Intelligent orthopedic joint rehabilitation monitoring method and system

Abstract

The invention discloses an intelligent orthopedic joint rehabilitation monitoring method and system, and the method comprises the steps: obtaining a three-dimensional spatial-temporal feature vector of joint motion according to joint motion data, muscle electric signals and pressure distribution information collected by a multi-modal sensor; based on the three-dimensional spatio-temporal feature vector, key parameters of a joint function state are extracted by using a spatio-temporal diagram convolutional neural network, and a joint stability quantitative index is obtained; according to the joint stability quantitative index and the patient pain feedback data, a rehabilitation stage quantitative index is obtained; generating a safety training instruction set according with the current rehabilitation stage based on the rehabilitation stage quantitative index; and driving the wearable exoskeleton to execute rehabilitation training according to the safety training instruction set, and synchronously monitoring a neuromuscular activation signal through a brain-computer interface to verify the consistency of a training action and neural control. According to the embodiment of the invention, accurate, continuous and self-adaptive monitoring and management of the whole process of joint rehabilitation can be realized.

Description

Technical Field

[0001] This invention belongs to the field of medical technology, specifically a smart orthopedic joint rehabilitation monitoring method and system. Background Technology

[0002] In the rehabilitation process following orthopedic joint surgery, traditional rehabilitation monitoring methods mainly rely on regular doctor checkups and patient subjective feedback, which suffers from problems such as untimely monitoring, discontinuous data, and inaccurate assessment. While existing wearable devices can collect some motion data, they generally lack the ability to comprehensively monitor multidimensional joint biomechanical parameters (such as angle, pressure, and torque) and have difficulty distinguishing the biomechanical differences between active and passive movements. Furthermore, insufficient multimodal fusion analysis of rehabilitation data leads to an inability to accurately assess rehabilitation progress and predict the risk of complications. Summary of the Invention

[0003] The purpose of this invention is to provide a smart orthopedic joint rehabilitation monitoring method and system to overcome the shortcomings of the existing technology and achieve accurate, continuous, and adaptive monitoring and management of the entire joint rehabilitation process.

[0004] One embodiment of this application provides a smart orthopedic joint rehabilitation monitoring method, the method comprising: Based on joint motion data, muscle electromyography signals, and pressure distribution information collected by multimodal sensors, dynamic biomechanical modeling is performed using a spatiotemporal feature fusion algorithm to obtain a three-dimensional spatiotemporal feature vector of joint motion. Based on the aforementioned three-dimensional spatiotemporal feature vector, key parameters of joint functional state are extracted using a spatiotemporal graph convolutional neural network. The coordinated motion pattern of joint ligaments and bones is captured through an adaptive topology structure to obtain quantitative indicators of joint stability. Based on the joint stability quantification index and patient pain feedback data, the dynamic evolution coefficient of the rehabilitation stage is calculated by the rehabilitation entropy model. The rehabilitation stage quantification index is obtained by nonlinear weighting of biomechanical parameters and physiological response parameters. Based on the quantitative index of the rehabilitation stage, a personalized rehabilitation action sequence is constructed using an adversarial generative network. The medical compliance of the action sequence is verified by a discriminator, and a safe training instruction set that conforms to the current rehabilitation stage is generated. The wearable exoskeleton is driven to perform rehabilitation training according to the safety training instruction set, and neuromuscular activation signals are monitored simultaneously through the brain-computer interface to verify the consistency between training movements and neural control.

[0005] Optionally, the step of performing dynamic biomechanical modeling based on joint motion data, electromyographic signals, and pressure distribution information collected by multimodal sensors, using a spatiotemporal feature fusion algorithm, to obtain a three-dimensional spatiotemporal feature vector of joint motion includes: Based on the joint motion data, a six-degree-of-freedom trajectory Kalman filter is performed to obtain the noise-reduced joint motion trajectory. Based on electromyographic signals and pressure distribution information, a time-domain phase alignment algorithm is used to generate a biomechanical coupling feature matrix. By fusing joint motion trajectories with biomechanical coupling feature matrices and processing them through tensor cross decomposition, a three-dimensional spatiotemporal feature vector is output.

[0006] Optionally, based on the three-dimensional spatiotemporal feature vector, the key parameters of joint functional state are extracted using a spatiotemporal graph convolutional neural network, and the coordinated motion patterns of joint ligaments and bones are captured through an adaptive topology structure to obtain quantitative indicators of joint stability, including: Based on the three-dimensional spatiotemporal feature vectors, dynamic skeletal topology construction is performed to generate an adaptive biomechanical graph structure; Based on an adaptive biomechanical graph structure, hierarchical spatiotemporal graph convolution processing is used to extract ligament cooperative motion features; By integrating ligament synergistic motion characteristics and processing them with attention-weighted pooling, a set of key parameters for joint stability is obtained. Based on a set of key parameters for joint stability, radial basis function mapping is performed in conjunction with clinical evaluation criteria to output quantitative indicators of joint stability.

[0007] Optionally, based on the joint stability quantification index and patient pain feedback data, the dynamic evolution coefficient of the rehabilitation stage is calculated using the rehabilitation entropy model. The rehabilitation stage quantification index is obtained by nonlinearly weighting the results of fusing biomechanical parameters and physiological response parameters, including: Based on the quantitative index of joint stability, time-series differential processing is performed to obtain the rate of change of biomechanical parameters; Based on patient pain feedback data, BioBERT feature extraction was used to generate standardized pain feature vectors. By integrating the rate of change of biomechanical parameters and the pain feature vector, and processing them through dual-entropy coupling calculation, dynamic evolution coefficients are output. Based on the dynamic evolution coefficient, adaptive weighting of parameters is performed to generate biomechanical-physiological response fusion weights; Based on the fusion weights and clinical thresholds, a quantitative index for the rehabilitation stage is output through S-shaped function transformation.

[0008] Optionally, based on the quantitative index of the rehabilitation stage, a personalized rehabilitation action sequence is constructed using an adversarial generative network. The medical compliance of the action sequence is verified by a discriminator, and a safe training instruction set conforming to the current rehabilitation stage is generated, including: Based on the quantitative index of the rehabilitation stage, the medical knowledge base is retrieved and processed to generate a safety boundary matrix for joint range of motion. Based on the safety boundary matrix, an LSTM-Transformer hybrid network is used to generate the initial action sequence; Based on the initial action sequence, the joint angle space-constrained action is obtained through inverse kinematic mapping. Based on joint angle spatial constraints, a three-dimensional biomechanical discriminant is used to verify and screen out medically compliant movements. By integrating medical compliance procedures with diverse reward mechanisms, and through instruction coding processing, a set of safety training instructions is output.

[0009] Optionally, the step of driving the wearable exoskeleton to perform rehabilitation training according to the safety training instruction set, while simultaneously monitoring neuromuscular activation signals through a brain-computer interface to verify the consistency between training movements and neural control, includes: Based on the safety training instruction set, the motor signals are decoded and processed to drive the exoskeleton to perform training actions; Based on brain-computer interface signals, blind source separation processing is used to extract neuromuscular activation features; By integrating actual motion trajectories and neuromuscular activation features, and processing them through mutual information entropy calculation, a neural control matching degree is generated. Based on the neural control matching degree, reinforcement learning is triggered to dynamically adjust the processing and output a neural-mechanical synchronization verification report.

