An upper limb rehabilitation training system

Abstract

The application provides an upper limb rehabilitation training system and relates to the technical field of medical rehabilitation training.The system comprises: a healthy side high-specificity nerve signal acquisition module that acquires nerve electrical signals from a plurality of brachial plexus branches of a healthy side upper limb; an abstract motion intention real-time decoding module that inputs extracted space-time features into a pre-trained abstract intention mapping model and outputs abstract motion intention parameters; an adaptive closed-loop stimulation encoder of an embedded time sequence multi-mode nerve regulation module that generates electrical stimulation instructions for a plurality of brachial plexus branches based on abstract motion intention parameters and real-time compound action potentials; and a contralateral multi-modal nerve regulation and feedback module that performs electrical stimulation on a plurality of brachial plexus branches of a contralateral upper limb to induce limb movement according to the electrical stimulation instructions, while collecting real-time compound action potential signals of the contralateral upper limb and feeding back the signals.The application can effectively improve the control accuracy and safety of the affected limb, accurately monitor the degree of nerve activation in real time, and effectively deal with the problem of spasticity interference.

Description

Technical Field

[0001] This invention relates to the technical field of medical rehabilitation training, and in particular to an upper limb rehabilitation training system. Background Technology

[0002] Stroke, traumatic brain injury and other diseases are the main causes of permanent unilateral upper limb dysfunction caused by central nervous system disorders. Traditional rehabilitation methods such as functional electrical stimulation and robot-assisted training mainly act on the affected limb, and their efficacy has a bottleneck for patients with severe damage to the central efferent pathway.

[0003] Currently, the medical community has explored two fundamentally different technical approaches to reconstruct the conduction pathway between the brain and the paralyzed limb. The first is biological neural reconstruction, represented by the contralateral C7 nerve transfer. This technique involves irreversible anatomical damage. Although it can restore gross motor function on the affected side, it sacrifices function on the contralateral side, resulting in insufficient recovery of fine motor function on the affected side and the inability to monitor neural activation in real time. The second is machine-based neural bypass, represented by cortical brain-computer interfaces. This requires craniotomy to implant electrodes, is highly invasive, and lacks tactile feedback, leading to imprecise control. As a compromise, surface electromyography (SEM) non-invasive control is easily interfered with and cannot meet the needs of fine control. Peripheral neural interfaces are not suitable for patients with central efferent damage. Existing neuromodulation systems have not achieved true neural-level closed loops and cannot accurately monitor neural activation in real time. In addition, ECAP technology, which has been proven to accurately assess neural responses, has not yet been applied to trans-organ neural function reconstruction systems. Therefore, new solutions are urgently needed in clinical practice. In particular, nearly 40% of patients with central hemiplegia also have spastic paralysis (spastic paralysis), characterized by abnormally high muscle tone and hyperactive tendon reflexes. Traditional electrical stimulation rehabilitation training faces significant challenges in patients with spastic paralysis: direct application of restorative motor stimulation may induce or exacerbate spasticity, leading to poor rehabilitation outcomes or even injury. Clinically, physical therapists usually need to first perform manual stretching to reduce muscle tone before conducting electrical stimulation training, which is cumbersome and dependent on professional personnel. Summary of the Invention

[0004] To address the problems of low precision in limb control, poor safety, difficulty in real-time accurate monitoring of nerve activation levels, and spasticity interference in rehabilitation training for patients with spastic paralysis, this invention proposes an upper limb rehabilitation training system that can effectively improve the precision and safety of limb control, accurately monitor nerve activation levels in real-time, and effectively solve the clinical problem of spasticity interference in rehabilitation training for patients with spastic paralysis.

[0005] To achieve the above-mentioned technical effects, the technical solution of the present invention is as follows: An upper limb rehabilitation training system, the system comprising: The healthy side high-specificity nerve signal acquisition module is used to acquire nerve electrical signals from several brachial plexus branches of the healthy upper limb; The abstract motion intent real-time decoding module is used to receive the neural electrical signals, extract the spatiotemporal features of the neural electrical signals, input the spatiotemporal features into a pre-trained abstract intent mapping model, and output abstract motion intent parameters. An adaptive closed-loop stimulation encoder is used to receive the abstract motor intention parameters and real-time compound action potentials, and based on the abstract motor intention parameters and real-time compound action potentials, to feed back and optimize the generation of electrical stimulation commands for several brachial plexus branches, and has a built-in sequential multi-mode neuromodulation module. The affected side multimodal neural modulation and feedback module is used to electrically stimulate several brachial plexus branches of the affected upper limb according to the electrical stimulation command to induce limb movement, and at the same time collect the real-time compound action potential signal of the affected side and feed it back to the adaptive closed-loop stimulation encoder.

[0006] Preferably, the healthy side high-specificity nerve signal acquisition module consists of a high-specificity electrode implanted on several brachial plexus branches of the patient's healthy upper limb and an integrated miniaturized preamplifier and processing unit, wherein the contacts of the high-specificity electrode are connected to the input terminal of the miniaturized preamplifier and processing unit.

[0007] Preferably, the abstract intent mapping model is a deep learning model, and the training of the abstract intent mapping model is as follows: Offline collection of neural signal data and synchronous motion capture data when users perform various standard actions; using the neural signal data and motion capture data to train the deep learning model to obtain a trained abstract intent mapping model.

[0008] Preferably, the step of generating electrical stimulation commands for several brachial plexus branches based on the abstract motor intention parameters and real-time compound action potentials includes: S31. Take the abstract motion intention parameter as the control target and the real-time composite action potential as the state variable; S32. Input the control target and the state quantity into the built-in electrical stimulation command prediction model of the adaptive closed-loop stimulation encoder, and use the electrical stimulation command prediction model to simulate the effect of multiple candidate multi-channel stimulation parameter sequences. S33. The effects are filtered by an optimizer to select electrical stimulation commands for several brachial plexus branches that are closest to the control target and meet the stimulation safety limits.

[0009] Preferably, the affected side multimodal neuromodulation and feedback module includes a stimulation submodule and a feedback submodule; the stimulation submodule is used to electrically stimulate several brachial plexus branches of the affected upper limb according to the electrical stimulation command to induce limb movement; the feedback submodule is used to collect real-time compound action potential signals of the affected side for feedback after the stimulation pulse is emitted.

[0010] Preferably, the stimulation submodule uses multi-contact stimulation electrodes implanted on several branches of the brachial plexus on the affected side of the patient; the feedback submodule uses independent recording contact electrodes integrated on the same array as the multi-contact stimulation electrodes.

[0011] Preferably, it also includes a cross-body wireless connection and energy management module, which includes an in-body relay / coordinator and an external processing and control unit; The in vivo relay / coordinator is used to receive the neural electrical signals, transmit the neural electrical signals to the external processing and control unit, receive electrical stimulation commands from several brachial plexus branches output by the external processing and control unit, and forward them to the affected side multimodal neuromodulation and feedback module. The external processing and control unit integrates an algorithm for running the real-time decoding module of the abstract motion intent and the adaptive closed-loop stimulation encoder, which is used to send electrical stimulation commands to several brachial plexus branches to the in vivo relay / coordinator.

[0012] Preferably, the in vivo relay / coordinator communicates bidirectionally with the external processing and control unit via a medical band wireless link and performs wireless power transmission.

[0013] Preferably, the healthy side high-specificity neural signal acquisition module is connected to the affected side multimodal neural modulation and feedback module through the in vivo relay / coordinator.

