An orthopedic patient rehabilitation self-checking method and system

Abstract

This invention discloses a method and system for self-assessment of orthopedic patient rehabilitation. The method includes: collecting multimodal biomechanical data of patients performing standardized rehabilitation movements to generate a spatiotemporally synchronized motion-mechanical coupling dataset; inputting the motion-mechanical coupling dataset into a variational autoencoder to form a three-dimensional biomechanical feature matrix; inputting the three-dimensional biomechanical feature matrix into a dynamic graph neural network to output dual-channel prediction results of joint stability risk probability and tissue healing abnormality probability; spatiotemporally aligning the dual-channel prediction results with a personalized knowledge graph constructed from the patient's historical rehabilitation data to generate a quantitative risk level; generating a dynamic heatmap of anatomical structure based on the quantitative risk level, and outputting an interactive three-dimensional report containing the location of key rehabilitation bottlenecks. Using this invention, a three-dimensional visual interactive self-assessment of orthopedic patient rehabilitation can be achieved, improving the objectivity, accuracy, and clinical applicability of rehabilitation assessment.

Description

Technical Field

[0001] This invention belongs to the field of medical technology, specifically a method and system for self-examination of orthopedic patients' rehabilitation. Background Technology

[0002] In the process of orthopedic postoperative rehabilitation, the assessment of patients' functional recovery usually relies on doctors' manual examination and subjective judgment, which has problems such as inconsistent assessment standards and insufficient quantitative indicators. Traditional rehabilitation assessment methods mostly use single-dimensional measurement methods, such as joint range of motion measurement or muscle strength testing, which are difficult to comprehensively reflect the coordinated recovery status of the patient's skeletal-muscle-nervous system. In existing technologies, although some systems have introduced biomechanical detection methods such as motion capture and surface electromyography, there is a lack of effective multimodal data fusion algorithms, resulting in the inability to accurately quantify key indicators such as joint stability and tissue healing progress. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for self-examination of orthopedic patient rehabilitation, in order to overcome the shortcomings of the prior art, realize three-dimensional visual interactive self-examination of orthopedic patient rehabilitation, and improve the objectivity, accuracy and clinical applicability of rehabilitation assessment.

[0004] One embodiment of this application provides a self-examination method for orthopedic patient rehabilitation, the method comprising: Collect multimodal biomechanical data of patients performing standardized rehabilitation movements to generate a spatiotemporally synchronized motion-mechanics coupled dataset; The motion-mechanics coupled dataset is input into a variational autoencoder, and the joint range vector, muscle load vector, and neural control coordination vector are separated through deentanglement representation learning to form a three-dimensional biomechanical feature matrix. The three-dimensional biomechanical feature matrix is ​​input into a dynamic graph neural network, and the connection weights of the graph nodes are dynamically adjusted according to the individual patient's rehabilitation stage. The dual-channel prediction results of joint stability risk probability and tissue healing abnormality probability are output. The dual-channel prediction results are spatiotemporally aligned with a personalized knowledge graph constructed from the patient's historical rehabilitation data. A nonlinear time warping algorithm is used to calculate the deviation between the current rehabilitation status and the ideal rehabilitation trajectory, generating a quantitative risk level. Based on the quantified risk level, a dynamic heat map of the anatomical structure is generated, a biomechanical load streamline animation is superimposed on the damaged tissue area, and an interactive 3D report containing the location of key rehabilitation bottlenecks is output.

[0005] Optionally, the step of collecting multimodal biomechanical data of patients performing standardized rehabilitation movements to generate a spatiotemporally synchronized motion-mechanical coupling dataset includes: Based on the patient's joint angular displacement sequence obtained by the optical motion capture system, the rigid body transformation of the skeleton is calculated to obtain high-precision motion trajectory data. Based on the raw electromyographic signals collected by the surface electromyography sensor array, power frequency interference was removed by wavelet packet decomposition to obtain denoised muscle activation time series data. Based on motion trajectory data and muscle activation timing data, a dynamic time warp algorithm is used to align timestamps and generate a spatiotemporally synchronized skeletal-muscle joint timing stream. Based on the joint temporal flow of the skeletal-muscle system, the three-dimensional force vectors measured by the plantar pressure plate are fused, and the joint torque vectors are calculated through the inverse dynamics model to output the motion-mechanics coupling dataset.