[0010] Another embodiment of this application provides a smart orthopedic joint rehabilitation monitoring system, the system comprising: The fusion module is used to perform dynamic biomechanical modeling based on joint motion data, electromyographic signals and pressure distribution information collected by multimodal sensors, and obtain three-dimensional spatiotemporal feature vectors of joint motion through spatiotemporal feature fusion algorithms. The extraction module is used to extract key parameters of joint functional state based on the three-dimensional spatiotemporal feature vector using a spatiotemporal graph convolutional neural network, and capture the coordinated motion pattern of joint ligaments and bones through an adaptive topology structure to obtain quantitative indicators of joint stability. The calculation module is used to calculate the dynamic evolution coefficient of the rehabilitation stage based on the joint stability quantification index and patient pain feedback data through the rehabilitation entropy model, and obtain the rehabilitation stage quantification index by nonlinear weighting through the fusion of biomechanical parameters and physiological response parameters. The module is used to construct personalized rehabilitation action sequences based on the quantitative index of the rehabilitation stage, using an adversarial generative network, verifying the medical compliance of the action sequences through a discriminator, and generating a set of safe training instructions that conforms to the current rehabilitation stage. The monitoring module is used to drive the wearable exoskeleton to perform rehabilitation training according to the safety training instruction set, and simultaneously monitor neuromuscular activation signals through the brain-computer interface to verify the consistency between training actions and neural control.

[0011] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0012] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0013] Compared with existing technologies, this invention provides a smart orthopedic joint rehabilitation monitoring method. Based on joint motion data, electromyographic signals, and pressure distribution information collected by multimodal sensors, a three-dimensional spatiotemporal feature vector of joint motion is obtained. Based on this three-dimensional spatiotemporal feature vector, a spatiotemporal graph convolutional neural network is used to extract key parameters of joint functional status, resulting in a quantitative index of joint stability. Based on the quantitative index of joint stability and patient pain feedback data, a quantitative index of the rehabilitation stage is obtained. Based on the quantitative index of the rehabilitation stage, a safe training instruction set conforming to the current rehabilitation stage is generated. The wearable exoskeleton is driven to perform rehabilitation training according to the safe training instruction set, while simultaneously monitoring neuromuscular activation signals through a brain-computer interface to verify the consistency between training movements and neural control. This enables precise, continuous, and adaptive monitoring and management of the entire joint rehabilitation process. Attached Figure Description

[0014] Figure 1 Hardware structure block diagram of a computer terminal for a smart orthopedic joint rehabilitation monitoring method provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a smart orthopedic joint rehabilitation monitoring method provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an intelligent orthopedic joint rehabilitation monitoring system provided in an embodiment of the present invention. Detailed Implementation

[0015] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0016] This invention first provides a smart orthopedic joint rehabilitation monitoring method, which can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0017] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a smart orthopedic joint rehabilitation monitoring method provided in an embodiment of the present invention. Figure 1As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0018] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions that, when executed, cause the processor to perform any intelligent orthopedic joint rehabilitation monitoring method.

[0019] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0020] Internal memory provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to execute any intelligent orthopedic joint rehabilitation monitoring method.

[0021] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0022] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0023] See Figure 2 The present invention provides an intelligent orthopedic joint rehabilitation monitoring method, which may include the following steps: S201: Based on the joint motion data, muscle electromyography signals and pressure distribution information collected by multimodal sensors, dynamic biomechanical modeling is performed through a spatiotemporal feature fusion algorithm to obtain the three-dimensional spatiotemporal feature vector of joint motion. Specifically, based on the joint motion data, a six-degree-of-freedom trajectory Kalman filter can be performed to obtain the noise-reduced joint motion trajectory; In the inertial measurement unit (IMU) worn by the patient, six key nodes of the knee joint (medial and lateral femoral condyles, tibial plateau, etc.) are deployed, acquiring raw motion data every 10 milliseconds. Each IMU contains a triaxial accelerometer (measuring linear acceleration), a triaxial gyroscope (measuring angular velocity), and a triaxial magnetometer (measuring azimuth), collectively forming a six-degree-of-freedom (6-DoF) motion capture system. The raw data contains sensor noise (such as gyroscope drift error ±0.5° / second), which is dynamically reduced using a Kalman filter. Prediction phase: Based on the knee joint biomechanical model (such as the flexion-extension range of 0°-130°), the current position (such as the predicted flexion angle of 45.2°) is calculated according to the posture quaternion (a three-dimensional rotational representation) of the previous moment.

[0024] Update Phase: Compare actual sensor readings (e.g., accelerometer readings of 0.8g vertical acceleration) with predicted values, and calculate confidence weights (Q-value and R-value parameters) using the covariance matrix. For example, the gyroscope noise covariance is set to 0.03, the accelerometer covariance is set to 0.1, and the fusion ratio is dynamically adjusted (angular velocity weight 70%, acceleration weight 30%).

[0025] Output of the joint trajectory after noise reduction: The trajectory of a single flexion and extension movement is restored to a smooth curve (the noise amplitude is reduced from ±3° to ±0.2°), and the timestamp accuracy is at the millisecond level.

[0026] Based on electromyographic signals and pressure distribution information, a time-domain phase alignment algorithm is used to generate a biomechanical coupling feature matrix. Electromyography (EMG) signals were acquired using 8-channel electrodes attached to the quadriceps and hamstring muscles (sampling rate 2000 Hz), while pressure distribution information was obtained from a plantar pressure sensing insole (256 sensing units, resolution 1.5 cm). 2 ( / unit) synchronous acquisition. Due to the millisecond-level time delay between the two types of signals (e.g., electromyography signal 50 milliseconds before heel touches the ground), temporal phase alignment is required: Key event markers: Electromyographic signals were detected using the slope threshold method (amplitude exceeding the baseline by 200 μV) with the EMG onset as the reference (e.g., the moment of activation of the vastus lateralis).

[0027] The pressure signal is based on the heel strike moment (pressure value surge > 20 kPa) and is located using a peak detection algorithm.

[0028] Dynamic Time Warping (DTW): Calculate the time offset between the electromyographic burst point and heel contact with the ground (e.g., -50 ms) and establish a nonlinear mapping path.

[0029] Stretch / compress signal segments in 5-millisecond windows (e.g., delay electromyographic signals by 50 milliseconds) to ensure alignment of biomechanical events (e.g., synchronizing quadriceps activation with weight-bearing phase).

[0030] The aligned data generates a biomechanical coupling feature matrix: each row represents a time point (time resolution 5 milliseconds), and each column includes 32 features such as electromyographic amplitude (μV), pressure center trajectory (COP-X / Y coordinates), and muscle activation duration (e.g., rectus femoris activation 120 milliseconds).

[0031] By fusing joint motion trajectories with biomechanical coupling feature matrices and processing them through tensor cross decomposition, a three-dimensional spatiotemporal feature vector is output.

[0032] The joint motion trajectories (six degrees of freedom attitude angles) are integrated with the biomechanical feature matrix into a three-dimensional spatio-temporal data cube (3D Spatio-Temporal Tensor): Dimension 1: Timeline (200 slices, covering a 1-second action cycle); Dimension 2: Spatial Axis (6 joint nodes × 3 orientation angles = 18 spatial parameters); Dimension 3: Biomechanical axis (32-dimensional electromyography-pressure characteristics); Extracting cross-dimensional correlation features using Tensor Cross-Decomposition: Higher-Order Singular Value Decomposition (HOSVD): The data cube is decomposed into a core tensor and three factor matrices (time factor, space factor, and biomechanical factor).

[0033] For example, in the core tensor, it was found that "when the knee is flexed at 60°, the peak heel pressure is strongly correlated with the activation intensity of the vastus lateralis muscle" (correlation coefficient 0.92).

[0034] Feature fusion and dimensionality reduction: The top 10 largest singular values ​​(reflecting feature importance) in the core tensor are retained and projected into a low-dimensional space.

[0035] Fusion rule: Remove minor components with a contribution rate of less than 5% (such as the weak correlation between ankle internal rotation angle and gastrocnemius muscle activation).

[0036] Output a 3D spatiotemporal feature vector (128-dimensional vector). Encoding example: Dimensions 1-30: Principal components of motion trajectory (e.g., rate of change of buckling angle); Dimensions 31-90: Electromyographic-pressure coupling patterns (e.g., the vastus medialis activation delay index). Dimensions 91-128: Cross-joint coordination parameters (such as hip-knee phase difference).

