A rehabilitation training system and method for hepatolenticular degeneration myotonia in combination of traditional Chinese and western medicine

Abstract

The application belongs to the field of intelligent rehabilitation medicine and traditional Chinese medicine informatization, and relates to a liver degeneration muscle tension disorder combined rehabilitation training system and method of traditional Chinese and western medicine; the system comprises a multi-modal data acquisition module, a traditional Chinese and western medicine collaborative digital twin prediction module, a self-adaptive mutation multi-objective optimization decision module and a rehabilitation training execution module. The self-adaptive mutation multi-objective optimization decision module takes a rehabilitation scheme vector as an optimization variable and adopts an improved differential evolution algorithm for optimization. The algorithm mutation operation contains two modes: a hidden structure syndrome topological potential-based guided mutation, which evolves along a syndrome transmission path according to real-time four diagnostic data and copper metabolism correction factors; and an escape mutation based on a neural burst index early warning, which forcibly corrects the scheme to a safety domain when the index exceeds a standard. The application embeds the evolution law of traditional Chinese pathogenesis and the neural electrical physiological early warning signal in the evolution search process, so as to realize safe, dynamic and individualized rehabilitation scheme generation.

Description

Technical Field

[0001] This invention belongs to the field of intelligent rehabilitation medicine and information technology of traditional Chinese medicine, specifically relating to a rehabilitation training system and method combining traditional Chinese and Western medicine for Wilson's disease dystonia. Background Technology

[0002] Wilson's disease is an autosomal recessive genetic disorder caused by mutations in the ATP7B gene, leading to abnormal copper metabolism. Dystonia is its most common and disabling neurological manifestation. The accumulation of copper toxins in the patient's body triggers dynamic evolution of TCM syndromes, with common syndromes including phlegm-blood stasis syndrome, damp-heat retention syndrome, liver qi stagnation syndrome, liver and kidney yin deficiency syndrome, and spleen and kidney yang deficiency syndrome. Current rehabilitation models have significant limitations: Western medicine rehabilitation training lacks consideration of the patterns of TCM syndrome progression; TCM rehabilitation heavily relies on physician experience and is difficult to quantify precisely; rehabilitation plans are mostly static prescriptions, unable to be dynamically adjusted in real time according to the patient's copper metabolism level, syndrome evolution trend, and the risk of neurophysiological outbursts in dystonia.

[0003] Existing studies have utilized latent structure models (such as the LTM-EAST algorithm) to mine the syndrome elements and identification criteria of Wilson's disease, revealing that the lesion involves the liver, brain, spleen, and kidneys, and that pathological factors include phlegm turbidity, blood stasis, qi stagnation, damp-heat, and yin deficiency. They have also summarized the symptom thresholds of common syndromes such as phlegm-blood stasis syndrome and damp-heat accumulation syndrome. However, these studies only remain at the level of analyzing syndrome differentiation patterns. There is currently no technology to transform the syndrome elements and their topological relationships revealed by latent structure models into real-time guidance signals that can be embedded in rehabilitation program optimization algorithms. Furthermore, they have not combined these with copper metabolism safety constraints and neurophysiological early warning to form a closed-loop intelligent rehabilitation decision-making system. Simultaneously, existing standard difference evolutionary algorithms, when used for optimizing rehabilitation parameters, lack constraints specific to the pathophysiological boundaries of Wilson's disease, and the resulting solutions may induce dystonia crisis or contradict traditional Chinese medicine treatment principles. Therefore, there is an urgent need for an intelligent optimization system that deeply integrates pathogenesis patterns with neurophysiological early warning. Summary of the Invention

[0004] The technical problem this invention aims to solve is that existing rehabilitation optimization techniques cannot incorporate the evolution of the latent structural syndrome of Wilson's disease and the risk of neurogenic dystonia into the search process, resulting in poor rationality and insufficient safety of the generated solutions.