[0014] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: This invention proposes an upper limb rehabilitation training system. Firstly, by acquiring highly specific neural signals from the healthy side and decoding abstract motor intentions, it achieves high-fidelity, interference-resistant conversion of motor intentions from the healthy side, avoiding invasive procedures such as craniotomy and anatomical damage, thus effectively improving safety. Secondly, using an adaptive closed-loop stimulation encoder, based on the abstract motor intention parameters and real-time compound action potentials, it feeds back and optimizes the generation of electrical stimulation commands for several brachial plexus branches, forming a precise closed-loop control at the neural level. This allows for real-time monitoring and optimization of neural activation while stimulating the affected side's nerves, significantly improving the precision of control over the movement of the affected limb. Furthermore, through the temporal multi-modal neural modulation module built into the adaptive closed-loop stimulation encoder, it achieves an intelligent rehabilitation paradigm of first reducing tension and then promoting movement, effectively solving the clinical problem of spasticity interference in the rehabilitation training of patients with spastic paralysis. Finally, according to the electrical stimulation commands, it electrically stimulates several brachial plexus branches of the affected upper limb to induce limb movement, while simultaneously acquiring real-time compound action potential signals from the affected side and feeding them back to the adaptive closed-loop stimulation encoder, achieving precise monitoring of the degree of neural activation. Attached Figure Description

[0015] Figure 1 This is a structural block diagram of an upper limb rehabilitation training system proposed in an embodiment of the present invention; Figure 2 This refers to the electrical stimulation command proposed in this embodiment of the invention for generating several brachial plexus branches. Flowchart; Figure 3 This diagram illustrates the connection relationship between the implanted part and the external processing and control part in an upper limb rehabilitation training system proposed in this embodiment of the invention. Detailed Implementation

[0016] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. It is understandable to those skilled in the art that some well-known details may be omitted from the accompanying drawings; To facilitate understanding of this embodiment, the prior art information of this embodiment is first introduced as follows: Biological neural regeneration technology, exemplified by the "contralateral C7 nerve transfer," involves severing the C7 nerve root that innervates the upper limb on the contralateral side, relocating its proximal end, and anastomosing it with the brachial plexus on the affected side. After a lengthy process of nerve regeneration (usually exceeding one year) and arduous brain function remodeling training, some patients can achieve simultaneous control of both upper limbs by the contralateral cerebral hemisphere, thereby restoring gross motor function and basic grasping function of the shoulder, elbow, and wrist joints on the affected side. This procedure represents a significant innovation in microsurgery, bringing hope to patients with central hemiplegia. However, its limitations are also substantial: firstly, this procedure involves irreversible anatomical destruction, permanently sacrificing some sensory and motor function of the contralateral upper limb, although this can often be compensated for; secondly, the fineness of functional recovery on the affected side is limited, making independent and dexterous finger movements difficult to achieve; and finally, the long rehabilitation period and uncertain recovery outcomes pose a significant challenge to patients.

[0017] The development of machine neural bypass technology, also known as brain-computer interface (BCI) technology, involves decoding neural signals from the brain's motor cortex to directly control external devices or stimulate affected limbs to achieve functional replacement or rehabilitation. This approach bypasses the point of injury and, theoretically, can achieve direct and rapid control of the affected limb. However, existing cortical BCIs face significant challenges: achieving high-precision decoding typically requires craniotomy for electrode implantation, which is highly invasive, carries certain surgical risks, and faces signal attenuation and biocompatibility issues after long-term implantation. More importantly, current technologies can hardly provide natural tactile and proprioceptive feedback, resulting in non-closed-loop and imprecise control.

[0018] As a compromise, non-invasive control schemes based on surface electromyography (EMG) signals are widely used. However, these signals are susceptible to interference and fatigue, and cannot decode complex intentions, failing to meet the needs of precise, multi-degree-of-freedom intuitive control. While emerging peripheral neural interfaces have shown excellent performance in intuitively controlled prostheses, research on them has largely focused on cases where the ipsilateral central efferent pathway is intact (such as bionic hands suitable for amputation stumps), and they are not applicable to patients with central efferent pathway damage such as stroke or traumatic brain injury.

[0019] Furthermore, while existing functional electrical stimulation or neuromodulation systems attempt to introduce closed-loop control, their feedback signals largely rely on external sensors (such as pressure pads and accelerometers) or surface electromyography signals, failing to achieve a truly "neural-level closed loop." These secondary signals are significantly affected by factors such as muscle fatigue, electrode displacement, and skin condition, and exhibit delays of hundreds of milliseconds, making real-time, precise monitoring of neural activation states impossible. In recent years, evoked compound action potentials (ECAPs) have been proven to be the most direct physiological indicator for assessing neural responses to electrical stimulation, and their application in fields such as spinal cord stimulation has demonstrated revolutionary closed-loop control effects. However, currently, there is no technology to apply them to transbody neural function reconstruction systems.

[0020] Furthermore, nearly 40% of patients with central hemiplegia also have spastic paralysis (spastic paralysis), characterized by abnormally high muscle tone and hyperactive tendon reflexes. Traditional electrical stimulation rehabilitation training faces significant challenges in patients with spastic paralysis: direct motor restorative stimulation may induce or exacerbate spasticity, leading to poor rehabilitation outcomes or even injury. Clinically, physical therapists usually need to first perform manual stretching to reduce muscle tone before conducting electrical stimulation training, which is cumbersome and dependent on professional personnel.

[0021] Therefore, a significant technological gap exists between "functionally sacrificed, slowly regenerating bio-grafting" and "high-risk, feedback-deficient cortical decoding." Clinically, a novel solution is urgently needed: it should be less invasive than craniotomy, achieve more precise and faster functional recovery than C7 cervical vertebral transfer, possess the potential for neural-level closed-loop feedback based on ECAP, and effectively address spasticity in patients with spastic paralysis, thus providing a superior "neural bypass" for patients with central upper limb paralysis.

[0022] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] Example 1 See Figure 1 This embodiment proposes an upper limb rehabilitation training system. The core objective of the system is to construct a novel neural bypass that does not rely on damaged central pathways and enables intuitive fine motor control and neural feedback. The system is used for patients with central unilateral upper limb paralysis and includes: The healthy side high-specificity nerve signal acquisition module is used to minimally invasively acquire high-fidelity, high-spatial-resolution nerve electrical signals from five brachial plexus branches of the healthy upper limb; the five brachial plexus branches are the musculocutaneous nerve, axillary nerve, radial nerve, ulnar nerve, and median nerve.

[0024] Preoperative assessment and individualized electrode selection methods: Before system implantation, a standardized assessment of the patient is required to determine the brachial plexus branch combination for which electrodes should be implanted. The assessment employs a three-tiered, progressive approach: The first level is standardized motor function assessment. Key movements of the affected upper limb, including shoulder abduction, elbow flexion, elbow extension, wrist extension, finger extension, wrist flexion, fist clenching, palm opposition, finger separation, and finger synthesis, are assessed one by one using a six-point scoring system from 0 to 4 (0 points for no contraction, 4 points for normal). Those with a score ≤2 proceed to the next level of assessment.

[0025] The second layer involves needle electromyography (EMG). For muscles initially screened for abnormalities, motor unit potential recruitment and denervation potentials (fibrillation potentials, positive spikes) are detected. The judgment criteria are as follows: If there is no recruitment and no denervation potential, it is considered central paralysis with intact peripheral pathways, making it an indication for implantation; if there is no recruitment but denervation potentials are present, it suggests concurrent peripheral nerve damage, and implantation is not permitted; if there is partial recruitment and no denervation potentials, selective implantation is possible.

[0026] The third step is to confirm the integrity of the nerve. For nerve branches that meet the indications for implantation, the amplitude of the complex muscle action potential (≥50% of the healthy side) is measured by nerve conduction testing to confirm that the peripheral nerve conduction function is intact.

[0027] Based on the results of the three-tiered assessment, the combination of nerve branches requiring electrode implantation was determined one by one. Electrode implantation was performed only on nerve branches that simultaneously met the three criteria of "loss of motor function, loss of central drive, and intact peripheral pathways," thereby avoiding unnecessary surgical trauma caused by implanting all five nerve branches.

[0028] The healthy side high-specificity nerve signal acquisition module consists of a high-specificity electrode implanted on several branches of the brachial plexus in the patient's healthy upper limb and an integrated miniaturized preamplifier and processing unit. The contacts of the high-specificity electrode are connected to the input end of the miniaturized preamplifier and processing unit, and the output end is connected to the cross-body wireless connection and power management module.

[0029] The highly specific electrode uses flexible biocompatible materials, such as polydimethylsiloxane (PDMS) and polyimide, as a substrate, on which multi-contact microelectrodes, such as platinum-iridium alloy, are integrated. The electrode is designed as a high-density cuff electrode that can surround the above five brachial plexus branches. Each brachial plexus branch corresponds to an independent cuff electrode array to ensure that independent signals from different functional branches can be recorded.

[0030] The miniaturized preamplifier and processing unit is integrated with the highly specific electrode in the same flexible package for preliminary amplification, filtering (e.g., bandpass filtering 300 Hz - 5 kHz to extract action potentials) and digitization of microvolt-level neural signals to suppress noise and reduce the amount of data before wireless transmission.

