Intelligent planning system for rehabilitation training of joint dysfunction of burn patients

Abstract

The present application relates to the technical field of rehabilitation training planning, in particular to a burn patient joint dysfunction rehabilitation training intelligent planning system, which comprises a multi-dimensional state perception module, a scar tension evaluation module, an intelligent planning generation module, a dynamic self-adaptive adjustment module and a rehabilitation efficiency evaluation module.In the present application, by establishing a state vector oriented to wound recovery and joint state, introducing nonlinear modeling technology to realize accurate evaluation of the constraint relationship between wound tension and range of motion, combining with reinforcement learning algorithm to construct an adaptive training strategy generation mechanism, through optimizing the combination of action type, intensity and frequency to improve the training targeting, and introducing fuzzy control method to realize real-time correction of physiological response and action deviation in the execution process, the training sequence can automatically adjust according to the change of patient state, a prediction and feedback system oriented to training effectiveness progress is built, and closed-loop optimization of the rehabilitation path is realized.

Description

Technical Field

[0001] This invention relates to the field of rehabilitation training planning technology, and in particular to an intelligent planning system for rehabilitation training of burn patients with joint dysfunction. Background Technology

[0002] The field of rehabilitation training planning technology encompasses a comprehensive technical system covering the entire process of assessing and analyzing patients' physiological functional impairments, developing training methods, executing training procedures, and providing feedback on training effects. This field includes core aspects such as patient condition data collection and identification, individual rehabilitation goal setting, parametric modeling of rehabilitation training movements, optimized planning of rehabilitation pathways, generation of control commands for rehabilitation equipment, and evaluation of rehabilitation effects. The overall technical system integrates kinematic parameter analysis, biomechanical modeling, human-computer interaction control, and rehabilitation process monitoring technologies. By implementing quantitative, individualized, and visualized training task planning throughout the rehabilitation cycle, it improves the scientific rigor and standardization of rehabilitation training, and is applicable to various disease scenarios, including neurological rehabilitation, orthopedic rehabilitation, and burn rehabilitation.

[0003] Traditional intelligent planning systems for joint dysfunction rehabilitation training in burn patients are designed to address joint mobility limitations caused by scar contractures and soft tissue adhesions following burns. These systems collect parameters such as the patient's joint range of motion, wound location and area, scar formation period, and age to develop rehabilitation training plans. Typically, these systems use a combination of manual assessment and paper records to determine training goals. Rehabilitation therapists then arrange training content based on their experience, such as passive joint stretching, active range of motion training, and resistance training, and adjust the plan based on periodic observations. Traditional methods rely heavily on the therapist's subjective judgment to set training intensity, frequency, and movement trajectories, lacking data-driven support. The training process relies on verbal instructions or written diagrams, and changes in the patient's posture and joint angles during training are often not recorded by the system, resulting in a lack of objective evidence for training optimization. Summary of the Invention

[0004] To address the technical problems existing in the prior art, embodiments of the present invention provide an intelligent planning system for rehabilitation training of burn patients with joint dysfunction. The technical solution is as follows:

[0005] On the one hand, an intelligent planning system for rehabilitation training of joint dysfunction in burn patients is provided, the system comprising: The multidimensional state perception module collects activity data of burn patients, including joint range of motion, muscle tension, wound healing and vital signs data. After Kalman filtering to remove noise, physiological functions and wound state features are extracted to construct a multidimensional patient state vector. The scar tension assessment module, based on the multidimensional patient state vector, uses a random forest model to analyze the nonlinear mapping between wound traction and joint movement, calculates the limit of movement threshold and pain tolerance boundary, and constructs a set of safety constraint parameters. The intelligent planning and generation module inputs the multidimensional patient state vector and safety constraint parameter set into the deep Q-network model to optimize the action type, intensity, frequency and interval duration with the goal of maximizing the recovery score, and generates an initial rehabilitation training matrix. The dynamic adaptive adjustment module loads the initial rehabilitation training matrix, monitors the actual trajectory and physiological feedback signals, uses fuzzy logic to calculate the trajectory deviation value and pain response level, corrects the resistance coefficient and movement amplitude, and generates an execution training sequence.

[0006] As a further aspect of the present invention, the multidimensional patient state vector includes joint kinematic feature vectors, wound healing state index, and physiological function benchmark parameters; the safety constraint parameter set includes maximum pain-free joint angle, wound tension limit threshold, and safe movement torque boundary; the initial rehabilitation training matrix includes movement type encoding sequence, training intensity distribution weight, and time frequency configuration parameters; and the execution training sequence includes real-time resistance control gain, dynamic amplitude adjustment command, and adaptive movement step length sequence.

[0007] As a further aspect of the present invention, the multi-dimensional state perception module includes: The physiological signal acquisition submodule collects activity data of burn patients, including joint range of motion, muscle tension, wound healing stage data and vital signs. It calls the Kalman filter algorithm to perform state estimation and error covariance update on the acquired multi-source time series data, filters out high-frequency environmental noise and motion artifact interference, and generates a denoised physiological parameter dataset. The data normalization processing submodule acquires the denoised physiological parameter dataset, analyzes the distribution characteristics of vital signs indicators and sets a benchmark range, calculates the offset of joint range of motion and muscle tension relative to the benchmark range, uses the max-min normalization method to map heterogeneous data to a unified numerical range, combines the timestamp information of data acquisition to perform spatiotemporal registration of multi-source data and interpolation to fill missing values, and establishes a standardized multidimensional feature matrix. The feature vector construction submodule calls the standardized multidimensional feature matrix, uses a sliding window to extract local data fragments, calculates the mean, variance and rate of change statistics of each indicator in the window, identifies key mutation points that characterize the trend of physiological function recovery and the evolution of wound healing status, and cascades and combines the extracted statistical features and mutation features according to the preset feature dimension priority order to generate a multidimensional patient state vector.