[0005] To address the above problems, this invention provides a rehabilitation training system and method combining traditional Chinese and Western medicine for Wilson's disease dystonia.

[0006] Firstly, a rehabilitation training system integrating traditional Chinese and Western medicine for Wilson's disease dystonia includes a multimodal data acquisition module, a digital twin prediction module combining traditional Chinese and Western medicine, an adaptive variable multi-objective optimization decision-making module, and a rehabilitation training execution module.

[0007] The multimodal data acquisition module collects patients' Western medicine motor function data, objective data from the four diagnostic methods of traditional Chinese medicine, and copper metabolism biochemical indicators in real time.

[0008] The adaptive variability multi-objective optimization decision module uses the rehabilitation plan parameter vector as the optimization variable. With the efficacy prediction support provided by the integrated traditional Chinese and Western medicine digital twin prediction module, it adopts an improved differential evolution algorithm to perform multi-objective optimization and generate personalized rehabilitation plans.

[0009] The improved differential evolution algorithm employs at least one of the following two mutation methods in its mutation operation:

[0010] Method 1: Based on the potential energy-guided variation of the hidden structure syndrome topology, the target syndrome state is determined by a pre-trained hidden structure model according to the patient's real-time diagnostic data. The potential energy gradient of the current rehabilitation plan vector in the pre-constructed syndrome topology map is calculated, and the potential energy gradient vector is scaled according to the copper metabolism correction factor and applied to the current rehabilitation plan vector to generate a variation plan. This mechanism ensures that the iterative search of the rehabilitation plan always follows the optimal path that conforms to the laws of syndrome transmission in traditional Chinese medicine, while avoiding the clinical risks caused by drastic fluctuations in copper metabolism.

[0011] Method 2: Based on the escape variation of the nerve burst warning in dystonia, the nerve burst index in the surface electromyography signal is calculated in real time; when the nerve burst index is lower than or equal to the set warning threshold, escape is not triggered; when the nerve burst index exceeds the set warning threshold, the escape intensity is dynamically increased according to the degree of exceedance, and the current rehabilitation plan vector is perturbed towards the preset safe treatment direction vector. The higher the degree of danger, the greater the escape amplitude, so as to generate a variation scheme to escape to the safe domain.

[0012] In each generation of evolution, each sub-problem individual calculates a neuroburst index based on the current patient state. If the index exceeds a threshold, an escape mutation mechanism is forcibly selected; otherwise, a mutation mechanism (including mode one and standard difference mutation mechanism) is selected from the strategy pool according to adaptive probability. This enables dynamic, safe, and individualized optimization of the integrated traditional Chinese and Western medicine rehabilitation program.

[0013] A further improvement is made to the multimodal data acquisition module, which includes:

[0014] A high-density surface electromyography (EMG) acquisition array is used to extract the synergistic contraction rate of antagonistic muscle groups and the pathological synchronous discharge index.

[0015] Inertial motion sensors and depth vision cameras are used for joint kinematics assessment.

[0016] The digital tongue diagnosis instrument and pulse diagnosis sensor acquire quantitative features of tongue and pulse images, and run a hidden structure model to output the probability of syndromes.

[0017] The blood biochemistry test data interface receives serum ceruloplasmin, non-ceruloplasmin-bound copper, and 24-hour urinary copper levels.

[0018] In a further improvement, the syndrome topology graph is constructed based on a hidden tree structure model, where nodes represent different syndrome states and edge weights reflect the correlation between symptoms and the difficulty of transformation between syndromes; the potential energy function is the cumulative pathogenesis intervention cost of the shortest syndrome transmission path from the current syndrome node to the target syndrome node; and the potential energy gradient direction is the steepest descent direction of the potential energy function.

[0019] In a further improvement, the copper metabolism correction factor is dynamically determined based on the patient's real-time 24-hour urinary copper excretion rate and serum non-ceruloplasmin-bound copper concentration. When the copper metabolism index exceeds the safe range, the copper metabolism correction factor decreases, thereby automatically reducing the mutation search step size.