[0031] The abstract motion intent real-time decoding module is used to receive the neural electrical signals, extract the spatiotemporal features of the neural electrical signals, input the spatiotemporal features into a pre-trained abstract intent mapping model, and output abstract motion intent parameters. The extraction of spatiotemporal features of the neural electrical signals includes further processing of the input multi-channel (5 branches × 8-32 contacts / branch) neural signals. The extracted spatiotemporal features include, but are not limited to, the instantaneous firing frequency of each channel, the synchronous / asynchronous modes between multiple channels, and the power spectral density of specific frequency bands. Specifically, the system analyzes the co-activation patterns among the five brachial plexus branches. For example, during strong gripping, a large number of motor units on the ulnar and median nerves are synchronously activated, exhibiting high-amplitude, high-frequency burst signals; while during fine pinching, the median nerve displays specific patterns of low-amplitude but highly temporally ordered signals.

[0032] The abstract intent mapping model is a deep learning model, such as a model based on a long short-term memory network or a Transformer architecture. The training of the abstract intent mapping model is as follows: Offline collection of neural signal data and synchronous motion capture data (as labels for "abstract intentions") of users performing various standard actions (such as grasping, holding, pinching, and independent extension of fingers with different strengths and speeds) is used to train the deep learning model, thereby learning the mapping relationship from the features of 5 neural signals to abstract motion parameters, and obtaining a trained abstract intention mapping model.

[0033] During online execution, real-time feature vectors extracted from spatiotemporal features are input into the trained abstract intention mapping model. The output of the abstract intention mapping model is the decoded abstract motion intention parameters. These abstract motion intention parameters are low-dimensional vectors, such as: action type encoding (e.g., [1,0,0] represents "strong grip", [0,1,0] represents "fine pinch", [0,0,1] represents "finger extension"), target joint angle vectors (e.g., metacarpophalangeal joint: 45°, proximal interphalangeal joint: 30°, wrist joint: 10°), expected end force vectors (e.g., grip force: 2N, pinch force: 1.5N), and movement urgency, etc.

[0034] The purpose of the abstract motion intent real-time decoding module is to convert the low-level execution signals related to specific situations, which are collected by the healthy side high-specificity neural signal acquisition module from five branches of the brachial plexus, into high-level abstract intentions that describe the motion target and are independent of the situation, i.e., abstract motion intent parameters.

[0035] An adaptive closed-loop stimulation encoder is used to receive the abstract motor intention parameters and real-time compound action potentials, and based on the abstract motor intention parameters and real-time compound action potentials, to feed back and optimize the generation of electrical stimulation commands for several brachial plexus branches. Based on the abstract motion intention parameters and real-time compound action potentials, the predictive control algorithm of the electrical stimulation command prediction model generates electrical stimulation commands for several brachial plexus branches in each control cycle through feedback optimization. The electrical stimulation command prediction model is a simplified "neuromuscular-skeleton" dynamic model built into the encoder, which includes: a neural activation model describing the relationship between stimulation parameters (current, pulse width, frequency, contact combination) and ECAP amplitude; a muscle contraction model describing the relationship between ECAP amplitude and muscle contraction force; and a skeletal dynamic model describing the relationship between muscle force and joint movement. (See also...) Figure 2 The feedback optimization process includes: S31. The abstract motion intention parameter of the current cycle is used as the control target, and the real-time composite action potential of the current cycle is used as the state quantity; the real-time composite action potential is received in real time through the feedback input interface, which also receives other sensory feedback. S32. Input the control target and the state quantity into the built-in electrical stimulation command prediction model of the adaptive closed-loop stimulation encoder, and use the electrical stimulation command prediction model to simulate the effects produced by multiple sets of candidate multi-channel stimulation parameter sequences; in S32, based on the electrical stimulation command prediction model, the algorithm simulates the possible effects of multiple sets of candidate 5-channel stimulation parameter sequences (including the stimulation pulse width, frequency, amplitude of each nerve branch and the activation combination of each electrode contact) within a limited time window in the future.

[0036] S33. The effects are filtered by an optimizer to select electrical stimulation commands for several brachial plexus branches that are closest to the control target and meet the stimulation safety limits. Finally, the set of these electrical stimulation commands is used as the optimal stimulation sequence. The first control command of this optimal stimulation sequence is output to the multimodal neuromodulation and feedback module on the affected side, and this process is repeated in the next cycle. The purpose of the adaptive closed-loop stimulation encoder is to dynamically generate and optimize the electrical stimulation commands driving the five nerve branches on the affected side based on the decoded abstract motor intention parameters and the real-time compound action potential feedback from the affected side, so as to achieve a precise mapping from abstract intention to five-channel stimulation.

[0037] The affected side multimodal neural modulation and feedback module is used to electrically stimulate several brachial plexus branches of the affected upper limb according to the electrical stimulation command to induce limb movement, and at the same time collect the real-time compound action potential signal of the affected side and feed it back to the adaptive closed-loop stimulation encoder.

[0038] The purpose of the affected-side multimodal neuromodulation and feedback module is to perform electrical stimulation to induce movement and collect real-time compound action potential signals from the affected side as core feedback to achieve precise closed-loop control. It includes a stimulation submodule and a feedback submodule. The stimulation submodule is used to perform electrical stimulation on several brachial plexus branches of the affected upper limb according to the electrical stimulation command to induce limb movement. The feedback submodule is used to collect real-time compound action potential signals from the affected side for feedback after the stimulation pulse is delivered.

[0039] It also includes a cross-body wireless connectivity and power management module, which aims to achieve stable, low-latency, and low-power data communication and power supply between the healthy and affected implants, and between the implant and the external unit.

[0040] The cross-body wireless connection and energy management module includes an in-body relay / coordinator and an external processing and control unit; The in vivo relay / coordinator is used to receive the neural electrical signals, transmit the neural electrical signals to the external processing and control unit, receive electrical stimulation commands from several brachial plexus branches output by the external processing and control unit, and forward them to the affected side multimodal neuromodulation and feedback module. The in vivo relay / coordinator described in this embodiment is a miniaturized implantable unit containing a microprocessor, a dual-band wireless transceiver chip, and a rechargeable battery. It is responsible for collecting nerve electrical signals from several brachial plexus branches of the healthy upper limb and forwarding them to the external unit. At the same time, it receives 5-channel electrical stimulation commands from the external processing and control unit and forwards them to the multimodal neuromodulation and feedback module on the affected side.

[0041] The external processing and control unit integrates an algorithm for running the real-time decoding module of the abstract motion intent and the adaptive closed-loop stimulation encoder, which is used to send electrical stimulation commands to several brachial plexus branches to the in vivo relay / coordinator.

[0042] The external processing and control unit described in this embodiment is a wearable device containing a high-performance processing chip, a main control wireless module, and a large-capacity battery. It charges the internal device wirelessly and processes the core algorithm. Its communication protocol adopts a highly reliable, low-latency customized medical frequency band wireless protocol to ensure that the total latency of the control loop is less than 150 milliseconds to meet the real-time interaction requirements.

[0043] The affected side multimodal neuromodulation and feedback module also includes a signal conditioning and ECAP processing circuit and an ECAP feature extraction unit. The signal conditioning and ECAP processing circuit amplifies and bandpass filters the weak real-time composite action potential signal, and then relies on a fast recovery circuit to quickly restore the linear state of acquisition after a strong stimulation pulse, eliminating stimulation artifacts and obtaining a pure real-time composite action potential signal. The recorded microvolt-level real-time composite action potential (ECAP) signals are amplified and filtered (typically within a bandpass range of 100Hz-5kHz to preserve the main frequency components of the ECAP). The recording channel is equipped with a fast recovery circuit to ensure linear operation within <100μs after a strong stimulus pulse, preventing stimulus artifacts from drowning out the weak ECAP signal. The ECAP feature extraction unit calculates key features of the real-time composite action potential signal after the stimulus pulse in real time, including: N1-P1 peak-to-peak amplitude, latency, area under the curve, and waveform morphology. These key ECAP features are encoded and uploaded in real time to the external processing and control unit via a cross-body wireless connection and energy management module, serving as the core input for closed-loop control.

[0044] The stimulation submodule is used to electrically stimulate several brachial plexus branches of the affected upper limb according to the electrical stimulation command to induce limb movement. The stimulation submodule uses multi-contact stimulation electrodes implanted on several brachial plexus branches corresponding to the affected side of the patient. The multi-contact stimulation electrodes include an array of multi-contact stimulation electrodes implanted on five nerve branches (musculocutaneous nerve, axillary nerve, radial nerve, ulnar nerve, and median nerve) corresponding to the affected side. Its design is similar to the high-specificity electrode of the high-specificity nerve signal acquisition module on the healthy side, but the contacts focus more on safe current injection and are connected to an independent programmable constant current / constant voltage stimulation generator.

