate movement trajectories, and ultimately, grip failure, severely limiting the practicality of the FES system and the patient's user experience.

[0004] To address the low accuracy of open-loop control, some studies have attempted to introduce brain-computer interface (BCI) systems to recognize patients' motor intentions and construct closed-loop systems by combining them with feedback mechanisms. However, in practical applications, some existing closed-loop systems still face specific technical challenges: 1. Insufficient feedback information dimensions lead to inaccurate evaluation: For example, in scenarios requiring precise gripping of a cup or pen, motion control needs to simultaneously meet multiple requirements, such as finger joint posture (from IMU sensors), contact force (from force sensors), and the activation state of the target muscle (from EMG). However, most existing systems only use a single sensor (e.g., only using IMU to monitor angle), which cannot comprehensively evaluate the execution effect of complex movements, resulting in inaccurate error signals generated by the system and difficulty in achieving precise grip force control. 2. Lack of adaptive adjustment strategies leads to slow convergence speed: When the system detects motion errors, many closed-loop systems use simple algorithms such as fixed-gain PID for adjustment. This single strategy cannot distinguish the nature of the error (whether it is insufficient force, incorrect posture, or electrode detachment), making it difficult to take differentiated and efficient adjustment measures based on the magnitude of the error. Especially when facing complex force variations within 50 N, it is prone to problems such as slow oscillation control or convergence speed. 3. Lack of proactive fatigue management leads to poor sustainability: When patients perform prolonged grasping tasks (such as rehabilitation training or continuous operations), nerve electrical stimulation easily causes rapid fatigue of the target muscles. Existing systems typically only passively attempt to increase the stimulation current after sensing a significant decrease in grasping force. However, by this time the muscles are already fatigued, and continued high-current stimulation may accelerate damage, resulting in effective movement time usually only lasting 5-8 minutes, which cannot meet the needs of prolonged functional use.

[0005] Therefore, there is an urgent need for a closed-loop neuromodulation method that can deeply integrate multi-dimensional biofeedback information, has the ability to adjust parameters in a graded and adaptive manner, and can actively predict and compensate for muscle fatigue, so as to significantly improve the accuracy, adaptability and sustainability of upper limb grasping function reconstruction in patients with spinal cord injury. Summary of the Invention

[0006] The purpose of this application is to provide a method and system for reconstructing upper limb grasping function based on a closed-loop brain-spinal cord interface using multimodal biofeedback, in order to solve the problems mentioned in the background art.

[0007] This application discloses a closed-loop brain-spinal cord interface method for upper limb grasping function reconstruction based on multimodal biofeedback, including the following steps:

[0008] S1. System initialization and calibration: Establish an individualized brain signal decoding model and determine the safe range of stimulation parameters;

[0009] S2. Brain Signal Acquisition and Intention Decoding: Motor cortex neural signals are acquired via intracranial electrodes, and the desired motor state is decoded using the aforementioned brain signal decoding model. The At least include the expected joint angle and expected grasp strength ;

[0010] S3, Initial stimulus execution: according to the... Initial stimulation parameters are generated, and nerve electrical stimulation is applied to the target nerve or muscle via a nerve electrical stimulator.

[0011] S4. Multimodal feedback acquisition: Simultaneously acquire at least three types of feedback signals, including: joint angle signals acquired through the inertial measurement unit, gripping force signals acquired through the force sensor or pressure sensor, and electromyography signals acquired through the surface electromyography sensor.

[0012] S5. Comprehensive Status Assessment and Error Calculation: The joint angle signal, the gripping force signal, and the electromyographic signal are weighted and fused to generate the actual motion state. and calculate the With the Error vector between ;

[0013] S6. Muscle Fatigue Monitoring and Prediction: Perform frequency domain analysis on the electromyographic signals to calculate the fatigue index. and according to the Determine fatigue compensation strategies;

[0014] S7. Adaptive Stimulation Parameter Adjustment: Based on the above... The stimulus parameters are updated using a graded adjustment strategy, and the fatigue compensation strategy is superimposed to adjust the stimulus parameters to generate new stimulus parameters. The graded adjustment strategy includes:

[0015] When the error magnitude is less than the first threshold At that time, only the current intensity is adjusted. ;

[0016] When the error magnitude is greater than or equal to And less than the second threshold At the same time, adjust the current intensity. Stimulation frequency and pulse width ;

[0017] When the error magnitude is greater than or equal to At that time, switch the electrode activation combination;

[0018] S8. Safety check and stimulation execution: After performing a safety check on the new stimulation parameters, nerve electrical stimulation is applied to the target nerve or muscle through the nerve electrical stimulator.

[0019] S9, Closed-loop iteration: Repeatedly execute steps To the steps Until the task is completed or the termination condition is triggered, it continues.

[0020] In a preferred embodiment, the actual motion state It is a multi-dimensional vector, including actual joint angles. Actual grasp strength and actual muscle activation Each dimension is obtained through standardized processing and weighted fusion of the corresponding sensor signals:

[0021]

[0022]

[0023]

[0024] in, Standardized signals acquired by joint angle sensors; Standardized signals acquired by force sensors; Standardized signals acquired by surface electromyography (EMG) sensors; weighting coefficients , , The value is dynamically adjusted based on current task requirements and sensor signal-to-noise ratio, with a range of [value missing]. .

[0025] In a preferred embodiment, the fatigue index The calculation is based on the frequency domain characteristics of electromyographic signals, using the following formula:

[0026]

[0027] in, , , These are weighting coefficients; and These are the median frequency and the average power frequency at the current moment, respectively. and These are the baseline median frequency and the baseline average power frequency, respectively. This is the root mean square value of the current electromyographic signal; This represents the current current intensity. and These are the baseline root mean square value and the baseline current intensity, respectively.

[0028] In a preferred embodiment, the fatigue compensation strategy includes:

[0029] · when Increase the stimulation frequency by 5-10 Hz.

[0030] · when At the same time, it activates the synergistic muscle groups for joint stimulation;

[0031] · when When this occurs, a rest mode is triggered, pausing the task or reducing the task intensity by 50%, lasting for 20-30 seconds before being reassessed.

[0032] In a preferred embodiment, the brain signal decoding model employs a hierarchical decoding architecture, including:

[0033] • First layer: A classifier is used to identify discrete high-level motion intentions, including grasping, releasing, holding, and reaching;

[0034] • Second layer: Decoding the above using a regression model The continuous motion parameters include at least the desired grip strength. and desired joint angle It also dynamically adjusts the mapping range and weights of continuous motion parameters based on the current task type.

[0035] In a preferred embodiment, the intracranial electrode is a cortical electroencephalogram (EEG) electrode or an intravascular stent electrode.

[0036] In a preferred embodiment, the first threshold The second threshold is set to 10% of the normalization error. Set to 30% of the normalization error.

[0037] In a preferred embodiment, in step S7:

[0038] • When the error size is less than At that time, the adjustment amount of current intensity ,in This is the proportionality coefficient. For the The force error component in the middle;

[0039] • When the error magnitude is greater than or equal to and less than At that time, the adjustment amount of current intensity Adjustment amount of stimulation frequency Pulse width adjustment amount ,in , , This is the proportionality coefficient. For the Angular error components in For the The activation error component in.

[0040] In a preferred embodiment, the range of the stimulation parameters is: current intensity Between 5mA and 50mA, stimulation frequency Between 20Hz and 100Hz, pulse width Between 100μs and 500μs.

[0041] In a preferred embodiment, the target nerve or muscle is an epidural nerve of the C5-C7 segment of the cervical spinal cord or a peripheral nerve of the forearm.

[0042] In a preferred embodiment, the update period of the closed-loop iteration is 50ms to 100ms.

[0043] In a preferred embodiment, step S1, establishing the individualized brain signal decoding model, includes: collecting motor cortical neural signals from the patient when performing various motor actions as training data, and training a classifier and a regression model to establish the brain signal decoding model.

[0044] In a preferred embodiment, in step S1, determining the range of safe stimulation parameters includes: gradually increasing the current intensity starting from the minimum safe current to determine the motion threshold current and the comfort upper limit current, and setting the range of safe stimulation parameters between the motion threshold current and the comfort upper limit current.

[0045] In a preferred embodiment, in step S8, the safety check includes: checking whether the new stimulation parameters are within the range of safe stimulation parameters, and checking whether the joint angle and grip strength exceed the physiological safety range.

[0046] In a preferred embodiment, the frequency domain analysis employs a fast Fourier transform with a window length of 500 ms and a window overlap rate of 50%.

[0047] In a preferred embodiment, in step S4, the synchronous acquisition ensures the time alignment of the joint angle signal, the gripping force signal, and the electromyographic signal through hardware timestamps, with a time alignment error of less than 1ms.

[0048] In a preferred embodiment, in step S7, the magnitude of a single parameter adjustment does not exceed 20% of the current parameter value.

[0049] In a preferred embodiment, the method is applicable to the reconstruction of upper limb grasping dysfunction caused by spinal cord injury.

[0050] This application also discloses a closed-loop brain-spinal cord interface upper limb grasping function reconstruction system based on multimodal biofeedback, including:

[0051] The initialization and calibration module is used to establish an individualized brain signal decoding model and determine the safe range of stimulus parameters;

[0052] The brain signal acquisition and decoding module includes intracranial electrodes for acquiring motor cortex neural signals and decoding the desired motor state using the brain signal decoding model. The At least include the expected joint angle and expected grasp strength ;

[0053] A nerve electrical stimulator, used according to the above Generate initial stimulation parameters and apply electrical nerve stimulation to the target nerve or muscle;

[0054] The multimodal feedback acquisition module includes an inertial measurement unit, a force sensor or pressure sensor, and a surface electromyography sensor, for simultaneously acquiring at least three types of feedback signals, including: joint angle signals, gripping force signals, and electromyography signals.

[0055] The integrated state assessment and error calculation module is used to weight and fuse the joint angle signal, the gripping force signal, and the electromyographic signal to generate the actual motion state. and calculate the With the Error vector between ;

[0056] The muscle fatigue monitoring and prediction module is used to perform frequency domain analysis on the electromyographic signals to calculate the fatigue index. and according to the Determine fatigue compensation strategies;

[0057] The adaptive stimulus parameter adjustment module is used to adjust the parameters according to the... The stimulus parameters are updated using a graded adjustment strategy, and the fatigue compensation strategy is superimposed to adjust the stimulus parameters to generate new stimulus parameters. The graded adjustment strategy includes:

[0058] • When the error size is less than the first threshold At that time, only the current intensity is adjusted. ;

[0059] • When the error magnitude is greater than or equal to And less than the second threshold At the same time, adjust the current intensity. Stimulation frequency and pulse width ;

[0060] • When the error magnitude is greater than or equal to At that time, switch the electrode activation combination;

[0061] The safety check module is used to perform a safety check on the new stimulation parameters;

[0062] The nerve stimulator is also used to apply nerve electrical stimulation to the target nerve or muscle based on new stimulation parameters after being checked by the safety inspection module.

[0063] The closed-loop control module is used to control the brain signal acquisition and decoding module, the neurostimulator, the multimodal feedback acquisition module, the comprehensive state assessment and error calculation module, the muscle fatigue monitoring and prediction module, the adaptive stimulation parameter adjustment module, and the safety check module to repeatedly perform the above operations until the task is completed or the termination condition is triggered.

[0064] This application constructs a closed-loop brain-spinal cord interface upper limb grasping function reconstruction method based on multimodal biofeedback, which effectively solves the technical problems of low control precision, poor adaptability and limited continuous movement ability in the prior art, and achieves significant technical results.