[0020] In a further improvement, the neural burst index is obtained by performing time-frequency analysis on high-density surface electromyography signals, extracting the synchronous discharge energy of antagonistic muscle groups in a preset frequency band, and calculating the correlation with the synergistic contraction envelope; an increase in the index reflects an increased risk of worsening dystonia; the escape variation is not triggered when the neural burst index is below or equal to the warning threshold, but is only triggered when it exceeds the warning threshold, and the escape amplitude is positively correlated with the degree of exceeding.

[0021] In a further improvement, the adaptive mutation multi-objective optimization decision module is configured as follows:

[0022] When the neural burst index is greater than the warning threshold, escape mutation method 2 is forcibly selected; otherwise, based on the probability of successfully generating non-dominated solutions in the history of each mutation method, method 1 or standard deviation mutation method is adaptively selected.

[0023] In a further improvement, the digital twin prediction module incorporates a graph-time fusion network, using acupoints on the human body's meridians as graph nodes. It captures the topological effects of acupoint stimulation through graph convolution, and combines the time-series motion signals with a Transformer encoder to output multi-objective prediction values, including the amount of improvement in motor function, the amount of improvement in quality of life, the amplitude of copper metabolism fluctuations, and the deviation of TCM syndromes.

[0024] Secondly, a rehabilitation training method combining traditional Chinese and Western medicine for Wilson's disease dystonia includes the following steps:

[0025] S1. Acquire real-time Western medicine motor function data, objective data of the four diagnostic methods of traditional Chinese medicine, and copper metabolism indicators of patients through the multimodal acquisition module;

[0026] S2. Input the candidate rehabilitation program vectors into a pre-trained graph temporal fusion network to predict multidimensional therapeutic effects;

[0027] S3. In the adaptive variation multi-objective optimization decision module, with the goal of maximizing exercise improvement and quality of life, and minimizing copper metabolism fluctuations and syndrome deviations, the improved differential evolution algorithm is executed to iteratively search for the optimal rehabilitation plan.

[0028] The improved differential evolution algorithm performs at least one of the following steps during the mutation operation:

[0029] Based on the target syndrome state determined by the implicit structure model, the potential gradient of the current rehabilitation plan vector in the syndrome topology is calculated, and the current plan is perturbed after adjusting the gradient using the copper metabolism correction factor.

[0030] The neural burst index is calculated based on surface electromyography signals. When the index exceeds the warning threshold, the current rehabilitation plan vector is adjusted towards safe treatment, generating an escape variant plan.

[0031] In a further improvement, the improved differential evolution algorithm described in step S3 calculates the neural burst index in real time when generating mutated individuals. If the index exceeds the threshold, escape mutation is performed; otherwise, either method one or standard differential mutation strategy is dynamically selected.

[0032] In a further improvement, the hidden structure model is pre-trained based on multicenter diagnostic data of patients with Wilson's disease. The resulting syndrome identification criteria and topological structure are used to determine the target syndrome state and calculate the potential energy gradient online. During online inference, 64 real-time collected symptoms from the four diagnostic methods are input into the hidden tree model. Bayesian posterior inference is used to calculate the probability distribution of each hidden variable, and the current syndrome type and target syndrome are determined according to the identification criteria threshold.

[0033] This invention, for the first time, incorporates the pathogenesis of Wilson's disease and the risk of dystonia outbreaks into the search process of an evolutionary algorithm through two major mutation mechanisms: topological potential guidance of latent structural symptoms and neural warning escape. Compared with existing technologies, it has the following outstanding advantages:

[0034] 1. The variations of the present invention strictly follow the syndrome transmission path revealed by the hidden structure, avoiding schemes that violate the principles of traditional Chinese medicine.

