Personalized training guidance system for stroke neurorehabilitation

Abstract

The application relates to the field of intelligent medical treatment and rehabilitation engineering, in particular to a stroke neural rehabilitation personalized training guidance system, which comprises the following steps: a core architecture construction step: establishing a communication connection of a central processing platform and a data sensing interface, a personalized strategy engine and a feedback execution interface; a state analysis and strategy generation step: determining a neural drive stage based on a multi-modal rehabilitation data set, generating a neural drive enhancement strategy for a low neural drive stage, and generating a motor learning optimization strategy for a high neural drive stage; and a closed-loop feedback execution step: converting the strategy into multi-sensory feedback control instructions to drive external equipment to execute personalized induction on patients. The application effectively solves the training dead zone problem of severe paralyzed patients and significantly improves the upper limb motor function score and the electromyographic response amplitude.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical and rehabilitation engineering, specifically a personalized training guidance system for neurological rehabilitation of stroke. Background Technology

[0002] In the medical informatics application scenario of stroke neurorehabilitation, the rehabilitation guidance system, as a typical human-computer interactive medical information platform, focuses on collecting multi-dimensional health data from patients and transforming it into executable rehabilitation training strategies using algorithmic models. Existing rehabilitation information processing systems typically employ signal discrimination logic based on fixed thresholds, combined with a central control unit for closed-loop control of assistive devices.

[0003] Existing rehabilitation information technologies have significant limitations in the data perception and decision generation stages. Firstly, regarding data processing algorithms for patient intent recognition, current solutions generally employ triggering mechanisms based on linear amplitude comparison. When dealing with severely paralyzed patients, the bioelectrical signals generated by their neural drives are extremely weak and often drowned out by background noise. Traditional filtering and discrimination algorithms, lacking nonlinear enhancement mechanisms, easily discard these weak intent data points with rehabilitation value as invalid noise, leading to perception dead zones in the system and preventing the triggering of subsequent training responses. This severely impacts the medical system's ability to cover critically ill patients.

[0004] Existing training strategies rely on a relatively simplistic logic, often passively guiding patients based on the principle of minimizing errors. This information-based processing model fails to adequately consider the need for motion error perception during neural remodeling. In particular, when patients enter a high-neural drive phase, if the system continuously provides precise assistance without a differentiated error processing mechanism, medical monitoring indicators may plateau. This makes it difficult for patients to correct erroneous muscle coordination patterns through system feedback, reducing the targeted nature of information-based rehabilitation guidance.

[0005] Existing system architectures still fall short in terms of multimodal data fusion depth and real-time monitoring of physiological load, making it difficult to automatically prevent secondary injuries caused by fatigue while ensuring training intensity. Therefore, establishing a closed-loop rehabilitation information processing system that can accurately extract weak feature signals, possess hierarchical strategy control capabilities, and take physiological safety into account is a pressing technical problem that needs to be solved in the field of intelligent medicine and rehabilitation engineering. Summary of the Invention

[0006] The purpose of this invention is to provide a personalized training guidance system for neurorehabilitation of stroke, in order to solve the problems mentioned in the background art. Specifically, the technical solution of this invention includes:

[0007] A central processing platform, which is communicatively connected to a data sensing interface, a personalized strategy engine, and a feedback execution interface;

[0008] The data sensing interface is used to acquire a multimodal rehabilitation dataset reflecting the patient's movement intentions and performance from external sensing devices;

[0009] The personalized strategy engine includes:

[0010] The state analysis unit is used to determine whether the patient's current rehabilitation training stage is a low neural drive stage or a high neural drive stage based on the multimodal rehabilitation dataset and a preset neuro-motor state model.

[0011] The strategy generation unit is used to generate a neural drive enhancement strategy containing random resonance parameters in response to the low neural drive phase; and to generate a motion learning optimization strategy containing error amplification parameters in response to the high neural drive phase.

[0012] The feedback execution interface is used to convert the neural drive enhancement strategy or motor learning optimization strategy into multi-sensory feedback control commands, driving external feedback devices to perform personalized induction on the patient.

[0013] Preferably, the process by which the state analysis unit determines the rehabilitation training stage includes:

[0014] Signal intensity features and motion trajectory features are extracted from the multimodal rehabilitation dataset;

[0015] The signal intensity feature is compared with a preset neural activation baseline range. If the signal intensity feature is within the neural activation baseline range and the motion trajectory feature shows disordered fluctuations, it is determined to be a low neural drive stage.

[0016] If the signal strength feature remains above the preset activation threshold and the motion trajectory feature exhibits directional displacement, it is determined to be in the high neural drive stage.

[0017] Preferably, the process by which the strategy generation unit generates neural-driven enhancement strategies includes:

[0018] Based on the characteristic data of the low neural drive stage, the difficulty barrier parameters of the current rehabilitation task are calculated.

