Hand function rehabilitation closed-loop training system based on multi-dimensional evaluation model

Abstract

The invention discloses a hand function rehabilitation closed-loop training system based on a multi-dimensional evaluation model, and relates to the technical field of hand rehabilitation, a data set acquisition module collects key physiological parameters such as a joint pressure coupling coefficient, and quantifies the biomechanical relevance between joint movement and pressure distribution; the LSTM neural network synchronously generates an action completion degree score and a joint pressure coupling coefficient prediction value, and dual functions of rehabilitation quality evaluation and abnormity early warning are realized; a confidence index is generated based on a feature consistency factor of a score fluctuation standard deviation and a prediction-actual measurement deviation, and the credibility of an evaluation result is accurately identified; switching response modes in real time according to a confidence level grading interval, and dynamically adjusting training difficulty and auxiliary force parameters; a confidence quantitative model and a rehabilitation strategy engine are deeply coupled to form an'evaluation-verification-regulation 'integrated framework, so that the rehabilitation safety and effectiveness are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of hand rehabilitation, in particular to a hand function rehabilitation closed-loop training system based on a multi-dimensional evaluation model. BACKGROUND

[0002] With the intelligent development trend of neural rehabilitation training systems, more and more hand rehabilitation devices integrate multi-modal sensors to collect physiological data of hands during rehabilitation, and use machine learning algorithms to evaluate the action completion quality of patients and judge the rehabilitation stage. Motion evaluation methods based on deep learning, such as LSTM neural networks, have good performance in time series feature extraction and have been gradually introduced for patient rehabilitation stage classification and training effect scoring.

[0003] However, in the existing evaluation system, the score results are mostly based on static or single-cycle, lacking dynamic modeling mechanism for the stability and reliability of the score data, and being easily disturbed by abnormal fluctuations or single errors, affecting the rationality of the training strategy and the timeliness of the system response.

[0004] The existing rehabilitation system has not formed a judgment mechanism for the confidence of the evaluation results during execution, which is specifically manifested as: lacking quantitative analysis means for continuous fluctuations of action completion degree scores, and not building a feature consistency verification channel for the coupling coefficient of predicted values and actual joint pressure, so that the system is difficult to determine whether the current score result is reliable in time. When there are action execution abnormalities, prediction instability or sensor data deviation, the system often cannot respond accurately, leading to misleading rehabilitation training, lagging auxiliary strategies and even inducing secondary injury risk. Such evaluation model lacking "self-evaluation ability" has become an important bottleneck restricting the closed-loop control performance of the current rehabilitation system. SUMMARY

[0005] In view of the deficiencies in the prior art, the present application provides a hand function rehabilitation closed-loop training system based on a multi-dimensional evaluation model.

[0006] In order to achieve the above purpose, the technical scheme of the present application is as follows:

[0007] The hand function rehabilitation closed-loop training system based on a multi-dimensional evaluation model comprises:

[0008] A data set acquisition module is configured to acquire a physiological parameter data set in a hand rehabilitation training process, wherein the physiological parameter data set at least includes a joint pressure coupling coefficient for quantifying the synchronization of joint bending and corresponding contact surface pressure;

[0009] A neural network processing module is configured to input the physiological parameter data set into a pre-trained LSTM neural network, and output an action completion degree score generated based on a weighted scoring rule and a predicted value of the joint pressure coupling coefficient.

[0010] a confidence calculation module configured to extract a standard deviation of the current motion completion score and scores of consecutive predetermined positioning training cycles, calculate a feature consistency factor of deviation between the actual value and the predicted value of the joint pressure coupling coefficient, and generate a confidence index based on the standard deviation and the feature consistency factor;

[0011] a rehabilitation strategy adjustment module configured to select a response mode according to a numerical interval of the confidence index, and adjust a rehabilitation training difficulty parameter and an auxiliary force parameter according to the selected response mode.

[0012] Compared with the prior art, the present application has the following beneficial effects:

[0013] 1. The confidence index is generated by the fusion calculation of the score fluctuation factor (standard deviation quantifying stability) and the feature consistency factor (deviation between the predicted value and the actual value), which is used to determine whether the current evaluation result has sufficient reliability, effectively reduces the problem of false feedback caused by accidental error scores, and makes the evaluation more robust.

[0014] 2. The system can judge the prediction accuracy of itself by comparing the predicted subsequent key feature values with the sensor measured values, and has a self-diagnosis and adjustment mechanism, which helps to avoid common problems in long-term training process such as score drift and model instability, and adapts to the dynamic changes of ability in different rehabilitation stages. BRIEF DESCRIPTION OF DRAWINGS

[0015] The disclosure of the present application will be described with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of the present application, and the same reference numerals are used to refer to the same components in the drawings. Among them:

[0016] Figure 1 is a system module diagram of the present application. DETAILED DESCRIPTION

[0017] It is easy to understand that, according to the technical scheme of the present application, a person skilled in the art can propose a plurality of structure modes and implementation modes which can be replaced with each other without changing the essential spirit of the present application. Therefore, the following specific embodiments and drawings are only exemplary description of the technical scheme of the present application, and should not be regarded as the whole or as the limitation or restriction of the technical scheme of the present application.

[0018] SUMMARY

[0019] In the traditional existing hand rehabilitation training system, the physiological parameter data collected by the multi-modal sensor is processed by the LSTM neural network to generate the action completion score. However, the scoring mechanism relies on the static output of a single training cycle, and a standard deviation quantification model for continuous scoring fluctuations has not been established, resulting in the system's inability to identify abnormal fluctuations during action execution. At the same time, the deviation between the joint pressure coupling coefficient predicted by the neural network and the actual measured value lacks a feature consistency factor calculation channel, making the model prediction error unable to be dynamically evaluated, and the system is difficult to judge the credibility of the current score. This defect directly leads to the lack of confidence basis for the response mode selection of the rehabilitation strategy adjustment module, and the update of the training difficulty parameter and the auxiliary force parameter lags behind the actual rehabilitation state change.