[0035] 2. The dual protection of copper metabolism correction factors and neural early warning escape mechanisms eliminates the risk of copper metabolism disorders and dystonia crises induced by rehabilitation training.

[0036] 3. All guidance signals are calculated based on the patient's real-time multimodal data, truly realizing dynamic rehabilitation planning of "one person, one policy, one time, one policy". Detailed Implementation

[0037] The present invention will be further described in detail below through specific embodiments.

[0038] Example 1

[0039] A rehabilitation training system combining traditional Chinese and Western medicine for Wilson's disease dystonia consists of a four-layer structure.

[0040] 1. Multimodal data acquisition layer;

[0041] include:

[0042] 1) A high-density surface electromyography (EMG) acquisition array (sampling rate 2000Hz) is attached to key muscle groups such as the biceps brachii, triceps brachii, and quadriceps femoris to calculate the synergistic contraction rate of antagonistic muscle groups and the pathological synchronous discharge energy in the 20-50Hz frequency band in real time.

[0043] 2) Nine-axis inertial motion sensor and depth vision camera, used to quantify joint range of motion, abnormal postures of dystonia, etc.;

[0044] 3) A digital tongue diagnosis instrument and pulse diagnosis pressure sensor array acquire quantitative values ​​of tongue color, tongue coating thickness / color, and pulse time-domain and frequency-domain characteristics, forming 64 items of symptoms for the four diagnostic methods.

[0045] (Including red tongue, yellow greasy coating, spontaneous sweating, bitter taste in the mouth, halitosis, wiry and slippery pulse, etc.);

[0046] 4) Blood biochemistry data interface, receiving copper metabolism indicators such as serum ceruloplasmin, non-ceruloplasmin-bound copper, and 24-hour urinary copper.

[0047] 2. Integrated Traditional Chinese and Western Medicine Digital Twin Prediction Layer

[0048] The core is a graph-series fusion network. The human meridians and acupoints are constructed as a graph structure, where nodes represent acupoints and edges represent meridian connections and synergistic relationships.

[0049] For each rehabilitation program parameter vector X = [electroacupuncture point combination and intensity, massage technique force / frequency, exoskeleton auxiliary torque, training duration, traditional Chinese medicine fumigation temperature, ...];

[0050] The network captures the topological effects of acupoint stimulation through graph convolutional layers, and then fuses the motion temporal signals with the four diagnostic features through a Transformer encoder to predict four optimization target indicators: maximizing the improvement in motor function, maximizing the improvement in quality of life, minimizing the fluctuation of copper metabolism, and minimizing the deviation of TCM syndrome.

[0051] 3. Online Inference Process of Implicit Structure Model

[0052] Before making optimization decisions, a pre-trained hidden structure model is needed to assess the patient's current symptom state in real time. The specific process is as follows:

[0053] The 64 real-time collected symptoms from the four diagnostic methods (e.g., dark red tongue = 1 / 0, yellow and greasy tongue coating = 1 / 0, wiry and slippery pulse = 1 / 0, etc.) are input into the trained LTM-EAST hidden tree model. This LTM-EAST hidden tree model constructs a multi-layer hidden tree with latent variables as nodes and symptoms as leaf nodes.

[0054] For each patient sample, the posterior probability distribution of each latent variable is calculated using the EM algorithm or Bayesian posterior inference: ;

[0055] : No. Latent variables (such as "tendency of phlegm and blood stasis" and "tendency of internal damp-heat");

[0056] 64 binarized four diagnostic symptoms.

[0057] Then, based on the identification criteria thresholds derived from the previous comprehensive clustering (e.g., for phlegm-stasis syndrome: pale tongue ≥20.3, thin and greasy tongue coating ≥20.3, wiry and slippery pulse ≥21, etc.), the syndrome type of the current patient is determined. Simultaneously, combining the mutual information or symptom association weights between syndrome nodes in the hidden tree structure, the target syndrome state for the next stage of treatment (i.e., the next node in the syndrome transmission path) is determined. The optimal rehabilitation parameter vector corresponding to this target syndrome is denoted as... .