[0037] S202, Based on the three-dimensional spatiotemporal feature vector, the key parameters of joint functional state are extracted using a spatiotemporal graph convolutional neural network, and the coordinated motion pattern of joint ligaments and bones is captured through an adaptive topology structure to obtain a quantitative index of joint stability. Specifically, dynamic skeletal topology construction can be performed based on three-dimensional spatiotemporal feature vectors to generate an adaptive biomechanical graph structure; After loading the 128-dimensional spatiotemporal feature vector, the system launches the dynamic skeletal topology construction engine. This engine first analyzes the biomechanical meaning of each dimension of the vector: the first 30 dimensions correspond to the six-degree-of-freedom motion trajectory of the skeleton (e.g., the femoral rotational angular velocity of 15 degrees per second), the middle 60 dimensions describe the muscle activation state (e.g., the rectus femoris discharge frequency of 35 peak times per second), and the last 38 dimensions encode the microscopic strain of the ligaments (e.g., the real-time tension value of the anterior cruciate ligament of 120 Newtons). Based on orthopedic anatomy atlases, the knee joint is divided into 12 key biomechanical nodes, including the medial and lateral femoral condyles (node ​​IDs: F1, F2), the tibial plateau (T1, T2), and the patellar center (P0).

[0038] The core of the adaptive mechanism lies in real-time response to changes in joint state: Dynamic calibration of node position: When the patient flexes the knee to more than 60 degrees, the system detects the posterior displacement of the tibial plateau (e.g., the Z-axis coordinate of node T1 is displaced by +3.8 mm from the initial position), and automatically triggers the node position iteration algorithm (PIA) to update the node's three-dimensional coordinates at a frequency of 5 milliseconds (accuracy ±0.1 mm).

[0039] Virtual ligament edge generation: Biomechanical action edges are created based on the anatomical attachment points of the ligaments. For example, the anterior cruciate ligament connects the femoral node F2 and the tibial node T1, with an initial weight of 0.8 (weight range 0.0-1.0, reflecting the proportion of mechanical contribution). If abnormal shear forces are detected during movement (e.g., tibial anterior displacement exceeding 5 mm), a compensatory virtual edge is added (e.g., the hamstring-posterior tibial edge, weight 0.6). Each edge stores three types of parameters: baseline tension threshold (e.g., 150 Newtons for the anterior cruciate ligament), real-time strain coefficient (current value 1.2 times the baseline value), and biomechanical priority (set to the highest level 9 for the anterior cruciate ligament). The final output is an adaptive biomechanical graph structure containing 12 nodes and 15-28 dynamic edges, with the topology refreshed every 100 milliseconds.

[0040] Based on an adaptive biomechanical graph structure, hierarchical spatiotemporal graph convolution processing is used to extract ligament cooperative motion features; The Hierarchical Spatio-Temporal Graph Convolutional Network (HST-GCN) consists of three parallel processing layers: Spatial Convolution Layer: The mechanical parameters of adjacent ligament edges are aggregated around the skeletal node. The convolution kernel size is set to 3×3 (covering 1 central node and 8 adjacent nodes), and the calculation rule is a weighted fusion of ligament tension and direction vector. For example, when calculating the convolution output of the femoral node F1: The tension values ​​of adjacent anterior cruciate ligaments were collected at 125 Newtons (weight 0.85). The lateral collateral ligament side tension value is 80 Newtons (weight 0.75). Output characteristic value = (125 × 0.85 + 80 × 0.75) × tension normalization coefficient 0.012; Output 16-dimensional primary spatial features to characterize the synergistic strength of local ligament groups (e.g., the synchronization coefficient between the quadriceps tendon and the patellar ligament is 0.78).

[0041] Temporal Recurrent Layer: A gated recurrent unit (GRU) is used to process continuous temporal data. Given the spatial characteristics of the past 10 frames (1-second time window), the GRU contains 128 memory units. For example: The memory unit stores the "peak tension pattern of the anterior cruciate ligament during the gait support phase" (reaching 140 Newtons at 0.3 seconds). When new data is input (current frame tension 128 Newtons), update the cell state and predict the next frame (e.g., peak is expected in 0.35 seconds). Output 32-dimensional temporal features, including dynamic parameters such as "ligament tension change rate" (current per second + 12 Newtons) and "coordination delay time" (e.g., hamstring activation lags quadriceps by 50 milliseconds).

[0042] Cross-Hierarchy Fusion Layer: Spatial features (16-dimensional) and temporal features (32-dimensional) are concatenated into a 48-dimensional vector, which is then compressed and mapped through a fully connected layer (containing 24 neurons).

[0043] A 24-dimensional ligament coordination motion feature set is generated, with core parameters including: ligament group synchronization index (0.0-1.0, healthy value >0.8); skeletal displacement coordination variance (unit: square millimeter, outlier >1.5); and dynamic stability margin (predicting the probability of joint dislocation risk, e.g., 0.05 represents a 5% risk).

[0044] By integrating ligament synergistic motion characteristics and processing them with attention-weighted pooling, a set of key parameters for joint stability is obtained. The attention-based weighted pooling module performs importance ranking and fusion of 24-dimensional features: Attention score generation: Input features are processed through a two-layer fully connected network (layer 1: 24→16 dimensions, layer 2: 16→24 dimensions) to generate attention scores. The activation functions used are ReLU (Rectified Linear Unit) and Softmax (Normalized Exponential Function).

[0045] For example, when the knee is flexed at 45 degrees: the anterior cruciate ligament related features (dimensions 12-15) score is 0.92 (due to bearing 60% of the load); the patellar ligament features (dimensions 8-11) score is 0.35 (low load state).

[0046] Postoperative rehabilitation special case: patients with medial collateral ligament reconstruction had their relevant characteristic scores forcibly increased to 1.2 times (to compensate for the loss of biomechanical perception sensitivity).

[0047] Weighted fusion execution: Weighted summation of feature values ​​by score: Key parameter = Feature value 1 × Score 1 + Feature value 2 × Score 2 + ...

[0048] An activity suppression mechanism is introduced: if a certain feature value exceeds the safety threshold (such as anterior cruciate ligament strain > 1.8 times the baseline value), its attention score is automatically increased to 1.5 times to strengthen the alarm.

[0049] Parameter set generation: Output a 6-dimensional set of key parameters for joint stability, specifically including: Ligament tension balance index (percentage difference between left and right sides, healthy value <15%); Skeletal trajectory deviation variance (unit: square millimeter, normal <0.8); Neuromuscular delay (unit: milliseconds, threshold 200 milliseconds); Dynamic stability coefficient (0.0-1.0, below 0.7 triggers an alarm).

[0050] Based on a set of key parameters for joint stability, radial basis function mapping is performed in conjunction with clinical evaluation criteria to output quantitative indicators of joint stability.

[0051] Clinical standard embedding: The system integrates with the International Knee Documentation Committee Scale (IKDC) for evaluating knee function, with four preset clinical stability thresholds, as shown in Table 1. Table 1 ; Radial basis function calculation: The clinical grade center point is used as the basis function center (e.g., the center point coordinates of "good" grade [0.77, 0.75 square millimeters, 125 milliseconds]).

[0052] Calculate the Euclidean distance between the key parameter set and the center point: Distance = √[(ligament index - 0.77)] 2 + (Offset variance -0.75) 2 + (Delay time - 125) 2 ].

[0053] The similarity is converted using the Gaussian kernel function: similarity = e^(-distance). 2 / (2×radius) 2 The radius can be 0.25.

[0054] Quantitative indicator synthesis: The similarity of the six parameters was calculated separately and then weighted averaged (the weights were set by the doctor: ligament index 50%, deviation variance 30%, and delay time 20%).

[0055] The final output is a joint stability quantification index with a continuous value from 0 to 1.0: Example 1: Patient's ligament index 0.82 (good), deviation variance 0.6 (good), delay 130 milliseconds (good) → overall score 0.78; Example 2: Patient ligament index 0.45 (poor), deviation variance 1.5 (moderate), delay 180 milliseconds (moderate) → overall score 0.42.