[0065] To address the problems of existing open-loop control systems, such as their inability to perceive actual motion execution effects and their susceptibility to muscle fatigue and external interference leading to low control accuracy, this application establishes an individualized brain signal decoding model and defines initialization steps for safe stimulus parameter ranges, laying an individualized foundation for subsequent closed-loop control. By acquiring motor cortical neural signals through intracranial electrodes and decoding them to obtain the desired motion state, including the desired joint angle and desired gripping force, a clear control objective is provided for the system. More importantly, this application simultaneously acquires at least three types of feedback signals, including joint angle signals, gripping force signals, and electromyographic signals, and weightedly fuses these multimodal feedback signals to generate the actual motion state. This allows for a comprehensive characterization of complex grasping movements from three dimensions: posture, force, and muscle activation. Compared to single-sensor feedback, this multimodal fusion evaluation provides more complete and accurate motion state information, offering a reliable reference for subsequent error calculation and helping to reduce misjudgments and unnecessary parameter adjustments caused by incomplete information.

[0066] To address the shortcomings of existing systems, such as their inability to intelligently adjust stimulation parameters and poor adaptability, this application achieves error-driven adaptive control by calculating the error vector between the desired and actual motor states and updating stimulation parameters according to the magnitude of the error using a tiered adjustment strategy. Specifically, when the error is small, only the current intensity is fine-tuned to avoid over-response; when the error is moderate, the current intensity, stimulation frequency, and pulse width are adjusted simultaneously to accelerate convergence; and when the error is large, the electrode activation combination is switched to fundamentally change the muscle recruitment pattern. This tiered adjustment strategy enables the system to take differentiated response measures based on the actual degree of deviation, ensuring control stability with small errors and rapid correction capability with large errors, thus improving closed-loop response speed and control accuracy. Through a continuous, iterative closed-loop process of brain signal acquisition, stimulation execution, feedback acquisition, error calculation, and parameter adjustment, the system can track changes in the desired motor state in real time and dynamically adjust the stimulation output, facilitating precise responses to the patient's motor intentions.

[0067] To address the issues of rapid muscle fatigue caused by electrical stimulation and the inability of existing systems to predict and compensate for fatigue, leading to poor sustained exercise capacity, this application employs frequency domain analysis of electromyography (EMG) signals to calculate a fatigue index. Based on this index, a fatigue compensation strategy is determined, enabling proactive fatigue management. The fatigue index is calculated based on the median frequency, average power frequency, and deviation from the baseline value of the EMG signal, reflecting the actual fatigue state of the muscles. When the fatigue index falls within different threshold ranges, the system employs differentiated compensation strategies such as increasing the stimulation frequency, activating synergistic muscle groups, or triggering a rest mode, proactively intervening before a significant decline in exercise quality. By combining error-based graded adjustment strategies with fatigue-based compensation strategies, the system can both correct execution errors in real time and prevent fatigue accumulation, thus helping to extend effective exercise time and improve training tolerance.

[0068] In terms of actual motion state assessment, this application integrates joint angle signals, gripping force signals, and electromyographic signals through a weighted fusion formula, where the weighting coefficients are dynamically adjusted according to the current task requirements and the sensor signal-to-noise ratio. This task-adaptive weighted fusion mechanism enables the system to optimize the contribution weights of each modality signal based on the different task's emphasis on angle, force, and muscle activation, thus improving the relevance and accuracy of comprehensive motion state assessment. For example, for fine gripping tasks, the force feedback weight is increased to enhance force control precision, while for large-range motion tasks, the angle feedback weight is increased to optimize trajectory tracking capabilities.

[0069] In terms of fatigue monitoring, this application adopts a fatigue index calculation formula based on median frequency and average power frequency, comprehensively considering the decline in electromyographic spectral characteristics and the efficiency change in the ratio of electromyographic amplitude to current intensity, which can quantify the degree of muscle fatigue in multiple dimensions. When the fatigue index is in different threshold ranges, corresponding compensation strategies are adopted. For mild fatigue, the stimulation frequency is increased to utilize the muscle frequency response characteristics; for moderate fatigue, synergistic muscle groups are activated to distribute the load; and for severe fatigue, forced rest is implemented to protect the muscles, which helps to implement targeted interventions at different fatigue stages.

[0070] In terms of brain signal decoding, this application adopts a hierarchical decoding architecture. The first layer identifies high-level motor intentions such as grasping, releasing, holding, and reaching. The second layer decodes continuous motor parameters such as desired grasping force and desired joint angle, and dynamically adjusts the mapping range and weight of the parameters according to the task type. This combination of hierarchical decoding and task-adaptive mapping ensures that the decoding results have both the accuracy of intention recognition and the precision of parameter extraction, which helps to improve the matching degree between the desired motor state and the patient's true intention.

[0071] Regarding the adjustment of stimulation parameters, this application specifies a first threshold as a specific proportion of the normalized error and a second threshold as a larger proportion of the normalized error, and employs differentiated parameter adjustment strategies for different error ranges. When the error is small, the adjustment amount of current intensity is proportional to the force error component; when the error is moderate, the adjustment amounts of current intensity, stimulation frequency, and pulse width are proportional to the force error component, angle error component, and activation error component, respectively. This proportional relationship between parameter adjustment amounts and error components allows the adjustment of stimulation parameters to specifically correct deviations in corresponding dimensions, helping to improve the efficiency and specificity of parameter optimization.

[0072] Regarding safety, this application limits stimulation parameters to specific safety ranges. Clear upper and lower limits are set for current intensity, stimulation frequency, and pulse width. The safety check procedure verifies whether new stimulation parameters are within the safety range, and checks whether joint angles and grip strength exceed physiological safety limits. This multi-layered safety check mechanism helps prevent overstimulation and abnormal movements, ensuring patient safety.

[0073] Regarding system real-time performance, this application sets the update cycle of the closed-loop iteration within a shorter time range, enabling the system to perform the perception-decision-execution cycle at a higher frequency. This helps to respond promptly to changes in motion state and updates to patient intentions, thereby improving the system's real-time control performance.

[0074] Regarding data synchronization, this application uses hardware timestamps to ensure the time alignment of joint angle signals, gripping force signals, and electromyographic signals. The time alignment error is controlled within a very small range, which helps to ensure that multimodal signals correspond to the motion state at the same moment and improves the accuracy of weighted fusion results.

[0075] Regarding the smoothness of parameter adjustment, this application sets a single parameter adjustment range not exceeding a specific proportion of the current parameter value, which helps to avoid muscle spasms or patient discomfort caused by sudden changes in stimulation parameters, thereby improving the stability and comfort of system operation.

[0076] In summary, this application achieves precise reconstruction of upper limb grasping function in spinal cord injury patients within a closed-loop control framework through the organic combination and synergistic cooperation of multimodal feedback fusion, error grading adjustment, and fatigue prediction compensation techniques. Compared to existing open-loop or simple closed-loop systems, this application helps improve motion control accuracy, enhances the system's adaptability to dynamic changes, prolongs continuous motion time, and improves patient comfort and safety, providing an effective technical solution for the clinical rehabilitation treatment of upper limb grasping dysfunction caused by spinal cord injury.

[0077] 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

[0078] Figure 1 This is a flowchart illustrating the upper limb grasping function reconstruction method based on multimodal biofeedback using a closed-loop brain-spinal cord interface according to the first embodiment of this application.

[0079] Figure 2 This is a schematic diagram of the structure of the upper limb grasping function reconstruction system based on multimodal biofeedback and a closed-loop brain-spinal cord interface according to the second embodiment of this application. Detailed Implementation

[0080] 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.

[0081] Explanation of some concepts:

[0082] Brain-Computer Interface (BCI): A communication system that does not rely on peripheral nerves and muscles. It identifies the user's motor intentions or cognitive state by collecting and decoding brain neural signals (such as EEG, cortical EEG, or neuronal action potentials), thereby controlling external devices or realizing neural function reconstruction.

[0083] Functional Electrical Stimulation (FES) is a neurorehabilitation technique that artificially activates paralyzed or impaired muscles to contract by applying brief electrical pulses to peripheral nerves or muscles, in order to restore or assist in performing specific functional movements (such as grasping, standing, walking, etc.).

[0084] Closed-Loop Control: A control system architecture that uses sensors to collect the output state of the controlled object in real time, compares it with the desired state to generate an error signal, and adjusts the control input according to the error signal, forming a "input-output-feedback-adjustment" loop to achieve precise and stable control.

[0085] Intracranial electrodes are electrode devices implanted within the skull, directly contacting the cerebral cortex or placed under the dura mater, used to acquire high spatiotemporal resolution neural electrical signals. The intracranial electrodes in this application include types such as electrocorticography (ECoG) electrodes and endovascular stent electrodes.

[0086] Desired Movement State: A set of parameters representing the target movement that the patient intends to perform, obtained by decoding neural signals from the patient's motor cortex. These parameters include continuous or discrete motion descriptors such as desired joint angles, desired grip strength, and desired movement speed.

[0087] Actual Movement State: A set of motion parameters that the patient's limbs are currently performing, collected and fused in real time by multimodal sensors, which truly reflects the actual posture, force, and muscle activation state of the limbs under the action of nerve electrical stimulation.

[0088] Multimodal biofeedback is a feedback mechanism that simultaneously acquires biosignals from different types of sensors (such as kinematic signals from inertial measurement units, dynamic signals from force sensors, and physiological electrical signals from surface electromyography sensors) and integrates information from multiple dimensions to comprehensively evaluate the system's state.

[0089] Inertial Measurement Unit (IMU): A miniature sensor device that integrates accelerometers, gyroscopes, and magnetometers. It calculates the attitude angle and trajectory of an object in real time by measuring acceleration, angular velocity, and magnetic field direction in three-dimensional space.

[0090] Surface electromyography (sEMG) is a signal of muscle electrical activity collected by electrodes attached to the skin surface. It reflects the overall electric field changes of the action potential of motor units during muscle contraction and can be used to assess muscle activation, contraction intensity, and fatigue status.

[0091] Weighted Fusion: An information fusion method that assigns appropriate weight coefficients to data from different sources and combines them according to specific mathematical formulas (such as weighted summation, weighted average, etc.) to generate a comprehensive fusion result. The weight coefficients are usually dynamically adjusted according to factors such as the reliability of each data source and task requirements.

[0092] Error Vector: A vector consisting of the deviations between the desired state and the actual state in various dimensions. It contains information about the magnitude, direction, and components of the error in each dimension and is the core signal for adjusting the driving parameters in closed-loop control.

[0093] Hierarchical Adjustment Strategy: This strategy divides the control into multiple levels based on the magnitude and type of error, employing different adjustment methods. For small errors, fine-tuning is used to avoid excessive disturbance; for medium errors, multi-parameter joint adjustment is used to accelerate convergence; and for large errors, strategy-level changes are used to achieve rapid correction.

[0094] Fatigue Index (FI): A comprehensive index for quantitatively assessing the degree of muscle fatigue. It is calculated by analyzing the frequency domain characteristics (such as median frequency and average power frequency) and time domain characteristics (such as the root mean square value to stimulus intensity ratio) of electromyographic signals. The numerical range is usually between 0 and 1, where 0 represents no fatigue and 1 represents severe fatigue. It should be noted that the original calculated value of the fatigue index FI(t) may exceed the range of [0,1]. In actual implementation, the system processes FI(t) as follows: 1. If FI(t) < 0, it is truncated to 0 (indicating no fatigue); 2. If FI(t) > 1, it is truncated to 1 (indicating extreme fatigue); 3. Under normal circumstances, FI(t) is between 0 and 1, linearly reflecting the degree of fatigue.