[0020] For example, in a rehabilitation device containing a pressure sensor array and a joint angle sensor, the LSTM network generates a score based on single-cycle joint bending time series data, but does not extract the standard deviation of the score for five consecutive training cycles, resulting in a sudden drop in a single score caused by temporary fatigue of the patient being misjudged as overall ability degradation, triggering unreasonable difficulty parameter reduction. At the same time, the data drift caused by poor contact of the pressure sensor is not detected by the feature consistency factor, and the deviation rate of the predicted joint pressure coupling coefficient from the actual measured value by the optical capture system reaches 25%, but the system still maintains the original auxiliary force parameter based on the wrong prediction value, causing compensatory overload of the joint.

[0021] If the above problems are not solved, abnormal scoring fluctuations will lead to frequent misadjustment of rehabilitation strategies, causing mismatch between training intensity and actual recovery stage of the patient, prolonging the rehabilitation period. The continuous deviation between the predicted value and the actual value will accumulate model errors, reducing the accuracy of joint pressure coupling analysis, causing the auxiliary force parameter setting to deviate from the safety threshold. Ultimately, the system lacks a confidence judgment mechanism and cannot trigger a redundant verification process when the sensor data is abnormal or the model is unstable, resulting in the risk of action execution misdirection in the rehabilitation training process, affecting the effect of neuromuscular function reconstruction.

[0022] In the face of the above problems, the present application first analyzes the response lag problem caused by the lack of dynamic credibility evaluation mechanism in the existing system, and finds that the core contradiction lies in the double lack of scoring fluctuation identification and prediction deviation detection. For the problem of continuous scoring abnormal fluctuations, the present application attempts to build a standard deviation calculation model based on time series, and quantifies the stability by extracting the dispersion degree of scores for multiple training cycles, but finds that a single standard deviation index cannot distinguish between patient ability degradation and temporary fatigue scenarios. In view of this, the present application further introduces the prediction value and actual value deviation analysis of the joint pressure coupling coefficient, and establishes a model prediction reliability evaluation channel through the feature consistency factor, but how to effectively integrate the two still exists technical obstacles. After multiple scheme verification, the present application finally selects to weight and integrate the scoring fluctuation factor and the feature consistency factor to form a comprehensive confidence index, providing a quantitative basis for subsequent strategy adjustment.

[0023] To this end, as Figure 1 shown, the present application proposes a hand function rehabilitation closed-loop training system based on a multi-dimensional evaluation model, comprising:

[0024] A data set acquisition module is configured to acquire a physiological parameter data set in a hand rehabilitation training process, wherein the physiological parameter data set at least includes a joint pressure coupling coefficient for quantifying the synchronization of joint bending and corresponding contact surface pressure;

[0025] The physiological parameter data set refers to a multi-dimensional data set generated during the hand rehabilitation training process, which can be collected by a sensor, and can be achieved by synchronous collection of a pressure sensor, a joint angle sensor, and an electromyography sensor, for quantifying the details of hand movement and the mechanical state of the patient during training. The joint pressure coupling coefficient is a quantitative index for representing the synchronization relationship between the change of joint bending angle and the change of corresponding contact surface pressure, which can be achieved by calculating the ratio of pressure change rate to joint angle change rate, and reflects the coordination of hand movement.

[0026] Specifically, the data set acquisition module comprises:

[0027] A data acquisition unit is configured to acquire hand pressure distribution data, joint bending time series data, and electromyography signals. The data acquisition unit acquires the pressure distribution time series data of each region of the hand through a pressure sensor array, acquires the change sequence of the metacarpophalangeal joint bending angle through a joint angle sensor, and acquires the forearm muscle group electrical signal through a surface electromyography electrode.

[0028] A pressure bending data processing unit is configured to time-synchronize the hand pressure distribution data and the joint bending time series data, calculate the joint pressure coupling coefficient, and calculate the stability index according to the sequence of the joint pressure coupling coefficient within a sliding time window. The pressure bending data processing unit uses a time stamp alignment algorithm to synchronize the pressure data and the joint angle data at a millisecond level of precision, eliminating the phase offset caused by the difference in sensor sampling frequency.

[0029] An electromyography data processing unit is configured to calculate a neuromuscular coordination index based on the integral value of the electromyography signal and the pressure work of the hand pressure distribution data. The electromyography data processing unit performs band-pass filtering on the original electromyography signal, and then performs signal integration operation using a moving time window.

[0030] The ratio of the pressure change rate to the joint angle change rate is calculated as the joint pressure coupling coefficient after the pressure distribution data and the joint bending data are time-synchronized, reflecting the coordination degree of the mechanical and kinematic parameters in action execution. The stability index is generated by weighted fusion of the standard deviation of the coupling coefficient, the short-time slope change rate, and the peak fluctuation amplitude in the sliding window, for quantifying the stability level in the action repetition process. The ratio of the integral value of the electromyographic signal to the pressure work is normalized to form the neuromuscular coordination index, representing the matching degree of muscle activation efficiency and mechanical output. Through the synchronous calculation of multi-dimensional parameters, the mechanical coordination, stability, and neural control efficiency of the hand rehabilitation action are comprehensively evaluated.

[0031] The neural network processing module is configured to input the joint pressure coupling coefficient, the stability index, and the neuromuscular coordination index into a pre-trained LSTM neural network, and output an action completion degree score generated based on a weighted scoring rule and a predicted value of the joint pressure coupling coefficient.