[0008] As a further aspect of the present invention, the process of performing spatiotemporal registration and missing value interpolation completion of multi-source data by combining the timestamp information of data acquisition specifically involves: extracting the timestamp sequence of each indicator in the denoised physiological parameter dataset; selecting the time axis corresponding to the indicator with the highest sampling frequency as the reference time axis; traversing the timestamps of each other indicator and mapping them to the nearest neighbor time point on the reference time axis to align the multi-source data in the time dimension; for the time points with missing values ​​on the reference time axis after alignment, constructing a local time window, selecting the existing valid data points in the window as reference nodes, and using the cubic spline interpolation function to calculate the values ​​of the missing time points to complete the data. The process of identifying key mutation points that characterize the trend of physiological function recovery and the evolution of wound healing status specifically involves: calculating the first-order difference sequence of adjacent time step values ​​for each dimension feature sequence in the standardized multidimensional feature matrix; applying a moving average filter to the first-order difference sequence to smooth high-frequency random fluctuations, resulting in a smoothed difference sequence; calculating the absolute value of each data point in the smoothed difference sequence and comparing the absolute value with a preset mutation detection threshold; when the absolute values ​​of a continuous number of data points all exceed the mutation detection threshold, it is determined that a state transition has occurred at the beginning of the continuous interval, and the beginning time is marked as a key mutation point.

[0009] As a further aspect of the present invention, the scar tension assessment module includes: The nonlinear mapping construction submodule extracts joint kinematic features and wound healing status index from the multidimensional patient state vector, constructs a regression prediction structure composed of multiple decision trees, inputs the extracted features into the regression prediction structure for parallel deduction, traverses the leaf nodes of the decision tree, quantifies the predicted values ​​of skin surface tension corresponding to various joint angles, aggregates the prediction results of multiple trees and calculates the average tension response curve, and establishes a wound tension distribution mapping table describing the nonlinear relationship between joint position and skin tension response. The critical threshold calculation submodule calls the wound tension distribution mapping table, loads the preset skin tissue fracture strength benchmark and patient pain sensitivity coefficient, retrieves the critical movement position where the tension change rate exceeds the fracture strength benchmark in the mapping table, simultaneously locates the joint angle node when the pain response value reaches the upper limit of tolerance, performs intersection calculation on the critical movement position and the joint angle node, determines the safe operation range, and generates the boundary value of extreme activity and pain tolerance. The constraint parameter generation submodule introduces a preset safety buffer coefficient to shrink and correct the boundary values ​​for the extreme activity and pain tolerance boundary values, reserves tissue elasticity space, converts the corrected boundary values ​​into the maximum allowable deflection angle in the joint coordinate system, calculates the maximum external load that does not trigger pain reflex within the deflection angle range, integrates angle and load limit data, and constructs a set of safety constraint parameters.

[0010] As a further aspect of the present invention, the intelligent planning generation module includes: The state-action space construction submodule constructs a digital state observation vector representing the current rehabilitation environment based on the multidimensional patient state vector. It calls the safety constraint parameter set to perform boundary constraint verification on the preset full set of action space, identifies high-risk actions that exceed the safety angle or torque limit, generates an action mask vector to shield against illegal operations, encapsulates the state observation vector and the action mask vector, and establishes a reinforcement learning input state set. The strategy optimization decision submodule imports the reinforcement learning input state set into a deep Q-network model, calculates the expected value Q value of each candidate rehabilitation action through neural network forward propagation, applies action mask vector to filter the output Q value vector, eliminates dangerous actions, performs value maximization optimization in the remaining effective action space, locks the action node with the highest expected return, and generates the optimal rehabilitation action index code. The parameter matrix generation submodule decodes the optimal rehabilitation action index code, maps and obtains the corresponding specific action modality, resistance intensity level, number of repetitions per set and rest duration between sets, and arranges and combines each parameter according to the temporal logic of rehabilitation training to establish structured data including the dimensions of action type, intensity, frequency and time parameters, and generates the initial rehabilitation training matrix.

[0011] As a further aspect of the present invention, the dynamic adaptive adjustment module includes: The real-time feedback monitoring submodule loads the initial rehabilitation training matrix as a standard reference benchmark, uses a high-frequency position sensor and electromyography acquisition device to acquire the patient's actual movement coordinate sequence and physiological response data in real time, performs spatiotemporal alignment and Euclidean distance calculation between the actual coordinates and the ideal trajectory defined in the matrix, quantifies the degree of spatial deviation in the action execution process, simultaneously extracts the spectral energy characteristics of the electromyography signal and matches it with a preset pain quantification table, and generates a trajectory deviation feature vector and an immediate pain response level. The fuzzy control decision submodule establishes a two-dimensional fuzzy input variable set including error rate and pain intensity based on the trajectory deviation feature vector and the instantaneous pain response level. It calls the preset membership rules to map the input variables to the fuzzy domain, performs fuzzy inference operations according to the expert rule base, determines the trigger weight of the control strategy, uses the centroid method to defuzzify the fuzzy inference results, quantifies the physical adjustment parameters used for real-time intervention of rehabilitation equipment, and generates resistance control correction coefficient and amplitude adjustment gain. The adaptive sequence construction submodule calls the resistance control correction coefficient and amplitude adjustment gain to perform multiplicative weighted adjustment on the preset resistance intensity parameters in the initial rehabilitation training matrix. It uses the adjustment gain to dynamically scale and correct the boundary coordinates of the movement amplitude, limiting the range of motion. Based on the corrected parameters, it updates the control command queue of the underlying actuator in real time, and reassembles and encapsulates the discrete control commands according to the timestamp order to generate an execution training sequence.

[0012] As a further aspect of the present invention, the process of establishing a two-dimensional fuzzy input variable set including error rate and pain intensity specifically involves extracting the Euclidean norm of the trajectory deviation feature vector as a position error scalar, mapping the instantaneous pain response level to a preset normalized numerical range to characterize pain intensity, and configuring five levels of fuzzy linguistic variables including negative large, negative small, zero, positive small, and positive large and corresponding membership rules for the position error scalar and pain intensity respectively. The process of performing fuzzy inference calculations based on the expert rule base specifically involves setting adjustment rules in the expert rule base. When the position error scalar is greater than a preset deviation threshold and the pain intensity is at a high response level, a negative large-scale adjustment rule for the resistance parameter is triggered to generate a fuzzy output set that reduces resistance. When the position error scalar is within the allowable range and the pain intensity is at a low response level, an adjustment rule that maintains or slightly increases resistance is triggered. The process of using the centroid method to defuzzify the fuzzy inference results specifically involves: performing a union aggregation on the fuzzy set of the trigger rule output to obtain the total output fuzzy surface; performing an integral operation along the universe axis of the total output fuzzy surface; calculating the geometric centroid coordinates of the region enclosed by the membership rule curve and the horizontal axis; determining the values ​​corresponding to the geometric centroid coordinates as the precise control quantities; and assigning them to the resistance control correction coefficient and the amplitude adjustment gain, respectively.