[0019] With the goal of maximizing the activation probability across the potential barrier, the optimal stochastic resonance parameters are calculated in a preset parameter space using an optimization algorithm.

[0020] The random resonance parameters are used to configure the noise characteristics in the feedback signal to induce a coordinated response in the nervous system.

[0021] Preferably, the process for calculating the optimal stochastic resonance parameters is as follows:

[0022] A mapping model is established that is positively correlated with the difficulty barrier parameter and correlated with a nonlinear function of noise intensity;

[0023] During parameter optimization, the information entropy of the strategy output is evaluated in real time. If the rate of change of information entropy exceeds the preset convergence threshold, the current stochastic resonance parameter combination is locked.

[0024] Preferably, the process by which the strategy generation unit generates a motion learning optimization strategy includes:

[0025] Calculate the deviation vector of the motion trajectory features extracted from the multimodal rehabilitation dataset relative to the standard rehabilitation trajectory;

[0026] Determine whether the direction of the deviation vector deviates from the target direction of the standard rehabilitation trajectory;

[0027] If the direction deviates, the first set of error amplification parameters is used to perform a nonlinear transformation on the deviation vector to generate a reinforcement learning signal for feedback.

[0028] If the direction is favorable, the second set of parameters is used to normalize the deviation vector.

[0029] Preferably, the system further includes a training load monitoring module for performing the following operations:

[0030] Based on the multimodal rehabilitation dataset, fatigue index reflecting the patient's physiological load is calculated;

[0031] The fatigue index is compared with a preset load threshold.

[0032] If the fatigue index indicates that the load exceeds the load threshold, a control command is sent to the strategy generation unit.

[0033] In response to the control command, the strategy generation unit pauses the generation of the motion learning optimization strategy and switches to generating a protective strategy oriented towards error reduction and assistance.

[0034] Preferably, the process of generating multi-sensory feedback control instructions by the feedback execution interface includes:

[0035] Get the current strategy parameters;

[0036] Based on a pre-defined multi-channel fusion model, dynamic weights are assigned to the visual feedback channel and the tactile feedback channel.

[0037] The random resonance parameters in the neural drive enhancement strategy are mapped to the dynamic change frequency of visual feedback or the intensity modulation frequency of tactile feedback.

[0038] The error amplification parameters and deviation vectors in the aforementioned motion learning optimization strategy are mapped to spatial guidance vectors for the pose of virtual objects or tactile force fields in the visual scene.

[0039] Preferably, the stochastic resonance parameter in the neural drive enhancement strategy is used to generate synthetic noise data with non-Gaussian statistical properties and specific spectral attenuation characteristics to simulate an information carrier suitable for neural modulation.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1. This invention constructs a millisecond-level closed-loop control system for perception-decision-execution and a multi-channel fusion computing mechanism, effectively solving the dead zone problem of severely paralyzed patients being unable to trigger training due to insufficient muscle strength; unlike traditional solutions that rely solely on a single electromyography amplitude threshold, this solution simultaneously collects electromyography signals, limb movement data, eye-tracking data, and tactile sensing data, and jointly analyzes the statistical characteristics of bioelectrical signals and kinematic signals. It can keenly capture weak intention states with non-random low-frequency fluctuations in the spectrum, thus accurately identifying the patient's active movement attempts even when the signal strength is extremely low, ensuring the timeliness and accuracy of rehabilitation intervention;

[0042] 2. This invention introduces a neural drive enhancement strategy based on the principle of stochastic resonance, which effectively enhances weak subthreshold nerve impulses. By calculating the difficulty barrier parameters of the rehabilitation task and using an optimization algorithm to generate the optimal noise intensity, this method converts noise into a functional energy, enabling weak signals that are originally insufficient to trigger muscle contraction to cross the activation threshold with a high probability under the assistance of noise. This successfully induces a coordinated response of the nervous system without increasing the physical burden on the patient, thus breaking through the energy barrier of synaptic transmission.

[0043] 3. This invention employs an asymmetric error amplification mechanism and a physiological load circuit breaker strategy, which significantly improves the efficiency of motor learning and ensures training safety. During the patient's high neural drive phase, by artificially amplifying the motor error that deviates from the target, the human brain's instinct to minimize prediction error forces the nervous system to recruit more motor units to correct the movement, thus solving the problem of adaptive stagnation in the later stages of rehabilitation. At the same time, the system quantifies fatigue by monitoring the shift in the frequency domain characteristics of electromyographic signals, and can automatically switch to a high-assistance protection mode in the early stages of muscle physical fatigue, preventing the erroneous compensatory movements caused by fatigue from being solidified by the nervous system.