[0032] The pre-trained LSTM neural network refers to a long short-term memory neural network model trained by historical rehabilitation data, which can be implemented by using a multi-hidden layer structure combined with a time reversal propagation algorithm, for extracting time-dependent features in physiological parameters and generating a predicted value.

[0033] The confidence calculation module is configured to extract the standard deviation of the current action completion degree score and the scores of the consecutive predetermined positioning times of training cycles, and calculate a feature consistency factor of the deviation between the actual value and the predicted value of the joint pressure coupling coefficient, and generate a confidence index based on the standard deviation and the feature consistency factor.

[0034] The action completion degree score refers to a quantitative evaluation result of the current action execution quality of the patient, which can be implemented by weighted fusion of the coordination, stability, and efficiency sub-scores, and provides a basis for adjusting the rehabilitation strategy. The standard deviation refers to a measure of the dispersion degree of the current action completion degree score and the scores of the consecutive predetermined positioning times of training cycles, which can be calculated by using a sample standard deviation formula, for evaluating the volatility and stability of the score result. The feature consistency factor refers to a quantitative index of the deviation degree between the actual measurement value and the predicted value of the joint pressure coupling coefficient, which can be implemented by calculating the normalized value of the difference between the two, reflecting the prediction reliability of the model. The confidence index refers to a comprehensive credibility evaluation index of the score volatility and the prediction consistency, which can be implemented by weighted fusion of the standard deviation and the feature consistency factor, for dynamically judging the credibility of the current evaluation result. The response mode refers to a set of different system response strategies divided according to the confidence index, which can be implemented by presetting different parameter adjustment rules corresponding to different threshold intervals, to ensure that the system takes adaptive control measures at different confidence levels.

[0035] The rehabilitation strategy adjustment module selects a response mode according to the numerical interval of the confidence index, and adjusts the rehabilitation training difficulty parameter and the auxiliary force parameter according to the selected response mode.

[0036] The rehabilitation training difficulty parameter is a controllable variable for controlling the complexity of the training task, which can be specifically realized by setting the joint range of motion threshold or resistance level of the target action, and is used to match the current ability level of the patient. The auxiliary force parameter is a controllable variable for controlling the auxiliary force applied to the hand of the patient by the rehabilitation device, which can be specifically realized by adjusting the output intensity of the motor torque or the air pressure driving system, and is used to provide mechanical support suitable for the current rehabilitation stage.

[0037] The core innovation of the present application lies in constructing a dynamic confidence evaluation mechanism in the closed-loop training system, realizing quantitative judgment of the reliability of the rehabilitation evaluation result by fusing the dual verification method of continuous cycle score fluctuation analysis and prediction-measured consistency verification, and triggering a graded response strategy based on the confidence index, thereby solving the regulation lag problem caused by the lack of dynamic monitoring of evaluation reliability in the existing system.

[0038] As a preferred embodiment, the scheme of the present application is implemented as follows:

[0039] The data set acquisition module uses a pressure sensor array and a joint angle sensor to collect hand pressure distribution and joint bending data. The neural network processing module uses a three-layer LSTM network structure with 128 hidden layer neurons. The confidence calculation module extracts the score standard deviation of the continuous 5 training cycles and calculates the feature consistency factor. The rehabilitation strategy adjustment module sets three confidence intervals: high, medium and low, which correspond to three response modes: maintenance, fine-tuning and reset.

[0040] In actual application, the patient performs finger flexion and extension training. The data set acquisition module collects pressure and angle data in real time. The neural network processing module outputs the score and the predicted value once per second. The confidence calculation module calculates the confidence index once every 5 seconds. The rehabilitation strategy adjustment module selects the response mode according to the confidence index, and adjusts the training difficulty and the size of the auxiliary force. For example, when the confidence index is in the low interval, the system triggers the reset mode, suspends the training and reacquires the data.

[0041] Through the above scheme, the present application realizes dynamic quantification of the reliability of the evaluation result. The system can timely identify abnormal fluctuations in the score and prediction deviation, and avoid unreasonable strategy adjustment caused by single error. At the same time, the response mechanism based on the confidence index improves the sensitivity of the system to sensor failure and model instability, and reduces the risk of misdirection and overload in rehabilitation training. This closed-loop control method with “self-evaluation ability” effectively improves the accuracy and safety of hand function rehabilitation training.

[0042] The application further proposes generating a confidence index based on the standard deviation and the feature consistency factor, which includes:

[0043] Extracting the standard deviation of the current action completion score and the consecutive predetermined position training cycle score, and calculating a score fluctuation factor for quantifying the stability of the training effect;

[0044] Obtaining a feature consistency factor, and determining whether the feature consistency factor is lower than a feature consistency threshold set for the current rehabilitation stage; the feature consistency threshold is dynamically adjusted according to the rehabilitation stage.

[0045] If the determination result is yes, a redundant sensor data verification process is started; the redundant sensor data verification process realizes cross-validation of data by introducing an optical motion capture system.

[0046] If the determination result is no, the current prediction configuration is maintained;

[0047] Based on the score fluctuation factor and the feature consistency factor, a confidence index is obtained by fusion calculation through a weighting coefficient. The weighting coefficient is dynamically configured according to the classification result of the rehabilitation stage, and the feature consistency factor is given a higher weight in the early stage of rehabilitation, and the weight proportion of the score fluctuation factor is increased in the later stage.