[0013] As a further aspect of the present invention, the system further includes: The rehabilitation efficacy assessment module gathers the executed training sequences, combines them with joint function recovery data, inputs them into a long short-term memory network model to analyze the evolution of activity and muscle strength, calculates rehabilitation progress indicators and predicts recovery trends, and generates rehabilitation progress assessment results. The rehabilitation progress assessment results include joint function recovery scores, training compliance metrics, and predicted recovery trends for the next stage.

[0014] As a further aspect of the present invention, the rehabilitation efficacy assessment module includes: The data collection and serialization submodule acquires the execution training sequence over multiple periods, synchronously retrieves joint range of motion and muscle strength detection data within the corresponding time window, performs temporal alignment and synchronous mapping of action commands and physiological response data based on timestamps, removes invalid period data with breakpoints and performs linear interpolation to complete the data, constructs a multi-dimensional sliding window according to a preset time step, converts discrete records into continuous time series samples, and establishes a historical rehabilitation time series dataset. The trend prediction and analysis submodule inputs the historical rehabilitation time series dataset into the long short-term memory network model, uses the forget gate mechanism to remove low-relevance historical state information, updates the cell state through the input gate, captures the nonlinear evolution law of joint range of motion and muscle strength over time, performs regression calculation on the hidden layer features based on the output gate, predicts the peak joint angle and muscle strength index of the next rehabilitation cycle, and generates a predicted value of joint function evolution trend. The efficacy quantification assessment submodule calls the predicted value of joint function evolution trend, dynamically compares it with the preset standard rehabilitation path curve, calculates the Euclidean distance of the actual recovery trajectory relative to the standard path to quantify the deviation of rehabilitation progress, calculates the functional recovery rate by combining the change slope of the trend prediction value, and integrates the progress deviation, recovery rate and predicted functional indicators to construct the rehabilitation progress assessment result.

[0015] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: This invention, based on the dynamic acquisition of joint range of motion, muscle tension, and physiological signals, achieves effective restoration and normalization of data in high-noise environments through a fusion filtering algorithm. It establishes a state vector oriented towards wound recovery and joint status, introduces nonlinear modeling technology to accurately assess the constraint relationship between wound tension and range of motion, and constructs an adaptive training strategy generation mechanism using reinforcement learning algorithms. By optimizing the combination of movement types, intensity, and frequency, it enhances training targeting. Simultaneously, it introduces fuzzy control methods to achieve real-time correction of physiological responses and movement deviations during execution. The training sequence can be automatically adjusted according to changes in the patient's state, establishing a prediction and feedback system guided by training progress, thus achieving closed-loop optimization of the rehabilitation pathway. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the intelligent planning system for rehabilitation training of joint dysfunction in burn patients provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the system framework of the present invention; Figure 3 This is a flowchart of the multi-dimensional state perception module in this invention; Figure 4 This is a flowchart of the scar tension assessment module in this invention; Figure 5 This is a flowchart of the intelligent planning generation module in this invention; Figure 6 This is a flowchart of the dynamic adaptive adjustment module in this invention; Figure 7 This is a flowchart of the rehabilitation efficacy assessment module in this invention. Detailed Implementation

[0018] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0019] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0020] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0021] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0022] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0023] like Figure 1-2 As shown, this embodiment of the invention provides an intelligent planning system for rehabilitation training of joint dysfunction in burn patients. The system includes a multi-dimensional state perception module, a scar tension assessment module, an intelligent planning generation module, a dynamic adaptive adjustment module, and a rehabilitation efficacy assessment module. The multidimensional state perception module collects activity data of burn patients, including joint range of motion, muscle tension, wound healing stage data and vital signs. It calls the Kalman filter algorithm to denoise and normalize the collected data, extracts features that represent physiological functions and wound status, and constructs a multidimensional patient state vector. The scar tension assessment module, based on a multidimensional patient state vector, calls a random forest model to analyze the nonlinear mapping relationship between the degree of wound traction and the range of motion of the joint, calculates the limit of activity threshold and pain tolerance boundary to avoid secondary tearing of the wound, and generates a set of safety constraint parameters including safety angle and torque limits. The intelligent planning and generation module inputs the multidimensional patient state vector and safety constraint parameter set into the deep Q-network model, and optimizes the strategy with the goal of maximizing the joint function recovery score. It combines and optimizes the movement type, intensity, frequency and interval duration to generate the initial rehabilitation training matrix. The dynamic adaptive adjustment module loads the initial rehabilitation training matrix, monitors the actual movement trajectory and physiological feedback signals in real time, calls the fuzzy logic control algorithm to calculate the trajectory deviation value and real-time pain response level, and corrects the movement resistance coefficient and movement amplitude accordingly to generate an execution training sequence that adapts to the immediate state. The rehabilitation efficacy assessment module gathers multi-cycle execution training sequences and joint function recovery data, inputs them into a long short-term memory network model to analyze the evolution of joint range of motion and muscle strength, calculates rehabilitation progress indicators and predicts the functional recovery trend of the next stage, and generates rehabilitation progress assessment results.

[0024] The multidimensional patient state vector includes joint kinematic feature vectors, wound healing status index, and physiological function baseline parameters. The safety constraint parameter set includes the maximum pain-free joint angle, wound tension limit threshold, and safe movement torque boundary. The initial rehabilitation training matrix includes movement type encoding sequence, training intensity distribution weight, and time frequency configuration parameters. The executed training sequence includes real-time resistance control gain, dynamic amplitude adjustment instructions, and adaptive movement step length sequence. The rehabilitation progress assessment results include joint function recovery score, training compliance metric, and predicted recovery trend for the next stage.