[0045] The feedback submodule is used to acquire the real-time compound action potential signal of the affected side after the stimulation pulse is delivered, and transmit the real-time compound action potential signal to the signal conditioning and processing circuit; the feedback submodule uses an independent recording contact electrode integrated on the same array as the multi-contact stimulation electrode; the independent recording contact electrode includes: a. ECAP recording electrodes: Independent recording contacts integrated on the same array as the stimulating electrodes, specifically designed to record compound action potentials evoked by electrical stimulation. The electrode array for each nerve branch contains 4-8 ECAP recording contacts, distributed at the proximal and distal ends of the stimulating contacts, spaced 3-5 mm apart, to capture ECAP signals propagating bidirectionally along the nerve trunk at the optimal distance.

[0046] b. Sensory nerve recording electrodes: used to record spontaneously occurring sensory afferent signals.

[0047] c. Biomechanical sensors: Miniaturized implantable sensors used to directly measure muscle contraction force, joint angles, etc.

[0048] This system's adaptive closed-loop stimulation encoder also incorporates a temporal multimodal neuromodulation module. This module provides temporal multimodal neuromodulation functionality to address the clinical problem of spastic paralysis (spastic paralysis) in nearly 40% of central hemiplegic patients. The system maintains a three-state machine: Normal Mode → Tonic Mode → Recovery Mode → Normal Mode. The temporal multimodal neuromodulation module includes: The spasticity detection submodule analyzes the resting-state signals of the ECAP (Electronic Electrode Capture) contacts on the affected side to monitor abnormal nerve discharge patterns in real time. Detection indicators include: the amplitude of resting-state ECAP baseline fluctuations, spontaneous nerve discharge frequency, and characteristic frequency band energy related to muscle tone. When the detected abnormal discharge frequency exceeds a preset threshold (e.g., >10Hz) or the ECAP baseline fluctuation is >20%, a spastic state is determined.

[0049] The triggering condition for the tension-reducing stimulation mode (high-frequency blocking mode) is the detection of spasticity or manual activation by the user / doctor before the start of rehabilitation training. This mode applies high-frequency, sub-motor threshold electrical stimulation to the nerve branches (median nerve, ulnar nerve) that innervate spastic muscle groups (such as forearm flexors). The specific stimulation parameters are: frequency 500-1500Hz (preferably 800-1200Hz), intensity below the motor threshold (usually 0.5-2mA, not causing visible muscle contraction), pulse width 30-100μs, and stimulation duration 30-120 seconds or until the spasticity is relieved. Its mechanism of action is that high-frequency stimulation can produce a "nerve block" or "synaptic inhibition" effect, rapidly reducing the overexcitability of α motor neurons. The closed-loop termination condition is that when the ECAP baseline fluctuation returns to the normal range (<10%) and the spontaneous discharge frequency is <5Hz, it automatically switches to the next mode.

[0050] The switching condition for the exercise recovery stimulation mode (low-frequency functional mode) is that the de-tension phase is completed and the muscle tone has returned to normal levels. After switching, low-frequency, adjustable-intensity electrical stimulation based on intention decoding is used. The specific stimulation parameters are: frequency 20-50Hz (preferably 20-30Hz), intensity above the exercise threshold (usually 2-10mA, which can induce clear muscle contraction), and pulse width 100-300μs. The control logic is to adjust the stimulation parameters in real time based on ECAP feedback according to the model predictive control algorithm, thereby driving the affected side to complete the expected rehabilitation training movements.

[0051] The state transition logic of the three-state sequential state machine is normal mode → tension reduction mode → recovery mode → normal mode. When rehabilitation training begins, the system defaults to normal mode. When a spasm is detected, it automatically enters tension reduction mode. After the tension reduction phase is completed, it automatically switches to recovery mode to execute the training task. After the training task is completed, it returns to normal mode and enters low-power standby mode. If a spasm is detected again during the training process, the recovery mode is immediately interrupted and the tension reduction mode is re-entered.

[0052] The temporal multimodal neuromodulation module uses full-process ECAP neural signal feedback to achieve closed-loop verification and adaptive adjustment, and matches different feedback indicators for different functional stages: in the tension reduction stage, the ECAP baseline fluctuation level and the frequency of abnormal neural discharges are used as the basis for closed-loop regulation, while in the motor recovery stage, the amplitude of the ECAP signal induced by electrical stimulation is used to dynamically track the target reference value, thereby completing precise closed-loop control throughout the entire process.

[0053] Based on the above modules, the method for functional reconstruction according to the present invention includes the following steps: S1: Personalized calibration and model training. After the system is implanted in the user, the user is guided to perform a series of standard actions. The electrical signals of the five brachial plexus branches on the healthy side and the real-time compound action potential signal response and real motion data on the affected side are recorded simultaneously. These data are used to train and abstract the decoding model of the real-time decoding module for motion intention and the electrical stimulation command prediction model of the adaptive closed-loop stimulation encoder. S2: Signal Acquisition. When the patient moves the unaffected upper limb according to specific motor commands, the motor pathway on the unaffected side is activated, generating characteristic neural electrical signals in the five brachial plexus branches on the unaffected side. These signals are acquired and transmitted by the unaffected side high-specificity neural signal acquisition module.

[0054] S3: Abstract Intent Decoding. The abstract motion intent real-time decoding module receives signals in real time, extracts the neural electrical signals of 5 nerves, and outputs abstract motion intent parameters through a pre-trained abstract intent mapping model.

[0055] S4: ECAP-based Adaptive Stimulus Generation and Delivery. The adaptive stimulation encoder receives abstract motion intention parameters and converts them into target ECAP amplitudes. Simultaneously, it reads the actual ECAP feedback from the affected side in real time. Using the controller, the optimal 5-channel stimulation parameter combination to compensate for the gap is calculated by comparing the target ECAP with the actual ECAP using the predictive control algorithm of the electrical stimulation command prediction model, and an electrical stimulation command is generated.

[0056] S5: Precise stimulation and movement execution on the affected side. Electrical stimulation commands are wirelessly transmitted back to the in-body relay / coordinator and then sent to the corresponding contact electrodes of the five nerve branches in the multimodal neuromodulation and feedback module on the affected side. The multi-contact stimulation electrodes release precisely controlled electrical pulses to activate the target nerve fibers. After the stimulation pulses are released, the system immediately switches to recording mode, capturing the induced ECAP signal through independent recording contact electrodes as the basis for the next adjustment.

[0057] S6: ECAP Feedback Acquisition and Closed-Loop Adjustment. During action execution, the ECAP generated by each stimulus is captured, processed, and uploaded to the external unit in real time, forming a real-time composite action potential signal, which serves as the basis for the adaptive encoder to adjust the stimulus parameters in the next control cycle. This "stimulation-ECAP acquisition-adjustment" cycle operates continuously within milliseconds, forming a precise bioelectronic closed loop with ECAP as the "nerve pulse".

[0058] S7: Sequential Multimodal Control (Specific Process for Spastic Paralysis Patients). When the system detects spasticity on the affected side (determined by resting-state ECAP baseline fluctuations or abnormal discharge frequencies), the following sub-processes are automatically executed: S7.1: Tonic de-stress phase: Apply high-frequency blocking stimulation (800-1200Hz, submotor threshold) to the nerve branches innervating the spastic muscle groups for 30-120 seconds until the ECAP baseline returns to normal. S7.2: Exercise recovery phase: Automatically switch to low-frequency functional stimulation mode (20-30Hz), and execute the rehabilitation training process of S2-S6. S7.3: Real-time monitoring: Continuously monitor the spasticity state during training. If it recurs, re-enter S7.1.