[0095] Median Frequency (MDF): The median frequency of the electromyographic (EMG) signal power spectrum, defined as the frequency point that divides the power spectrum area into two equal halves. During muscle fatigue, the decreased conduction velocity of muscle fibers causes the EMG signal to shift to lower frequencies, manifested as a decrease in the median frequency.

[0096] Mean Power Frequency (MF): The weighted average frequency of the power spectrum of electromyographic signals, calculated by summing the products of frequency and power and dividing by the total power. Similar to the median frequency, the mean power frequency also shows a decreasing trend during muscle fatigue and is an important indicator for assessing fatigue.

[0097] Fast Fourier Transform (FFT): An efficient algorithm for calculating the Discrete Fourier Transform, capable of converting time-domain signals to the frequency domain, revealing the frequency components and power distribution of the signal, and widely used in the frequency domain analysis of electromyographic signals.

[0098] Fatigue Compensation Strategy: Active interventions for muscle fatigue, including graded compensation methods such as increasing stimulation frequency to delay fatigue by utilizing the frequency-force relationship, activating synergistic muscle groups to share the load, and triggering rest modes to allow muscles to recover.

[0099] Electrode Activation Pattern: The set of electrodes selected for simultaneous activation in a multi-electrode stimulation system, along with their stimulation parameters. By switching different activation patterns, the distribution of stimulated muscle groups and nerves can be altered, resulting in different movement patterns or the distribution of fatigue load.

[0100] A neural electrical stimulator is a medical electronic device that can generate programmable electrical pulse sequences and output them to nerve or muscle tissue through electrodes. The current intensity, frequency, pulse width, and other parameters of each electrode channel can be independently controlled.

[0101] Safe Stimulation Parameter Range: The boundaries of stimulation parameters determined through individualized testing for a specific patient, including the working range between the motor threshold (the minimum stimulus that produces visible contraction) and the comfort upper limit (the maximum stimulus that the patient can tolerate), ensuring that the stimulation is both effective and safe and comfortable.

[0102] Motor cortex: A cortical region located in the precentral gyrus of the frontal lobe of the brain, responsible for generating and controlling voluntary movement. It contains motor representation areas of various parts of the body, and its neuronal activity patterns encode motor intentions and motor parameter information.

[0103] Spinal cord injury (SCI): Damage to the spinal cord tissue caused by trauma, disease, or other reasons, resulting in motor, sensory, and autonomic dysfunction below the level of injury. It is classified according to the degree of injury as complete injury (complete loss of function below the level of injury) and incomplete injury (preservation of some function).

[0104] The following is a brief summary of some of the innovative aspects of this application:

[0105] In summary, existing technologies for addressing the clinical challenge of upper limb grasping dysfunction caused by spinal cord injury generally suffer from multiple technical bottlenecks, including insufficient execution precision due to open-loop control, the inability of a single feedback dimension to comprehensively represent complex movement states, and continuous decline in motor ability caused by muscle fatigue. This application creatively constructs a closed-loop brain-spinal cord interface based on multimodal biofeedback, thereby realizing the desired movement state... Refined decoding, actual motion state Multi-dimensional fusion evaluation and error vector The hierarchical adaptive adjustment strategy driven by the system is organically coupled to form a technical solution with inherent correlation and synergistic effect.

[0106] Specifically, this application does not simply physically superimpose brain-computer interfaces, functional electrical stimulation, and feedback sensors. Instead, it establishes an error vector calculation mechanism between the desired and actual motion states, enabling the system to identify the magnitude, type, and distribution characteristics of the error in a multi-dimensional space. Based on this, the system determines the error magnitude relative to a first threshold. Second threshold Based on the positional relationship, a differentiated parameter adjustment strategy is adopted: when the error is within... In the following small disturbance range, the system only responds to the current intensity Implement local fine-tuning to avoid control oscillations caused by overresponse; when the error crosses... But it has not yet been achieved. When the deviation is in the medium range, the system needs to simultaneously monitor the current intensity. Stimulation frequency and pulse width The stimulus parameters across three dimensions are coordinated and adjusted, with the adjustment amount of each parameter corresponding to the error vector. Force error component Angular error components and activation error components Establish a proportional correspondence; however, when the error exceeds... When the system enters a region of significant deviation, it determines that simple parameter amplitude adjustments are no longer effective in correcting the deviation and instead switches the electrode activation combination to fundamentally change the muscle recruitment pattern. This error-grading response mechanism allows the system to adopt an adjustment strategy adapted to different error states, ensuring both stable convergence under small errors and rapid reconstruction capability under large errors.

[0107] More importantly, this application combines error-based reactive adjustment with fatigue index-based adjustment. The proactive compensation is used in a synergistic, additive manner. Fatigue Index The calculation does not rely on a single physiological indicator, but rather on the median frequency of the electromyographic signals. Average power frequency The deviations from their respective baseline values, along with the efficiency index of the ratio of electromyographic amplitude to stimulation current, are weighted and fused to form a multi-dimensional quantitative assessment of muscle fatigue. When the fatigue index falls within different threshold ranges, the system executes corresponding compensation strategies: for mild fatigue, the frequency of stimulation is increased to utilize the frequency-force characteristics of muscles to slow the fatigue process; for moderate fatigue, synergistic muscle groups are activated to distribute the load; and for severe fatigue, a rest mode is forcibly triggered to protect muscle tissue. This fatigue prediction and compensation mechanism intervenes in the temporal dimension before a significant decline in movement quality, forming a dual guarantee of spatiotemporal coordination with error-based spatial adjustments.

[0108] Furthermore, in the multimodal feedback fusion process, this application incorporates the joint angle signals acquired by the inertial measurement unit. The gripping force signal collected by the force sensor Electromyography signals acquired by surface electromyography sensors Hardware-level time alignment is achieved through a time synchronization module, followed by task-adaptive weighting coefficients. , , The dynamic adjustment mechanism assigns differentiated contribution weights to each modal feedback signal based on the different emphases required by the current task regarding angular accuracy, force accuracy, and muscle activation coordination. This weighted fusion strategy integrates the overall motion state... The evaluation is no longer limited to the partial information of a single sensor or a rigid combination of fixed weights. Instead, it can dynamically optimize the information fusion method according to the task characteristics and sensor signal-to-noise ratio, thereby providing a more accurate and robust representation of motion state and a reliable reference benchmark for subsequent error calculation.

[0109] In summary, this application achieves high-accuracy characterization of motion state through multimodal feedback fusion, balances rapid convergence and stability in parameter optimization through an error grading adjustment strategy, and ensures long-term operational sustainability through fatigue prediction and compensation mechanisms. These three aspects are not independent but form a tightly coupled and synergistic relationship within a closed-loop control framework: accurate motion state assessment provides reliable input for error calculation, intelligent grading adjustment strategy ensures rapid error convergence, and forward-looking fatigue compensation guarantees long-term system stability. This organic integration of multi-level, multi-dimensional, and multi-timescale technical features constitutes the essential difference between this application and existing open-loop or simple closed-loop systems, which cannot be easily obtained by those skilled in the art through conventional improvements or simple combinations of existing technologies.

[0110] 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.

[0111] The first embodiment of this application relates to a closed-loop brain-spinal cord interface upper limb grasping function reconstruction method based on multimodal biofeedback, the process of which is as follows: Figure 1 As shown, the method includes the following steps:

[0112] The patient in this embodiment is a 32-year-old male who suffered a complete C6 segment spinal cord injury due to a traffic accident, resulting in complete loss of motor function below the level of injury, but retaining the ability to generate motor intentions. Furthermore, six months after the injury, the patient's neurological function was stable, meeting the indications for implantation of a brain-spinal cord interface system. The hardware platform used in the system includes: an intracranial cortical EEG electrode array, a multi-channel neurostimulator, an inertial measurement unit, a force sensor, a surface electromyography sensor, and an embedded computing platform.

[0113] Step 100: System Initialization and Calibration

[0114] When the system is first started, individualized calibration is required, which is fundamental to ensuring the effectiveness of subsequent closed-loop control. Because there are significant individual differences in brain signal characteristics, muscle responsiveness, and neural excitability among different patients, fixed control parameters cannot be applied to all patients. Therefore, the core task of step 100 is to establish an individualized brain signal decoding model for the patient and determine a safe and effective range of stimulation parameters for that patient.

[0115] Step 110: First, an individualized brain signal decoding model is established. Specifically, under the guidance of a physician, the patient is asked to imagine or attempt to perform various motor actions, including four basic high-level motor intentions: grasping, releasing, holding, and reaching, as well as grasping at different strengths (light grasping, moderate grasping, and forceful grasping) and hand movements in different postures (wrist flexion and extension). During the patient's performance of these imagined movements, cortical EEG electrodes implanted in the motor cortex are continuously collected to acquire neural signals from the motor cortex. Each motor action is repeated 10-15 times, each lasting 3-5 seconds, for a total acquisition time of approximately 10 minutes. After preprocessing, the acquired neural signals are used to extract temporal features (such as signal amplitude and variance) and frequency features (such as... Frequency band power, high (Frequency band power), forming a high-dimensional feature vector.

[0116] Furthermore, the collected labeled data was divided into training and testing sets in a 7:3 ratio. For the first layer of the hierarchical decoding architecture, a support vector machine was used as a classifier to train it to recognize four discrete high-level motion intentions (grasp, release, hold, and reach). After training, the classification accuracy was verified on the test set, reaching 92% in this embodiment, which meets the requirements for clinical application. For the second layer of the hierarchical decoding architecture, a recurrent neural network was used as a regression model to train it to decode continuous motion parameters, including the expected grip strength. and desired joint angle The output of the regression model is normalized, and the mapping range and weights of the continuous motion parameters are dynamically adjusted according to the current task type (such as fine grasping or strength grasping tasks). For example, for fine grasping tasks (such as holding a pen). It is mapped to a range of 0.5N to 5N, with a force resolution of 0.1N; while for force-gripping tasks (such as holding a water cup). It is mapped to a range of 5N to 50N, with a velocity resolution of 1N. Through this task-adaptive mapping, the decoding output can more accurately match the actual task requirements.

[0117] Step 120: Next, establish the biofeedback baseline parameters. The patient remains relaxed and still, with the forearm resting naturally on the armrest and the hand in a neutral position. Baseline data is collected for 30 seconds in this state. The joint angle baseline at rest is recorded using an inertial measurement unit. In this embodiment, the wrist joint is in a neutral position at 0°, and the metacarpophalangeal joints are naturally slightly flexed at approximately 15°. Resting electromyographic (EMG) signals of the forearm flexor and extensor muscle groups are collected using a surface electromyography (SEMG) sensor, and the root mean square (RMS) value of the resting EMG is calculated. In this embodiment, it is approximately 3.2 μV, which represents the background electrical activity level of the muscle in a fully relaxed state.

[0118] More specifically, a fast Fourier transform is performed on the resting electromyographic signal to calculate the baseline median frequency. and baseline average power frequency In this embodiment, the frequencies are 95Hz and 102Hz, respectively. These two frequency domain parameters will serve as the reference baseline for subsequent fatigue monitoring. The baseline current intensity will also be recorded. In this embodiment, the initial test current value is determined in step 130, the test phase.