[0025] Specifically, such as Figure 2 , 3 As shown, the multi-dimensional state perception module includes: The physiological signal acquisition submodule collects activity data of burn patients, including joint range of motion, muscle tension, wound healing stage data and vital signs. It calls the Kalman filter algorithm to perform state estimation and error covariance update on the acquired multi-source time series data, filters out high-frequency environmental noise and motion artifact interference, and generates a denoised physiological parameter dataset. The flexible electronic skin sensor array and multi-channel electromyography (EMG) instrument, deployed at the burn patient's affected area (right elbow joint and surrounding skin), were activated. The sensor array... The sampling frequency captures minute deformations and resistance changes on the skin surface, converting the physical deformation signals into analog voltage signals; the electromyography (EMG) instrument simultaneously captures the biopotential signals generated by the depolarization of subcutaneous muscle fibers in the biceps and triceps brachii. An analog-to-digital converter converts these analog signals into a raw digital signal sequence. Subsequently, a Kalman filter algorithm program pre-programmed in the processing unit is invoked. This program initializes the state estimation vector. ,in Set as the initial measurement angle , Set as Initialize the error covariance matrix. identity matrix Measure the noise covariance matrix Set according to the sensor calibration parameters Process noise covariance matrix Set as After entering the iterative loop, the algorithm executes the time update step, utilizing the state transition matrix based on Newton's laws of motion. (in ), to estimate the prior state value at the current time. Next, a measurement update step is performed to collect the current actual sensor measurements. Calculate measurement residuals ,in Given the observation matrix. Calculate the Kalman gain matrix. If high-frequency noise is mixed into the measurement data, it will cause... If the relative increase, then Decrease, and tend to trust the model's predictions more. Utilize the calculated... and The prior state estimate is corrected to obtain the posterior state estimate. This value represents the denoised physiological parameter after filtering out random noise and sudden motion artifacts (e.g., the corrected joint angle). Simultaneously update the error covariance matrix. This process is performed point-by-point on multi-source time series data, ultimately outputting a denoised physiological parameter dataset.

[0026] The data normalization processing submodule acquires a denoised physiological parameter dataset, analyzes the distribution characteristics of vital signs indicators and sets a baseline range, calculates the offset of joint range of motion and muscle tension relative to the baseline range, uses the max-min normalization method to map heterogeneous data to a unified numerical range, combines the timestamp information of data acquisition to perform spatiotemporal registration of multi-source data and interpolation to complete missing values, and establishes a standardized multidimensional feature matrix. The process of performing spatiotemporal registration and missing value interpolation completion of multi-source data by combining the timestamp information of data collection is as follows: extract the timestamp sequence of each indicator in the denoised physiological parameter dataset, and select the time axis corresponding to the indicator with the highest sampling frequency as the reference time axis; traverse the timestamps of each other indicator and map them to the corresponding nearest neighbor time point on the reference time axis to perform multi-source data alignment in the time dimension; for the time points with missing values ​​on the reference time axis after alignment, construct a local time window, select the existing valid data points in the window as reference nodes, and use the cubic spline interpolation function to calculate the values ​​of the missing time points to perform data completion. The baseline ranges for various indicators are retrieved from a pre-defined clinical medical standards database. For example, the baseline range for elbow flexion range in healthy adults is [insert range here]. The baseline for resting muscle tone is Calculate the offset of the current patient's joint range of motion data relative to the baseline range. Set the maximum flexion angle acquired to... Then the offset For muscle tone data, if the currently measured tension is... Relative to the upper limit of the baseline The offset is Then, a maximum-minimum normalization operation is performed. For the joint angle data, the maximum value within the current observation period is searched. and minimum value For a measurement at a certain moment The normalization calculation formula is: This operation maps all heterogeneous data to... Interval. Based on this, perform spatiotemporal registration. Extract the time axis of the electromyographic signal with the highest sampling frequency ( ( ) is used as the reference time axis. For temperature signals with a low sampling frequency ( Iterate through its timestamps and map them to the point on the reference timeline with the smallest absolute time difference. For points on the reference timeline with missing temperature values ​​(such as...), ), constructing both the front and back A local window containing a set of valid data points. The valid data points within the window are... etc. Using cubic spline interpolation functions Calculate the missing point values. Substituting into the solved equation, we obtain the interpolated and completed numerical values. Establish a standardized multidimensional feature matrix.

[0027] The feature vector construction submodule calls a standardized multidimensional feature matrix, uses a sliding window to extract local data fragments, calculates the mean, variance and rate of change statistics of each indicator within the window, identifies key mutation points that characterize the trend of physiological function recovery and the evolution of wound healing status, and cascades and combines the extracted statistical features and mutation features according to the preset feature dimension priority order to generate a multidimensional patient state vector. The process of identifying key mutation points that characterize the trend of physiological function recovery and the evolution of wound healing status is as follows: For each dimension feature sequence in the standardized multidimensional feature matrix, calculate the first-order difference sequence of its adjacent time step values; apply a moving average filter to the first-order difference sequence to smooth high-frequency random fluctuations, obtaining a smoothed difference sequence; calculate the absolute value of each data point in the smoothed difference sequence and compare the absolute value with a preset mutation detection threshold; when the absolute values ​​of the data points of a continuous target number all exceed the mutation detection threshold, it is determined that a state transition has occurred at the beginning time of the continuous interval, and the beginning time is marked as a key mutation point. Set the length of the sliding window (correspond ), step size Extract a local data segment, and calculate the mean for the normalized sequence of joint angles within the window. ,variance and rate of change Simultaneously, the key mutation point identification process is initiated. For the electromyographic signal sequence in the feature matrix, the first-order difference sequence of adjacent time step values ​​is calculated. .right Application window size is A moving average filter is used to obtain a smoothed difference sequence. Calculate the absolute value. and with preset mutation detection threshold (This threshold is based on the resting noise level) Compare using the multiplier setting. When consecutive absolute value of data points ( When ), determine A state transition occurs at a certain moment (e.g., the onset of a muscle spasm). The initial moment... Marked as key mutation points, and mutation magnitude extracted. Following a preset priority order: [mutation magnitude, mean, variance, rate of change], features are concatenated and combined to generate a multidimensional patient state vector. .