[0058] 4. Adaptive Mutation Multi-Objective Optimization Decision Layer

[0059] A decomposition-based multi-objective evolutionary algorithm framework (MOEA / D) is employed to aggregate the four objectives—improved exercise, enhanced quality of life, minimized copper metabolism fluctuations, and minimized syndrome deviation—into multiple scalar quantum problems. Each sub-problem maintains an individual, namely, a rehabilitation plan vector. .

[0060] The core of this invention lies in the design of two variant modes that are linked to the pathological mechanism of Wilson's disease.

[0061] (1) An example of an implementation of method one: operator one;

[0062] Let the current individual be (No. The generation (A vector of rehabilitation plans). The target syndrome state obtained through online inference using the hidden structure model, and its corresponding optimal rehabilitation parameter vector, are denoted as... The syndrome topology graph uses the latent variables and latent classes output by the LTM-EAST model as nodes and symptom relevance as edges, constructing a weighted directed graph. Edge weights... This represents the comprehensive cost of transitioning from syndrome node u to node v, determined by the degree of symptom overlap, the degree of pathogenesis consistency, and the difficulty of clinical treatment.

[0063] In some embodiments, it is assumed that the syndrome topology graph contains three nodes: A (phlegm-blood stasis syndrome), B (liver qi stagnation syndrome), and C (spleen deficiency with dampness syndrome). The edge weights are: The cost of the transformation from phlegm and blood stasis to liver qi stagnation; The cost of the transformation from liver qi stagnation to spleen deficiency and dampness accumulation; The high cost of phlegm and blood stasis directly transforming into spleen deficiency and dampness is inconsistent with the clinical progression pattern.

[0064] The patient's current condition is accurately identified as syndrome node A using the hidden structure model. Based on the sequential progression pattern of Wilson's disease syndrome in Traditional Chinese Medicine, node B is determined as the optimal target syndrome. Potential function. Defined as: the cumulative pathogenesis intervention cost of the shortest syndrome transmission path from the current rehabilitation program's mapped syndrome node to the intervention target syndrome node. In this example, the shortest transmission path is from A to B, therefore the potential energy value is: .

[0065] For the potential energy gradient, in actual calculations, The potential energy gradient is approximated using the finite difference method. This involves applying small perturbations to each dimension of the rehabilitation plan vector and calculating the rate of change of the potential energy function. The potential energy function is defined as the cumulative cost of the shortest path from the current plan's mapped symptom node to the target symptom node; the direction is the path with the fastest potential energy decrease. The numerical approximation can be obtained by calculating the finite difference of the potential energy function with respect to each scheme dimension, and its direction points to the path where the potential energy decreases the fastest (in this case, the direction from A to B).

[0066] Copper metabolism corrector For a diagonal matrix, the first... Diagonal coefficients It is negatively correlated with 24-hour urinary copper and non-ceruloplasmin-bound copper concentrations; when urinary copper levels are elevated... Normal copper removal period .

[0067] For example, when 24-hour urinary copper exceeds 800 μg / 24h (the upper limit of safety), k d Take 0.5 to reduce the search step size; during the normal copper removal period, k d If we set it to 1.0, the mutation operation will generate the donor vector according to the following formula. ;

[0068] ;

[0069] in,

[0070] For the first The generation A vector of rehabilitation plans;

[0071] This is an adaptive scaling factor; it is dynamically updated according to the success rate of population iteration, and its value range is [0.2, 0.8].

[0072] Correct the diagonal matrix for safe copper metabolism;

[0073] The potential energy gradient;

[0074] The optimal rehabilitation parameter vector for the target syndrome;

[0075] It is a very small constant to prevent the denominator from being 0;

[0076] This is the Gaussian perturbation coefficient; used to slightly increase the diversity of the solution search, with a fixed value of 0.01.