[0059] The upper limb rehabilitation training system proposed in this embodiment constructs a complete, intelligent, and biomimetic closed loop of neurofunctional rehabilitation training, from the coordinated movement of the five nerves on the healthy side to the precise execution of the five nerves on the affected side, and finally to the ECAP nerve response feedback on the affected side. It also integrates a temporal multimodal neuromodulation function specifically for patients with spastic paralysis, achieving a fundamental breakthrough over existing technologies. The upper limb rehabilitation training system proposed in this embodiment differs from existing technologies in that it constructs a novel neurofunctional rehabilitation training pathway of "cross-peripheral nerve signals - abstract intention decoding - adaptive closed-loop stimulation." Its key points can be broken down into the following four interrelated and progressively layered elements, which work together to achieve the purpose of this invention and obtain superior results, as follows: The first key point is the selection of peripheral nerve signals from the healthy side as the core control source. The difference lies in the following: For cortical brain-computer interface technology, this system avoids the high-risk path of invading the cranial cavity and directly collecting signals from the cortex, shifting the signal acquisition point from the "source of intention generation" (cerebral cortex) to the "terminus of intention execution" (peripheral nerves). For the surface electromyography (EMG) scheme, this system collects the electrical signals of the nerves themselves, rather than the secondary electrical signals (EMG) generated by muscle contraction. Therefore, the signals are more direct, purer, and have higher information density, and are not significantly affected by muscle fatigue and skin condition. For traditional peripheral nerve interfaces, existing peripheral nerve interfaces (such as those used in prostheses) mainly collect signals from the ipsilateral stump nerves. However, this system innovatively proposes to collect signals from healthy, functionally intact peripheral nerves to control the affected limb, realizing a conceptual leap from "ipsilateral control" to "cross-body control." This key point is that when a patient with unilateral upper limb paralysis freely performs specific movements of the unaffected upper limb, the motor intention descends through the intact corticospinal tract and ultimately generates corresponding, highly specific motor command electrical signals on the five brachial plexus branches on the unaffected side (musculocutaneous nerve, axillary nerve, radial nerve, ulnar nerve, and median nerve). These signals encode precise spatiotemporal patterns of movement. Acquiring these signals is equivalent to obtaining a high-quality "execution blueprint" of how to move at the final stage of the neural pathway. This is more stable and easier to decode specific movements than cortical signals, and more direct and interference-resistant than electromyographic signals, laying a high-quality signal foundation for achieving fine control. Furthermore, in practical application, patients do not need to consciously imagine; they can achieve rehabilitation training for the corresponding affected upper limb simply through natural limb movements of the unaffected upper limb. This simplifies rehabilitation training, significantly reduces the cognitive load on patients (especially elderly patients), and provides a new and efficient "neural bypass" solution.

[0060] The second key point is to achieve reverse decoding from peripheral execution signals to abstract motor intentions. The difference lies in the fact that most existing direct stimulation and brain-computer interfaces deal with what stimulus pattern should be output at the moment or what predefined action the user wants to perform. This system faces a unique challenge: the peripheral nerve signals on the healthy side encode specific joint angles and muscle strength as "concrete execution instructions," containing specific contextual information such as current limb posture and load. If directly mirrored to the affected side, the action will fail due to the different physical state of the affected side. Therefore, this system introduces a crucial abstract intention decoder. Its task is not to simply transmit signals, but to extract the user's fundamental motor goal, independent of the specific physical state, from the specific context-dependent signals. This key point draws on the hierarchical theory of motor control. Higher centers (such as the brain) issue abstract descriptions of motor goals (such as "move the hand to a certain position" or "grasp with a specific force"), while the peripheral nerves execute specific implementation schemes that take into account the integration of proprioceptive feedback during the descending process. The decoder in this system, trained on a large amount of motion-signal pairing data using deep learning and other algorithms, has learned the inverse mapping relationship of inferring abstract motion targets from specific execution signals. For example, regardless of whether the healthy hand picks up a pen from a table or a ball from the ground, as long as the user's abstract intention is a three-finger pinch, the decoder can recognize and output this same intention code from different muscle activation patterns. This is equivalent to transforming the "how to do" on the healthy side into a general "what to do," making it possible to subsequently regenerate new, specific instructions suitable for the current state on the affected side. This is the core conversion hub for achieving cross-body adaptive control.

[0061] The third key point is the construction of an adaptive closed-loop stimulation based on the affected side's ECAP signal. The difference lies in the fact that existing functional electrical stimulation or some brain-computer interface stimulation schemes are mostly open-loop, meaning they perform stimulation in a fixed pattern according to preset or decoded instructions, unable to adjust in real time based on the performance of the action. A few closed-loop systems also rely heavily on external sensors (such as cameras, pressure pads) or surface electromyography signals. These secondary signals have inherent defects such as large delays, susceptibility to interference, and inability to directly reflect the neural activation state. This system, for the first time, proposes using the affected side's ECAP signal as the core feedback source, constructing a neural-level adaptive closed-loop control system. This is not merely a simple replacement of the feedback signal, but a fundamental change in the control paradigm—from monitoring the action result to monitoring the neural activation process. In this key point, ECAP is the sum potential generated by the synchronous discharge of nerve fiber groups after electrical stimulation. Its amplitude directly reflects the number of activated nerve fibers, its latency reflects the nerve conduction velocity, and its waveform morphology contains information about the activation pattern of the nerve bundle. The core advantage of using ECAP as the feedback signal is: 1. Directness and Real-Time Response: ECAP is the first response of a nerve to a stimulus, which can be recorded within 0.5-2 ms after stimulation, making it faster and more direct than muscle contraction or biomechanical signals. This allows the system to evaluate and adjust the effect of the stimulus before the action is completed.

[0062] 2. Physiological Fidelity: ECAP originates directly from the nerve fibers themselves and is unaffected by peripheral factors such as muscle fatigue, changes in electrode-tissue interface impedance, and skin condition, making it the gold standard for evaluating stimulation effectiveness. This signal purity is the physical basis for achieving precise closed-loop control.

[0063] 3. Stable neural activation window: Drawing on the advanced closed-loop spinal cord stimulation principle, this system can monitor the fluctuations in stimulation effect caused by factors such as changes in body position and electrode micro-movements in real time by setting the target ECAP amplitude, and automatically adjust the stimulation intensity to keep the neural activation level constant within the treatment window.

[0064] 4. Closed-loop verification of intent-response: In this scheme, ECAP is not only a feedback of stimulus effect, but also a "grounded verification" of abstract motor intent on the affected side. The working principle of the adaptive stimulus encoder is as follows: Target reception: Obtain the current motion target (such as the expected neural activation level corresponding to "target grip strength 70%)" from the "abstract intention decoder"; Perceive the current status: Obtain the current actual neural activation level from ECAP feedback (e.g., the actual ECAP amplitude is only 50% of the target value). Calculate and adjust errors: Compare the difference between the target ECAP and the actual ECAP, and optimize the stimulation parameters (current intensity, pulse width, frequency, and contact combination) for the next cycle in real time through the built-in neuromuscular dynamics model. Achieving stability: Through continuous fine-tuning, the ECAP amplitude is made infinitely close to the target value, thereby ensuring that the neural activation level and motor intention are precisely matched.

[0065] 5. Robust control against interference: This mechanism enables the system to automatically compensate for various types of interference. Changes in the electrode-nerve interface: such as the formation of a fibrous capsule leading to increased impedance, the system can maintain the ECAP target by increasing the current.

[0066] Muscle fatigue: When the muscle contraction force generated by the same stimulus decreases, ECAP can provide an early warning and the system can adjust its strategy in a timely manner.

[0067] Changes in body position: such as changes in joint angle that cause changes in the relative position of nerves and electrodes, ECAP feedback can capture and compensate for these changes in a timely manner.

[0068] Multi-level closed-loop integration: ECAP, as "low-level feedback" (neural activation level), together with "intermediate-level feedback" (sensory level) from sensory nerve input and "high-level feedback" (motor execution level) from biomechanical sensors, constitutes a multi-level, multi-timescale closed-loop control system. ECAP is responsible for millisecond-level rapid adjustment to ensure the accuracy of neural activation; sensory feedback is responsible for tens of millisecond-level adjustment to ensure the adaptability of movement; and biomechanical feedback is responsible for hundreds of millisecond-level adjustment to ensure the quality of task completion.

[0069] The fourth key point is the temporal multimodal neuromodulation strategy—an intelligent rehabilitation paradigm of "first reducing tension, then promoting movement." The difference lies in the fact that existing functional electrical stimulation systems are mostly single-mode stimulations, unable to dynamically adjust the stimulation strategy according to the patient's muscle state (such as spasticity). For patients with spastic paralysis, directly applying motor rehabilitation stimulation may induce or exacerbate spasticity, leading to poor rehabilitation results or even injury. Clinically, physical therapists usually need to first perform manual stretching to reduce muscle tone before electrical stimulation training, a cumbersome procedure that relies on professional personnel. This solution proposes and implements temporal multimodal neuromodulation based on ECAP feedback for the first time, seamlessly connecting high-frequency blocking stimulation with low-frequency functional stimulation to form an automated and intelligent rehabilitation closed loop of "detection-tension reduction-motor promotion."