[0028] Specifically, such as Figure 2 , 4 As shown, the scar tension assessment module includes: The nonlinear mapping construction submodule extracts joint kinematic features and wound healing status index from the multidimensional patient state vector, constructs a regression prediction structure composed of multiple decision trees, inputs the extracted features into the regression prediction structure for parallel inference, traverses the leaf nodes of the decision tree, quantifies the predicted values ​​of skin surface tension corresponding to various joint angles, aggregates the prediction results of multiple trees and calculates the average tension response curve, and establishes a wound tension distribution mapping table describing the nonlinear relationship between joint position and skin tension response. Based on the multidimensional patient state vector, joint kinematic features (angles) are extracted. angular velocity ) and wound healing status index (red component of wound) Moisture content ). Build by A gradient boosting regression tree (GBDT) structure composed of decision trees. Extracted features are input into the regression prediction structure and extrapolated in parallel. For example, the first... Decision trees are based on splitting rules (such as "angle") "and moisture content" ) Classify the samples to leaf nodes The predicted skin surface tension value stored in this node is ;No. Tree output .polymerization Calculate the weighted average of the predicted results for each tree. Traverse the feasible region of the joints. ,by Virtual inputs were generated for the step length, the healing index was kept constant, the tension prediction values ​​corresponding to various joint angles were quantified, and a wound tension distribution mapping table was established (see Table 1).

[0029] Table 1. Wound Tension Distribution Mapping Table (Partial Example) Joint angle ( ) Predicting skin tension (N / cm) Rate of change of tension (N / cm / deg) 80.0 5.2 0.10 90.0 8.5 0.33 100.0 14.8 0.63 110.0 22.1 0.73 This mapping table describes the nonlinear relationship between joint position and skin tension response in the current healing state.

[0030] The critical threshold calculation submodule calls the wound tension distribution mapping table, loads the preset skin tissue fracture strength benchmark and patient pain sensitivity coefficient, retrieves the critical movement position where the tension change rate exceeds the fracture strength benchmark in the mapping table, simultaneously locates the joint angle node when the pain response value reaches the upper limit of tolerance, performs intersection calculation on the critical movement position and the joint angle node, determines the safe operation range, and generates the boundary value of extreme activity and pain tolerance. Load the preset skin tissue fracture strength benchmark (Based on the characteristics of granulation tissue during the healing period) and the patient's pain sensitivity coefficient Retrieve tension changes from the mapping table. As shown in Table 1, when the joint angle reaches... At that time, predict tension When the angle reaches At that time, predict tension The tension is determined by linear interpolation. Critical motion position Simultaneously, based on the pain response model Calculate the pain score. Let the upper limit of pain tolerance be the VAS score. .when At that time, the calculation yielded ;when hour, The calculated pain response value reached... Joint angle nodes For the critical motion position With joint angle node Perform an intersection operation and take the minimum value to determine the upper limit of the safe operating range. Generate boundary values ​​for extreme activities and pain tolerance. .

[0031] The constraint parameter generation submodule introduces a preset safety buffer coefficient to shrink and correct the boundary values ​​for extreme activities and pain tolerance, reserves tissue elasticity space, converts the corrected boundary values ​​into the maximum allowable deflection angle in the joint coordinate system, calculates the maximum external load that does not trigger pain reflex within the deflection angle range, integrates angle and load limit data, and constructs a set of safety constraint parameters.

[0032] For boundary values upper limit Introduce a preset safety buffer coefficient The boundary values ​​are shrunk and corrected, and the maximum permissible deflection angle after correction is calculated. Reserved The tissue elastic space. Next, the maximum externally applied load is calculated. Based on the lever principle and the patient's forearm length... Combined with the maximum muscle strength estimate that does not trigger a pain reflex Calculate the maximum allowable torque within the deflection angle range. (0.8 is the reduction factor). In hour, Integration perspective With torque Construct a set of safety constraint parameters.

[0033] Specifically, such as Figure 2 , 5As shown, the intelligent planning generation module includes: The state-action space construction submodule constructs a digital state observation vector representing the current rehabilitation environment based on a multi-dimensional patient state vector. It calls a set of safety constraint parameters to perform boundary constraint verification on the preset full set of action spaces, identifies high-risk actions that exceed the safety angle or torque limits, generates action mask vectors to shield against violations, encapsulates the state observation vectors and action mask vectors, and establishes a reinforcement learning input state set. Constructing a digital state observation vector representing the current environment based on a multidimensional patient state vector. Set the current Call the set of security constraint parameters (upper limit). ) on the entire action space Perform boundary constraint verification. For actions... The angle after prediction execution is ,because This was determined to be a violation of regulations. Regarding the action... Prediction angle The action is deemed safe. An action mask vector is generated. (0 represents masking). The state vector... With mask Encapsulate and build a set of input states for reinforcement learning.

[0034] The strategy optimization decision-making submodule imports the reinforcement learning input state set into a deep Q-network model, calculates the expected value Q value of each candidate rehabilitation action through neural network forward propagation, applies action mask vector to filter the output Q value vector, eliminates dangerous actions, performs value maximization optimization in the remaining effective action space, locks the action node with the highest expected return, and generates the optimal rehabilitation action index code. Deep Q-networks compute the expected value Q-vector of candidate actions through forward propagation. , respectively corresponding Apply action mask vector Perform filtering: (Set the value of the violation to a minimum negative number). Perform a value-maximizing optimization within the remaining effective action space. For the corresponding action "Hold", lock the action node and generate the optimal rehabilitation action index code. .

[0035] The parameter matrix generation submodule decodes the optimal rehabilitation movement index code, maps and obtains the corresponding specific movement modality, resistance intensity level, number of repetitions per set, and rest duration between sets. Based on the temporal logic of rehabilitation training, it arranges and combines each parameter to establish structured data including movement type, intensity, frequency, and time parameter dimensions, and generates the initial rehabilitation training matrix.

[0036] Decoding the optimal rehabilitation movement index code According to the mapping table, this index corresponds to the action mode = equal-length resistance, and the resistance strength level = level 2 (corresponding to...). Each set consists of 5 repetitions, with a 30-second rest period between sets. This follows the logical sequence of rehabilitation training (warm-up). train (Relax), if currently in the training phase, use this parameter directly. Establish structured data: row vectors. Generate the initial rehabilitation training matrix.