[0044] 4. This invention establishes a parameter convergence logic and multi-sensory feedback mapping model based on information entropy, endowing the system with extremely high robustness and physical interaction accuracy. By introducing an entropy convergence index and a completeness logic branch, the system effectively prevents program crashes caused by oscillations and logical dead zones during parameter optimization, ensuring the stability of control variables. In addition, this scheme dynamically maps abstract strategy parameters to the jitter frequency of the visual scene and the spatial guidance vector of the tactile force field, effectively activating the sensory system using unconventional feedback methods, and significantly improving the signal-to-noise ratio of the sensorimotor cortex. Attached Figure Description

[0045] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0046] Figure 1 This is a structural diagram of the system of the present invention; Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0048] Example 1:

[0049] Please see Figure 1 The personalized training guidance system for neurological rehabilitation of stroke includes a central processing platform, which has communication connections to a data sensing interface, a personalized strategy engine, and a feedback execution interface.

[0050] The data sensing interface is used to acquire multimodal rehabilitation datasets reflecting the patient's motor intentions and performance from external sensing devices;

[0051] Personalized strategy engine, including:

[0052] The state analysis unit is used to determine whether the patient's current rehabilitation training stage is a low neural drive stage or a high neural drive stage based on a multimodal rehabilitation dataset and a preset neuro-motor state model.

[0053] The strategy generation unit is used to generate a neural drive enhancement strategy containing stochastic resonance parameters in response to a low neural drive phase, and to generate a motion learning optimization strategy containing error amplification parameters in response to a high neural drive phase.

[0054] Feedback execution interface is used to transform neural drive enhancement strategies or motor learning optimization strategies into multi-sensory feedback control commands to drive external feedback devices to perform personalized induction on patients.

[0055] This embodiment constructs a millisecond-level closed-loop control system for perception, decision-making, and execution. The central processing platform, serving as the core control hub of the system, is implemented in this embodiment as an industrial control computer equipped with a real-time operating system. The system acquires multimodal rehabilitation datasets reflecting the patient's movement intentions and performance from external sensing devices through a data sensing interface. This interface simultaneously collects four types of key data:

[0056] First, electromyography (EMG) signals are collected through a high-density surface EMG sensor array, with a sampling rate set at 2000Hz. Second, data on the position, velocity, and acceleration of the limb extremities are obtained through an inertial measurement unit. Third, gaze point trajectory and pupil change data are collected in real time through an infrared eye-tracking module integrated into the virtual reality headset. Fourth, contact pressure and vibration perception threshold data are measured through a tactile sensor array distributed on the surface of the rehabilitation robot handle. These newly added sensory channels provide the necessary physical input foundation for subsequent multi-channel fusion calculations.

[0057] The state parsing unit in the personalized strategy engine is based on a multimodal rehabilitation dataset and determines the patient's current rehabilitation training stage according to a preset neuro-motor state model. The strategy generation unit dynamically generates control laws based on the determination results. The feedback execution interface transforms neural drive enhancement strategies or motor learning optimization strategies into multisensory feedback control commands to drive external feedback devices, including virtual reality headsets and end-effector traction rehabilitation robots, to perform personalized guidance for the patient.

[0058] This embodiment effectively solves the dead zone problem of severely paralyzed patients being unable to trigger training due to insufficient muscle strength by smoothly switching between two distinct control mechanisms—signal enhancement and error induction—based on the patient's real-time neuromotor state. To verify the beneficial technical effects of this embodiment, a 4-week clinical comparative experiment was conducted at a cooperative rehabilitation center. Thirty stroke hemiplegic patients were randomly divided into an experimental group using this system and a control group receiving conventional rehabilitation. The results showed that the average improvement in the Fugl-Meyer Upper Limb Motor Function Score (FMA-UE) of the experimental group reached [value missing]. The score was significantly higher than that of the control group. Furthermore, electromyography (EMG) data analysis showed that the average amplitude of the EMG response in the experimental group increased during the low neural drive phase. This confirms the significant efficacy of this system in resolving dead zone issues and inducing neural remodeling.

[0059] Example 2:

[0060] The process by which the state analysis unit determines the rehabilitation training stage includes:

[0061] Extract signal intensity features and motion trajectory features from a multimodal rehabilitation dataset;

[0062] The signal intensity features are compared with the preset neural activation baseline range. If the signal intensity features are within the neural activation baseline range and the motion trajectory features fluctuate randomly, it is determined to be a low neural drive stage.

[0063] If the signal strength characteristic is consistently higher than the preset activation threshold and the motion trajectory characteristic shows directional displacement, it is determined to be in the high neural drive stage.