[0056] S203, Based on the joint stability quantification index and patient pain feedback data, the dynamic evolution coefficient of the rehabilitation stage is calculated through the rehabilitation entropy model, and the rehabilitation stage quantification index is obtained by nonlinear weighting through the fusion of biomechanical parameters and physiological response parameters. Specifically, the rate of change of biomechanical parameters can be obtained by performing time-series differential processing based on the quantitative indicators of joint stability. The system loads joint stability quantification metrics recorded over time (one sampling point every 0.1 seconds, for a total of 100 points / cycle), and calculates the dynamic change trend using a time-series differential processor. Data preprocessing: Sliding window smoothing (window width 0.5 seconds) is used to eliminate transient disturbances (such as occasional data jitter caused by a patient's cough), and the data fluctuation amplitude is reduced by 60% after smoothing.

[0057] Mark key biomechanical event points (such as the heel strike time in the gait cycle) and align multi-cycle data based on these events.

[0058] Central difference differential calculation: The instantaneous rate of change is calculated using the Central Difference Method: Rate of change = [Current value - Previous value] / Time interval, where the time interval is fixed at 0.1 seconds (i.e., sampling frequency of 10 Hz). For example, the stability index is 0.85 at 0.3 seconds and 0.82 at 0.4 seconds → Rate of change = (0.82 - 0.85) / 0.1 = -0.3 / second.

[0059] For outliers (such as single-point mutations exceeding ±30%), an error correction mechanism is activated: the outlier is replaced by the average of the two points before and after the mutation.

[0060] Extraction of rate of change features: Output three types of biomechanical parameter change rates: short-term volatility (absolute value of the maximum change rate within 10 seconds, e.g., 0.45 / second); long-term trend slope (linear fitting of 1 minute of data, slope unit: change per minute, e.g., -0.12 / minute); and periodic consistency variance (stability of the change rate within the same action cycle, variance threshold 0.05).

[0061] Based on patient pain feedback data, BioBERT feature extraction was used to generate standardized pain feature vectors. Patients submit pain feedback (text description + visual analog scale score) via a mobile app, and the BioBERT processor performs multimodal feature standardization. Text semantic parsing: Input text description (e.g., "stabbing pain on the outside of the knee when bending the knee to 60 degrees"), and extract key information using the pre-trained medical language model BioBERT (Biomedical Bidirectional Encoder Representations from Transformers): Identify medical terms ("stinging" mapped to neuropathic pain code P345.6); Locate the area of ​​pain ("outer side of the knee" corresponds to anatomical zone - Zone 17); Associated action trigger points (“Knee flexion at 60 degrees” is associated with biomechanical phase angle); Output a 128-dimensional text feature vector (containing semantic encodings such as pain type, intensity, and location).

[0062] Scoring data quantification: Visual Analogue Scale (VAS) scores (0-10) are normalized to the 0-1.0 range (e.g., VAS 7 score → 0.7). Emotional features are extracted using a convolutional neural network, combined with emoticon feedback.

[0063] Multimodal fusion standardization: splicing text features (128 dimensions), VAS value (1 dimension), and sentiment features (32 dimensions) into a total of 161-dimensional vectors.

[0064] The pain feature vector is compressed to 64 dimensions through a fully connected layer. Key dimensions include: Dimension 23: Neuropathic pain intensity index (0.82); Dimension 41: Probability of pain triggered by knee flexion at 60 degrees (0.95); Dimension 58: Heatmap encoding of pain area (Zone-17 weight 0.9).

[0065] By integrating the rate of change of biomechanical parameters and the pain feature vector, and processing them through dual-entropy coupling calculation, dynamic evolution coefficients are output. The dual-entropy coupling engine merges two types of heterogeneous data into a rehabilitation progress assessment metric: Data dimension alignment: Biomechanical rate of change (3D) and pain feature vector (64D) are unified into a 32-dimensional space through a fully connected mapping layer (67 input layer neurons, 32 output layer neurons).

[0066] Residual connections are used to preserve original features (such as the rate of change slope value, which is directly passed to the 5th dimension of the output layer).

[0067] Dual-entropy joint computation: Shannon Entropy: Calculates the uncertainty of pain characteristics, reflecting the degree of confusion in pain description.

[0068] Example: The entropy value of "sometimes throbbing pain and sometimes stabbing pain" is 0.92 (high uncertainty), while the entropy value of "persistent dull pain" is 0.35 (low uncertainty).

[0069] Approximate Entropy: Analyzes the pattern complexity of biomechanical rate of change to assess joint stability.

[0070] Example: The rate of change sequence [0.1, -0.2, 0.3, -0.4] has an approximate entropy of 0.87 (high complexity → joint instability).

[0071] Coupling rule: Evolution coefficient = Shannon entropy × 0.7 + approximate entropy × 0.3 (weights are set by clinical experience).

[0072] Dynamic threshold correction: If the pain intensity exceeds the threshold (standardized value > 0.8), the biomechanical entropy weight is forcibly reduced to 0.1 (to avoid the risk of training in pain).

[0073] Output dynamic evolution coefficient from 0 to 1.0: 0.75: rapid recovery period; 0.5-0.75: stable recovery period; <0.5: potential risk period.

[0074] Based on the dynamic evolution coefficient, adaptive weighting of parameters is performed to generate biomechanical-physiological response fusion weights; The adaptive weighted controller dynamically assigns parameter weights based on the rehabilitation stage. Rehabilitation stage determination: When the dynamic evolution coefficient is >0.75, it is determined to be the functional reconstruction period, and the weight of biomechanical parameters increases; when the coefficient is <0.5, it is determined to be the inflammatory response period, and the weight of pain parameters dominates.

[0075] Weighting Calculation Strategy: Basic weighting formula: Biomechanical weight = Sigmoid (evolution coefficient × 5); Pain weight = 1 - Biomechanical weight. The Sigmoid function compresses the input to the 0-1 range.

[0076] Example: Evolution coefficient 0.6 → Biomechanical weight = Sigmoid(3)≈0.95 → Pain weight =0.05.

[0077] Clinical rule infusion: Doctors set special rules: pain weight must be ≥0.7 within 3 days after surgery (even if the evolution coefficient is high); when joint swelling is detected (abnormal pressure sensor signal), the pain weight increases by 30%; output fusion weight vector (2-dimensional): [biomechanical weight, pain weight], such as [0.85, 0.15].

[0078] Based on the fusion weights and clinical thresholds, a quantitative index for the rehabilitation stage is output through S-shaped function transformation.

[0079] The Rehabilitation Index Generator integrates weights and clinical criteria to generate the final assessment value. Parameter fusion calculation: Input biomechanical parameter change rate (3D) and pain feature vector (64D), and weight them according to the fusion weight: Fusion value = Σ(biomechanical parameter_i × biomechanical weight) + Σ(pain feature_j × pain weight).

[0080] Example: Biomechanical weight 0.8, pain weight 0.2 → Fusion value = (rate of change slope × 0.8) + (neural pain index × 0.2).

[0081] Clinical threshold mapping: Load the clinical threshold matrix for rehabilitation staging. The matrix information is shown in Table 2 for example: Table 2 ; S-shaped function normalization: The Sigmoid function is applied to compress the fusion value to the 0-1 range: recovery index = 1 / [1 + e^(-steepness coefficient × (fusion value - 0.5))], where the steepness coefficient can be set to 12 to control the oversensitivity.

[0082] Example: Fusion value 0.6 → Recovery index = 1 / (1+e^(-12×0.1))≈0.77 (functional period).

[0083] Dynamically corrected output: If the "nocturnal resting pain" flag is activated in the pain characteristics (value > 0.7), the rehabilitation index is forcibly reduced by 20%. The final output is a quantitative index of the rehabilitation stage ranging from 0 to 1.0, with an accuracy of ±0.01.