[0077] This is Gaussian noise related to serum free copper concentration. It can be taken as a linear function of serum free copper concentration, for example, Unit: μmol / L, specific coefficients are clinically calibrated.

[0078] The above formula is a specific mathematical expression of "Method 1" (referred to as Operator 1). It should be understood that any variation method that adjusts the search direction and step size based on the symptom topological potential gradient and using the copper metabolism correction factor falls within the protection scope of Method 1.

[0079] (2) An example of an implementation of method two: operator two;

[0080] Dystonia Neuroblastic Index By analyzing high-density electromyography signals, the following calculations were performed: a short-time Fourier transform was performed on the signal to extract the average power in the 20-50 Hz frequency band, and the cross-correlation coefficient of the power envelope of the antagonistic muscle groups was calculated. The product of the two factors constitutes the neuroblastic index of dystonia.

[0081] Warning threshold It is manually set by the physician based on the patient's baseline copper metabolism status and medical history.

[0082] State determination rule 1: Safe steady state; when When this condition is met, it indicates that the current neurophysiological state is within a safe range and does not trigger escape mutation. The algorithm then performs optimization search according to the conventional strategy selection mechanism.

[0083] State determination rule 2: Risk stress state; when This indicates an increased risk of worsening dystonia, triggering escape mutation mechanisms. (Safe treatment direction unit vector) The parameters of the preset low-risk program can be changed in different directions, such as increasing the stimulation intensity of acupoints that dredge meridians (e.g., Hegu and Taichong), reducing the resistance training load, and increasing the proportion of traditional Chinese medicine sedation fumigation.

[0084] The escape mutation formula is:

[0085] ;

[0086] in, The donor vector generated for escape mutation (the first) The generation (individuals)

[0087] This is the vector of the current iterative rehabilitation plan;

[0088] The escape intensity coefficient is the maximum escape correction step size of the control algorithm; it is calibrated according to the upper and lower limits of the rehabilitation equipment parameters, and its value range is [0.1, 0.5].

[0089] The neurophysiological state corresponding to the current rehabilitation program is the neuroburst index.

[0090] Preset early warning thresholds for the risk of dystonia attacks;

[0091] A unit vector for pre-defined safe treatment directions in clinical settings.

[0092] The quantitative analysis table of the above formula behavior is shown in Table 1:

[0093] Table 1

[0094]

[0095] The core feature of the above formula is:

[0096] 1) Based on the truncation function constraint, strictly satisfy The escape term is set to zero, so no escape mutation is triggered. The zero-interference algorithm performs conventional optimization iteration; it has a threshold-triggered unidirectional triggering characteristic.

[0097] 2) Satisfy After the trigger condition is met, the escape correction strength varies with the risk difference. Non-linear, smooth, and incremental; the higher the risk of a neural burst, the greater the risk mitigation adjustment in the rehabilitation plan, which aligns with the logic of clinical graded risk control; it also exhibits a positive risk-correlation response characteristic.

[0098] 3) The escape correction range under all operating conditions is expressed as a coefficient. The upper limit threshold is used to avoid excessive algorithm correction and prevent rehabilitation plans from deviating significantly from the clinically effective treatment range; it also has escape amplitude saturation constraint characteristics.

[0099] During the algorithm iteration process, before each generation of subproblem individuals undergoes mutation operation, the corresponding physiological indicators of the individual are solved first. :

[0100] If the judgment : Forcefully invoke the operator two escape mutation formula to complete individual mutation;

[0101] If the judgment Based on the historical non-dominated solution production success rate of two types of strategies, namely Operator 1 and Standard Deviation Mutation, the mutation operator is adaptively and optimally called.