[0064] This embodiment further specifies the process of the state analysis unit in determining the rehabilitation training stage; this process aims to quantify the patient's neurological and motor state; the preset neuro-motor state model is constructed as a multi-dimensional feature discrimination space, whose input vector includes signal intensity features. and motion trajectory characteristics The discrimination boundary is determined by the neural activation baseline. Disorder fluctuation threshold Activation threshold and directional consistency operator Together, they constitute the state resolution unit, which determines the patient's rehabilitation stage based on the region where the input vector falls within the discrimination space. The specific process is as follows:

[0065]

[0066] in, Instantaneous amplitude of the preprocessed electromyographic signal : Number of sampling points within the sliding time window; here Defined as the root mean square value of the signal, to ensure its physical units are consistent with the reference voltage in subsequent steps. To maintain consistency, both are in the dimension of voltage;

[0067] Motion trajectory characteristics The standard deviation algorithm is used to calculate the dispersion of limb movements. The calculation formula is as follows:

[0068]

[0069] in, : The spatial coordinates of the limb's extremity Average position within the sliding window. Number of trajectory sampling points;

[0070] To ensure the clarity and executability of the decision logic, this embodiment specifically defines the key thresholds and vector parameters involved in the formula: the lower limit of the neural activation baseline range. and upper limit The average amplitude of the electromyographic signal in the patient's resting state was collected respectively. Double Multiples; Disorder fluctuation threshold Set as Used to filter out subtle tremors not controlled by the mind; activation threshold Set to 15% of the patient's maximum voluntary contraction MVC voltage value; vector Defined as the current instantaneous velocity vector of the limb's extremity. Defined as a normalized direction vector pointing from the current limb position to the rehabilitation target point;

[0071] The signal intensity features are compared with a preset neural activation baseline; in response to the signal intensity features Within the baseline range, i.e. and motion trajectory characteristics It exhibits disordered fluctuations, that is The system determines this to be a low-neural drive phase; this logic aims to identify weak intention states that appear to have no signal but whose spectrum shows non-random low-frequency fluctuations; in response to a signal strength characteristic that remains above a preset activation threshold, i.e. Regardless of the current state of the movement trajectory, the system determines it to be in a high neural drive stage. At this time, the system confirms that the patient has active drive ability, and will subsequently match the corresponding optimization strategy according to the correctness of the movement direction.

[0072] To prevent the pseudocode compiler from triggering circuit breakers in logical dead zones, this embodiment adds a completeness logic branch: when the patient's state falls into the intermediate energy zone. At this time, the system executes the state maintenance and default reset strategy, that is, it prioritizes maintaining the judgment result of the previous moment, and if there is no historical state, it defaults to the low neural drive stage; this closed-loop logic ensures that the variable of the current stage can return a valid value under any input scenario, avoiding the program crash caused by the subsequent policy generation unit having no input.

[0073] This embodiment, by jointly analyzing the statistical characteristics of bioelectrical and kinematic signals, can keenly capture the patient's extremely weak movement attempts, providing a precise timing for subsequent random resonance intervention. This avoids the shortcomings of traditional systems that fail to respond due to weak signals, ensuring the timeliness and accuracy of rehabilitation intervention.

[0074] Example 3:

[0075] The process by which the strategy generation unit generates neurally driven enhancement strategies includes:

[0076] Based on the characteristic data of the low neural drive stage, the difficulty barrier parameters of the current rehabilitation task are calculated.

[0077] With the goal of maximizing the activation probability across the potential barrier, the optimal stochastic resonance parameters are calculated in a preset parameter space using an optimization algorithm.

[0078] Random resonance parameters are used to configure the noise characteristics in the feedback signal to induce a coordinated response in the nervous system.

[0079] This embodiment of the system calculates the difficulty barrier parameters of the current rehabilitation task based on characteristic data from the low neural drive stage. This parameter quantifies the energy deficit required for the patient to produce effective movement, in order to avoid the risk of hypertonia leading to... This leads to the computational risk of negative potential barriers. In this embodiment, a non-negativity constraint function is introduced, and the computational model is modified as follows:

[0080]

[0081] in, Standard electromyography amplitude reference value, unit: ; This is the root mean square value of the current real-time electromyography signal, in units of: , Normalization coefficient, unit: ;

[0082] To ensure the specific computability of the above parameters and to solve the problem of dimensional matching in exponentiation, this embodiment clarifies that... and Determination method: Standard electromyographic amplitude reference value This is obtained by collecting the root mean square (RMS) value of electromyography (EMG) signals from the patient's unaffected limb or a healthy subject performing the same standard rehabilitation movements without assistance during the system initialization phase. For example, a value of [value missing] is used. Normalization coefficient Set as That is, the reciprocal of the microvolt; through The product of the inverse dimension and the voltage difference is physically eliminated. and The voltage unit maps the voltage difference in physical units to a strictly dimensionless barrier value suitable for the stochastic resonance algorithm.