[0048] In the confidence calculation process, first, the score sequence of consecutive multiple training cycles is extracted, and the standard deviation thereof is calculated as a score fluctuation factor. At the same time, the actual value and the predicted value of the joint pressure coupling coefficient are difference calculated, and the feature consistency factor is obtained by sliding window statistics. When the feature consistency factor is lower than the threshold value of the current stage, the redundant sensor data verification process is started. The score fluctuation factor and the feature consistency factor are multiplied by the weighting coefficient corresponding to the stage, and then superimposed to generate a comprehensive confidence index. The index is used for response mode selection of the subsequent rehabilitation strategy adjustment module, to ensure that data verification is preferentially performed when data is abnormal, and to avoid the risk of misjudgment due to failure of a single sensor.

[0049] Through the above technical solutions, the application realizes dynamic judgment of the reliability of the evaluation result. Therefore, the system can timely identify score fluctuations and prediction deviations, and improve the accuracy and safety of rehabilitation training. Further, by introducing a redundant sensor data verification process, the robustness of the system to abnormal situations is enhanced, and misjudgment due to failure of a single data source is avoided. This confidence evaluation mechanism based on multiple indicators effectively improves the closed-loop regulation ability of the rehabilitation system, and provides more accurate and personalized rehabilitation training programs for patients.

[0050] The application further proposes starting a redundant sensor data verification process, which includes:

[0051] Synchronously acquire the joint angle true value, which is acquired by an optical motion capture system; the joint angle true value of the optical motion capture system as an independent data source is used to verify the accuracy of the joint bending time series data collected by the pressure sensor.

[0052] Calculate the deviation rate of the joint bending time series data and the joint angle true value; the deviation threshold is dynamically set according to the rehabilitation stage.

[0053] When the deviation rate exceeds the preset deviation threshold, trigger the data reacquisition process to exclude transient interference, reacquire the joint bending time series data, and recalculate the joint pressure coupling coefficient.

[0054] If the feature consistency factor is lower than the feature consistency threshold set for the current rehabilitation stage for three consecutive times and the deviation rate is lower than the preset deviation threshold, adjust the output layer weight of the LSTM neural network. If the data reacquisition is not triggered for three consecutive verifications, it is determined that there is a systematic error in the neural network prediction, and the prediction accuracy needs to be improved by adjusting the output layer weight. The weight adjustment process introduces a fine-tuning data set accumulation mechanism to ensure the stability of the parameter correction.

[0055] After synchronously collecting the joint angle data by the optical system and the pressure sensor, the data reliability is judged by calculating the frame difference rate of the time series data. When the difference rate exceeds the preset threshold, the system automatically discards the current pressure sensor data and reacquires it, eliminating transient noise interference of the sensor. If the difference rate does not exceed the limit for three consecutive verifications but the feature consistency factor is still lower than the threshold, it indicates that the output layer weight configuration of the neural network does not match the data distribution of the current rehabilitation stage. At this time, the fine-tuning data set collects the current period input feature vector and the prediction error, and when the sample size reaches the preset threshold, the weight adjustment amount is calculated according to the formula, and the output layer parameters are gradually corrected to adapt to the patient's rehabilitation progress. This process ensures the robustness and adaptability of the evaluation system through the dual mechanisms of physical measurement data verification and model parameter dynamic adjustment.

[0056] As a preferred embodiment, the scheme of the application is implemented as follows:

[0057] Starting the redundant sensor data verification process includes the following steps:

[0058] First, synchronously acquire the joint angle true value. The joint angle true value is acquired by an optical motion capture system. Specifically, an optical motion capture system composed of multiple infrared cameras can be used, and reflective marker points are attached to key positions on the patient's hand. Through multi-angle imaging, the three-dimensional motion trajectory of the hand joint is reconstructed, thereby obtaining high-precision joint angle true value data.

[0059] Secondly, the deviation rate of joint bending time series data and joint angle true value is calculated. The joint bending time series data collected by the flexible sensor is compared with the joint angle true value obtained by the optical system, and the deviation percentage of the two at the corresponding time point is calculated to obtain the deviation rate index.

[0060] Further, when the deviation rate exceeds the preset deviation threshold, the joint bending time series data is reacquired, and the joint pressure coupling coefficient is recalculated. For example, the deviation threshold can be set to 5%, and when the deviation rate exceeds 5%, the reacquisition process is triggered. During reacquisition, the position of the flexible sensor can be adjusted or the sensor can be recalibrated to improve data accuracy.

[0061] Finally, if the feature consistency factor is lower than the feature consistency threshold set for the current rehabilitation stage for three consecutive times and the deviation rate is lower than the preset deviation threshold, the output layer weight of the LSTM neural network is adjusted. Specifically, the feature consistency factor and the deviation rate data of three consecutive training periods can be recorded. When the feature consistency factor is lower than the threshold (such as 0.8) for three consecutive times and the deviation rate is lower than the preset threshold (such as 5%), the adjustment process of the output layer weight of the LSTM neural network is triggered.

[0062] Through the above technical solutions, the application realizes dynamic verification of the reliability of sensor data in the hand function rehabilitation training system. Therefore, the system can timely discover and correct sensor data deviation, and improve the accuracy and stability of the evaluation model. Further, by introducing an optical motion capture system as a redundant verification means, a high-precision calibration reference is provided for the flexible sensor, effectively reducing the influence of sensor drift and nonlinear error on the evaluation results. At the same time, the mechanism of dynamically adjusting the neural network weight based on continuous multiple verification results enables the system to have adaptive learning ability, and can continuously optimize the evaluation model according to the changes in the patient's rehabilitation process. This closed-loop verification and adaptive adjustment method significantly improves the reliability and intelligent level of the hand function rehabilitation training system.