[0070] Theoretical Feasibility: Following central nervous system injury, inhibitory input to α-motor neurons decreases, leading to an abnormally hyperactive response to the stretch reflex. High-frequency stimulation (500-1500Hz) can rapidly reduce muscle tone through mechanisms such as enhancing GABAergic inhibitory synaptic transmission, inducing nerve fiber conduction blockade, and activating Renshaw cell recurrent inhibitory pathways. This protocol utilizes ECAP recording points to monitor spontaneous nerve activity in the resting state, achieving objective, real-time, and quantitative assessment of spasticity. The "first reduce tension, then promote movement" sequence has significant clinical implications: functional electrical stimulation under low muscle tone results in less resistance to joint movement and smoother motion; it avoids the risk of stimulation-induced spasticity, improving training safety; patients can learn correct movement patterns in a "cleaner" neuromuscular state, which is beneficial for positive remodeling of neural plasticity; if spasticity recurs during training, the system can intervene immediately, forming a dynamic protective mechanism of "training-monitoring-intervention".

[0071] In summary, the four key points of this system constitute a complete technical chain. The first key point provides a high-quality, cross-body control signal source; the second key point solves the core problem of command generalization in cross-body control, elevating specific signals to abstract intentions; the third key point utilizes the ECAP neural biomarker on the affected side to construct an intelligent closed loop that enables the abstract intentions to be accurately and adaptively implemented on the affected side; and the fourth key point effectively solves the clinical problem of spasticity interference in the rehabilitation training of patients with spastic paralysis. These four key points are interconnected and indispensable, enabling this invention to achieve more precise and faster functional recovery than C7 nerve transfer surgery with relatively low invasiveness (compared to craniotomy), and possesses control naturalness and environmental adaptability that surpasses existing brain-computer interface and functional electrical stimulation technologies, thereby achieving the invention's purpose and obtaining superior technical effects. This invention also integrates a time-sequential multimodal neuromodulation function, achieving an intelligent rehabilitation paradigm of "first reducing tension, then promoting movement" for patients with spastic paralysis through the seamless connection of high-frequency blocking stimulation and low-frequency functional stimulation.

[0072] Example 2 This embodiment proposes an upper limb rehabilitation training system, see [link / reference]. Figure 3 The system is divided into two parts: an implantable part and an external processing and control part. The implantable part is connected to the power supply module via percutaneous communication to the external processing and control part. The implanted component comprises a healthy-side high-specificity neural signal acquisition module, an affected-side multimodal neural modulation and feedback module, and a trans-body wireless connection and power management module. The healthy-side high-specificity neural signal acquisition module consists of highly specific electrodes implanted on five target nerve branches (musculocutaneous nerve, axillary nerve, radial nerve, ulnar nerve, and median nerve) in the patient's healthy upper limb, along with an integrated miniaturized pre-amplification and processing unit. The affected-side multimodal neural modulation and feedback module consists of multi-contact stimulation / recording integrated electrodes implanted on the corresponding five nerve branches on the patient's affected side. The in-vivo relay / coordinator module within the trans-body wireless connection and power management module is a miniaturized electronic module encapsulated in a biocompatible shell, containing a microprocessor, a wireless transceiver unit, and a rechargeable battery.

[0073] The external processing control section consists of an external processing and control unit, which integrates the algorithms for running the abstract motion intent real-time decoding module and the adaptive closed-loop stimulation encoder. The in-vivo relay / coordinator communicates bidirectionally with the external processing and control unit via a medical band wireless link and transmits wireless power.

[0074] The healthy side high-specificity neural signal acquisition module is connected to the affected side multimodal neural modulation and feedback module through the in vivo relay / coordinator; the connection medium is a micro-lead. The following is a detailed description of the implementation method of each core module of the upper limb rehabilitation training system proposed in this embodiment: For the high-specificity electrode arrays of the healthy-side high-specificity neural signal acquisition module and the affected-side multimodal neural modulation and feedback module, a flexible polymer (such as polyimide or PDMS) is used as the substrate, on which platinum-iridium alloy microelectrode contacts are fabricated using microfabrication technology. For the healthy-side high-specificity neural signal acquisition module with five brachial plexus branches, each brachial plexus branch is equipped with an independent slotted C-shaped cuff electrode with 8-32 independent recording contacts distributed on the inner wall to achieve differentiated acquisition of signals from different functional bundles within each nerve trunk.

[0075] For the ipsilateral multimodal neuromodulation and feedback module involving the five branches of the brachial plexus on the affected side, the electrodes for each branch employ a flat, multi-contact parafascicular design to simultaneously perform highly selective electrical stimulation and neural signal recording. Specifically, the electrodes include two types of functional contacts: Multi-contact stimulation electrodes: 8-16, distributed in the middle of the electrode, for safe current injection; Independent recording contact electrodes: 4-8, distributed at the proximal and distal ends of the stimulation contact, with a spacing of 3-5 mm, to capture ECAP signals propagating bidirectionally along the nerve trunk at the optimal distance; The positions of the independent recording contact electrodes are precisely designed: the proximal independent recording contact electrodes capture the ECAP propagating retrogradely to the spinal cord, and the distal independent recording contact electrodes capture the effect potential propagating anterogradely to the muscle; both can be used for closed-loop control. Implantation: Through microsurgery, a meticulous dissection is performed at the target nerve (such as the radial, ulnar, or median nerve), and the electrode is carefully implanted and fixed to the epineurium. The electrode lead is led through a subcutaneous tunnel to the implantation site of the relay / coordinator in the body, usually located below the clavicle on the affected side.

[0076] For the real-time decoding module of abstract motion intent, its algorithm runs on an external processing and control unit. The training process of the personalized abstract intent mapping model is as follows: First, after the system is implanted and healed, the patient undergoes several days of calibration training under the guidance of a doctor. Second, the user is guided to attempt a series of standard movements of the unaffected upper limb (such as grasping, holding, pinching, and extending the hand), while the system simultaneously records raw neural signals from five nerve branch electrodes in the high-specificity neural signal acquisition module on the unaffected side. Then, using this raw neural signal pairing data, a deep neural network is trained. For example, a one-dimensional convolutional neural network is used to extract spatiotemporal features, followed by fully connected layers for classification and regression. The network learns to map the co-activation patterns of the five nerve branches to abstract motor intentions. This deep neural network learns to map neural signal patterns to two outputs as follows: Output by category: Action type, such as: rest, clench fist, thumb to finger, etc.; Regression output: continuous abstract motion parameters, such as a 0-1 scalar representing "grip strength" or a target value representing "hand opening angle"; When used online, the collected real-time neural signals are preprocessed in the same way and then input into the pre-trained network model as an abstract intent mapping model, which can output the currently decoded abstract motion intent parameters every 50-100 milliseconds.

[0077] For the adaptive closed-loop stimulation encoder, its algorithm runs on the external processing and control unit. The adaptive closed-loop stimulation encoder is used for feedback signal acquisition, and it receives feedback from three channels in real time: ECAP feedback: Signals obtained from the ECAP recording contacts of electrodes on the five nerve branches on the affected side are processed as follows: Stimulus artifact suppression: Fast recovery circuit and adaptive template subtraction are used to eliminate the residual effects of strong stimulus pulses; ECAP detection and feature extraction: Within a preset time window (0.5-3ms after stimulation), the ECAP waveform is detected, and features such as the peak-to-peak amplitude of N1-P1, latency, and area under the curve are calculated in real time. Quality assessment: The signal-to-noise ratio of each ECAP record is assessed, and outliers contaminated by noise are removed; Sensory feedback: Natural sensory input signals obtained from the sensory recording contacts of the electrodes on the affected side; Biomechanical feedback: Data such as grip strength and joint angles obtained from implanted sensors; The core of the control algorithm for the adaptive closed-loop stimulus encoder adopts a model predictive control framework based on the ECAP objective, as follows: a. Electrical Stimulation Command Prediction Model: This model incorporates a simplified neuromuscular-skeleton dynamic model for the affected side, used to predict the ECAP response and subsequent motor effects produced by different stimulation patterns. The model includes: a neural activation model describing the relationship between stimulation parameters (current, pulse width, frequency, contact combination) and ECAP amplitude; a muscle contraction model describing the relationship between ECAP amplitude and muscle contraction force; and a skeletal dynamics model describing the relationship between muscle force and joint movement. b. Rolling optimization: In each control cycle (e.g., 10ms): i. Obtain the current target intent (e.g., target grip force = 5N) from the abstract motion intent real-time decoding module and convert it into the corresponding target ECAP amplitude (through a pre-calibrated "intent-ECAP mapping table"); ii. Obtain the current actual ECAP amplitude and other state quantities from the affected side's multimodal neuromodulation and feedback module; iii. Calculate the ECAP error: ΔECAP = Target ECAP - Actual ECAP; iv. The electrical stimulation command prediction model simulates and predicts the effects of multiple candidate stimulation sequences (adjusting stimulation current, frequency, and contact combination) within a short time period (e.g., 50ms), with a focus on predicting their impact on ECAP. v. The optimization algorithm selects the stimulation sequence that makes the predicted ECAP closest to the target ECAP, with the least energy consumption and the highest stimulation comfort. c. Output and Execution: The first-step stimulation parameters of the selected sequence are sent wirelessly to the affected side's multimodal neuromodulation and feedback module for execution. This process is repeated in the next cycle to achieve dynamic closed-loop adjustment based on ECAP.