[0083] With the goal of maximizing the activation probability across the potential barrier, an optimization algorithm is used to calculate the optimal stochastic resonance parameters within a preset parameter space. The stochastic resonance parameters are defined as the ratio of physical noise intensity to the system's baseline noise intensity, and are dimensionless. Given that the activation probability function varies with noise intensity... Due to the monotonically increasing characteristic, and to prevent the optimization results from diverging to non-physical infinity, the preset parameter space is strictly limited as follows:

[0084]

[0085] in, The upper bound is a dimensionless safety bound, for example, 100, corresponding to a physical output voltage of 5V. Under this constraint, in order to solve the problem of numerical overflow caused by the original probability formula when τ is extremely small, i.e., P>1.0, this embodiment defines the cross-barrier activation probability P as the normalized probability within a single time window. The formula is modified as follows:

[0086]

[0087] in, The difficulty barrier parameter is no longer a negative constraint and is dimensionless. : Normalized noise intensity to be optimized, dimensionless;

[0088] By introducing dimensional coefficients make sure Dimensionless, and in conjunction with The non-negativity constraint ensures that the calculated Always strictly adhere to Within the interval, this provides a valid mathematical input for subsequent information entropy calculation; the calculated random resonance parameters are used to configure the noise characteristics in the feedback signal to induce a coordinated response in the nervous system;

[0089] This embodiment utilizes the principle of random resonance to convert the calculated optimal noise intensity into a functional energy, enabling weak nerve impulses that would otherwise be insufficient to trigger muscle contraction to cross the activation threshold with the aid of noise. This successfully induces a coordinated response of the nervous system without increasing the physical burden on the patient, thus achieving effective enhancement of subthreshold signals.

[0090] Example 4:

[0091] The process of calculating the optimal stochastic resonance parameters also includes:

[0092] Establish a mapping model that shows the cross-barrier activation probability is positively correlated with the difficulty barrier parameter and related to a nonlinear function of noise intensity;

[0093] During parameter optimization, the information entropy of the strategy output is evaluated in real time. If the rate of change of information entropy exceeds the preset convergence threshold, the current stochastic resonance parameter combination is locked.

[0094] Specifically, this embodiment establishes a mapping model that is positively correlated with the difficulty barrier parameter and related to a nonlinear function of noise intensity across potential barriers; this is to ensure that patients can effectively handle higher difficulty barriers. It can obtain sufficient assistance at the same time, while preventing the target probability from overflowing. The interval caused the algorithm to fail. This model uses a saturated growth function to define the target activation probability of the system. That is, the higher the potential barrier, the greater the activation probability of the target set by the system, but it is always less than 1. The model formula is corrected as follows:

[0095]

[0096] in, To control the decay constant of the saturation velocity, the system adjusts the noise intensity by executing the following discrete-time iterative update law. This makes the actual activation probability approach the target value:

[0097]

[0098] This step size is selected based on empirical considerations of the balance between system response speed and overshoot, to ensure probabilistic convergence within 10 control cycles.

[0099] in, The number of iterations. To adjust the step size, for example, take To avoid confusion with the convergence metric notation defined later, the step size parameter is defined here as follows: ; This represents the calculated probability under the current noise intensity; this explicit mathematical operator ensures that the system can automatically calculate the convergent noise parameters; to make the objective function practically operable, this embodiment sets... To adapt Value at Variations within a range; setting a minimum probability This ensures that the system maintains basic auxiliary activity even in low-barrier tasks; during parameter optimization, the information entropy of the policy output is evaluated in real time. Information entropy To quantify the uncertainty of the system's activation state, this embodiment uses an activation probability-based approach. The binary Shannon entropy is defined and its calculation formula is as follows:

[0100]

[0101] The formula uses the real-time activation probability calculated in the preceding steps. As input, when near The maximum entropy value indicates that the system is in a sensitive region of stochastic resonance. To accurately determine the convergence state and ensure logical consistency, this embodiment specifies the rate of change of information entropy as an entropy convergence index that reflects the degree of steady-state approximation of the system. This indicates that the uncertainty fluctuations of the system have decayed to an extremely low level, and the system is determined to have entered a steady state. This indicator is constructed by normalizing the time derivative of entropy through exponential mapping, in accordance with the control logic of locking when the threshold is exceeded, which is convergence. The calculation formula is as follows:

[0102]

[0103] in, The absolute value of the time derivative of information entropy, i.e., the rate of entropy change per unit time, is expressed in bits per second. To adjust the scaling constant for sensitivity, it is set to 0.01 bit / s to ensure dimensional consistency of the exponential term and to normalize the input; when the system tends to stabilize, the rate of change of entropy... Tend to This makes the indicators tending towards the maximum value When this indicator Exceeding the preset convergence threshold For example, set to This means that the uncertainty fluctuations of the system have decayed to an extremely low level, the system is determined to have entered a steady state, and the current combination of stochastic resonance parameters is locked.

[0104] This embodiment ensures strict consistency between the algorithm logic and the constraints by redefining the target probability model and convergence index, effectively preventing excessive noise from drowning out the effective signal, and ensuring the accuracy and safety of the neural modulation process.

[0105] Example 5:

[0106] The process by which the policy generation unit generates motion learning optimization policies includes:

[0107] Calculate the deviation vector of motion trajectory features extracted from the multimodal rehabilitation dataset relative to the standard rehabilitation trajectory;

[0108] Determine whether the direction of the deviation vector deviates from the target direction of the standard rehabilitation trajectory;

[0109] If the direction deviates, the first set of error amplification parameters is called to perform a nonlinear transformation on the deviation vector to generate a reinforcement learning signal for feedback.