[0063] The application further proposes adjusting the output layer weight of the LSTM neural network, which includes:

[0064] extracting the input feature vector composed of the joint pressure coupling coefficient, the stability index, and the neuromuscular coordination index in the current period LSTM neural network and the prediction error value of the output layer ;

[0065] adding the input feature vector and the prediction error value to the fine-tuning data set, and triggering the weight adjustment when the sample number of the fine-tuning data set reaches the preset sample threshold;

[0066] Based on the fine-tuning dataset, the output layer weight adjustment matrix is calculated according to the formula:

[0067] wherein, is the weight adjustment amount, is a learning rate coefficient determined according to the current rehabilitation stage classification result, is the number of fine-tuning dataset samples;

[0068] The LSTM output layer weight is updated as:

[0069] wherein, is the updated weight, is the original weight.

[0070] After each training cycle, the input feature vector and the prediction error value are stored in the fine-tuning dataset, and when the number of samples accumulates to a preset threshold, the weight adjustment process is triggered. The learning rate coefficient is dynamically adjusted according to the rehabilitation stage, for example, a higher coefficient is used in the early rehabilitation stage to speed up convergence, and a lower coefficient is used in the later stage to improve stability. The weight adjustment amount is determined by the average error gradient of the batch samples, avoiding the noise interference of a single sample.

[0071] The update of the new weight is realized by subtracting the weight adjustment amount from the original weight , ensuring that the weight adjustment process conforms to the gradient descent optimization principle. This method effectively suppresses weight mutations by accumulating enough samples before adjustment, while combining stage-based learning rate control to improve the adaptability and stability of the model in rehabilitation training.

[0072] Through the above technical solutions, the present application realizes the dynamic self-adaptive adjustment of the output layer weight of the LSTM neural network. By introducing the fine-tuning dataset and the learning rate adjustment mechanism based on the rehabilitation stage, the adaptability and prediction accuracy of the model to new data are improved. At the same time, the weight adjustment process considers the sample quantity threshold and the maximum adjustment amplitude limit, ensuring the stability and reliability of the model. This adaptive adjustment mechanism can effectively cope with individual differences and stage changes in the patient's rehabilitation process, improving the flexibility and accuracy of the hand function assessment model.

[0073] The present application further includes the following before updating the output layer weight of the LSTM neural network:

[0074] The weight adjustment amplitude index is calculated based on the ratio of the Frobenius norm of to ​ :

[0075] ; determine whether it is greater than the maximum weight adjustment ratio allowed in the current rehabilitation stage ;

[0076] If the result is yes, discard the current fine-tuning dataset, and maintain the original weight .

[0077] After calculating the weight adjustment amount, first calculate the matrix norm corresponding to the adjustment amount, and calculate the proportion with the norm of the original weight matrix. The proportion value is used to quantify the relative amplitude of the weight change. If the proportion exceeds the preset threshold, it indicates that there is a significant difference between the error distribution of the current fine-tuning dataset and the original training data, and forced updating may cause model overfitting or prediction deviation. At this time, the system will discard the current fine-tuning dataset to avoid introducing noise data to interfere with the model. At the same time, maintaining the original weight can ensure that the model maintains stable prediction performance during rehabilitation training, preventing motion score or coupling coefficient prediction from being inaccurate due to weight mutation. This mechanism balances model adaptability and stability by dynamically constraining the weight update amplitude, ensuring the reliability of rehabilitation strategy adjustment.

[0078] Through the above technical solutions, the application can effectively control the adjustment amplitude of the output layer weight of the LSTM neural network, and prevent the model from losing stability due to a single large amplitude adjustment. At the same time, by dynamically setting the maximum weight adjustment ratio, adaptive adjustment for different rehabilitation stages is realized, ensuring the stability and reliability of the model throughout the rehabilitation process. In addition, discarding the fine-tuning dataset that causes excessive adjustment can avoid the negative impact of abnormal data on the model, further improving the robustness of the system.

[0079] The application further proposes a calculation method of the joint pressure coupling coefficient, which includes:

[0080] Taking the rising edge of the electromyographic signal as the time reference point, align the timestamps of the hand pressure distribution data and the joint bending timing data;

[0081] Calculate the pressure change amount , the joint angle change amount in a single action cycle , and calculate the pressure change rate and the joint angle change rate ;

[0082] Take the ratio of the pressure change rate and the joint angle change rate as the joint pressure coupling coefficient:

[0083] .

[0084] The rising edge of the electromyographic signal serves as a time reference point, enabling the capture of the starting moment of muscle activation, thereby eliminating initial delay differences in sensor data collection. After timestamp alignment of the pressure distribution data and joint flexion timing data, the pressure value sequence and joint angle sequence within a single action cycle are extracted through a sliding window, and the difference between their maximum and minimum values is calculated as the pressure change , joint angle change . The pressure change rate is obtained by dividing the pressure change by the action cycle , and the joint angle change rate is obtained by dividing the joint angle change by the action cycle . Finally, the ratio of the two is taken as the joint pressure coupling coefficient , where a ratio greater than 1 indicates that the pressure change rate is higher than the joint flexion rate , and vice versa, indicating that joint flexion dominates the action process. Through the above steps, the synergistic relationship between pressure and joint flexion can be accurately quantified, providing reliable input for subsequent scoring and prediction.

[0085] Through the above technical solution, the present application realizes the accurate quantification of joint pressure and angle change during hand rehabilitation training. Thus, the patient's strength control ability and joint coordination during action execution can be accurately evaluated. This quantitative method considers the timing relationship between pressure change and joint movement, providing more comprehensive and dynamic evaluation indicators. Further, this method introduces electromyographic signals as a time reference to ensure the accuracy of data synchronization, thereby improving the reliability of the evaluation results. This technical solution provides an objective basis for developing personalized rehabilitation strategies and adjusting training difficulty, which helps to improve the pertinence and effectiveness of rehabilitation training.