[0084] Clinical scenario: Patient on the 10th day after anterior cruciate ligament reconstruction surgery Temporal differential processing: Joint stability index increased from 0.38 to 0.42 (5-minute data) → rate of change +0.08 / min.

[0085] Pain feature extraction: Feedback text "swelling pain at the lower edge of the patella when knee is flexed at 40 degrees" → BioBERT outputs pain area Zone-05, intensity 0.75.

[0086] Dual-entropy coupling calculation: Pain description has low Shannon entropy (0.25, consistent description); biomechanical approximation has high entropy (0.81, significant joint movement) → evolution coefficient = 0.25 × 0.7 + 0.81 × 0.3 = 0.41; Adaptive weighting: Evolution coefficient 0.41 → Biomechanical weight = Sigmoid (2.05) ≈ 0.88 → Pain weight = 0.12.

[0087] Recovery index generation: Fusion value = 0.08 × 0.88 + 0.75 × 0.12 = 0.16 → Sigmoid mapping output index 0.29 (acute phase); triggers safety training instruction set downgrade.

[0088] S204. Based on the quantitative index of the rehabilitation stage, a personalized rehabilitation action sequence is constructed using an adversarial generative network. The medical compliance of the action sequence is verified by a discriminator, and a safe training instruction set that conforms to the current rehabilitation stage is generated. Specifically, based on the quantitative index of the rehabilitation stage, the medical knowledge base can be retrieved and processed to generate a safety boundary matrix for joint range of motion; After receiving the rehabilitation stage quantitative index (range 0-1.0), the system activates the medical knowledge base retrieval engine. This engine links to clinical guideline databases (such as the AAOS Orthopedic Rehabilitation Guidelines), individual patient records (age, surgical procedure, comorbidities), and real-time biomechanical data. Index mapping to clinical staging: When the quantitative index is ≤0.3, it is determined to be the acute phase, and the postoperative 0-2 week safety rules are applied: the upper limit of the knee flexion angle is 30 degrees, and the weight-bearing limit is 10% of body weight (about 7 kg).

[0089] An index of 0.3-0.6 corresponds to the subacute phase, with the flexion angle expanding to 90 degrees and the load increasing to 30% of body weight (approximately 21 kg).

[0090] When the index is >0.6, the system enters the functional period, allowing full range of motion but limiting rotational torque (e.g., internal rotation torque threshold of 5 N·m).

[0091] Multi-dimensional boundary generation: Construct a three-dimensional security boundary matrix, with the following dimensions: Angle boundaries (e.g., knee flexion 0-45 degrees); Load boundaries (e.g., patellofemoral joint pressure < 150% of body weight); Velocity limits (e.g., knee extension angular velocity ≤ 15 degrees per second).

[0092] Dynamic correction mechanism: If the pressure sensor detects a swelling signal (pressure distribution variation coefficient > 0.25), the matrix automatically tightens the angle boundary by 20%.

[0093] Example of a matrix structure: ; Based on the safety boundary matrix, an LSTM-Transformer hybrid network is used to generate the initial action sequence; LSTM-Transformer hybrid generator integrates temporal modeling and global attention mechanisms: LSTM (Long Short-Term Memory) layer: Input the security boundary matrix (12-dimensional parameters) and set 128 memory cells.

[0094] Capturing motion continuity: For example, when standing up from a seated position, LSTM ensures that the knee flexion angle smoothly transitions from 90 degrees to 0 degrees (with a change of ≤3 degrees per frame) to avoid joint impact.

[0095] Output a 32-dimensional primary motion sequence (including timestamps, joint angles, and velocity baselines).

[0096] Transformer coding layer: Introducing a self-attention mechanism to analyze multi-joint collaborative relationships: Calculate the attention scores for the hip-knee-ankle joints (e.g., with hip flexion, the knee joint score is 0.92 and the ankle joint score is 0.75). Weighted generation of coordinated movement instructions (such as "hip flexion of 30 degrees must be accompanied by knee flexion of 15 degrees").

[0097] Positional encoding marks the timing of actions to prevent sequence errors (e.g., knee extension must be followed by heel strike).

[0098] Action sequence optimization: The output of LSTM and Transformer (32+64 dimensions) is integrated through residual connection, and the initial action sequence is generated through a fully connected layer: containing 20 action frames (each frame is 0.5 seconds long); each frame stores 6 parameters (hip / knee / ankle angle, velocity, load, duration).

[0099] Example sequence frames: Frame 1: Knee flexion 10 degrees (speed 5 degrees / second, load 15% body weight); Frame 5: Knee extension 20 degrees (speed 8 degrees / second, load 30% body weight).

[0100] Based on the initial action sequence, the joint angle space-constrained action is obtained through inverse kinematic mapping. The inverse kinematics mapper translates abstract motion instructions into executable instructions for the exoskeleton: Ergonomic model loading: The lower limb skeletal chain was constructed based on the patient's bone parameters (femur length L_femur=420mm, tibia length L_tibia=380mm).

[0101] Define mechanical constraints such as knee joint rotation center offset (Offset=5mm) and hip joint movement cone angle (ConeAngle=120 degrees).

[0102] Reverse solution process: Input the target foot displacement (e.g., forward 200mm), and solve for the joint angle combination: Step 1: Calculate the ankle dorsiflexion angle θ_ankle (range -20 to 30 degrees); Step 2: Use geometric constraints to inversely deduce the knee flexion angle θ_knee (numerical solution such as θ_knee=35 degrees); Step 3: Verify whether the hip joint angle θ_hip is within the safety matrix (e.g., θ_hip = 25 degrees < 45 degrees threshold).

[0103] Real-time collision detection: If the minimum distance between the thigh and calf is less than 50mm (anti-self-collision), trigger angle replanning.

[0104] Enhanced spatial constraints: Output joint angle spatial constraint actions, each action includes: three-dimensional joint angle (accuracy ±0.5 degrees); six-dimensional spatial coordinates (such as the XYZ coordinates of the patella center point); biomechanical effectiveness markers (such as the ligament tension safety value = 1).

[0105] Based on joint angle spatial constraints, a three-dimensional biomechanical discriminant is used to verify and screen out medically compliant movements. The three-dimensional biomechanical discriminator is a multi-layer neural network, which performs multi-physics simulation verification. Ligament tension verification module: Simulated stress on the anterior cruciate ligament: when the knee is flexed at 45 degrees, the tension threshold is ≤ 150% of body weight (approximately 100 kg).

[0106] If the test exceeds the limit (e.g., tension value 110 kg), the flag is set to 0 (violation).

[0107] Joint surface pressure analysis module: Calculation of tibiofemoral joint contact pressure based on finite element mesh (FEM mesh): The health threshold is <3MPa (e.g., 2.4MPa was measured when the patient weighed 70 kg). An alarm is triggered if there is an abnormal pressure distribution (such as the pressure on the inner platform being 150% greater than that on the outer platform).

[0108] Dynamic stability assessment: Calculate the rate of change of angular momentum: the threshold of trunk swing angular velocity per second ±10 degrees during the gait cycle.

[0109] Output a set of medical compliance actions, with the following selection criteria: All physical parameters are within the safety boundaries; Inter-frame abrupt changes in continuous motion < threshold (e.g., angle change rate ≤ 20 degrees per second); Energy efficiency > baseline value (oxygen consumption per unit of movement ≤ 0.8 times the average value).

[0110] By integrating medical compliance procedures with diverse reward mechanisms, and through instruction coding processing, a set of safety training instructions is output.

[0111] Instruction encoding engine integrates security requirements with training diversity: Diversity reward mechanism: Action mutation algorithm: Generates variants (such as speed ±10%, amplitude ±5%) based on compliant actions.

[0112] Novelty score: If the cosine similarity between the new action and the previous training sequence is less than 0.7, the reward coefficient is increased by 0.2.

[0113] Example: The basic knee flexion movement (30 degrees / second) leads to: Variant A: Slow speed control type (15 degrees / second); Variant B: Resistance Enhanced (increases resistance by 10 kg).