[0037] Specifically, such as Figure 2 , 6 As shown, the dynamic adaptive adjustment module includes: The real-time feedback monitoring submodule loads the initial rehabilitation training matrix as a standard reference benchmark, uses a high-frequency position sensor and electromyography acquisition device to acquire the patient's actual movement coordinate sequence and physiological response data in real time, performs spatiotemporal alignment and Euclidean distance calculation between the actual coordinates and the ideal trajectory defined in the matrix, quantifies the degree of spatial deviation in the action execution process, simultaneously extracts the spectral energy characteristics of the electromyography signal and matches it with a preset pain quantification table, and generates a trajectory deviation feature vector and an immediate pain response level. set up Ideal trajectory coordinates at any moment The actual coordinates are obtained using a high-frequency position sensor. After performing spatiotemporal alignment, calculate the Euclidean distance. Simultaneously extract electromyographic signals and calculate the mid-frequency band (...). Power spectral density integral value Matching a preset pain quantification scale ( Generate an immediate pain response level. (Mild pain). Output trajectory deviation feature vector With pain level .

[0038] The fuzzy control decision submodule establishes a two-dimensional fuzzy input variable set, including error rate and pain intensity, based on the trajectory deviation feature vector and the immediate pain response level. It calls the preset membership rules to map the input variables to the fuzzy domain, performs fuzzy inference operations according to the expert rule base, determines the trigger weight of the control strategy, uses the centroid method to perform defuzzification calculation on the fuzzy inference results, quantifies the physical adjustment parameters used for real-time intervention of rehabilitation equipment, and generates resistance control correction coefficients and amplitude adjustment gains. The process of establishing a two-dimensional fuzzy input variable set including error rate and pain intensity is as follows: extract the Euclidean norm of the trajectory deviation feature vector as the position error scalar, map the instantaneous pain response level to the preset normalized numerical range to represent the pain intensity, and configure five levels of fuzzy linguistic variables including negative large, negative small, zero, positive small, and positive large and corresponding membership rules for the position error scalar and pain intensity respectively. The process of performing fuzzy inference based on the expert rule base is as follows: setting adjustment rules in the expert rule base; when the position error scalar is greater than the preset deviation threshold and the pain intensity is at a high response level, triggering a negative large adjustment rule for the resistance parameter to generate a fuzzy output set that reduces resistance; when the position error scalar is within the allowable range and the pain intensity is at a low response level, triggering an adjustment rule that maintains or slightly increases resistance. The process of defuzzifying the fuzzy inference results using the centroid method is as follows: the fuzzy set of the trigger rule output is aggregated by union to obtain the total output fuzzy surface; integral operation is performed along the universe axis of the total output fuzzy surface; the geometric centroid coordinates of the region enclosed by the membership rule curve and the horizontal axis are calculated; the values ​​corresponding to the geometric centroid coordinates are determined as the precise control quantities and assigned to the resistance control correction coefficient and the amplitude adjustment gain, respectively. The Euclidean norm of the trajectory deviation eigenvector is extracted as a scalar of position error. Immediate pain response level Mapping to normalized interval get Establish a two-dimensional fuzzy input variable set. According to the membership rule, Belongs to "positive small" (membership degree) ), Belongs to "low" (membership degree) According to the expert rule base: "If the positional error is positively small and the pain intensity is low, then the resistance control correction coefficient is positively small (slightly increasing)." The activation strength of the trigger rule is... The centroid method is used to resolve fuzziness: the output fuzzy set (set as a triangular distribution, with the center at...) is... Integrate to calculate the geometric centroid coordinates. Determine the precise control quantity as follows: Generate drag control correction coefficient With amplitude adjustment gain (Painless, maintain the original range).

[0039] The adaptive sequence construction submodule calls the resistance control correction coefficient and amplitude adjustment gain to perform multiplicative weighted adjustment on the preset resistance intensity parameters in the initial rehabilitation training matrix. It uses the adjustment gain to dynamically scale and correct the boundary coordinates of the movement amplitude, limiting the range of motion. Based on the corrected parameters, it updates the control command queue of the underlying actuator in real time, and reassembles and encapsulates the discrete control commands according to the timestamp order to generate the execution training sequence. Call the resistance control correction factor With amplitude adjustment gain Preset resistance in the initial matrix. Perform multiplicative weighted adjustment: Using gain to define the motion amplitude boundary Scaling: The interval length remains unchanged. Based on the corrected parameters. Update the underlying control instructions. Reassemble the discrete instructions according to their timestamps: ; Generate and execute the training sequence.

[0040] Specifically, such as Figure 2 , 7 As shown, the rehabilitation efficacy assessment module includes: The data collection and serialization submodule acquires multi-cycle execution training sequences, synchronously retrieves joint range of motion and muscle strength detection data within the corresponding time window, performs temporal alignment and synchronous mapping of action commands and physiological response data based on timestamps, removes invalid cycle data with breakpoints and performs linear interpolation to complete the data, constructs a multi-dimensional sliding window according to a preset time step, converts discrete records into continuous time series samples, and establishes a historical rehabilitation time series dataset. Align action instructions based on timestamps ( ) and joint angle data ( ). Detected to There are data breakpoints (intervals) Remove invalid periods or perform linear interpolation if the interval is small: According to step size Constructing a multidimensional sliding window Establish a historical rehabilitation time series dataset.

[0041] The trend prediction and analysis submodule inputs the historical rehabilitation time series dataset into the long short-term memory network model, uses the forget gate mechanism to remove low-relevance historical state information, updates the cell state through the input gate, captures the nonlinear evolution law of joint range of motion and muscle strength over time, performs regression calculation on the hidden layer features based on the output gate, predicts the peak joint angle and muscle strength index of the next rehabilitation cycle, and generates a predicted value of joint function evolution trend. Input the historical rehabilitation time-series dataset into the LSTM model. Input gate. Process the current input features (angle sequence, muscle strength sequence). Forget gate. Remove historical information with low relevance. Cell state. Update memory. Output gate. Combination Output hidden layer features Regression layer Calculations to predict the next cycle (e.g.) Peak joint angle of (day) With muscle strength Generate predicted values ​​for the evolution trend of joint function.

[0042] The efficacy quantification assessment submodule calls the predicted value of joint function evolution trend, dynamically compares it with the preset standard rehabilitation path curve, calculates the Euclidean distance of the actual recovery trajectory relative to the standard path to quantify the deviation of rehabilitation progress, and calculates the functional recovery rate by combining the change slope of the trend prediction value. It integrates the progress deviation, recovery rate and predicted functional indicators to construct a rehabilitation progress assessment result including the current status score and future potential prediction.