[0086] The present application further proposes that the stability index is calculated according to the sequence of joint pressure coupling coefficients within a sliding time window, and the specific process is as follows:

[0087] Extract the change curve of joint pressure coupling coefficients within a preset sliding time window, and calculate the standard deviation of coupling coefficients within the time window ; the length of the sliding time window is dynamically adjusted according to the rehabilitation training stage, for example, set to 5 seconds in the initial rehabilitation stage and shortened to 3 seconds in the advanced stage.

[0088] Based on the time series trend changes within the sliding window, extract the short-time slope change rate and the peak fluctuation amplitude The short-time slope change rate is calculated by fitting a linear trend of the data points in the time window by the least square method, and the absolute value of the slope is calculated. The peak fluctuation amplitude is determined by identifying the difference between the maximum value and the minimum value of the coupling coefficient in the window.

[0089] The coupling coefficient standard deviation The short-time slope change rate The peak fluctuation amplitude The stability index is generated by fusing the above three indicators through a predefined weighting function.

[0090] The weighting function is wherein , , are empirical weight coefficients, and satisfy The empirical weight coefficients are pre-set according to the correlation between different indicators and rehabilitation effects in historical data, for example, the standard deviation weight is 0.5, the slope change rate weight is 0.3, and the peak fluctuation amplitude weight is 0.2.

[0091] In the sliding time window, the change curve of the joint pressure coupling coefficient is collected and stored in real time, and the coupling coefficient standard deviation is calculated to quantify the dispersion degree of the data, reflecting the overall fluctuation level of the coupling coefficient. The short-time slope change rate can capture the trend change of the coupling coefficient in the time series, such as rapid rising or falling motion abnormalities. The peak fluctuation amplitude is used to measure the extreme fluctuation range of the coupling coefficient in the window, identifying sudden motion deviations. The above three indicators respectively describe the stability characteristics of the coupling coefficient from three dimensions of statistical distribution, trend dynamics and extreme range, and the multi-dimensional features are mapped into a single stability index by weighted fusion, so that the rehabilitation strategy adjustment module can judge the stability state of the current training motion based on the index, thereby dynamically adjusting the training difficulty parameters. For example, when the stability index is lower than the threshold value, it indicates that the patient's hand motion coordination has decreased, and the system automatically reduces the training intensity to avoid excessive load.

[0092] Through the above technical solutions, the present application effectively solves the technical problem that the existing rehabilitation system cannot dynamically quantify the stability of the evaluation result, and realizes real-time monitoring of the joint pressure coupling fluctuation characteristics. Through the fusion analysis of the multi-dimensional features in the sliding window, the system can accurately identify abnormal fluctuations caused by sensor drift or patient motion instability, avoiding misjudgment caused by a single statistical indicator. The weighted fusion mechanism further improves the robustness of the stability index, providing a reliable basis for subsequent training strategy adjustment, thereby reducing the risk of auxiliary strategy lag caused by distorted evaluation results.

[0093] The application further proposes a calculation process of the neuromuscular coordination index, which comprises:

[0094] The collected electromyographic signals are integrated in a preset time window to obtain an integrated electromyographic value. The integration processing adopts a rectangular method or a trapezoidal method to cumulatively sum the amplitude of the electromyographic signals. The length of the integration time window is set to an integer multiple of the action period to avoid truncation error.

[0095] In the corresponding time window, based on the hand pressure distribution data, the pressure work amount per unit time is calculated in combination with the joint displacement change. The work amount is calculated by numerically integrating the product of pressure and displacement. The calculation of the pressure work amount needs to time-align the spatial distribution data of the pressure sensor array with the joint displacement trajectory. The displacement change is obtained by converting the joint bending angle and the preset joint length parameter. The normalization processing adopts the maximum and minimum value method or Z-score standardization to eliminate the influence of the dimension difference of different sensors on the coordination index.

[0096] The ratio of the integrated electromyographic value to the pressure work amount is taken as the neuromuscular coordination index, and the neuromuscular coordination index is normalized.

[0097] The integrated electromyographic value reflects the total energy consumption of muscle activation in the period. The pressure distribution data and the joint displacement change are synchronously collected. The product sum of pressure and displacement is accumulated in the time dimension to form the pressure work amount, which represents the effective mechanical work exerted by the hand action on the external environment. The ratio of the integrated electromyographic value to the pressure work amount quantifies the neuromuscular resources consumed per unit mechanical work, directly reflecting the action execution efficiency. The normalization processing eliminates the differences in electromyographic signal intensity and pressure sensor range of different patients, making the coordination index have cross-individual and cross-stage horizontal comparability. Through the integration operation in the time window, the influence of abnormal fluctuations of single-frame data on the evaluation result is effectively suppressed, and the robustness of neuromuscular coordination evaluation is improved.

[0098] Through the above technical solutions, the application can effectively quantify the matching degree between neuromuscular control efficiency and mechanical work. By calculating the normalized coordination index in real time, an objective physiological basis is provided for adjusting the difficulty of rehabilitation training. The scheme solves the coordination efficiency judgment deviation problem caused by separate evaluation of electromyographic signals and mechanical parameters in the prior art, so that the system can accurately identify the coupling state between the patient's nerve innervation ability and the movement execution effect, and avoid the risk of misjudgment caused by abnormal single parameter.

[0099] The application further proposes a processing process of the LSTM neural network, which comprises:

[0100] Based on the abstract feature vector output by the hidden layer of the pre-trained LSTM neural network, the abstract feature vector is divided into a coordination feature group, a stability feature group and an efficiency feature group; the coordination feature group is used to capture the correlation features of joint pressure and motion coordination, the stability feature group reflects the fluctuation characteristics in the motion execution process, and the efficiency feature group represents the relationship between energy consumption and motion efficiency.