[0078] Example 3 This embodiment further illustrates the workflow of the upper limb rehabilitation training system proposed in the above embodiments. The workflow is as follows: S1: Personalized Calibration and Model Training. Preoperatively, a three-layer assessment is conducted to determine the specific nerve branch combinations requiring electrode implantation (the following description uses a full 5-branch implantation as an example). Through microsurgery, electrodes for acquiring the electrical signals of the five brachial plexus branches on the healthy side are implanted into the musculocutaneous nerve, axillary nerve, radial nerve, ulnar nerve, and median nerve on the patient's healthy side; five integrated stimulation / recording electrodes on the affected side are implanted into the corresponding five brachial plexus branches on the affected side; and an in vivo relay / coordinator is implanted in the subclavian region on the affected side.

[0079] During the postoperative recovery period, patients underwent systematic calibration training under the guidance of physicians: performing a series of standard movements while simultaneously recording the electrical signals of the five brachial plexus branches on the healthy side and the ECAP response on the affected side. These data were used to train a decoding model, establishing a mapping relationship between "signal characteristics on the healthy side → abstract motor intention," and calibrating an electrical stimulation command prediction model that predicts "stimulation parameters → ECAP response."

[0080] S2: Signal Acquisition. When the patient performs an intended movement on the unaffected side, the unaffected upper limb produces natural movement under the control of the intact neural pathway. At this time, characteristic motor nerve electrical signals are generated on the five branches of the brachial plexus on the unaffected side (musculocutaneous nerve, axillary nerve, radial nerve, ulnar nerve, and median nerve), which are captured by their respective electrode arrays. The acquired signals are pre-amplified, filtered, and converted from analog to digital, and then wirelessly transmitted in real time by an in-body repeater to an external processing and control unit.

[0081] S3: Abstract Intent Decoding. The external processing and control unit extracts spatiotemporal features (including the firing frequency of each nerve branch and the co-activation pattern between nerve branches) from the received five neural electrical signals, and inputs the feature vectors into a pre-trained abstract intent mapping model. The abstract intent mapping model completes inference within milliseconds and outputs the abstract motion intent parameters for the current cycle, including: action type (e.g., clenching a fist, pinching, extending fingers), target joint angle (e.g., finger flexion angle), target force, etc. The decoded abstract motion intent parameters are then transmitted to the adaptive closed-loop stimulation encoder.

[0082] S4: ECAP-based adaptive stimulus generation and delivery. The adaptive stimulus encoder receives abstract intent parameters and converts them into target ECAP amplitudes corresponding to the five neural branches through a pre-calibrated "intent-ECAP mapping table" (e.g., median nerve target ECAP = 45 μV, ulnar nerve target ECAP = 30 μV).

[0083] Simultaneously, the adaptive stimulation encoder reads the actual ECAP feedback (ECAP amplitude after the most recent stimulation) from the five nerve branches on the affected side in real time. The ECAP error of each nerve branch is calculated as: ΔECAP = Target ECAP - Actual ECAP.

[0084] Based on the built-in "neuro-muscle-skeleton" dynamic model of the adaptive closed-loop stimulation encoder, the adaptive closed-loop stimulation encoder employs a model predictive control algorithm within a 10ms control cycle: Simulate the ECAP response that may be generated by multiple sets of candidate 5-channel stimulation parameter sequences (current, pulse width, frequency, and contact combination of each nerve branch); Select the optimal stimulus sequence that makes the predicted ECAP closest to the target ECAP; The first stimulation command of the sequence is sent wirelessly to the corresponding electrodes of the five nerve branches on the affected side.

[0085] S5: Precise stimulation and movement execution on the affected side. Stimulation electrodes on the five nerve branches on the affected side release precisely controlled electrical pulses according to the received instructions, activating the target nerve fibers, triggering coordinated muscle contraction, and enabling the affected limb to perform the expected movement.

[0086] After the stimulation pulse is delivered, the system eliminates stimulation artifacts within <100μs and immediately switches to recording mode. ECAP recording contacts located near the stimulation contacts of each neural branch capture the induced ECAP signal and extract key features such as ECAP amplitude and latency in real time.

[0087] The extracted ECAP features are uploaded to the in vitro processing and control unit in real time via an in vivo relay / coordinator.

[0088] S6: ECAP Feedback Acquisition and Closed-Loop Adjustment. After receiving the ECAP feedback, the external processing and control unit uses it as the input for the next control cycle, repeating the closed-loop adjustment cycle of S4→S5→S4, so that the actual ECAP amplitude of each nerve branch continuously tracks the target value, achieving millisecond-level dynamic adjustment.

[0089] When the system detects that the target action has been completed (such as successfully grasping a water cup and maintaining stability), or when the user stops the healthy side activity for more than a set threshold, the stimulation intensity is gradually reduced until it stops, and the system enters a low-power standby mode.

[0090] S7: Sequential Multimodal Control (Specific Process for Spastic Paralysis Patients). When the system detects spasticity on the affected side (determined by resting-state ECAP baseline fluctuations or abnormal discharge frequencies), the following sub-processes are automatically executed: S7.1: Tension reduction phase: Apply high-frequency blocking stimulation (800-1200Hz, submotor threshold) to the nerve branches innervating the spastic muscle groups for 30-120 seconds until the ECAP baseline returns to normal; S7.2: Exercise recovery phase: Automatically switch to low-frequency functional stimulation mode (20-30Hz) and execute the rehabilitation training process of S2-S6; S7.3: Real-time monitoring: Continuously monitor the spasticity state during training. If it recurs, re-enter S7.1.

[0091] In this embodiment, by acquiring and decoding motor nerve signals from the brachial plexus branches of the healthy upper limb and extracting abstract motor intentions, mirror stimulation is applied to the affected side. Adaptive closed-loop adjustment is then performed using the evoked compound action potential (ECAP) induced by electrical stimulation on the affected side as the core feedback signal, thereby establishing a "digital neural bypass" for patients with central unilateral upper limb paralysis. The core innovation of this technical solution lies in directly utilizing peripheral nerve signals as a high-fidelity signal source through the mirror anatomical correspondence of human peripheral nerves, achieving reverse extraction from execution signals to abstract intentions, and constructing a biomimetic closed-loop control based on ECAP neural-level feedback. In particular, this invention also proposes a temporal multimodal neural modulation strategy to achieve an intelligent rehabilitation paradigm of "first reducing tension, then promoting movement," effectively solving the clinical problem of spasticity interference in rehabilitation training for patients with spastic paralysis. Compared to existing surgical rehabilitation methods, this system demonstrates significant advantages in achieving precision control of the affected limb, safety, and acceptability, providing a novel solution for the reconstruction of upper limb intuitive and multi-degree-of-freedom motor functions.

[0092] Example 4 This embodiment illustrates the application scenario of an upper limb rehabilitation training system proposed in the above embodiments.

[0093] Scenario 1: Rehabilitation training for conventional hemiplegic patients A patient with complete right-hand paralysis due to left basal ganglia hemorrhage used this system: The patient had cuff electrodes implanted in all five branches of the brachial plexus (musculocutaneous nerve, axillary nerve, radial nerve, ulnar nerve, and median nerve) on both sides, and an in vivo relay / coordinator module was implanted in the right subclavian region.