[0110] If the direction is favorable, the second set of parameters is used to normalize the deviation vector.

[0111] This embodiment further specifies the process by which the strategy generation unit generates motion learning optimization strategies; this process, namely the error amplification mechanism, aims to stimulate motion learning through specific error processing during the high neural drive phase; and calculates the deviation vector of the motion trajectory features relative to the standard rehabilitation trajectory. This vector is derived from the difference between the patient's current actual position vector and the nearest point position vector of the standard trajectory. The standard rehabilitation trajectory is obtained by: during the training initialization phase, using the mirror algorithm to collect and record the motion path data of the patient's healthy limb performing the same action, or directly calling the standard kinematic template data of the action stored in the system database.

[0112] Determine whether the direction of the deviation vector deviates from the target direction of the standard rehabilitation trajectory; this step is achieved by calculating the projection relationship between the deviation vector and the target tangential direction; in response to directional deviation, the system calls the first set of error amplification parameters. In this embodiment, it is set The preferred value is A nonlinear transformation is performed on the bias vector to generate a reinforcement learning signal for feedback. ,Right now This operation creates a virtual high-potential field, forcing the patient to perceive a larger error than actually exists; in response to directional tendencies, the system calls the second set of parameters. In this embodiment, it is set The preferred value is The deviation vector is normalized to enhance patient confidence; the processed deviation signal is output to the feedback interface.

[0113] This embodiment employs an asymmetric error processing strategy. By artificially amplifying the error in the deviating direction, it leverages the human brain's instinct to minimize prediction errors, forcing the nervous system to recruit more motor units to correct the movement. This forced learning strategy significantly accelerates the remodeling efficiency of neural pathways and solves the problem of adaptive stagnation that patients are prone to in traditional rehabilitation.

[0114] Example 6:

[0115] The system also includes a training load monitoring module, which performs the following operations:

[0116] Based on a multimodal rehabilitation dataset, fatigue indexes reflecting patients' physiological load were calculated.

[0117] The fatigue index is compared with the preset load threshold.

[0118] If the fatigue index indicates that the load exceeds the load threshold, a control command is sent to the strategy generation unit.

[0119] In response to the control command, the strategy generation unit pauses the generation of motion learning optimization strategies and switches to generating protective strategies oriented towards error reduction and assistance.

[0120] This embodiment further supplements the system by adding a training load monitoring module. This module aims to prevent muscle damage or compensatory movements caused by high-intensity error amplification training. Based on a multimodal rehabilitation dataset, it calculates fatigue indicators reflecting the patient's physiological load. This index utilizes the frequency domain characteristics of electromyographic signals for offset quantization, as shown in the following formula:

[0121]

[0122] in, The median frequency of the power spectrum of the electromyographic signal within the current time window. : The baseline median frequency at the start of training;

[0123] Fatigue index With the preset load threshold A comparison is performed; to ensure that the judgment logic has a clear trigger boundary, this embodiment uses a preset load threshold. Set as That is, the corresponding median frequency decreases. The physiological fatigue threshold; the response of fatigue index to indicate that the load exceeds the threshold, i.e. The system sends a control command to the strategy generation unit. In response to the control command, the strategy generation unit pauses the generation of motion learning optimization strategies and switches to generating protective strategies oriented towards error reduction and assistance, and starts the robot-assisted mode. The protective strategy is specifically manifested as follows: the control law of the rehabilitation robot is switched to high-impedance position control mode, the PID position controller is called, the standard rehabilitation trajectory is used as the set value, and the maximum safety stiffness coefficient, such as 500 N / m, is used to output an auxiliary force to forcibly pull the affected limb back to the standard trajectory, thereby eliminating motion errors and unloading muscle load.

[0124] This embodiment constructs a physiologically safe circuit breaker mechanism. By monitoring the leftward shift of the electromyographic spectrum, a physiological gold standard, it can intervene in a timely manner at the early stage of muscle physical fatigue, switching the training mode from high challenge to high assistance. This not only protects the patient from sports injuries but also prevents the erroneous compensatory movements caused by fatigue from being solidified by the nervous system.

[0125] Example 7:

[0126] The process of generating multi-sensory feedback control commands through the feedback execution interface includes:

[0127] Get the current strategy parameters;

[0128] Based on a pre-defined multi-channel fusion model, dynamic weights are assigned to the visual feedback channel and the tactile feedback channel.

[0129] The stochastic resonance parameters in the neural drive enhancement strategy are mapped to the dynamic change frequency of visual feedback or the intensity modulation frequency of tactile feedback.

[0130] The error amplification parameters and deviation vectors in the motion learning optimization strategy are mapped to spatial guidance vectors of virtual object poses or tactile force fields in the visual scene.