[0102] After generating new individuals through mutation, parameter boundary constraints and population crossover operations are performed. The effectiveness of the new individual's treatment plan is evaluated using a rehabilitation digital twin model. The population neighborhood solution set and external archive solution set are updated based on the Pareto dominance criterion, and the adaptive selection probabilities of the three types of mutation operators are updated simultaneously. The algorithm iterates until the termination condition is met, outputting the optimal Pareto front solution set. Finally, physicians, considering clinical treatment preferences, select and implement rehabilitation treatment plans.

[0103] 5. Rehabilitation training implementation level

[0104] The final solution is parsed into equipment instructions, which drive the electroacupuncture device to output pulses of a specific frequency and intensity, control the massage robot to apply force along the path, adjust the resistance of the exoskeleton joint trainer, regulate the temperature and duration of the traditional Chinese medicine fumigation, and simultaneously update the virtual reality training scene to form a closed-loop integrated traditional Chinese and Western medicine rehabilitation training.

[0105] In the above embodiments, operator one and operator two are merely preferred mathematical expressions of the present invention; operator one and operator two given in the above specific embodiments are disclosed as preferred specific implementations of method one and method two, respectively, and numerical examples are provided to illustrate the calculation process of the symptom topological potential energy. Those skilled in the art, based on the teachings of the present invention, are fully capable of using other equivalent mathematical calculation methods to achieve the same technical effect. Any substitution of equivalent calculation methods by those skilled in the art based on the ideas of the present invention falls within the protection scope of the present invention.

[0106] The above are merely preferred embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A rehabilitation training system integrating traditional Chinese and Western medicine for Wilson's disease dystonia, comprising a multimodal data acquisition module, a digital twin prediction module combining traditional Chinese and Western medicine, an adaptive variable multi-objective optimization decision-making module, and a rehabilitation training execution module, characterized in that: The multimodal data acquisition module collects patients' Western medicine motor function data, objective data from the four diagnostic methods of traditional Chinese medicine, and copper metabolism biochemical indicators in real time. The adaptive variability multi-objective optimization decision module uses the rehabilitation plan parameter vector as the optimization variable. With the efficacy prediction support provided by the integrated traditional Chinese and Western medicine digital twin prediction module, it adopts an improved differential evolution algorithm to perform multi-objective optimization and generate personalized rehabilitation plans. The improved differential evolution algorithm employs at least one of the following two mutation methods in its mutation operation: Method 1: Based on the potential energy of the hidden structure syndrome topology, the target syndrome state is determined by the pre-trained hidden structure model according to the patient's real-time four diagnostic data. The potential energy gradient of the current rehabilitation plan vector in the pre-constructed syndrome topology map is calculated. The potential energy gradient vector is scaled according to the copper metabolism correction factor and then applied to the current rehabilitation plan vector to generate the variant plan. Method 2: Based on the escape variation of the nerve burst warning in dystonia, the nerve burst index in the surface electromyography signal is calculated in real time; when the nerve burst index is lower than or equal to the set warning threshold, escape is not triggered; when the nerve burst index exceeds the set warning threshold, the escape intensity is dynamically increased according to the degree of exceedance, and the current rehabilitation plan vector is perturbed towards the preset safe treatment direction vector. The higher the degree of danger, the greater the escape amplitude, so as to generate a variation scheme to escape to the safe domain.

2. The system according to claim 1, characterized in that, The multimodal data acquisition module includes: A high-density surface electromyography (EMG) acquisition array is used to extract the synergistic contraction rate of antagonistic muscle groups and the pathological synchronous discharge index. Inertial motion sensors and depth vision cameras are used for joint kinematics assessment. The digital tongue diagnosis instrument and pulse diagnosis sensor acquire quantitative features of tongue and pulse images, and run a hidden structure model to output the probability of syndromes. The blood biochemistry test data interface receives serum ceruloplasmin, non-ceruloplasmin-bound copper, and 24-hour urinary copper levels.