[0119] Step 130: Next, determine the safe range of stimulation parameters for this patient. The stimulation electrodes were pre-implanted via minimally invasive surgery into the epidural space of the C6 segment of the cervical spinal cord, near the anterior horn motor neurons that innervate the forearm flexor and extensor muscles. The stimulation parameters were determined using a progressive testing method. Starting with a minimum safe current of 5 mA, the current intensity was gradually increased in 2 mA increments, while maintaining a stimulation frequency of 50 Hz and a pulse width of 250 μs. At each current intensity, the patient's forearm muscles were observed for visible contraction, and the patient's subjective sensations were recorded.

[0120] Specifically, when the current intensity reached 12 mA, a slight but visible contraction of the forearm flexor muscles was observed for the first time, with slight flexion of the fingers. At this point, the motor threshold current was determined to be 12 mA. The current intensity was further increased, and when it reached 38 mA, the patient reported feeling a noticeable muscle contraction that remained within a comfortable and acceptable range, without pain or discomfort. Therefore, the upper limit of comfort current was determined to be 38 mA. Considering safety margins, the operating range of the current intensity for this patient was set between 12 mA and 35 mA. Similarly, the effective range of stimulation frequency was determined through testing to be 20 Hz to 100 Hz, and the pulse width range was determined to be 100 μs to 500 μs. These parameter ranges will serve as boundary constraints for subsequent closed-loop control adjustments of stimulation parameters, ensuring that stimulation is always kept within a safe and effective range.

[0121] Step 140: Finally, set the initial stimulation parameters. For a moderate grasping task (the target task in this example is to grasp a plastic cup containing 200ml of water, requiring a grasping force of approximately 10-15N), the initial stimulation parameters are set as follows: Current intensity The stimulation frequency is 23mA (approximately 60% of the comfort limit). 50Hz, pulse width The activation time was 250 μs, with the primary flexor muscle electrode group as the activation electrode set. These initial parameters were empirical values ​​based on the test results from the calibration phase, capable of generating muscle contractions close to the target force, providing a reasonable starting point for subsequent closed-loop adjustments.

[0122] It should be noted that the entire calibration process in step 100 takes about 15-20 minutes. The established individualized model and parameter range are stored in the system and do not need to be recalibrated every time they are used. However, it is recommended to perform a calibration update every 1-2 weeks to adapt to long-term changes in the patient's neuromuscular state.

[0123] Step 200: Brain Signal Acquisition and Intent Decoding

[0124] When the patient is ready to begin the grasping task, the system enters real-time operation mode. The core task of step 200 is to continuously acquire the patient's motor cortex neural signals and decode the patient's current desired motor state in real time. .

[0125] Step 210: Acquire motor cortical neural signals via intracranial electrodes. In one embodiment, the cortical EEG electrode array used in this embodiment comprises, for example, 64 microelectrode contacts, evenly distributed in the hand representation area of ​​the motor cortex at a spacing of, for example, 4×4 mm. The electrode sampling rate is set to 1000 Hz, with a bandwidth of 0.5 Hz to 250 Hz, capable of capturing rich neural activity information, including low-frequency local field potentials and high-frequency action potentials.

[0126] Specifically, the acquired raw neural signals first undergo preprocessing, including bandpass filtering (preserving the 0.5-250Hz frequency band), notch filtering (removing 50Hz power frequency interference), and artifact removal (removing physiological interference signals such as eye movement and electrocardiogram). The preprocessed signal is segmented into 500ms time windows with a 250ms overlap between windows to achieve smooth continuous decoding.

[0127] Step 220: Decode the preprocessed neural signals using the individualized brain signal decoding model established in Step 100. The decoding process adopts a hierarchical architecture. In the first layer, feature vectors are extracted from the neural signals in each 500ms time window and input into a support vector machine classifier. The classifier outputs the high-level motor intention category at the current moment; in this embodiment, the patient's current intention is identified as "grasping".

[0128] Furthermore, in the second layer, the same feature vectors are input into the recurrent neural network regression model to decode continuous motion parameters. The regression model outputs the expected grip strength. 14N, desired wrist joint angle The angle is -10° (indicating mild buckling), and the desired motion speed is 3° / s. These continuous parameters are adaptively mapped for the task, since the current task is a routine grasping task. Mapped to a range of 5N to 50N to ensure that the force resolution is suitable for the task requirements.

[0129] Step 230: Smooth the decoded output. Since neural signals inherently exhibit random fluctuations, directly using the raw decoded result may lead to jitter in the desired state, affecting the stability of subsequent control. Therefore, a moving average filter is used to smooth the continuous parameters, with a window length of 5 sampling points (corresponding to 250ms). The smoothed desired motion state. Includes desired joint angle =-10° and expected grip strength =14N, this state vector will serve as the target reference for subsequent stimulus execution and closed-loop control.

[0130] It should be noted that the entire process of step 200 is continuously updated in 50ms cycles, meaning the system decodes the patient's latest motor intention every 50ms to ensure rapid response to changes in the patient's intention. This update frequency is sufficient to track the dynamic characteristics of most voluntary movements without placing an excessive burden on the computing platform.

[0131] Step 300: Execution of Initial Stimulus

[0132] Obtain the desired motion state Then, the system needs to convert it into specific neural electrical stimulation parameters and apply them to the target nerve or muscle through a neural electrical stimulator to generate corresponding motor output.

[0133] Step 310: Based on the desired motion state Initial stimulus parameters are generated. This application generates parameters based on the mapping relationship between the desired state and stimulus parameters established in step 100 of the calibration phase. More specifically, this mapping relationship is based on the force-current relationship curve and angle-activation combination correspondence table obtained during the calibration process.

[0134] For example, regarding the expected grip strength =14N, the required current intensity is calculated based on the force-current relationship curve established in step 100 calibration stage. Approximately 24mA. For the desired joint angle... =-10°, requiring activation of the forearm flexor muscles while moderately inhibiting the antagonistic effect of the extensor muscles, therefore the primary flexor electrode group was selected as the activation combination. Stimulation frequency The frequency was set to 50Hz, a suitable frequency that produces stable fusion contraction without easily causing rapid fatigue. Pulse width. A pulse width of 250 μs was set, which effectively activated motor neurons. The generated initial stimulation parameters were { =24mA, =50Hz, =250μs, electronode_set=flexor main electrode group}.

[0135] Step 320: Apply electrical nerve stimulation to the target nerve or muscle using a neurostimulator. The neurostimulator is a multi-channel programmable device capable of independently controlling the stimulation parameters of each electrode channel. In this embodiment, the stimulation target is the anterior horn of the epidural spinal cord at the C6 segment of the cervical spinal cord, and precise stimulation is achieved through an implanted multi-contact electrode array.

[0136] Specifically, based on the generated stimulation parameters, the neurostimulator sends a biphasic square wave pulse sequence with a current intensity of 24 mA, a frequency of 50 Hz, and a pulse width of 250 μs to the flexor muscle master electrode group. The current pulses are transmitted to the spinal cord tissue through the electrodes, activating the α motor neurons that innervate the forearm flexor muscles, thereby causing muscle contraction. Due to the stimulation frequency of 50 Hz, the rapid superposition of individual muscle twitches forms a fused tetanic contraction, manifesting as a smooth, continuous grasping motion.

[0137] Step 330: While the stimulus is being delivered, the system records the timestamp of the stimulus initiation. This is crucial for the timing alignment of subsequent multimodal feedback signals. Because there is a certain electromechanical delay (typically 20-50 ms) between the stimulus delivery, muscle contraction, and sensor detection of the feedback signal, accurate timestamp recording ensures that the feedback signal is correctly correlated with the corresponding stimulus parameters.

[0138] After step 300 is completed, the nerve electrical stimulation begins to continuously act on the target nerve, and the patient's hand begins to make grasping movements. At this point, the system immediately proceeds to step 400 to begin collecting motor feedback signals.

[0139] Step 400: Multimodal Feedback Acquisition

[0140] The core task of step 400 is to synchronously acquire the actual movement state of the patient's limbs using multiple types of sensors, providing accurate feedback information for subsequent closed-loop control. One of the key innovations of this application is the use of at least three types of feedback signals to comprehensively characterize complex grasping movements from different dimensions.

[0141] Step 410: Acquire joint angle signals using an inertial measurement unit (IMU). The IMU is fixedly mounted on the back of the patient's hand and forearm, and can measure the hand's posture in three-dimensional space. This embodiment focuses on the angular changes of the wrist joint in the flexion and extension directions, and the sampling rate of the IMU is set to 100Hz.

[0142] Specifically, when nerve electrical stimulation causes contraction of the forearm flexor muscles, the hand begins to flex towards the palmar side, and the inertial measurement unit records the trajectory of the wrist joint angle change in real time. 300ms after the stimulation begins, the wrist joint angle reaches -8°, close to but not yet fully reaching the desired angle of -10°. After Kalman filtering to remove high-frequency noise, the angle data is output as a smooth angle time series.

[0143] Step 420: Acquire gripping force signals using force sensors or pressure sensors. Force sensors are arranged on the contact surfaces of the palm and fingers to measure the contact force between the fingers and the object being gripped. In one embodiment, a thin-film pressure sensor is installed on the palmar side of the thumb, index finger, and middle finger, with a sampling rate set to 50Hz.

[0144] Furthermore, when the fingers begin to touch and grasp the cup, the sensors detect pressure signals, and the system calculates the sum of the readings from each sensor as the total gripping force. 400ms after the stimulus begins, the following measurements were taken: The value was 11N, which is about 21% lower than the expected force of 14N. After the force data was low-pass filtered to remove transient pressure fluctuations, the output was a stable force time series.

[0145] Step 430: Acquire electromyographic signals using surface electromyography (EMG) sensors. Surface EMG electrodes are attached to the skin of the forearm flexor muscles (flexor carpi radialis) and extensor muscles (extensor carpi radialis) to record electrical activity during muscle activation. The EMG signal sampling rate is set to 1000 Hz to capture the high-frequency components of the EMG signal.

[0146] More specifically, the acquired raw electromyography (EMG) signals are bandpass filtered (20-500Hz), rectified by full-wave rectification, and then lowpass filtered to calculate the EMG envelope. The root mean square (RMS) value of the EMG signal within a 200ms time window is further calculated. As a quantitative indicator of muscle activation, the root mean square value of electromyography (EMG) of the flexor muscle group was measured 400 ms after the onset of stimulation. The electromyographic value was 18.5 μV, which is 5.8 times the resting baseline value of 3.2 μV, indicating that the muscle was effectively activated. At the same time, the root mean square value of the electromyographic value of the extensor muscle group remained at 4.1 μV, close to the resting level, indicating that the antagonist muscle was not accidentally activated.

[0147] Step 440: Achieve precise time synchronization of the three types of feedback signals. Since the inertial measurement unit, force sensor, and surface electromyography sensor have different sampling rates and data transmission delays, hardware timestamps must be used to ensure time alignment of the three types of signals. The system uses a unified hardware clock, embedding microsecond-level precision timestamps in each sensor data packet.

[0148] Specifically, after receiving data from each sensor, the data acquisition module aligns the data from different sources to a unified time grid based on the timestamps, with the time alignment error controlled within 1ms. For sensor data with a low sampling rate (such as 50Hz from a force sensor), a linear interpolation method is used to upsample to a unified time grid (20Hz, i.e., one data point every 50ms), ensuring that there is corresponding synchronous data available for the three types of signals in each control cycle.

[0149] Step 450: Standardize the three types of signals after synchronization. The joint angle signal is standardized to a normalized deviation relative to the resting baseline. In this embodiment, the current angle is -8° relative to the baseline of 0°, and is normalized to... =-0.8 (Assuming the joint angle working range is ±10°, the normalization formula is) = / 10).