[0131] This embodiment further specifies the process of generating multi-sensory feedback control instructions through the feedback execution interface. This process transforms abstract strategy parameters into concrete physical stimuli. Specifically, it acquires the current strategy parameters and assigns dynamic weights to the visual and tactile feedback channels based on a preset multi-channel fusion model. The multi-channel fusion model receives two normalized input variables: a visual attention index based on eye-tracking data. and tactile interaction index based on threshold and pressure data To address the issue of lost contact pressure parameter mapping, this embodiment reconstructs... The definition includes the pressure characteristics of Example 1; simultaneously, based on the pupillary change data defined in Example 1, the visual attention index... The corresponding calculation factors were added; Visual Attention Index The calculation formula is as follows:

[0132]

[0133] in, In the current time window ,like The cumulative time that the patient's gaze falls within the target area; The rate of change of pupil diameter. The maximum physiological rate of change, such as This item reflects the patient's cognitive load level; tactile interaction index The calculation introduces a division-by-zero protection mechanism:

[0134]

[0135] in, Defined as the reference value of vibration intensity output by the feedback device at the current moment, for example... , The tactile threshold, For example, a preset safety minimum value. , The contact pressure is measured in real time by the sensor array. For the pressure sensor range, such as and restrictions The weight allocation algorithm employs a dynamic complementary strategy to prevent issues arising when the patient's eyes are closed and there is no tactile input. and The formula introduces a smoothing term to address the anomaly of a denominator being zero caused by time-related errors. The corrections are as follows:

[0136]

[0137] in, It is a very small positive smoothing constant, for example This ensures the robustness of the algorithm under extreme conditions; For visual channel weights, For tactile channel weights; random resonance parameters in the neural drive enhancement strategy. This is mapped to the dynamic frequency of visual feedback or the intensity modulation frequency of tactile feedback; for example, the visual channel manifests as high-frequency positional jitter of the virtual hand. The tactile channel manifests as the random micro-vibration modulation frequency of the rehabilitation handle. ;

[0138] The error amplification parameters and deviation vectors in the motion learning optimization strategy are mapped to spatial guidance vectors of virtual object poses or tactile force fields in the visual scene; for example, the tactile channel generates a spatial guidance force field opposite to the deviation direction. To ensure the feasibility of the above mapping technology and to eliminate the dimensionless parameters in Example 3. To resolve unit conflicts and address the inconsistency in units in the original scheme, this embodiment modifies the definition of the mapping coefficients involved: visual channel coefficients. Set as For dimensionless parameters Frequency gain, tactile channel coefficient Set as For dimensionless parameters The frequency gain, to output a frequency control quantity that conforms to the specified parameters, force field coefficient Set as The driver hardware executes the above instructions, and the above coefficient values ​​are obtained based on experiments measuring human sensory thresholds.

[0139] This embodiment specifies the algorithm parameters as visual jitter, tactile micro-vibration, and spatial force field, and effectively activates the ion channels of the sensory system using an unconventional feedback method, significantly improving the signal-to-noise ratio of the sensorimotor cortex, thereby achieving a highly efficient neural regulation closed loop at the physical level.

[0140] Example 8:

[0141] The stochastic resonance parameter in the neural drive enhancement strategy is used to generate synthetic noise data with non-Gaussian statistical properties and specific spectral attenuation characteristics to simulate information carriers suitable for neural modulation.

[0142] This embodiment further specifies the stochastic resonance parameters in the neural-driven enhancement strategy. This process generates synthetic noise data with specific statistical properties to maximize biocompatibility. The generated base noise data must satisfy non-Gaussian statistical properties, such as a Laplace distribution, to increase the probability of large signals occurring, thus making it easier to trigger threshold crossing. The noise data undergoes spectral shaping to give it specific spectral attenuation characteristics, i.e., noise power spectral density. satisfy To avoid confusion with the median frequency symbol in the aforementioned embodiments, the characteristic used here is... Representing general frequency variables:

[0143]

[0144] in, Frequency independent variable Attenuation index, with a value range covering the pink noise region; specific recommended values ​​are provided. or range of values

[0145] The synthesized noise is output as a carrier of random resonance. The synthesized noise generated in this embodiment conforms to the characteristics of biological neurodynamics. Compared with traditional Gaussian white noise, it can induce a more significant random resonance effect with lower energy input. This not only improves the efficiency of neural regulation, but also significantly reduces the discomfort caused by long-term stimulation, demonstrating the system's deep adaptation to the neural characteristics of organisms.