[0101] The coordination feature group is input into a prediction sub-network to generate a predicted value of the joint pressure coupling coefficient; the prediction sub-network adopts a fully connected layer structure, the input of which is the vector of the coordination feature group, and the output is mapped to the 0-1 interval through a Sigmoid function to generate the joint pressure prediction value.

[0102] The prediction layer of the LSTM neural network calculates the joint pressure prediction value through a Sigmoid activation function;

[0103] The coordination feature group, the stability feature group and the efficiency feature group are respectively input into corresponding scoring sub-modules to obtain dimension-specific score results; the scoring sub-modules include three independent neural network branches, which respectively receive inputs of different feature groups and output dimension-specific scores.

[0104] Based on the classification result of the rehabilitation stage, the weighting parameters of each scoring sub-module are dynamically set, and multiple dimension-specific scores are weighted and fused according to a preset score synthesis rule to generate a target motion completion degree score. The weighting parameters are dynamically adjusted according to the classification result of the rehabilitation stage, for example, in the early rehabilitation stage, the weight of the coordination feature is set to 0.5, the weight of the stability is 0.3, and the weight of the efficiency is 0.2, and in the later stage, the weight of the coordination is reduced to 0.4, and the weight of the stability is increased to 0.4;

[0105] The abstract feature vector output by the hidden layer is first subjected to a sensitivity analysis module to calculate the correlation coefficient between each feature dimension and the score change, and high-sensitivity features are selected as the coordination feature group. The remaining features are matched with historical feature groups through a clustering algorithm, and if a feature has a Euclidean distance less than the efficiency feature group, it is assigned to the stability feature group. The prediction sub-network generates a predicted value based on the coordination feature, and the scoring sub-module fuses the dimension-specific scores according to the dynamic weight. For example, in the middle stage of rehabilitation, when the stability feature group has a large score fluctuation, the system automatically increases the stability weight and reduces the coordination weight, so as to more accurately reflect the actual ability of the patient.

[0106] The application also discloses that the division process of the abstract feature vector is dynamically adjusted based on the sensitivity of each feature dimension to the score result, including:

[0107] For each feature dimension of the hidden layer output, the sensitivity of the feature dimension to the change of the action completion score is evaluated to obtain a feature importance indicator; the feature importance indicator is obtained by calculating the Pearson correlation coefficient of the feature dimension and the score change, and features with an absolute value of the correlation coefficient greater than 0.5 are classified into a coordination feature group; the clustering distribution adopts a K-means algorithm to classify the data distribution of the historical feature group as a clustering center.

[0108] According to the feature importance indicator, feature dimensions higher than a set threshold are classified into the coordination feature group;

[0109] The remaining features are clustered and distributed according to the correlation with the historical stability features or efficiency features, and are classified into the stability feature group or the efficiency feature group, so as to realize adaptive updating of the feature group division strategy. The adaptive updating mechanism of feature division is realized by periodically recalculating the feature importance indicator and the clustering distribution result, so as to ensure that the feature group division is adjusted synchronously with the rehabilitation progress. This process avoids the score deviation caused by fixed feature division, and improves the adaptability of the score result to different rehabilitation stages through dynamic weighting.

[0110] Through the above technical solutions, the present application realizes dynamic optimization and distribution of neural network hidden features, effectively improves the adaptability of the action score model to feature changes in different rehabilitation stages. By evaluating the sensitivity of the feature dimension to the score result in real time, key coordination indicators can be automatically screened, and redundant features can be prevented from interfering with the score stability. Based on the dynamic adjustment of the weighting mechanism in the rehabilitation stage, the scoring system focuses on action specification evaluation in the early recovery stage, and enhances the consideration of movement stability in the functional strengthening stage, thereby improving the accuracy of training strategy adjustment. The adaptive updating mechanism of the feature group can automatically capture the changes of neuromuscular control patterns in the patient's rehabilitation process, and prevent the accumulation of model evaluation deviation caused by fixed feature division.

[0111] The technical scope of the present application is not limited to the contents in the above description, and those skilled in the art can make various modifications and changes to the above embodiments without departing from the technical idea of the present application, and these modifications and changes should all be within the protection scope of the present application.

Claims

1. A hand function rehabilitation closed-loop training system based on a multi-dimensional evaluation model, characterized in that: The method comprises the following steps: a data set acquisition module for acquiring physiological parameter data sets in the hand rehabilitation training process, the physiological parameter data sets at least including joint pressure coupling coefficients for quantifying the synchronization of joint flexion and corresponding contact surface pressure; a neural network processing module for inputting the physiological parameter data sets into a pre-trained LSTM neural network and outputting an action completion degree score generated based on a weighted scoring rule and a predicted value of the joint pressure coupling coefficient; a confidence calculation module for extracting the standard deviation of the current action completion degree score and the scores of consecutive predetermined position training cycles, calculating a feature consistency factor of the deviation between the actual value and the predicted value of the joint pressure coupling coefficient, and generating a confidence index based on the standard deviation and the feature consistency factor; a rehabilitation strategy adjustment module for selecting a response mode according to the numerical interval of the confidence index, and adjusting the rehabilitation training difficulty parameter and the auxiliary force parameter according to the selected response mode.

2. The multi-dimension evaluation model based hand function rehabilitation closed-loop training system according to claim 1, characterized in that: Generating a confidence index based on the standard deviation and the feature consistency factor comprises: extracting the standard deviation of the current action completion degree score and the scores of consecutive predetermined position training cycles, and calculating a score fluctuation factor; acquiring a feature consistency factor and determining whether the feature consistency factor is lower than a feature consistency threshold set for the current rehabilitation stage; if the determination result is yes, starting a redundant sensor data verification process; if the determination result is no, maintaining the current prediction configuration; based on the score fluctuation factor and the feature consistency factor, performing fusion calculation through a weighting coefficient to obtain a confidence index.