[0150] Furthermore, the gripping force signal is normalized to a ratio relative to the maximum force. In this embodiment, the current force of 11N is normalized to the maximum force of 50N. =0.22. The electromyographic signal is normalized to a multiple of the resting baseline; in this embodiment, the current... The ratio of 18.5 μV to the baseline of 3.2 μV is 5.8. To make it suitable for subsequent fusion calculations, a boundary adjustment process is used to map this ratio to a reasonable range, such as logarithmic compression. = (1 + ratio) / (1 + maximum expected ratio), or normalization mapping using the Sigmoid function, so that The final result falls within the [0,1] interval, which facilitates weighted fusion.

[0151] It should be noted that these standardized data will be weighted and fused in step 500. After step 400 is completed, the system obtains multi-dimensional measurements of the actual motor state of the patient's limbs after stimulation. This feedback information truly reflects the actual effect of nerve electrical stimulation and is the key to achieving closed-loop control.

[0152] Step 500: Comprehensive Status Assessment and Error Calculation

[0153] The core task of step 500 is to fuse the multimodal feedback signals collected in step 400, generate a unified representation of the actual motion state, and compare it with the expected motion state to calculate the error, thus providing a basis for subsequent adaptive adjustment.

[0154] Step 510: Weight and fuse the multimodal feedback signals to generate the actual motion state. The weighted fusion method proposed in this application can comprehensively utilize complementary information provided by different sensors, thereby improving the accuracy and robustness of state assessment.

[0155] Actual motion state It is a multi-dimensional vector, containing the actual joint angles. Actual grasp strength and actual muscle activation Three dimensions. The calculation formula for each dimension is as follows:

[0156]

[0157]

[0158]

[0159] in, The signal is collected and standardized by the joint angle sensor (inertial measurement unit); The signal acquired and standardized by the force sensor; The signal is collected and standardized by the surface electromyography sensor.

[0160] Weighting coefficient , , The weighting is dynamically adjusted based on current task requirements and sensor signal-to-noise ratio. For example, for the everyday grasping task (grasping a water cup) in this embodiment, the core requirement is accurate and stable gripping force, with the joint angles reaching a roughly appropriate posture. Muscle activation status is mainly used to assist in judgment and fatigue monitoring. Therefore, the weighting coefficient is set to... (Joint angle) (grip strength) (Electromyographic signals), this weighting highlights the importance of force control in this task.

[0161] Ultimately, the actual motion state is represented in vector form:

[0162]

[0163] This multidimensional vector form can fully preserve the independent information of each dimension, which facilitates subsequent calculation of errors by dimension.

[0164] For the everyday grasping task (grasping a water cup) in this embodiment, the core requirement is accurate and stable gripping force, with the joint angles reaching a roughly appropriate posture. Muscle activation status is mainly used to assist in judgment and fatigue monitoring. Therefore, the weighting coefficient is set to... =0.3 (joint angle) =0.5 (grip strength) =0.2 (electromyographic signal), this weighting emphasizes the importance of force control in this task. For example, in other tasks (such as reaching for an object), it might be adjusted to... =0.6, =0.2, =0.2, to emphasize the accuracy of the joint angle.

[0165] Step 520: Calculate the desired motion state With actual motion state Error vector between .

[0166] The error vector is the core driving signal of closed-loop control, and its magnitude and direction determine the strategy for subsequent parameter adjustment.

[0167] because and Both are multidimensional vectors, error vectors It is also a multidimensional vector, calculated one dimension at a time:

[0168]

[0169] Specifically, the error components for each dimension are:

[0170]

[0171]

[0172]

[0173] For this embodiment, the desired motion state Actual motion state Calculate the error in each dimension:

[0174] Angular error components: Force error component:

[0175] Specifically, a negative sign indicates that the actual angle is less than the expected angle (further buckling is needed), and a positive sign indicates that the actual force is less than the expected force (more force is needed). In addition, the activation error component can be calculated. This represents the difference between the actual degree of muscle activation and the desired degree of activation, which will not be elaborated on in this embodiment.

[0176] Step 530: Calculate the overall magnitude of the error to assess its severity. The magnitude of the error vector is calculated using Euclidean distance:

[0177]

[0178] To facilitate subsequent threshold determination, the error is normalized as a percentage relative to the desired state. For the angle error, the relative error is |-2°| / 10°=20%; for the force error, the relative error is |3N| / 14N≈21%. Taking the weighted average of the two (with the same weights as in step 510), the normalized comprehensive error is approximately 20.6%.

[0179] Step 540: Identify the main type of error. By comparing the relative magnitudes of errors in each dimension, determine whether the current error is force-dominated, angle-dominated, or a mixture. In this embodiment, the relative angle error of 20% is close to the relative force error of 21%, therefore it is determined to be a mixture error. This determination will be used to select an appropriate adjustment strategy in the parameter adjustment in step 700.

[0180] After step 500 is completed, the system obtains clear error information. This includes the magnitude of the error (20.6%), and the components of each dimension ( , The information includes the error type (mixed type) and provides clear guidance for subsequent adaptive parameter tuning.

[0181] Step 600: Muscle Fatigue Monitoring and Prediction

[0182] During continuous operation of closed-loop control, prolonged neural electrical stimulation can easily induce muscle fatigue, leading to a decline in the quality of motor output or even failure. The core task of step 600 is to monitor and predict muscle fatigue status in real time based on the frequency domain characteristics of electromyographic signals, and to actively generate compensation strategies before fatigue significantly affects movement. This is one of the important innovations of this application.

[0183] Step 610: Perform frequency domain analysis on the acquired electromyographic signals. Frequency domain analysis employs the Fast Fourier Transform (FFT) method, which converts the time-domain electromyographic signals to the frequency domain, revealing the frequency component distribution of the signal. In one embodiment, the FFT window length is set to 500 ms, and the window overlap rate is 50%, meaning that each window is shifted forward 250 ms before calculating the next Fourier transform, thereby achieving smooth and continuous updating of frequency domain features.

[0184] More specifically, for an electromyography signal with a sampling rate of 1000 Hz, a 500 ms window contains 500 sampling points, and the power spectral density distribution in the frequency range of 0 to 500 Hz is obtained after fast Fourier transform.

[0185] Step 620: Extract fatigue-related frequency domain characteristic parameters from the power spectrum. The two most important parameters are the median frequency and the median frequency. and average power frequency The median frequency is defined as the frequency point that divides the power spectrum area into two equal halves, i.e., the frequency at which the accumulated power reaches 50% of the total power. The average power frequency is defined as the first moment of the power spectrum, i.e., the weighted average of the product of frequency and power.

[0186] Specifically, in this embodiment, at a time 400ms after the stimulus begins (i.e., when feedback is acquired in step 400), the following is calculated: , These two values ​​are the same as the baseline values ​​established in step 100. , The values ​​are very close, indicating that the muscles are not yet fatigued. These frequency domain parameters will gradually change as the stimulation continues.

[0187] Step 630: Calculate the fatigue index The fatigue index is used to quantitatively assess muscle fatigue. It is calculated based on a combination of multiple physiological indicators using the following formula:

[0188] Formula (3)

[0189] in, , , As the weighting coefficient, it is set to in this embodiment. =0.4, =0.4, =0.2, this weighting is determined based on literature research and clinical experience, emphasizing the dominant role of frequency domain parameters in fatigue assessment. and These are the median frequency and the average power frequency at the current moment, respectively. and The baseline median frequency and the baseline average power frequency are established by step 100, respectively. This represents the root mean square value of the current electromyographic signal. This represents the current current intensity. and These are the baseline root mean square value and the baseline current intensity, respectively.

[0190] The physical meaning of Formula 3 can be understood as follows: When muscles are fatigued, the decrease in muscle fiber conduction velocity causes the electromyographic signal to shift to lower frequencies, which manifests as... and Decrease, therefore and Both of these increase with fatigue. Simultaneously, during fatigue, maintaining the same force output requires recruiting more motor units or increasing the firing frequency, but stimulation efficiency decreases. This is reflected in the electromyographic signal as a weakening of the effective electromyographic response produced under the same current. The ratio decreased relative to the baseline level, therefore This efficiency loss index increases with increasing fatigue. FI(t) is normalized and limited to the interval [0,1], with negative values ​​truncated to 0.

[0191] It should be noted that in the early stages of exercise, muscles are in an activated contraction state, and the stimulation efficiency ratio is low. A value much greater than 1 results in a negative value for the third term, reflecting the normal phenomenon that muscle activation efficiency is higher than the resting baseline. In practical implementation, Limited to the range [0,1], negative values ​​are truncated to 0. Therefore, in this embodiment at this moment... A value of 0 indicates that the muscles are in good condition and show no signs of fatigue. As stimulation continues, and It will gradually decrease, and the efficiency ratio will drop to less than 1. It will gradually increase and truly reflect the degree of fatigue.

[0192] Step 640: Based on the fatigue index Determine a fatigue compensation strategy. This application proposes a graded fatigue compensation strategy, based on... Different intervention measures are used to determine the numerical range.

[0193] In one specific embodiment, fatigue threshold , and The values ​​are set to 0.6, 0.75, and 0.85 respectively, corresponding to mild fatigue warning, moderate fatigue, and severe fatigue. Specific strategies include:

[0194] when At this point, it is determined to be a mild fatigue warning state. The compensation strategy adopted at this time is to increase the stimulation frequency. 5-10Hz. Increasing the stimulation frequency can utilize the frequency-force relationship of muscles to generate stronger muscle contraction tension under the same current intensity, thereby offsetting the decrease in strength caused by fatigue to some extent. At the same time, moderately increasing the frequency can also improve the synchronization of motor units and delay the fatigue process.

[0195] when At this point, the condition is assessed as moderate fatigue. Simply increasing the frequency is insufficient for effective compensation, requiring a more aggressive strategy: activating synergistic muscle groups for combined stimulation. More specifically, this involves shifting some of the stimulation load to other synergistic muscle groups that are not yet fatigued or have a milder level of fatigue. For example, if the primary flexor muscles (flexor carpi radialis) are moderately fatigued, then activating accessory flexors (flexor carpi ulnaris and flexor digitorum profundus) to participate in the grasping action, through multi-muscle group synergy to share the load, maintains motor output while providing partial recovery time for the primary muscle groups.

[0196] when At this point, the condition is determined to be severe fatigue. Continuing stimulation at this stage will not only fail to produce effective motor output but may also exacerbate fatigue or even cause muscle damage. Therefore, a rest mode is triggered, immediately pausing the task or reducing the task intensity by 50%, and reassessing the fatigue state after 20-30 seconds. During the rest period, the stimulation intensity is reduced to maintain a minimum activation level or stopped completely to allow the muscles to fully recover metabolically. Upon reassessment, if... If it drops below 0.5, resume normal exercise; if If the level remains above 0.7, continue resting or advise the patient to end the current training task.

[0197] In this embodiment, the time is 400ms after the start of stimulation. The readings are well below the mild fatigue warning threshold of 0.6, therefore no fatigue compensation strategy needs to be activated at present. However, the system will continue to monitor the situation. The changing trend, once detected When the warning threshold is approached, corresponding compensation measures will be prepared in advance.

[0198] Step 650: Perform fatigue trend prediction. In addition to assessing the current fatigue state, the system also records historical data. Data, for a recent period of time (e.g., 5 minutes) Linear regression analysis was performed on the changing trend. Extrapolation was used to predict the trend for the next 1-2 minutes. If the predicted value is about to exceed a threshold, an early warning is issued and the strategy is adjusted accordingly. This proactive fatigue management is a significant advantage of this application over passive response systems.