[0114] Instruction encoding protocol: A hierarchical instruction structure is adopted: Top level: Action type code (e.g., knee bend = KC01); Middle layer: Parameter set (angle = 30 degrees, speed = 5 degrees / second, load = 20 kg); The bottom layer: security check code (SHA-256 hash value to verify data integrity).

[0115] Example of generation instructions: KC01|ANG:30,VEL:5,LOAD:200|SHA:9f86d081... Dynamically optimized output: Real-time monitoring of patient fatigue index (e.g., a 30% decrease in electromyographic signal amplitude) automatically reduces the instruction set size by 50%. Final output: Safe training instruction set containing 5-20 movement instructions, each accompanied by a three-dimensional biomechanical safety certificate.

[0116] Clinical scenario: Patients after knee replacement surgery (recovery index 0.5) Security boundary generation: Search the knowledge base → Subacute phase: Knee flexion limit 90 degrees, weight limit 30 kg.

[0117] Action sequence generation: LSTM generates the basic sit-stand transition sequence (10 frames). Add a synergistic command to the Transformer (simultaneous flexion and extension of the hip and knee when standing up).

[0118] Inverse kinematic mapping: Objective: Raise the seat by 100mm → Solve for θ_hip=25°, θ_knee=50°.

[0119] Three-dimensional biomechanical verification: Ligament tension verification: Quadriceps tendon under stress of 65 kg (< threshold 80 kg) → Compliant; Joint surface pressure: Tibial plateau 2.1MPa (<3MPa) → Compliant.

[0120] Instruction set generation: Basic commands: SIT_STAND|ANG_HIP:25,ANG_KNEE:50,LOAD:25kg|SHA:7d793037... Variation: Resistance type (increases resistance by 5 kg, hash value changes).

[0121] S205, drive the wearable exoskeleton to perform rehabilitation training according to the safety training instruction set, and simultaneously monitor neuromuscular activation signals through the brain-computer interface to verify the consistency between training actions and neural control.

[0122] Specifically, the exoskeleton can be driven to perform training movements by decoding motor signals according to the safety training instruction set. After receiving the security training instruction set, the system first performs data deconstruction and verification through the instruction parsing engine. Each instruction adopts a layered encoding structure: Action type code (4-digit code, such as KC01 for knee bending action); Parameter group (key-value pair format, such as ANG=30,VEL=5,LOAD=10 indicating an angle of 30 degrees, a speed of 5 degrees per second, and a load of 10 kg); Security certificate (including a 64-bit SHA-256 hash checksum and a biomechanical security level label, such as Level A); The parsing engine performs four security checks: Hash verification: Calculate the SHA-256 digest value of the instruction (a 64-bit hexadecimal string such as 9f86d081...), and compare it with the value embedded in the certificate. The match rate must be 100%. Timeliness verification: Check the difference between the instruction generation timestamp and the actual execution time; if it exceeds 500 milliseconds, it is considered an expired instruction. Device compatibility matching: Retrieve the exoskeleton model database (e.g., the maximum knee flexion angle of model X-Joint v3.0 is 120 degrees) and filter out commands that exceed the limit; Security level screening: Only A-level (fully compliant) and B-level (boundary compliance requires monitoring) instructions are executed, and C-level high-risk instructions are automatically discarded.

[0123] After verification, the motor signal decoder initiates the physical control conversion: Angle mapping: Based on the exoskeleton gear reduction ratio (e.g., 12:1), the angle value is converted into the number of motor rotations (30 degrees → 0.25 rotations). Speed ​​conversion: PWM (Pulse Width Modulation) technology is used to convert speed parameters into duty cycle (5 degrees per second → 35% duty cycle, i.e., high level lasts for 3.5 milliseconds / cycle). Load control: Based on the current-torque curve (preset parameter: 1 kg = 0.075 amperes), the load value is mapped to the current threshold (10 kg → 0.75 amperes).

[0124] The generated motor drive signal is transmitted to the exoskeleton joint motor via a CAN bus (Controller Area Network, transmission rate 1 Mbps), and the motor encoder provides real-time feedback on the actual angle (accuracy ±0.1 degrees). If an execution deviation > ±5% is detected (e.g., a command of 30 degrees is measured as ≤28.5 degrees or ≥31.5 degrees), emergency braking is immediately triggered and error code E102 is reported.

[0125] Based on brain-computer interface signals, blind source separation processing is used to extract neuromuscular activation features; The brain-computer interface (BCI) uses a 64-channel dry electrode cap to acquire raw electroencephalogram (EEG) signals at a sampling rate of 2 kHz. Signal preprocessing includes a four-stage filtering chain. Power frequency notch filter: Eliminates 50 Hz mains interference (attenuation > 40 dB); Bandpass filtering: preserves the effective frequency band of 0.5-200 Hz (Theta wave to Gamma wave); Wavelet denoising: The Daubechies 4 wavelet basis was used for 5-level decomposition to filter out electromyographic artifacts; Independent Component Analysis (ICA): Separates interference sources such as eye movement and electrocardiogram (components with a variance contribution rate <0.3 are considered noise).

[0126] The preprocessed signal is input into the blind source separation processor, whose core is the JADE algorithm (joint approximate diagonalization): Calculate the fourth-order cumulant matrix and achieve approximate diagonalization of the matrix through Jacobi rotation; Eight neural signal sources were isolated (such as the μ rhythmic wave of the C3 channel in the motor cortex and the β wave of the S1 channel in the sensory cortex). Output source signal power spectral density plot (frequency resolution 0.5 Hz).

[0127] Neuromuscular activation features were extracted from three dimensions: Motor intention intensity: Analyze event-related desynchronization (ERD) in the motor cortex. When the power of the β wave (13-30 Hz) decreases by more than 40% for 100 milliseconds, it is marked as an active motor intention (intensity value = 1 - power residual rate). Muscle recruitment index: Detects the burst energy of gamma waves (30-80 Hz) in the primary motor cortex (M1 area). Effective recruitment is considered when the peak energy is greater than twice the baseline (index value = peak value / baseline). Nerve fatigue coefficient: Calculate the power ratio of alpha wave (8-13 Hz) to beta wave. When the ratio is greater than 1.5, a fatigue alarm is triggered (coefficient value = actual ratio / 1.5).

[0128] The final 32-dimensional feature vector is generated (e.g., [left leg flexion intention = 0.87, quadriceps recruitment = 0.92, fatigue coefficient = 0.35...]), and is updated every 200 milliseconds.

[0129] By integrating actual motion trajectories and neuromuscular activation features, and processing them through mutual information entropy calculation, a neural control matching degree is generated. The actual motion trajectory data comes from two types of sensors: Exoskeleton joint encoder (angle accuracy ±0.1 degrees, sampling rate 100 Hz). Inertial measurement unit (IMU, angular velocity accuracy ±0.5 degrees / second, acceleration accuracy ±0.01g).

[0130] The spatiotemporal alignment module performs two-level synchronization: Hardware-level synchronization: The BCI and exoskeleton clock are aligned via the PTP protocol (Precision Time Protocol, accuracy ±1 microsecond); Motion phase alignment: The data stream is segmented based on gait cycle events (such as heel strike trigger signals) to divide the biomechanical phases such as the standing phase and the swing phase.

[0131] Aligned data is input into the mutual information entropy calculation engine: Data discretization: Continuous angle values ​​are divided into 10-degree intervals (e.g., 0-10 degrees, 10-20 degrees, etc.), and neural features are normalized to the 0-1 interval and then divided into 20 equal levels; Joint probability construction: statistically analyze the frequency of occurrence of neural-mechanical combinations (e.g., the probability of detecting movement intention is 0.93 when the knee joint is in the 30-40 degree range). Mutual information entropy calculation: Based on the KL divergence (Kullback-Leibler Divergence) quantification of statistical dependence, it is used as the neural control matching degree output (range 0-1.0), and the grading criteria are as follows: 0.85-1.0: Excellent (neural signals precisely guide mechanical movement with a delay of <50 milliseconds); 0.60-0.85: Acceptable (acceptable neural conduction delay exists); <0.60: Disconnection (immediate intervention required).