3. The system according to claim 1, characterized in that, The syndrome topology is constructed based on a hidden tree structure model, where nodes represent different syndrome states and edge weights reflect the symptom correlation and transformation difficulty between syndromes; the potential energy function is the cumulative pathogenesis intervention cost of the shortest syndrome transmission path from the current syndrome node to the target syndrome node; the potential energy gradient direction is the steepest descent direction of the potential energy function.

4. The system according to claim 1, characterized in that, The copper metabolism correction factor is dynamically determined based on the patient's real-time 24-hour urinary copper excretion rate and serum non-ceruloplasmin-bound copper concentration. When the copper metabolism index exceeds the safe range, the copper metabolism correction factor decreases, thereby automatically reducing the mutation search step size.

5. The system according to claim 1, characterized in that, The neural burst index is obtained by performing time-frequency analysis on high-density surface electromyography signals, extracting the synchronous discharge energy of antagonistic muscle groups in a preset frequency band, and calculating it in conjunction with the correlation of the synergistic contraction envelope. An increase in the index reflects an increased risk of worsening dystonia; the escape variation is not triggered when the neuroburst index is below or equal to the warning threshold, but only when it exceeds the warning threshold, and the magnitude of the escape is positively correlated with the degree of exceeding.

6. The system according to claim 1, characterized in that, The adaptive mutation multi-objective optimization decision module is configured as follows: When the neural burst index is greater than the warning threshold, escape mutation method 2 is forcibly selected; otherwise, based on the probability of successfully generating non-dominated solutions in the history of each mutation method, method 1 or standard deviation mutation method is adaptively selected.

7. The system according to claim 1, characterized in that, The digital twin prediction module incorporates a graph-time fusion network, using acupoints on the human body's meridians as graph nodes. It captures the topological effects of acupoint stimulation through graph convolution, and combines it with a Transformer encoder to fuse temporal motion signals, outputting multi-objective prediction values, including the amount of improvement in motor function, the amount of improvement in quality of life, the amplitude of copper metabolism fluctuations, and the deviation of TCM syndromes.

8. A rehabilitation training method combining traditional Chinese and Western medicine for Wilson's disease dystonia, applied to the system described in any one of claims 1 to 7, characterized in that, Includes the following steps: S1. Acquire real-time Western medicine motor function data, objective data of the four diagnostic methods of traditional Chinese medicine, and copper metabolism indicators of patients through the multimodal acquisition module; S2. Input the candidate rehabilitation program vectors into a pre-trained graph temporal fusion network to predict multidimensional therapeutic effects; S3. In the adaptive variation multi-objective optimization decision module, with the goal of maximizing exercise improvement and quality of life, and minimizing copper metabolism fluctuations and syndrome deviations, the improved differential evolution algorithm is executed to iteratively search for the optimal rehabilitation plan. The improved differential evolution algorithm performs at least one of the following steps during the mutation operation: Based on the target syndrome state determined by the implicit structure model, the potential gradient of the current rehabilitation plan vector in the syndrome topology is calculated, and the current plan is perturbed after adjusting the gradient using the copper metabolism correction factor. The neural burst index is calculated based on surface electromyography signals. When the index exceeds the warning threshold, the current rehabilitation plan vector is adjusted towards safe treatment, generating an escape variant plan.

9. The method according to claim 8, characterized in that, The improved differential evolution algorithm described in step S3 calculates the neural burst index in real time when generating mutated individuals. If the index exceeds the threshold, escape mutation is performed; otherwise, either method one or standard differential mutation strategy is dynamically selected.

10. The method according to claim 8, characterized in that, The hidden structure model is pre-trained based on multicenter diagnostic data of patients with Wilson's disease. The resulting syndrome identification criteria and topology are used to determine the target syndrome state and calculate the potential gradient online. During online inference, 64 real-time collected symptoms from the four diagnostic methods are input into the hidden tree model. Bayesian posterior inference is used to calculate the probability distribution of each hidden variable, and the current syndrome type and target syndrome are determined according to the identification criteria threshold.

Images