[0199] After step 600 is completed, the system obtains the current fatigue state assessment and corresponding compensation strategy suggestions. This information will be combined with the error-driven parameter adjustment strategy in step 700 to achieve more comprehensive adaptive control.

[0200] Step 700: Adaptive Stimulus Parameter Adjustment

[0201] Step 700 is the core decision-making module of closed-loop control. Based on the error information calculated in step 500 and the fatigue state obtained in step 600, new stimulus parameters are generated. The hierarchical adjustment strategy proposed in this application can adopt differentiated parameter optimization methods according to the magnitude and type of error, which has a faster convergence speed and higher control accuracy compared with the traditional fixed gain control algorithm.

[0202] Step 710: Update the stimulus parameters using a tiered adjustment strategy based on the magnitude of the error. The system first determines the level of the current error. In this embodiment, the normalized comprehensive error is 20.6%, exceeding the first threshold. But less than the second threshold Therefore, it falls within the range of medium error and should be adjusted using a two-stage adjustment strategy.

[0203] Step 720: Execute the secondary adjustment strategy, that is, adjust the current intensity simultaneously. Stimulation frequency and pulse width Three parameters. This multi-parameter joint adjustment can optimize the stimulus effect from multiple dimensions, reducing errors faster than single-parameter adjustment. The specific adjustment amount is calculated as follows:

[0204] The formula for calculating the adjustment amount of current intensity is:

[0205] in This is a scaling factor, set to [value] in this embodiment. , Error vector The force error component in the equation. Substitute the values: This indicates that the current intensity needs to be increased by 3mA.

[0206] The formula for calculating the adjustment amount of the stimulation frequency is:

[0207] in This is a scaling factor, set to [value] in this embodiment. , Error vector The angular error component in the equation. Substitute the values: This indicates that the stimulation frequency needs to be reduced by 4Hz.

[0208] It should be noted that the frequency reduction is due to the current joint angle (-8°) being close to the desired value (-10°) and slightly insufficient; a slight reduction in frequency helps with fine-tuning and avoids overshoot. Furthermore, the proportional coefficient... The specific values ​​and symbols have been calibrated according to the patient's neuromuscular response characteristics during the individualized calibration process in step 100. Different patients may have different values ​​and symbols. The value used in this embodiment is the optimized result for this patient.

[0209] The formula for calculating the pulse width adjustment is:

[0210] in This is a scaling factor, set to [value] in this embodiment. , Error vector The activation error component in the equation. Assuming the current muscle activation level is slightly below the ideal value by about 5%, then... This indicates that the pulse width needs to be increased by 50 μs to enhance the nerve activation effect.

[0211] Apply these adjustments to the current parameters: New current intensity: New stimulation frequency: New pulse width: The electrode activation set remains unchanged: electrode_set = flexor main electrode set

[0212] Step 730: Further adjust the stimulus parameters using a fatigue compensation strategy. Based on the results of step 600, the current... There is no need to activate the fatigue compensation strategy. However, if at this time... If the muscle is within the mild fatigue warning range (0.6-0.75), the stimulation frequency needs to be increased by 5-10 Hz on top of the error adjustment. This superimposed adjustment reflects the multi-objective optimization design concept of this application: to correct the current error while taking into account the long-term muscle health status.

[0213] It should be noted that error-driven frequency adjustment is mainly used for fine-tuning motion trajectories within a small range, and its adjustment direction and magnitude depend on the specific error type and control strategy; while the frequency increase in fatigue compensation is based on the frequency-force relationship, maintaining or increasing muscle output force by increasing the stimulation frequency in a fatigued state. The two have different regulatory goals and mechanisms of action, and can be used in combination or executed independently according to priority; there is no logical contradiction.

[0214] Step 740: Apply safety limits and smoothing to the adjusted parameters. First, check if the parameters are within a safe range. In this embodiment, Within the safe range [12mA, 35mA], Within the safe range [20Hz, 100Hz], Within the safe range [100μs, 500μs], all parameters satisfy the safety constraints.

[0215] Furthermore, check that the magnitude of a single parameter adjustment does not exceed 20% of the current parameter value. Current adjustment magnitude. Frequency adjustment range Pulse width adjustment range All constraints are met. This limitation prevents muscle spasms or discomfort caused by sudden parameter changes, ensuring a smooth and comfortable adjustment process.

[0216] Step 750: Generate the final new stimulus parameters. After grading adjustments, fatigue compensation, and safety limitations, the new stimulus parameters generated in this embodiment are: { =27mA, =46Hz, =300μs, electrode_set=flexor main electrode set}. This set of parameters will be executed after a safety check in step 800.

[0217] After step 700 is completed, the system has intelligently generated optimized stimulation parameters based on the current error and physiological state, which can better achieve the desired motion state compared to the initial parameters.

[0218] Step 800: Safety Check and Stimulus Execution

[0219] Step 800 is the final execution stage of closed-loop control. Multiple safety verifications are performed on the new stimulation parameters generated in step 700 to ensure that the stimulation will not cause any harm to the patient. Then, the updated stimulation is implemented through the neurostimulator.

[0220] Step 810: Perform a parameter validity check on the new stimulus parameters. Verify that all parameters are within the safe stimulus parameter range determined in Step 100. The check includes:

[0221] Current intensity Within the safe range [12mA, 35mA], through

[0222] Stimulation frequency Within the safe range [20Hz, 100Hz], through

[0223] Pulse width Within the safe range [100μs, 500μs], through

[0224] All parameters have passed the validity check.

[0225] Step 820: Perform a physiological safety threshold check. In addition to the limitations of the stimulation parameters themselves, it is also necessary to check whether the actual motor output is within the physiological safety range to prevent excessive exercise from causing joint or soft tissue damage. The check includes:

[0226] Joint angle check: Current wrist joint angle -8°, the physiological safety range is [-60°, +60°], far from the extreme position, passing

[0227] Grip strength test: Current grip strength is 11N, maximum safe grip strength is 50N (considering the patient's long-term inactivity and weak muscle strength), not exceeding the limit, passed.

[0228] Electromyography (EMG) overactivation test: Current EMG signal The value is 18.5 μV, which is less than the overactivation threshold (typically set at 10 times the baseline, i.e., 32 μV).

[0229] Step 830: Monitor cumulative stimulus load. The system records the average stimulus intensity over the past 5 minutes to prevent long-term high-intensity stimulation from causing tissue damage or excessive fatigue. In this embodiment, since the exercise has just begun and the cumulative load is still low, monitoring is used. If the long-term average stimulus intensity is detected to exceed 80% of the comfort limit (i.e., exceeding 30.4 mA), and the fatigue index... If the value is also high (>0.7), the system will forcibly trigger the rest mode.

[0230] Step 840: Check the patient's active control. The system continuously monitors for emergency stop signals from the patient. Emergency stop can be achieved through specific brain signal patterns (such as strongly imagining the action of "stop," which is triggered upon decoder recognition) or an external button. In this embodiment, no stop signal was detected. If the patient feels uncomfortable or wishes to stop at any time, they can immediately activate the emergency stop function, and the system will terminate all stimulation within 50ms, ensuring the patient's autonomy and sense of security.

[0231] Step 850: Execute the updated stimulation parameters. After all safety checks are passed, the neural stimulator receives the new stimulation parameters. =27mA, =46Hz, =300μs, electrode_set=flexor main electrode group}, immediately switch to the new parameters for stimulation. More specifically, the parameter switching adopts a soft switching method, that is, gradually transitioning from the old parameters to the new parameters within one stimulation cycle (about 20ms), avoiding abrupt changes that may cause muscle discomfort. The updated stimulation pulse sequence continues to act on the C6 segment of the cervical spinal cord, causing the forearm flexor muscles to produce stronger contractions, propelling the hand further toward the desired state.

[0232] Step 860: Record stimulus execution logs. The system records key information such as the timestamp of this parameter update, old parameter values, new parameter values, current error, and fatigue index to a log file for subsequent effect evaluation and system optimization. This log data is of great value for identifying system problems, optimizing control algorithms, and adjusting individualized parameters.

[0233] After step 800 is completed, the new stimulus parameters have been safely implemented, and the patient's motor state begins to adjust towards the desired target. The system immediately proceeds to step 900 to determine whether to continue closed-loop iteration.

[0234] Step 900: Closed-loop iteration

[0235] Step 900 is the closed-loop control module, which determines whether to continue executing the loop from step 200 to step 800 to achieve continuous dynamic optimization.

[0236] Step 910: Update system status and timers. Record the current iteration number (this is the 1st iteration) and the cumulative running time (approximately 500ms). The update cycle for the closed-loop iteration is set to 50ms to 100ms. This embodiment uses a 50ms update cycle, meaning a complete "perception-decision-execution" loop is executed every 50ms. This frequency is sufficient to track the dynamic motion process while allowing ample processing time for the computing platform.

[0237] Step 920: Determine if the task is complete. Check if the desired motor state has been stably achieved. The criterion is that the normalized comprehensive error is less than the convergence threshold (e.g., 5%) and the state is maintained for more than 2 seconds. In this embodiment, the current error is 20.6%, which is much higher than the convergence threshold, so the task is not yet complete. The system will continuously adjust the stimulation parameters to gradually reduce the error in subsequent iterations. Usually, after 5-15 iterations (approximately 250-750ms), the error can converge to an acceptable range, and the patient's hand achieves the desired grip posture and strength, stably holding the water cup.

[0238] Step 930: Detect any change in patient intention. The patient's motor intention is continuously monitored through brain signal decoding in step 200. If the decoder detects a change in patient intention from "grasping" to "holding" or "releasing," the current task objective is updated.

[0239] More specifically, for the intention to "hold," the desired state remains unchanged, and the system enters a maintenance mode, making only minor adjustments to counteract external interference and fatigue. For the intention to "release," the desired grip strength decreases to near zero, and the desired joint angle returns to extension. The system adjusts the stimulation parameters accordingly, reducing the current intensity or switching to extensor activation to achieve the release action. In this embodiment, the patient's intention is to maintain a "grip," so no task switching is required.

[0240] Step 940: Check if the forced termination condition has been triggered. Forced termination conditions include: fatigue index. The following conditions may trigger a stop signal: a fatigue level exceeding 0.9 (extreme fatigue), a safety violation detected (such as excessive joint angle or abnormally high electromyography), a task duration exceeding the maximum (e.g., 30 minutes to prevent overtraining), or the patient issuing an emergency stop signal. In this embodiment, It is within the normal range, with no safety violations, a runtime of only 500ms, and no termination conditions were triggered.

[0241] Step 950: Decide whether to continue the loop. Based on the judgments from steps 920 to 940, in this embodiment, the task is not completed, the intention has not changed, and no termination condition has been triggered; therefore, it is decided to continue the loop. The system waits 50ms (until the next update cycle arrives) and then returns to step 200 to re-acquire brain signals. In subsequent iterations, because better stimulation parameters have just been implemented (…), It is expected that the patient's actual gripping force will increase from 11N to closer to 14N, and the joint angle will further approach -10°, with the error gradually decreasing.

[0242] Step 960: Long-term iterative operation. As the closed-loop system continues to run, it may experience the following typical scenarios:

[0243] Scenario 1: Normal convergence. After approximately 10 iterations (500ms), the error is reduced to within 5%, the patient's hand holds the water cup stably, and the system enters the maintenance phase. During the maintenance phase, closed-loop operation is maintained, but the parameter adjustments are very small, mainly to compensate for slow fatigue accumulation and external disturbances (such as the liquid sloshing inside the water cup changing the weight distribution).