[0043] Calling the predicted value of joint function evolution trend Compared with the preset standard rehabilitation pathway curve, in Standard value of the day Calculate schedule deviations. Combined with the measured values ​​from the previous period Calculate the recovery rate Normalize and weight each indicator (weights: deviation 0.4, rate 0.3, absolute value 0.3): (Normalization maximum score 1.0). Construct a system including the current state score. Assessment results of rehabilitation progress and future potential projections.

[0044] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An intelligent planning system for rehabilitation training of joint dysfunction in burn patients, characterized in that, The system includes: The multidimensional state perception module collects activity data of burn patients, including joint range of motion, muscle tension, wound healing and vital signs data. After Kalman filtering to remove noise, physiological functions and wound state features are extracted to construct a multidimensional patient state vector. The scar tension assessment module, based on the multidimensional patient state vector, uses a random forest model to analyze the nonlinear mapping between wound traction and joint movement, calculates the limit of movement threshold and pain tolerance boundary, and constructs a set of safety constraint parameters. The intelligent planning and generation module inputs the multidimensional patient state vector and safety constraint parameter set into the deep Q-network model to optimize the action type, intensity, frequency and interval duration with the goal of maximizing the recovery score, and generates an initial rehabilitation training matrix. The dynamic adaptive adjustment module loads the initial rehabilitation training matrix, monitors the actual trajectory and physiological feedback signals, uses fuzzy logic to calculate the trajectory deviation value and pain response level, corrects the resistance coefficient and movement amplitude, and generates an execution training sequence. The rehabilitation efficacy assessment module gathers the executed training sequences, combines them with joint function recovery data, inputs them into a long short-term memory network model to analyze the evolution of activity and muscle strength, calculates rehabilitation progress indicators and predicts recovery trends, and generates rehabilitation progress assessment results.

2. The intelligent planning system for rehabilitation training of joint dysfunction in burn patients according to claim 1, characterized in that, The multidimensional patient state vector includes joint kinematic feature vectors, wound healing status index, and physiological function baseline parameters. The safety constraint parameter set includes maximum pain-free joint angle, wound tension limit threshold, and safe movement torque boundary. The initial rehabilitation training matrix includes movement type encoding sequence, training intensity distribution weight, and time frequency configuration parameters. The executed training sequence includes real-time resistance control gain, dynamic amplitude adjustment instructions, and adaptive movement step length sequence. The rehabilitation progress assessment results include joint function recovery score, training compliance metric, and next stage recovery trend prediction value.

3. The intelligent planning system for rehabilitation training of joint dysfunction in burn patients according to claim 1, characterized in that, The multi-dimensional state perception module includes: The physiological signal acquisition submodule collects activity data of burn patients, including joint range of motion, muscle tension, wound healing stage data and vital signs. It calls the Kalman filter algorithm to perform state estimation and error covariance update on the acquired multi-source time series data, filters out high-frequency environmental noise and motion artifact interference, and generates a denoised physiological parameter dataset. The data normalization processing submodule acquires the denoised physiological parameter dataset, analyzes the distribution characteristics of vital signs indicators and sets a benchmark range, calculates the offset of joint range of motion and muscle tension relative to the benchmark range, uses the max-min normalization method to map heterogeneous data to a unified numerical range, combines the timestamp information of data acquisition to perform spatiotemporal registration of multi-source data and interpolation to fill missing values, and establishes a standardized multidimensional feature matrix. The feature vector construction submodule calls the standardized multidimensional feature matrix, uses a sliding window to extract local data fragments, calculates the mean, variance and rate of change statistics of each indicator in the window, identifies key mutation points that characterize the trend of physiological function recovery and the evolution of wound healing status, and cascades and combines the extracted statistical features and mutation features according to the preset feature dimension priority order to generate a multidimensional patient state vector.

4. The intelligent planning system for rehabilitation training of joint dysfunction in burn patients according to claim 3, characterized in that, The process of performing spatiotemporal registration and missing value interpolation of multi-source data by combining the timestamp information of data collection is specifically as follows: extract the timestamp sequence of each indicator in the denoised physiological parameter dataset, select the time axis corresponding to the indicator with the highest sampling frequency as the reference time axis; traverse the timestamp of each other indicator, map it to the nearest neighbor time point on the reference time axis, and perform alignment of multi-source data in the time dimension. For the time points where values ​​are missing on the baseline time axis after alignment, a local time window is constructed, and existing valid data points within the window are selected as reference nodes. The cubic spline interpolation function is used to calculate the values ​​of the missing time points to complete the data. The process of identifying key abrupt changes that characterize the trend of physiological function recovery and the evolution of wound healing status specifically involves calculating the first-order difference sequence of adjacent time step values ​​for each dimension feature sequence in the standardized multidimensional feature matrix; applying a moving average filter to the first-order difference sequence to smooth high-frequency random fluctuations, thereby obtaining a smoothed difference sequence. Calculate the absolute value of each data point in the smoothed difference sequence and compare the absolute value with a preset mutation detection threshold; When the absolute value of the number of consecutive target data points all exceeds the mutation detection threshold, it is determined that a state transition has occurred at the start time of the consecutive interval, and the start time is marked as a critical mutation point.

5. The intelligent planning system for rehabilitation training of joint dysfunction in burn patients according to claim 1, characterized in that, The scar tension assessment module includes: The nonlinear mapping construction submodule extracts joint kinematic features and wound healing status index from the multidimensional patient state vector, constructs a regression prediction structure composed of multiple decision trees, inputs the extracted features into the regression prediction structure for parallel deduction, traverses the leaf nodes of the decision tree, quantifies the predicted values ​​of skin surface tension corresponding to various joint angles, aggregates the prediction results of multiple trees and calculates the average tension response curve, and establishes a wound tension distribution mapping table describing the nonlinear relationship between joint position and skin tension response. The critical threshold calculation submodule calls the wound tension distribution mapping table, loads the preset skin tissue fracture strength benchmark and patient pain sensitivity coefficient, retrieves the critical movement position where the tension change rate exceeds the fracture strength benchmark in the mapping table, simultaneously locates the joint angle node when the pain response value reaches the upper limit of tolerance, performs intersection calculation on the critical movement position and the joint angle node, determines the safe operation range, and generates the boundary value of extreme activity and pain tolerance. The constraint parameter generation submodule introduces a preset safety buffer coefficient to shrink and correct the boundary values ​​for the extreme activity and pain tolerance boundary values, reserves tissue elasticity space, converts the corrected boundary values ​​into the maximum allowable deflection angle in the joint coordinate system, calculates the maximum external load that does not trigger pain reflex within the deflection angle range, integrates angle and load limit data, and constructs a set of safety constraint parameters.