3. The multi-dimension evaluation model based hand function rehabilitation closed-loop training system according to claim 2, characterized in that: Starting a redundant sensor data verification process comprises: synchronously acquiring joint angle true values, which are synchronously acquired by an optical motion capture system; calculating the deviation rate of the joint flexion time series data and the joint angle true values; when the deviation rate exceeds a preset deviation threshold, reacquiring joint flexion time series data and recalculating the joint pressure coupling coefficient; if the feature consistency factor is lower than the feature consistency threshold set for the current rehabilitation stage for three consecutive times and the deviation rate is lower than the preset deviation threshold, adjusting the output layer weight of the LSTM neural network.

4. The multi-dimension evaluation model based hand function rehabilitation closed-loop training system according to claim 3, characterized in that: Adjusting the output layer weight of the LSTM neural network comprises: extracting an input feature vector consisting of a joint pressure coupling coefficient, a stability index, and a neuromuscular coordination index in the current cycle LSTM neural network and a prediction error value of the output layer ; the input feature vector and a prediction error value adding a fine-tuning data set, triggering weight adjustment when the number of samples of the fine-tuning data set reaches a preset sample threshold based on a fine-tuning data set, calculating an output layer weight adjustment matrix according to a formula: ; wherein, is a weight adjustment amount, is a learning rate coefficient determined according to a current rehabilitation stage classification result, is a fine-tuning dataset sample number; updating the LSTM output layer weight to: wherein, is the updated weight, is the original weight.

5. The multi-dimension evaluation model based hand function rehabilitation closed-loop training system according to claim 4, characterized in that: The output layer weight of the LSTM neural network before updating further comprises: Based on With Frobenius norm ratio of the weight adjustment range index : ； determining whether the maximum weight adjustment ratio allowed for the current rehabilitation phase ; If the judgment result is yes, the current fine-tuning dataset is discarded, and the original weight is maintained .

6. The multi-dimension evaluation model based hand function rehabilitation closed-loop training system according to claim 1, characterized in that: The data set acquisition module comprises: a data acquisition unit for acquiring hand pressure distribution data, joint flexion time series data and electromyographic signals; a pressure flexion data processing unit for time-synchronizing the hand pressure distribution data and the joint flexion time series data, calculating the joint pressure coupling coefficient and the stability index; an electromyographic data processing unit for calculating the neuromuscular coordination index based on the integral value of the electromyographic signals and the pressure work.

7. The multi-dimension evaluation model based hand function rehabilitation closed-loop training system according to claim 6, characterized in that: The calculation of the joint pressure coupling coefficient comprises: aligning the time stamps of the hand pressure distribution data and the joint flexion time series data with the rising edge of the electromyographic signals as the time reference point; Computing a single action period The amount of pressure change in the interior The amount of joint angle change And computing the rate of pressure change And the rate of joint angle change ; The ratio of the rate of change of pressure to the rate of change of joint angle is taken as the joint pressure coupling coefficient: 。 8. The multi-dimension evaluation model based hand function rehabilitation closed-loop training system according to claim 6, characterized in that: The stability index is calculated according to the sequence of the joint pressure coupling coefficients in the sliding time window, and the specific process is as follows: extract a variation curve of the joint pressure coupling coefficient in a preset sliding time window, and calculate a standard deviation of the coupling coefficient in the time window ; extracting short-time slope change rate based on time series trend change within a sliding window and peak fluctuation amplitude ; standard deviation of coupling coefficients short-term slope change rate peak fluctuation amplitude a stability index is generated by fusing through a predefined weighting function The weighting function is : wherein, , , are empirical weight coefficients, and satisfy .

9. The multi-dimension evaluation model based hand function rehabilitation closed-loop training system according to claim 6, characterized in that: The calculation process of the neuromuscular coordination index comprises: The integral electromyogram value is calculated by integrating the collected electromyogram signal in a preset time window; In the corresponding time window, based on the hand pressure distribution data, combined with the change of joint displacement, the pressure work amount per unit time is calculated, and the work amount is calculated by numerical integration method to calculate the product sum of pressure and displacement; The ratio of the integral electromyogram value and the pressure work amount is taken as the neuromuscular coordination index, and the neuromuscular coordination index is normalized.

10. The multi-dimension evaluation model based hand function rehabilitation closed-loop training system according to claim 1, characterized in that: The processing process of the LSTM neural network is: Based on the abstract feature vector output by the hidden layer of the pre-trained LSTM neural network, the abstract feature vector is divided into coordination feature group, stability feature group and efficiency feature group; The coordination feature group is input into the prediction subnetwork to generate the predicted value of the joint pressure coupling coefficient; The prediction layer of the LSTM neural network calculates the joint pressure prediction value through the Sigmoid activation function; The coordination feature group, the stability feature group and the efficiency feature group are respectively input into the corresponding scoring submodules to obtain the dimension-by-dimension scoring results; Based on the classification results of the rehabilitation stage, the weighting parameters of each scoring submodule are dynamically set, the multiple dimension-by-dimension scores are weighted and fused according to the preset score synthesis rule to generate the target action completion degree score; The division process of the abstract feature vector is based on the sensitivity of each feature dimension to the scoring result and dynamically adjusts, including: For each feature dimension of the hidden layer output, the sensitivity of the feature dimension to the change of the action completion degree score is evaluated to obtain a feature importance index; According to the feature importance index, the feature dimensions higher than the set threshold are divided into the coordination feature group; The remaining features are clustered and allocated according to the correlation with historical stability features or efficiency features, and are divided into the stability feature group or the efficiency feature group, so as to realize adaptive update of the feature group division strategy.

Images