[0244] Scenario 2: Fatigue occurs. After 10 minutes of continuous exercise, the fatigue index... The intensity gradually increased to 0.65, triggering a mild fatigue warning. The system then implemented a fatigue compensation strategy on top of the error adjustment, increasing the stimulation frequency from 46Hz to 55Hz, successfully counteracting the effects of fatigue and maintaining a stable grip strength.

[0245] Scenario 3: Intent Switching. After holding the water glass for 5 seconds, the patient decides to put it back on the table. The brain signal decoder detects the intention switching from "holding" to "releasing," and updates the expected grip strength to 0N. The system quickly adjusts parameters, reducing the current intensity to 5mA (close to the threshold), and gradually releases the hand. The entire release process is completed smoothly within 1-2 seconds, and the water glass is safely placed on the table.

[0246] Scenario 4: Task Complete. After the water cup is returned, the patient rests and no longer shows any new motor intentions. The system detects no clear motor intention for 5 consecutive seconds, automatically enters standby mode, stops stimulation but continues to monitor brain signals, ready to respond to the next motor command.

[0247] Step 900 achieves dynamic and continuous closed-loop control through a cyclical iteration mechanism, enabling the system to automatically adapt to changes in patient intent, physiological state, and external interference, which is significantly superior to traditional open-loop fixed-mode stimulation systems.

[0248] This embodiment fully demonstrates the implementation process of a closed-loop brain-spinal cord interface upper limb grasping function reconstruction method based on multimodal biofeedback. Through the coordinated operation of nine main steps, the system achieves a complete closed loop from patient motor intention perception, neural electrical stimulation execution, multi-dimensional motor feedback acquisition, intelligent error correction, fatigue prediction and compensation to safety assurance. Compared with existing open-loop or simple closed-loop systems, this application has the following significant technical advantages:

[0249] First, control precision is significantly improved. Through multimodal feedback fusion and hierarchical adaptive adjustment strategies, the gripping force error is reduced from ±40% in the open-loop system to within ±10%, and the joint angle error is reduced from ±15° to ±5°, enabling patients to complete fine daily living tasks.

[0250] Second, the duration of exercise is significantly extended. Through fatigue monitoring and prediction mechanisms, the system can proactively adjust its strategies before fatigue significantly affects exercise, extending the effective exercise time from 5-8 minutes in traditional systems to 15-20 minutes, greatly improving its practicality.

[0251] Third, safety is fully guaranteed. Through multiple safety checks, parameter constraints, and patient-controlled permissions, the system did not exhibit any excessive stimulation or abnormal movement throughout the entire operation, ensuring a comfortable experience for the patient.

[0252] Fourth, the degree of naturalness is improved. Patients only need to generate movement intentions like healthy people, and the system automatically translates the intentions into precise movement outputs. All the technical complexities in between are completely transparent to the patient, reducing cognitive burden and learning difficulty.

[0253] This embodiment demonstrates the feasibility and effectiveness of the method proposed in this application, providing an advanced technical solution for the reconstruction of motor function in patients with spinal cord injuries. Of course, this embodiment is merely a specific application scenario of this application, and the scope of protection of this application is not limited thereto. Any equivalent transformations or modifications based on the technical concept of this application should be included within the scope of protection of this application.

[0254] The second embodiment of this application relates to a closed-loop brain-spinal cord interface upper limb grasping function reconstruction system based on multimodal biofeedback, the structure of which is as follows: Figure 2 As shown, this closed-loop brain-spinal cord interface upper limb grasping function reconstruction system based on multimodal biofeedback includes:

[0255] The initialization and calibration module is used to establish an individualized brain signal decoding model and determine the safe range of stimulus parameters;

[0256] The brain signal acquisition and decoding module includes intracranial electrodes for acquiring motor cortex neural signals and decoding the desired motor state using the brain signal decoding model. The At least include the expected joint angle and expected grasp strength ;

[0257] A nerve electrical stimulator, used according to the above Generate initial stimulation parameters and apply electrical nerve stimulation to the target nerve or muscle;

[0258] The multimodal feedback acquisition module includes an inertial measurement unit, a force sensor or pressure sensor, and a surface electromyography sensor, for simultaneously acquiring at least three types of feedback signals, including: joint angle signals, gripping force signals, and electromyography signals.

[0259] The integrated state assessment and error calculation module is used to weight and fuse the joint angle signal, the gripping force signal, and the electromyographic signal to generate the actual motion state. and calculate the With the Error vector between ;

[0260] The muscle fatigue monitoring and prediction module is used to perform frequency domain analysis on the electromyographic signals to calculate the fatigue index. and according to the Determine fatigue compensation strategies;

[0261] The adaptive stimulus parameter adjustment module is used to adjust the parameters according to the... The stimulus parameters are updated using a graded adjustment strategy, and the fatigue compensation strategy is superimposed to adjust the stimulus parameters to generate new stimulus parameters. The graded adjustment strategy includes:

[0262] • When the error size is less than the first threshold At that time, only the current intensity is adjusted. ;

[0263] • When the error magnitude is greater than or equal to And less than the second threshold At the same time, adjust the current intensity. Stimulation frequency and pulse width ;

[0264] • When the error magnitude is greater than or equal to At that time, switch the electrode activation combination;

[0265] The safety check module is used to perform a safety check on the new stimulation parameters;

[0266] The nerve stimulator is also used to apply nerve electrical stimulation to the target nerve or muscle based on new stimulation parameters after being checked by the safety inspection module.

[0267] The closed-loop control module is used to control the brain signal acquisition and decoding module, the neurostimulator, the multimodal feedback acquisition module, the comprehensive state assessment and error calculation module, the muscle fatigue monitoring and prediction module, the adaptive stimulation parameter adjustment module, and the safety check module to repeatedly perform the above operations until the task is completed or the termination condition is triggered.

[0268] 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.

[0269] The above embodiments have the following technical effects:

[0270] The above embodiments effectively solve the technical problems of low control precision, poor adaptability and limited continuous movement ability in the prior art by constructing a closed-loop brain-spinal cord interface upper limb grasping function reconstruction method based on multimodal biofeedback, and achieve significant technical results.

[0271] To address the issues of existing open-loop control systems' inability to perceive actual motion execution effects and their susceptibility to muscle fatigue and external interference leading to low control accuracy, the above embodiments establish an individualized brain signal decoding model and define an initialization step for safe stimulus parameter ranges, laying an individualized foundation for subsequent closed-loop control. By acquiring motor cortical neural signals through intracranial electrodes and decoding the desired motion state, including the desired joint angle and desired gripping force, a clear control objective is provided to the system. More importantly, the above embodiments simultaneously acquire at least three types of feedback signals, including joint angle signals, gripping force signals, and electromyographic signals, and weightedly fuse these multimodal feedback signals to generate the actual motion state. This allows for a comprehensive characterization of complex grasping movements from three dimensions: posture, force, and muscle activation. Compared to single-sensor feedback, this multimodal fusion evaluation provides more complete and accurate motion state information, offering a reliable reference for subsequent error calculation and helping to reduce misjudgments and unnecessary parameter adjustments caused by incomplete information.

[0272] To address the shortcomings of existing systems, such as their inability to intelligently adjust stimulation parameters and poor adaptability, the above embodiment calculates the error vector between the desired and actual motor states and updates the stimulation parameters using a tiered adjustment strategy based on the magnitude of the error, achieving error-driven adaptive control. Specifically, when the error is small, only the current intensity is fine-tuned to avoid over-response; when the error is moderate, the current intensity, stimulation frequency, and pulse width are adjusted simultaneously to accelerate convergence; and when the error is large, the electrode activation combination is switched to fundamentally change the muscle recruitment pattern. This tiered adjustment strategy enables the system to take differentiated response measures according to the actual degree of deviation, ensuring control stability with small errors and rapid correction capability with large errors, thus improving closed-loop response speed and control accuracy. Through continuous repetition of the closed-loop iterative process of brain signal acquisition, stimulation execution, feedback acquisition, error calculation, and parameter adjustment, the system can track changes in the desired motor state in real time and dynamically adjust the stimulation output, contributing to accurate response to the patient's motor intentions.

[0273] To address the issues of rapid muscle fatigue caused by electrical stimulation and the inability of existing systems to predict and compensate for fatigue, leading to poor sustained exercise capacity, the above embodiment calculates a fatigue index through frequency domain analysis of electromyography (EMG) signals and determines fatigue compensation strategies based on the fatigue index, achieving proactive fatigue management. The fatigue index is calculated based on the median frequency, average power frequency, and deviation from the baseline value of the EMG signals, reflecting the actual fatigue state of the muscles. When the fatigue index falls within different threshold ranges, the system employs differentiated compensation strategies such as increasing the stimulation frequency, activating synergistic muscle groups, or triggering a rest mode, proactively intervening and adjusting before a significant decline in exercise quality. By combining error-based graded adjustment strategies with fatigue-based compensation strategies, the system can both correct execution errors in real time and prevent fatigue accumulation, helping to extend effective exercise time and improve training tolerance.

[0274] In terms of actual motion state assessment, the above embodiments integrate joint angle signals, gripping force signals, and electromyographic signals through a weighted fusion formula, where the weighting coefficients are dynamically adjusted according to the current task requirements and the sensor signal-to-noise ratio. This task-adaptive weighted fusion mechanism enables the system to optimize the contribution weights of each modality signal based on the different task's emphasis on angle, force, and muscle activation, thus improving the relevance and accuracy of comprehensive motion state assessment. For example, for fine gripping tasks, the force feedback weight is increased to enhance force control precision, while for large-range motion tasks, the angle feedback weight is increased to optimize trajectory tracking capabilities.

[0275] In terms of fatigue monitoring, the above embodiments employ a fatigue index calculation formula based on median frequency and average power frequency, comprehensively considering the decline in electromyographic spectral characteristics and the efficiency change in the ratio of electromyographic amplitude to current intensity, enabling multi-dimensional quantification of muscle fatigue levels. When the fatigue index falls within different threshold ranges, corresponding compensation strategies are adopted: for mild fatigue, the stimulation frequency is increased to utilize the muscle frequency response characteristics; for moderate fatigue, synergistic muscle groups are activated to distribute the load; and for severe fatigue, forced rest is implemented to protect the muscles, facilitating targeted interventions at different fatigue stages.

[0276] In terms of brain signal decoding, the above embodiments employ a hierarchical decoding architecture. The first layer identifies high-level motor intentions such as grasping, releasing, holding, and reaching. The second layer decodes continuous motor parameters such as desired grasping force and desired joint angle, and dynamically adjusts the mapping range and weight of the parameters according to the task type. This combination of hierarchical decoding and task-adaptive mapping ensures that the decoding results possess both the accuracy of intention recognition and the precision of parameter extraction, helping to improve the matching degree between the desired motor state and the patient's true intention.

[0277] Regarding the adjustment of stimulation parameters, the above embodiments clearly define the first threshold as a specific proportion of the normalized error and the second threshold as a larger proportion of the normalized error, and employ differentiated parameter adjustment strategies in different error ranges. When the error is small, the adjustment amount of the current intensity is proportional to the force error component; when the error is moderate, the adjustment amounts of the current intensity, stimulation frequency, and pulse width are proportional to the force error component, angle error component, and activation error component, respectively. This proportional correspondence between the parameter adjustment amount and the error components allows the adjustment of stimulation parameters to specifically correct deviations in corresponding dimensions, helping to improve the efficiency and specificity of parameter optimization.

[0278] Regarding safety, the above embodiments limit stimulation parameters to specific safety ranges. Clear upper and lower limits are set for current intensity, stimulation frequency, and pulse width. A safety check procedure verifies whether new stimulation parameters are within the safety range, and checks whether joint angles and grip strength exceed physiological safety limits. This multi-layered safety check mechanism helps prevent overstimulation and abnormal movements, ensuring patient safety.