6. The intelligent planning system for rehabilitation training of joint dysfunction in burn patients according to claim 1, characterized in that, The intelligent planning generation module includes: The state-action space construction submodule constructs a digital state observation vector representing the current rehabilitation environment based on the multidimensional patient state vector. It calls the safety constraint parameter set to perform boundary constraint verification on the preset full set of action space, identifies high-risk actions that exceed the safety angle or torque limit, generates an action mask vector to shield against illegal operations, encapsulates the state observation vector and the action mask vector, and establishes a reinforcement learning input state set. The strategy optimization decision submodule imports the reinforcement learning input state set into a deep Q-network model, calculates the expected value Q value of each candidate rehabilitation action through neural network forward propagation, applies action mask vector to filter the output Q value vector, eliminates dangerous actions, performs value maximization optimization in the remaining effective action space, locks the action node with the highest expected return, and generates the optimal rehabilitation action index code. The parameter matrix generation submodule decodes the optimal rehabilitation action index code, maps and obtains the corresponding specific action modality, resistance intensity level, number of repetitions per set and rest duration between sets, and arranges and combines each parameter according to the temporal logic of rehabilitation training to establish structured data including the dimensions of action type, intensity, frequency and time parameters, and generates the initial rehabilitation training matrix.

7. The intelligent planning system for rehabilitation training of joint dysfunction in burn patients according to claim 1, characterized in that, The dynamic adaptive adjustment module includes: The real-time feedback monitoring submodule loads the initial rehabilitation training matrix as a standard reference benchmark, uses a high-frequency position sensor and electromyography acquisition device to acquire the patient's actual movement coordinate sequence and physiological response data in real time, performs spatiotemporal alignment and Euclidean distance calculation between the actual coordinates and the ideal trajectory defined in the matrix, quantifies the degree of spatial deviation in the action execution process, simultaneously extracts the spectral energy characteristics of the electromyography signal and matches it with a preset pain quantification table, and generates a trajectory deviation feature vector and an immediate pain response level. The fuzzy control decision submodule establishes a two-dimensional fuzzy input variable set including error rate and pain intensity based on the trajectory deviation feature vector and the instantaneous pain response level. It calls the preset membership rules to map the input variables to the fuzzy domain, performs fuzzy inference operations according to the expert rule base, determines the trigger weight of the control strategy, uses the centroid method to defuzzify the fuzzy inference results, quantifies the physical adjustment parameters used for real-time intervention of rehabilitation equipment, and generates resistance control correction coefficient and amplitude adjustment gain. The adaptive sequence construction submodule calls the resistance control correction coefficient and amplitude adjustment gain to perform multiplicative weighted adjustment on the preset resistance intensity parameters in the initial rehabilitation training matrix. It uses the adjustment gain to dynamically scale and correct the boundary coordinates of the movement amplitude, limiting the range of motion. Based on the corrected parameters, it updates the control command queue of the underlying actuator in real time, and reassembles and encapsulates the discrete control commands according to the timestamp order to generate an execution training sequence.

8. The intelligent planning system for rehabilitation training of joint dysfunction in burn patients according to claim 7, characterized in that, The process of establishing a two-dimensional fuzzy input variable set including error rate and pain intensity specifically involves extracting the Euclidean norm of the trajectory deviation feature vector as a position error scalar, mapping the instantaneous pain response level to a preset normalized numerical range to characterize pain intensity, and configuring five levels of fuzzy linguistic variables including negative large, negative small, zero, positive small, and positive large and corresponding membership rules for the position error scalar and pain intensity respectively. The process of performing fuzzy inference calculations based on the expert rule base specifically involves setting adjustment rules in the expert rule base. When the position error scalar is greater than a preset deviation threshold and the pain intensity is at a high response level, a negative large-scale adjustment rule for the resistance parameter is triggered to generate a fuzzy output set that reduces resistance. When the position error scalar is within the allowable range and the pain intensity is at a low response level, an adjustment rule that maintains or slightly increases resistance is triggered. The process of using the centroid method to defuzzify the fuzzy inference results specifically involves: performing a union aggregation on the fuzzy set of the trigger rule output to obtain the total output fuzzy surface; performing an integral operation along the universe axis of the total output fuzzy surface; calculating the geometric centroid coordinates of the region enclosed by the membership rule curve and the horizontal axis; determining the values ​​corresponding to the geometric centroid coordinates as the precise control quantities; and assigning them to the resistance control correction coefficient and the amplitude adjustment gain, respectively.

9. The intelligent planning system for rehabilitation training of joint dysfunction in burn patients according to claim 1, characterized in that, The rehabilitation efficacy assessment module includes: The data collection and serialization submodule acquires the execution training sequence over multiple periods, synchronously retrieves joint range of motion and muscle strength detection data within the corresponding time window, performs temporal alignment and synchronous mapping of action commands and physiological response data based on timestamps, removes invalid period data with breakpoints and performs linear interpolation to complete the data, constructs a multi-dimensional sliding window according to a preset time step, converts discrete records into continuous time series samples, and establishes a historical rehabilitation time series dataset. The trend prediction and analysis submodule inputs the historical rehabilitation time series dataset into the long short-term memory network model, uses the forget gate mechanism to remove low-relevance historical state information, updates the cell state through the input gate, captures the nonlinear evolution law of joint range of motion and muscle strength over time, performs regression calculation on the hidden layer features based on the output gate, predicts the peak joint angle and muscle strength index of the next rehabilitation cycle, and generates a predicted value of joint function evolution trend. The efficacy quantification assessment submodule calls the predicted value of joint function evolution trend, dynamically compares it with the preset standard rehabilitation path curve, calculates the Euclidean distance of the actual recovery trajectory relative to the standard path to quantify the deviation of rehabilitation progress, calculates the functional recovery rate by combining the change slope of the trend prediction value, and integrates the progress deviation, recovery rate and predicted functional indicators to construct the rehabilitation progress assessment result.

Images