[0132] For example, when a patient performs a knee extension movement, the motor cortex signal leads the actual movement by 80 milliseconds, with a matching degree of 0.73.

[0133] Based on the neural control matching degree, reinforcement learning is triggered to dynamically adjust the processing and output a neural-mechanical synchronization verification report.

[0134] The reinforcement learning dynamic adjuster is constructed as a Markov decision process (MDP), and the framework includes three main elements: State space: A 12-dimensional state vector encodes the following parameters: neural matching degree (discretized as high ≥0.8, medium 0.6-0.8, low <0.6); rehabilitation stage index (0-1.0 continuous value); real-time fatigue coefficient (0-2.0).

[0135] Motion Space: Defines five types of adjustment commands: A1: Reduce exoskeleton assistance by 10%; A2: Increase visual / auditory guidance intensity; A3: Shorten single training session duration by 20%; A4: Switch to low-complexity movements (such as sitting training instead of standing); A5: Maintain current parameters.

[0136] Reward function: Dynamic scoring rules: Matching degree increased by 0.1 → +5 points; Pain level increased by 1 level → -10 points; Energy consumption per unit movement decreased by 15% → +3 points.

[0137] The decision-making process uses the Q-learning algorithm (state-action value iteration): Initialize the Q-value table (12 states × 5 action matrix, initial value = 0); The strategy is updated every 5 minutes: learning rate α = 0.1, discount factor γ = 0.9; Example: When the state = [matching degree = medium, fatigue coefficient = 1.8], the Q value of choosing action A3 (shortening the duration) increases from 1.2 to 3.5, becoming the optimal solution.

[0138] The adjustment results are written into the neuro-mechanical synchronization verification report, which includes four parts: Synchronization performance summary: Matching degree time series curve (1 data point per minute); Dynamically adjust logs: such as "14:25 Matching degree 0.75 → Execute A1 (reduce assist by 10%)"; Abnormal event log: such as "At 14:30, a neural signal delay of 250 milliseconds was detected (code W207)"; Clinical recommendations include: "The current average matching degree is 0.78. It is recommended to add a mirror neuron training module."

[0139] Clinical scenario example: Patients undergoing knee replacement surgery undergoing sit-stand training. Instruction execution phase: Analyzing instructions: SIT_STAND|ANG_HIP=25,ANG_KNEE=50,LOAD=15kg|HASH:7d793037...→Motor output: 0.21 rotations (25 degrees) for the hip joint and 0.42 rotations (50 degrees) for the knee joint. Neural feature extraction: BCI detection showed 52% inhibition of β waves in the motor cortex (intensity value 0.93), and the peak energy of γ wave bursts was 3.8 times the baseline (recruitment index 0.95). Matching degree calculation: Neural signal leads mechanical action by 70 milliseconds → Mutual information entropy matching degree = 0.81; Dynamic adjustment of decision: Q-learning strategy: State [matching degree = high, fatigue coefficient = 0.4] → select A5 (maintaining parameter); Verification report output: Neuro-mechanical synchronization validation report: Patient ID: P20240812, Date: 2025-04-20; Average match score: 0.82 (Grade: Excellent); Adjustment record: 14:15 [Status: High Match + Low Fatigue] → A5 (Maintain Parameters); 14:30 [Status: Medium Fatigue] → A3 (Shorten Duration by 20%); Abnormal events: None; Recommendation for tomorrow: Load can be increased to 18kg (currently 15kg).

[0140] As can be seen, a three-dimensional spatiotemporal feature vector of joint motion is obtained based on joint motion data, electromyographic signals, and pressure distribution information collected by multimodal sensors. Based on the three-dimensional spatiotemporal feature vector, key parameters of joint functional state are extracted using a spatiotemporal graph convolutional neural network to obtain a quantitative index of joint stability. Based on the quantitative index of joint stability and patient pain feedback data, a quantitative index of the rehabilitation stage is obtained. Based on the quantitative index of the rehabilitation stage, a set of safe training instructions that conforms to the current rehabilitation stage is generated. The wearable exoskeleton is driven to perform rehabilitation training according to the set of safe training instructions, and neuromuscular activation signals are monitored simultaneously through a brain-computer interface to verify the consistency between training actions and neural control. This enables precise, continuous, and adaptive monitoring and management of the entire joint rehabilitation process.

[0141] Another embodiment of the present invention provides an intelligent orthopedic joint rehabilitation monitoring system, see [link to relevant documentation]. Figure 3 The system may include: The fusion module 301 is used to perform dynamic biomechanical modeling based on joint motion data, muscle electrical signals and pressure distribution information collected by multimodal sensors, and obtain a three-dimensional spatiotemporal feature vector of joint motion through a spatiotemporal feature fusion algorithm. The extraction module 302 is used to extract key parameters of joint functional state based on the three-dimensional spatiotemporal feature vector using a spatiotemporal graph convolutional neural network, and capture the coordinated motion pattern of joint ligaments and bones through an adaptive topology structure to obtain quantitative indicators of joint stability. The calculation module 303 is used to calculate the dynamic evolution coefficient of the rehabilitation stage based on the joint stability quantification index and the patient pain feedback data through the rehabilitation entropy model, and obtain the rehabilitation stage quantification index by nonlinear weighting through the fusion of biomechanical parameters and physiological response parameters. The construction module 304 is used to construct a personalized rehabilitation action sequence based on the quantitative index of the rehabilitation stage using an adversarial generative network, verify the medical compliance of the action sequence through a discriminator, and generate a safe training instruction set that conforms to the current rehabilitation stage. The monitoring module 305 is used to drive the wearable exoskeleton to perform rehabilitation training according to the safety training instruction set, and simultaneously monitor neuromuscular activation signals through the brain-computer interface to verify the consistency between training actions and neural control.

[0142] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0143] Specifically, in this embodiment, the storage medium can be configured to store a computer program for performing the following steps: S201: Based on the joint motion data, muscle electromyography signals and pressure distribution information collected by multimodal sensors, dynamic biomechanical modeling is performed through a spatiotemporal feature fusion algorithm to obtain the three-dimensional spatiotemporal feature vector of joint motion. S202, Based on the three-dimensional spatiotemporal feature vector, the key parameters of joint functional state are extracted using a spatiotemporal graph convolutional neural network, and the coordinated motion pattern of joint ligaments and bones is captured through an adaptive topology structure to obtain a quantitative index of joint stability. S203, Based on the joint stability quantification index and patient pain feedback data, the dynamic evolution coefficient of the rehabilitation stage is calculated through the rehabilitation entropy model, and the rehabilitation stage quantification index is obtained by nonlinear weighting through the fusion of biomechanical parameters and physiological response parameters. S204. Based on the quantitative index of the rehabilitation stage, a personalized rehabilitation action sequence is constructed using an adversarial generative network. The medical compliance of the action sequence is verified by a discriminator, and a safe training instruction set that conforms to the current rehabilitation stage is generated. S205, drive the wearable exoskeleton to perform rehabilitation training according to the safety training instruction set, and simultaneously monitor neuromuscular activation signals through the brain-computer interface to verify the consistency between training actions and neural control.