[0279] Regarding system real-time performance, the above embodiments set the update cycle of the closed-loop iteration within a short time range, enabling the system to perform the perception-decision-execution cycle at a high frequency. This helps to respond promptly to changes in motion state and updates to patient intentions, thereby improving the system's real-time control performance.

[0280] Regarding data synchronization, the above embodiments ensure the time alignment of joint angle signals, gripping force signals, and electromyographic signals through hardware timestamps. The time alignment error is controlled within a very small range, which helps to ensure that multimodal signals correspond to the motion state at the same moment and improves the accuracy of weighted fusion results.

[0281] Regarding the smoothness of parameter adjustment, the above embodiments set a specific proportion in which the magnitude of a single parameter adjustment does not exceed the current parameter value, which helps to avoid muscle spasms or patient discomfort caused by sudden changes in stimulation parameters, and improves the stability and comfort of system operation.

[0282] In summary, the above embodiments, through the organic combination and synergistic cooperation of multimodal feedback fusion, error grading adjustment, and fatigue prediction compensation techniques, achieve precise reconstruction of upper limb grasping function in spinal cord injury patients within a closed-loop control framework. Compared to existing open-loop or simple closed-loop systems, these embodiments help improve motion control accuracy, enhance the system's adaptability to dynamic changes, prolong continuous motion time, and improve patient comfort and safety, providing an effective technical solution for the clinical rehabilitation treatment of upper limb grasping dysfunction caused by spinal cord injury.

[0283] Furthermore, optionally, in the closed-loop control of this application, the error magnitude is calculated using the normalized Euclidean distance method. Specifically, for the error vector... The error components in each dimension are first normalized: angle error component. Divide by the angular working range (e.g., ±10° corresponds to a range of 20°), force error component Divide by the working force range (e.g., 0-50N corresponds to a range of 50N) to activate the error component. Divide by the activated working range. Then calculate the square root of the weighted sum of squares of the normalized components as the error magnitude:

[0284]

[0285] Among them, the weighting coefficient , , The weights used are the same as those used in multimodal fusion in step S5.

[0286] Explanation of typical values ​​for the fatigue threshold

[0287] In one specific embodiment, fatigue threshold , and Typical values ​​for these thresholds are 0.6, 0.75, and 0.85. These thresholds are set based on the fact that when the fatigue index... When the frequency reaches 0.6, the median frequency of the electromyographic signal decreases by about 10%. At this point, the muscle still has a strong reserve capacity, and increasing the stimulation frequency can effectively delay fatigue. When the median frequency reaches 0.75, it decreases by approximately 15-20%, making it difficult for a single muscle group to maintain the target output, requiring the intervention of synergistic muscle groups to share the load; when When the median frequency reaches 0.85, the median frequency drops by more than 25%, and continued stimulation may lead to muscle damage, thus requiring mandatory rest. Those skilled in the art can adjust these thresholds within the range of 0.5-0.95 based on the specific muscle characteristics and task requirements of each patient.

[0288] Furthermore, optionally, in the adaptive stimulus parameter adjustment in step S7, the scaling factor... , , , The method for determining the parameter is as follows: In the individualized calibration stage of step S1, the parameter-output response curve for the patient is established by gradually changing a single stimulus parameter and recording the corresponding changes in motion output. For example, for the current intensity-force relationship, the current intensity is changed in steps of 2mA, the steady-state force change is recorded, and the force-current sensitivity coefficient (unit: N / mA) is obtained by fitting the coefficient. Its reciprocal is the parameter-output response curve. The initial value. In a typical embodiment, The value range is 0.5-2.0 mA / N. The value range is 1-5Hz / °. The value range is 5-20 μs / %. These coefficients can be adaptively updated online based on the actual control effect during system operation.

[0289] Furthermore, optionally, in the above embodiments, the desired motion state Mainly includes the desired joint angle and expected grasp strength Two explicit decoding parameters. For the desired muscle activation level. The determination of this application adopts the following method:

[0290] Method 1: Based on force-activation mapping. The system establishes the desired grip force during the calibration phase in step 100. Compared with expected muscle activation The mapping relationship curve between them. Specifically, by collecting gripping force and corresponding electromyographic RMS values ​​under different stimulation parameters, the F-EMG mapping function is fitted and obtained:

[0291]

[0292] in This is the mapping function obtained through individualized calibration. In subsequent closed-loop control, it is based on the decoded mapping function... Direct calculation .

[0293] Method 2: Implicit Activation Error Assessment. In some implementation scenarios, the system does not explicitly calculate... Instead, it will activate the error component. Defined as the deviation of actual muscle activation efficiency from expected efficiency:

[0294]

[0295] in and The expected electromyographic-current ratio under the current intensity target is determined by calibration data. This definition ensures that activation error reflects the deviation from electrical stimulation efficiency, rather than the difference in absolute activation level.

[0296] Those skilled in the art can choose one of the above methods according to specific application requirements, or use other equivalent methods for determining the expected muscle activation. All such variations should be understood to fall within the protection scope of this application.

[0297] It should be noted that those skilled in the art should understand that the implementation functions of each module shown in the above-described embodiments of the closed-loop brain-spinal cord interface upper limb grasping function reconstruction system based on multimodal biofeedback can be understood by referring to the relevant descriptions of the aforementioned closed-loop brain-spinal cord interface upper limb grasping function reconstruction method based on multimodal biofeedback. The functions of each module shown in the above-described embodiments of the closed-loop brain-spinal cord interface upper limb grasping function reconstruction system based on multimodal biofeedback can be implemented by a program (executable instructions) running on a processor, or by specific logic circuits. If the above-described closed-loop brain-spinal cord interface upper limb grasping function reconstruction system based on multimodal biofeedback is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage media include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks. Therefore, the embodiments of this application are not limited to any specific hardware and software combination.

[0298] Accordingly, this application also provides a computer storage medium storing computer-executable instructions, which, when executed by a processor, implement the various method implementations of this application.

[0299] Furthermore, this application also provides a closed-loop brain-spinal cord interface upper limb grasping function reconstruction system based on multimodal biofeedback, including a memory for storing computer-executable instructions and a processor; the processor is used to implement the steps in the above-described method embodiments when executing the computer-executable instructions in the memory. The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The aforementioned memory can be read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or solid-state drive, etc. The steps of the methods disclosed in the embodiments of this application can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules in the processor.

[0300] It should be noted that in this patent application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In this patent application, if it refers to performing an action according to an element, it means performing the action at least according to that element, including two cases: performing the action only according to that element, and performing the action according to that element and other elements. Expressions such as "multiple," "repeatedly," and "various" include two, two times, two kinds, and more than two, more than two times, and more than two kinds.

[0301] 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 multi-modal biofeedback-based closed-loop brain-spinal interface upper limb grasp function reconstruction system, characterized in that, Comprise: An initialization and calibration module for establishing an individualized brain signal decoding model and determining a safe stimulation parameter range; The brain signal acquisition and decoding module comprises an intracranial electrode for acquiring motor cortex neural signals and decoding desired motion states using the brain signal decoding model , the brain signal decoding model at least comprises a desired joint angle and a desired grip strength ; Neurostimulator for generating initial stimulation parameters and applying neurostimulation to a target nerve or muscle in dependence of the generated initial stimulation parameters; A multi-modal feedback acquisition module including an inertial measurement unit, a force sensor or a pressure sensor, and a surface electromyography sensor for synchronously acquiring at least three types of feedback signals, including joint angle signals, grip force signals, and electromyography signals; The comprehensive state evaluation and error calculation module is configured to weight and fuse the joint angle signal, the grip strength signal and the electromyography signal to generate an actual motion state , and calculate an error vector between the actual motion state and the target motion state . The target motion state is determined by the target state evaluation module . a muscle fatigue monitoring and prediction module for performing a frequency domain analysis of the electromyography signal to calculate a fatigue index and determining a fatigue compensation strategy based on the determining a fatigue compensation strategy based on the The calculation of the fatigue index is based on the frequency domain features of the electromyography signal, using the following formula: wherein, , , is a weighting factor; and are the median frequency and the mean power frequency at the current time instant, respectively; and are the baseline median frequency and the baseline mean power frequency, respectively; is the root mean square value of the current electromyographic signal; is the current current intensity; and are the baseline root mean square value and the baseline current intensity, respectively. The fatigue compensation strategy includes: when the stimulation frequency is increased by 5-10 Hz; when the synergist muscle group combined stimulation is activated; when the rest mode is triggered, the task is suspended or the task intensity is reduced by 50%, and after 20-30 seconds of re-evaluation, the task is resumed. An adaptive stimulation parameter adjustment module is configured to update the stimulation parameter according to the size of the fatigue compensation strategy and superimpose the fatigue compensation strategy on the stimulation parameter to generate a new stimulation parameter. The hierarchical adjustment strategy includes: when the error size is smaller than a first threshold only adjust the current intensity ; when the error magnitude is greater than or equal to a second threshold , simultaneously adjusting current intensity , stimulation frequency , and pulse width ; when the error size is greater than or equal to switching the electrode activation combination; A safety check module for performing a safety check on the new stimulation parameters; Wherein the neuroelectric stimulator is further configured to implement neuroelectric stimulation on the target nerve or muscle according to the new stimulation parameters that have passed the safety check; A closed-loop control module for controlling the brain signal acquisition and decoding module, the neuroelectric stimulator, the multi-modal feedback acquisition module, the comprehensive state assessment and error calculation module, the muscle fatigue monitoring and prediction module, the adaptive stimulation parameter adjustment module, and the safety check module to repeatedly perform the above operations until the task is completed or a termination condition is triggered.

2. The system of claim 1, wherein, the actual motion state is a multi-dimensional vector including actual joint angles , actual grip strength and actual muscle activation , each dimension being obtained by standardizing and weighting fusing corresponding sensor signals respectively: wherein, is a normalized signal collected by the joint angle sensor; is a normalized signal collected by the force sensor; is a normalized signal collected by the surface electromyography sensor; weight coefficient , , According to the current task requirement and the sensor signal-to-noise ratio dynamic adjustment, the value range is .

3. The system of claim 1, wherein, The brain signal decoding model adopts a hierarchical decoding architecture, comprising: A first layer using a classifier to identify discrete high-level movement intentions, including grasping, releasing, holding, and reaching out; Second layer: decode continuous motion parameters in the using regression models, which at least include desired grasp force and desired joint angles and dynamically adjust the mapping range and weights of the continuous motion parameters according to the current task type.

4. The system of claim 1, wherein, The intracranial electrode is a cortical electroencephalogram electrode or an intravascular stent electrode.

5. The system of claim 1, wherein, the first threshold value the second threshold value is set to 10% of the normalized error the second threshold value is set to 30% of the normalized error.

6. The system of claim 1, wherein, The adaptive stimulation parameter adjustment module, when executing the hierarchical adjustment strategy: When the error size is smaller than the adjustment amount of the current intensity , wherein is a proportional coefficient, is the force error component in the error ​ when the error size is greater than or equal to and less than , the adjustment amount of the current intensity , the adjustment amount of the stimulation frequency , the adjustment amount of the pulse width , wherein , , is a proportional coefficient, is an angle error component in the , and is an activation error component in the .

7. The system of claim 1, wherein, The stimulation parameters range from: current intensity between 5 mA and 50 mA, stimulation frequency between 20 Hz and 100 Hz, pulse width between 100 μs and 500 μs.

Images