[0094] After calibration training, the patient picked up a glass of water with their left hand, and signals from five nerve branches on the left side were collected by the system. The system decoded the signal as "performing a grasping motion with moderate target force" and converted it into a target ECAP amplitude of 50μV.

[0095] The system generates initial 5-channel stimulation pulses and immediately captures the ECAPs induced in each neural branch. The controller compares the target ECAP with the actual ECAP and automatically adjusts the stimulation parameters for the next cycle.

[0096] During the gripping process, the cup's potential slippage is detected by the sensory nerves of the right hand, which is then integrated with ECAP feedback to further fine-tune the stimulation, increase grip strength, and stabilize the cup.

[0097] The core of the entire process is the millisecond-level closed-loop regulation driven by ECAP feedback, which ensures that the activation level of the five nerve branches remains constant near the target value regardless of changes in the patient's position or muscle fatigue, thus achieving smooth, stable, and reliable autonomous control.

[0098] Scenario 2: Sequential Multimodal Rehabilitation Training for Patients with Spinal Paralysis A patient with right-sided hemiplegia due to left-sided middle cerebral artery occlusion, accompanied by significant right upper limb flexor spasticity (modified Ashworth score 3). The patient used this system for rehabilitation training: Phase 1: Pre-training spasticity detection After the patient put on the system, the external relay automatically detected the resting ECAP signal on the affected side. The system detected that the baseline fluctuation amplitude of ECAP of the median nerve and ulnar nerve reached 35% (threshold 20%), and the spontaneous discharge frequency was 15Hz (threshold 10Hz), which was judged as a moderate spastic state.

[0099] Phase 2: Automatic tension reduction stimulation The system automatically entered the tension-reducing mode, applying high-frequency blocking stimulation to the median and ulnar nerves on the affected side: frequency 1000Hz, pulse width 60μs, intensity 1.2mA (below the motor threshold). After 45 seconds of stimulation, the ECAP baseline fluctuation decreased to 8%, and the spontaneous discharge frequency decreased to 4Hz, at which point the system determined that the spasticity had been relieved.

[0100] Phase 3: Exercise Recovery Training The system automatically switched to recovery mode. The patient attempted a "three-finger pinching" motion with their unaffected left hand. Electrodes from five peripheral nerves on the unaffected side acquired 80 channels of signals, and an external relay decoded the abstract intention of "pinching with moderate target force." Based on ECAP feedback, the system applied low-frequency stimulation (25Hz, 200μs pulse width, dynamically adjustable intensity 3-6mA) to the corresponding nerve branches on the affected side (partial bundles of the median and radial nerves), driving the affected right hand to complete the pinching motion. Due to reduced muscle tone, the smoothness of the movement significantly improved, and no spasticity was induced.

[0101] Phase 4: Real-time monitoring during training During the 20-minute training session, the system continuously monitored the ECAP baseline on the affected side. At the 12-minute mark, the ECAP baseline fluctuation was detected to rise again to 18% (close to the threshold). The system then automatically inserted a 20-second tonic stimulation before continuing the training to ensure that subsequent training was completed in a low spasticity state.

[0102] Effect Comparison: Compared with traditional open-loop electrical stimulation, this system enables safe training without spasticity in patients with spastic paralysis, improves the completion rate of movements, and significantly enhances the subjective comfort of patients.

[0103] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. An upper limb rehabilitation training system, characterized in that, The system includes: The healthy side high-specificity nerve signal acquisition module is used to acquire nerve electrical signals from several brachial plexus branches of the healthy upper limb; The abstract motion intent real-time decoding module is used to receive the neural electrical signals, extract the spatiotemporal features of the neural electrical signals, input the spatiotemporal features into a pre-trained abstract intent mapping model, and output abstract motion intent parameters. An adaptive closed-loop stimulation encoder is used to receive the abstract motor intention parameters and real-time compound action potentials, and based on the abstract motor intention parameters and real-time compound action potentials, to feed back and optimize the generation of electrical stimulation commands for several brachial plexus branches, and has a built-in sequential multi-mode neuromodulation module. The affected side multimodal neural modulation and feedback module is used to electrically stimulate several brachial plexus branches of the affected upper limb according to the electrical stimulation command to induce limb movement, and at the same time collect the real-time compound action potential signal of the affected side and feed it back to the adaptive closed-loop stimulation encoder.

2. The upper limb rehabilitation training system according to claim 1, characterized in that, The healthy side high-specificity nerve signal acquisition module consists of high-specificity electrodes implanted on several brachial plexus branches of the patient's healthy upper limb and an integrated miniaturized preamplifier and processing unit. The contacts of the high-specificity electrodes are connected to the input terminal of the miniaturized preamplifier and processing unit.

3. The upper limb rehabilitation training system according to claim 1, characterized in that, The abstract intent mapping model is a deep learning model, and the training of the abstract intent mapping model is as follows: Offline collection of neural signal data and synchronous motion capture data when users perform various standard actions; using the neural signal data and motion capture data to train the deep learning model to obtain a trained abstract intent mapping model.

4. The upper limb rehabilitation training system according to claim 1, characterized in that, The process of generating electrical stimulation commands for several brachial plexus branches based on the abstract motor intention parameters and real-time compound action potentials includes: S31. Take the abstract motion intention parameter as the control target and the real-time composite action potential as the state variable; S32. Input the control target and the state quantity into the built-in electrical stimulation command prediction model of the adaptive closed-loop stimulation encoder, and use the electrical stimulation command prediction model to simulate the effect of multiple candidate multi-channel stimulation parameter sequences. S33. The effects are filtered by an optimizer to select electrical stimulation commands for several brachial plexus branches that are closest to the control target and meet the stimulation safety limits.

5. The upper limb rehabilitation training system according to claim 1, characterized in that, The affected-side multimodal neuromodulation and feedback module includes a stimulation submodule and a feedback submodule. The stimulation submodule is used to electrically stimulate several brachial plexus branches of the affected upper limb according to the electrical stimulation command to induce limb movement. The feedback submodule is used to collect real-time compound action potential signals of the affected side for feedback after the stimulation pulse is emitted.

6. The upper limb rehabilitation training system according to claim 5, characterized in that, The stimulation submodule uses multi-contact stimulation electrodes implanted on several branches of the brachial plexus on the affected side of the patient; the feedback submodule uses independent recording contact electrodes integrated on the same array as the multi-contact stimulation electrodes.

7. The upper limb rehabilitation training system according to claim 1, characterized in that, It also includes a cross-body wireless connectivity and energy management module, which includes an in-body relay / coordinator and an external processing and control unit; The in vivo relay / coordinator is used to receive the neural electrical signals, transmit the neural electrical signals to the external processing and control unit, receive electrical stimulation commands from several brachial plexus branches output by the external processing and control unit, and forward them to the affected side multimodal neuromodulation and feedback module. The external processing and control unit integrates an algorithm for running the real-time decoding module of the abstract motion intent and the adaptive closed-loop stimulation encoder, which is used to send electrical stimulation commands to several brachial plexus branches to the in vivo relay / coordinator.

8. The upper limb rehabilitation training system according to claim 7, characterized in that, The in-vivo relay / coordinator communicates bidirectionally with the external processing and control unit via a medical band wireless link and transmits wireless power.

9. The upper limb rehabilitation training system according to claim 7, characterized in that, The healthy side high-specificity neural signal acquisition module is connected to the affected side multimodal neural modulation and feedback module through the in vivo relay / coordinator.

10. The upper limb rehabilitation training system according to claim 1, characterized in that, The temporal multimodal neuromodulation module includes: The spasm detection submodule is used to analyze the resting state signal of the contact point by analyzing the compound action potential signal of the affected side, and to monitor the abnormal discharge pattern of the nerve in real time. When the abnormal discharge frequency exceeds the preset threshold or the baseline fluctuation of the compound action potential signal exceeds the threshold, it is determined to be a spasm state. The tension-reducing stimulation submodule is used to apply high-frequency, submotor threshold electrical stimulation to the nerve branches that innervate spastic muscle groups; The motor recovery stimulation submodule is used to switch to a low-frequency functional stimulation mode after tension reduction is completed, and adjusts the stimulation parameters in real time based on the feedback of compound action potential signals. A timing state machine is used to automatically control the switching between normal mode, tension reduction mode, and recovery mode.

Images