[0144] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0145] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0146] Specifically, in this embodiment, the processor can be configured to perform the following steps via a computer program: S201: Based on the joint motion data, muscle electromyography signals and pressure distribution information collected by multimodal sensors, dynamic biomechanical modeling is performed through a spatiotemporal feature fusion algorithm to obtain the three-dimensional spatiotemporal feature vector of joint motion. S202, Based on the three-dimensional spatiotemporal feature vector, the key parameters of joint functional state are extracted using a spatiotemporal graph convolutional neural network, and the coordinated motion pattern of joint ligaments and bones is captured through an adaptive topology structure to obtain a quantitative index of joint stability. S203, Based on the joint stability quantification index and patient pain feedback data, the dynamic evolution coefficient of the rehabilitation stage is calculated through the rehabilitation entropy model, and the rehabilitation stage quantification index is obtained by nonlinear weighting through the fusion of biomechanical parameters and physiological response parameters. S204. Based on the quantitative index of the rehabilitation stage, a personalized rehabilitation action sequence is constructed using an adversarial generative network. The medical compliance of the action sequence is verified by a discriminator, and a safe training instruction set that conforms to the current rehabilitation stage is generated. S205, drive the wearable exoskeleton to perform rehabilitation training according to the safety training instruction set, and simultaneously monitor neuromuscular activation signals through the brain-computer interface to verify the consistency between training actions and neural control.

[0147] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A smart orthopedic joint rehabilitation monitoring method, characterized in that, The method includes: Based on joint motion data, muscle electromyography signals, and pressure distribution information collected by multimodal sensors, dynamic biomechanical modeling is performed using a spatiotemporal feature fusion algorithm to obtain a three-dimensional spatiotemporal feature vector of joint motion. Based on the aforementioned three-dimensional spatiotemporal feature vector, key parameters of joint functional state are extracted using a spatiotemporal graph convolutional neural network. The coordinated motion pattern of joint ligaments and bones is captured through an adaptive topology structure to obtain quantitative indicators of joint stability. Based on the joint stability quantification index and patient pain feedback data, the dynamic evolution coefficient of the rehabilitation stage is calculated by the rehabilitation entropy model. The rehabilitation stage quantification index is obtained by nonlinear weighting of biomechanical parameters and physiological response parameters. Based on the quantitative index of the rehabilitation stage, a personalized rehabilitation action sequence is constructed using an adversarial generative network. The medical compliance of the action sequence is verified by a discriminator, and a safe training instruction set that conforms to the current rehabilitation stage is generated. The wearable exoskeleton is driven to perform rehabilitation training according to the safety training instruction set, and neuromuscular activation signals are monitored simultaneously through the brain-computer interface to verify the consistency between training movements and neural control.

2. The method according to claim 1, characterized in that, The process involves using joint motion data, electromyographic signals, and pressure distribution information collected by multimodal sensors to perform dynamic biomechanical modeling through a spatiotemporal feature fusion algorithm, resulting in a three-dimensional spatiotemporal feature vector of joint motion, including: Based on the joint motion data, a six-degree-of-freedom trajectory Kalman filter is performed to obtain the noise-reduced joint motion trajectory. Based on electromyographic signals and pressure distribution information, a time-domain phase alignment algorithm is used to generate a biomechanical coupling feature matrix. By fusing joint motion trajectories with biomechanical coupling feature matrices and processing them through tensor cross decomposition, a three-dimensional spatiotemporal feature vector is output.

3. The method according to claim 2, characterized in that, Based on the three-dimensional spatiotemporal feature vector, key parameters of joint functional state are extracted using a spatiotemporal graph convolutional neural network. An adaptive topology is used to capture the coordinated motion patterns of joint ligaments and bones, resulting in quantitative indicators of joint stability, including: Based on the three-dimensional spatiotemporal feature vectors, dynamic skeletal topology construction is performed to generate an adaptive biomechanical graph structure; Based on an adaptive biomechanical graph structure, hierarchical spatiotemporal graph convolution processing is used to extract ligament cooperative motion features; By integrating ligament synergistic motion characteristics and processing them with attention-weighted pooling, a set of key parameters for joint stability is obtained. Based on a set of key parameters for joint stability, radial basis function mapping is performed in conjunction with clinical evaluation criteria to output quantitative indicators of joint stability.

4. The method according to claim 3, characterized in that, Based on the joint stability quantification index and patient pain feedback data, the dynamic evolution coefficient of the rehabilitation stage is calculated using the rehabilitation entropy model. A quantitative index of the rehabilitation stage is obtained by nonlinearly weighting the fusion of biomechanical parameters and physiological response parameters, including: Based on the quantitative index of joint stability, time-series differential processing is performed to obtain the rate of change of biomechanical parameters; Based on patient pain feedback data, BioBERT feature extraction was used to generate standardized pain feature vectors. By integrating the rate of change of biomechanical parameters and the pain feature vector, and processing them through dual-entropy coupling calculation, dynamic evolution coefficients are output. Based on the dynamic evolution coefficient, adaptive weighting of parameters is performed to generate biomechanical-physiological response fusion weights; Based on the fusion weights and clinical thresholds, a quantitative index for the rehabilitation stage is output through S-shaped function transformation.

5. The method according to claim 4, characterized in that, Based on the quantitative index of the rehabilitation stage, a personalized rehabilitation action sequence is constructed using an adversarial generative network. The medical compliance of the action sequence is verified by a discriminator, and a safe training instruction set conforming to the current rehabilitation stage is generated, including: Based on the quantitative index of the rehabilitation stage, the medical knowledge base is retrieved and processed to generate a safety boundary matrix for joint range of motion. Based on the safety boundary matrix, an LSTM-Transformer hybrid network is used to generate the initial action sequence; Based on the initial action sequence, the joint angle space-constrained action is obtained through inverse kinematic mapping. Based on joint angle spatial constraints, a three-dimensional biomechanical discriminant is used to verify and screen out medically compliant movements. By integrating medical compliance procedures with diverse reward mechanisms, and through instruction coding processing, a set of safety training instructions is output.

6. The method according to claim 5, characterized in that, The process of driving the wearable exoskeleton to perform rehabilitation training according to the safety training instruction set, while simultaneously monitoring neuromuscular activation signals through a brain-computer interface to verify the consistency between training movements and neural control, includes: Based on the safety training instruction set, the motor signals are decoded and processed to drive the exoskeleton to perform training actions; Based on brain-computer interface signals, blind source separation processing is used to extract neuromuscular activation features; By integrating actual motion trajectories and neuromuscular activation features, and processing them through mutual information entropy calculation, a neural control matching degree is generated. Based on the neural control matching degree, reinforcement learning is triggered to dynamically adjust the processing and output a neural-mechanical synchronization verification report.

7. A smart orthopedic joint rehabilitation monitoring system, characterized in that, The system includes: The fusion module is used to perform dynamic biomechanical modeling based on joint motion data, electromyographic signals and pressure distribution information collected by multimodal sensors, and obtain three-dimensional spatiotemporal feature vectors of joint motion through spatiotemporal feature fusion algorithms. The extraction module is used to extract key parameters of joint functional state based on the three-dimensional spatiotemporal feature vector using a spatiotemporal graph convolutional neural network, and capture the coordinated motion pattern of joint ligaments and bones through an adaptive topology structure to obtain quantitative indicators of joint stability. The calculation module is used to calculate the dynamic evolution coefficient of the rehabilitation stage based on the joint stability quantification index and patient pain feedback data through the rehabilitation entropy model, and obtain the rehabilitation stage quantification index by nonlinear weighting through the fusion of biomechanical parameters and physiological response parameters. The module is used to construct personalized rehabilitation action sequences based on the quantitative index of the rehabilitation stage, using an adversarial generative network, verifying the medical compliance of the action sequences through a discriminator, and generating a set of safe training instructions that conforms to the current rehabilitation stage. The monitoring module is used to drive the wearable exoskeleton to perform rehabilitation training according to the safety training instruction set, and simultaneously monitor neuromuscular activation signals through the brain-computer interface to verify the consistency between training actions and neural control.

8. The system according to claim 7, characterized in that, The fusion module is specifically used for: Based on the joint motion data, a six-degree-of-freedom trajectory Kalman filter is performed to obtain the noise-reduced joint motion trajectory. Based on electromyographic signals and pressure distribution information, a time-domain phase alignment algorithm is used to generate a biomechanical coupling feature matrix. By fusing joint motion trajectories with biomechanical coupling feature matrices and processing them through tensor cross decomposition, a three-dimensional spatiotemporal feature vector is output.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-6 when it is run.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-6.

Images