[0146] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A personalized training and guidance system for neurorehabilitation of stroke, characterized in that, It includes a central processing platform, which is communicatively connected to a data-aware interface, a personalized policy engine, and a feedback execution interface; The data sensing interface is used to acquire a multimodal rehabilitation dataset reflecting the patient's movement intentions and performance from external sensing devices; The personalized strategy engine includes: The state analysis unit is used to determine whether the patient's current rehabilitation training stage is a low neural drive stage or a high neural drive stage based on the multimodal rehabilitation dataset and a preset neuro-motor state model. The strategy generation unit is used to generate a neural drive enhancement strategy containing random resonance parameters in response to the low neural drive phase; and to generate a motion learning optimization strategy containing error amplification parameters in response to the high neural drive phase. The feedback execution interface is used to convert the neural drive enhancement strategy or motor learning optimization strategy into multi-sensory feedback control commands to drive external feedback devices to perform personalized induction on the patient. The process by which the strategy generation unit generates neural-driven enhancement strategies includes: Based on the characteristic data of the low neural drive stage, the difficulty barrier parameter of the current rehabilitation task is calculated using the following calculation model: ； in, For the unit The standard reference value for electromyography amplitude; For the unit The current real-time root mean square value of electromyography (EMG) signal; For the unit The normalized coefficient; With the goal of maximizing the activation probability across the potential barrier, the optimal stochastic resonance parameters are calculated in a preset parameter space using an optimization algorithm. The stochastic resonance parameters are used to configure the noise characteristics in the feedback signal to induce a coordinated response in the nervous system; the cross-barrier activation probability Defined as the normalized probability within a single time window, the formula is as follows: ； in, The difficulty barrier parameter has non-negative constraints and dimensionless properties. The normalized noise intensity to be optimized is a dimensionless property. The process of calculating the optimal stochastic resonance parameters: A mapping model is established that is positively correlated with the difficulty barrier parameter and correlated with a nonlinear function of noise intensity; The target activation probability of the system is defined using a saturated growth function. The model formula is as follows: ； in, The decay constant is used to control the saturation velocity; The system adjusts the noise intensity by executing the following discrete-time iterative update law. This makes the actual activation probability approach the target value: ； in, The number of iterations. To adjust the step size, The calculated probability is given by the current noise level. During parameter optimization, the information entropy of the strategy output is evaluated in real time. This embodiment uses an activation probability-based approach. The binary Shannon entropy is defined and its calculation formula is as follows: ； The rate of change of information entropy is concretized into an entropy convergence index that reflects the degree of steady-state approximation of the system. The calculation formula is as follows: ； in, The absolute value of the time derivative of information entropy, which characterizes the rate of entropy change per unit time and is expressed in bits per second; The scale constant is used to adjust the sensitivity; When this indicator If the convergence threshold is exceeded, the system determines that it has entered a steady state and then locks the current combination of stochastic resonance parameters.

2. The system according to claim 1, characterized in that, The process by which the state analysis unit determines the rehabilitation training stage includes: Signal intensity features and motion trajectory features are extracted from the multimodal rehabilitation dataset; The signal intensity feature is compared with a preset neural activation baseline range. If the signal intensity feature is within the neural activation baseline range and the motion trajectory feature shows disordered fluctuations, it is determined to be a low neural drive stage. If the signal strength feature remains above the preset activation threshold and the motion trajectory feature exhibits directional displacement, it is determined to be in the high neural drive stage.

3. The system according to claim 1, characterized in that, The process by which the strategy generation unit generates a motion learning optimization strategy includes: Calculate the deviation vector of the motion trajectory features extracted from the multimodal rehabilitation dataset relative to the standard rehabilitation trajectory; Determine whether the direction of the deviation vector deviates from the target direction of the standard rehabilitation trajectory; If the direction deviates, the first set of error amplification parameters is used to perform a nonlinear transformation on the deviation vector to generate a reinforcement learning signal for feedback. If the direction is favorable, the second set of parameters is used to normalize the deviation vector.

4. The system according to claim 3, characterized in that, The system also includes a training load monitoring module for performing the following operations: Based on the multimodal rehabilitation dataset, fatigue index reflecting the patient's physiological load is calculated; The fatigue index is compared with a preset load threshold. If the fatigue index indicates that the load exceeds the load threshold, a control command is sent to the strategy generation unit. In response to the control command, the strategy generation unit pauses the generation of the motion learning optimization strategy and switches to generating a protective strategy oriented towards error reduction and assistance.

5. The system according to claim 1, characterized in that, The process by which the feedback execution interface generates multi-sensory feedback control commands includes: Get the current strategy parameters; Based on a pre-defined multi-channel fusion model, dynamic weights are assigned to the visual feedback channel and the tactile feedback channel. The random resonance parameters in the neural drive enhancement strategy are mapped to the dynamic change frequency of visual feedback or the intensity modulation frequency of tactile feedback. The error amplification parameters and deviation vectors in the aforementioned motion learning optimization strategy are mapped to spatial guidance vectors for the pose of virtual objects or tactile force fields in the visual scene.

6. The system according to claim 1, characterized in that, The stochastic resonance parameter in the neural drive enhancement strategy is used to generate synthetic noise data with non-Gaussian statistical properties and specific spectral attenuation characteristics to simulate an information carrier suitable for neural modulation.

Images