[0006] Optionally, the step of inputting the motion-mechanics coupled dataset into a variational autoencoder, and learning to separate the joint range vector, muscle load vector, and neural control coordination vector through deentanglement representation to form a three-dimensional biomechanical feature matrix includes: Based on the action-mechanics coupled dataset, spatiotemporal features are extracted using a convolutional encoder, and the initial encoding of the latent space is output. Based on the initial encoding of the latent space, the entangled subspace is separated by adversarial regularization constraints to obtain three types of latent variables: orthogonalized joint mobility, muscle load, and neural coordination. Based on the three types of latent variables, the joint kinematic parameters, electromyographic envelope signals and motor coordination indices are reconstructed through a multi-branch decoder, and the unentangled feature vector set is output. Based on the unentangled feature vector set, it is stacked into a tensor along the time axis, and a three-dimensional biomechanical feature matrix is ​​generated by normalizing the feature dimension.

[0007] Optionally, the step of inputting the three-dimensional biomechanical feature matrix into a dynamic graph neural network, dynamically adjusting the connection weights of graph nodes according to the individual patient's rehabilitation stage, and outputting a dual-channel prediction result of joint stability risk probability and tissue healing abnormality probability includes: Based on the three-dimensional biomechanical feature matrix, an initial graph structure is constructed with joints as nodes and biomechanical associations as edges, and a static topological graph is output. Based on the patient's recovery stage label, the edge weight update matrix of the static topology graph is generated through the meta-learning controller, and the updated dynamic adaptive graph is output. Based on the dynamic adaptive graph, a gated graph convolutional network is used to aggregate the features of neighboring nodes and output a spatiotemporal embedding vector. Based on the spatiotemporal embedding vector, the joint stability risk score and tissue healing abnormality score are calculated through a dual-channel fully connected layer, and the original probability distribution is output. Based on the original probability distribution, the bias is adjusted using a Bayesian uncertainty calibration module, and the calibrated dual-channel prediction results are output.

[0008] Optionally, the step of spatiotemporally aligning the dual-channel prediction results with a personalized knowledge graph constructed from the patient's historical rehabilitation data, and calculating the deviation between the current rehabilitation status and the ideal rehabilitation trajectory using a nonlinear time warping algorithm to generate a quantitative risk level, includes: Based on the patient's historical rehabilitation data, the joint-tissue association topology is extracted using a graph embedding algorithm, and a personalized knowledge graph embedding vector is output. Based on the dual-channel prediction results and the knowledge graph embedding vector, an attention mechanism is used to project them into a unified semantic space, and cross-modal alignment features are output. Based on cross-modal alignment features, a nonlinear time warping algorithm is used to match the temporal path between the current state and the ideal rehabilitation trajectory, and output the dynamic time warping distance. Based on the dynamic time curvature distance, it is mapped to a preset risk threshold range and a four-level quantitative risk level is generated, which includes low, medium, high and critical.

[0009] Optionally, the step of generating a dynamic heatmap of the anatomical structure based on the quantified risk level, overlaying a biomechanical load streamline animation on the damaged tissue area, and outputting an interactive 3D report containing the location of key rehabilitation bottlenecks includes: Based on the quantitative risk level and anatomical atlas, a color gradient heatmap of the joint region is generated through a multi-scale rendering engine, and a risk visualization base map is output. Based on biomechanical load streamline animation, the direction of muscle force transmission is simulated through a particle system, and a dynamic streamline overlay layer is output. Based on the risk visualization base map and dynamic streamline overlay, a keyframe interpolation algorithm is used to locate the rehabilitation bottleneck area and output the coordinate set of the abnormal area. Based on the coordinate set of abnormal areas, the system integrates the patient's three-dimensional skeletal model with biomechanical data to generate a rotatable, scalable, and interactive three-dimensional rehabilitation examination report.

[0010] Another embodiment of this application provides a self-testing system for orthopedic patient rehabilitation, the system comprising: The acquisition module is used to collect multimodal biomechanical data of patients when performing standardized rehabilitation movements, and generate a spatiotemporally synchronized motion-mechanical coupling dataset. The separation module is used to input the motion-mechanics coupled dataset into the variational autoencoder, and learn to separate the joint range vector, muscle load vector and neural control coordination vector through deentanglement representation learning to form a three-dimensional biomechanical feature matrix. The adjustment module is used to input the three-dimensional biomechanical feature matrix into the dynamic graph neural network, dynamically adjust the connection weights of the graph nodes according to the individual patient's rehabilitation stage, and output a dual-channel prediction result of joint stability risk probability and tissue healing abnormality probability. The alignment module is used to align the dual-channel prediction results with a personalized knowledge graph constructed from the patient's historical rehabilitation data in a spatiotemporal manner. It calculates the deviation between the current rehabilitation status and the ideal rehabilitation trajectory using a nonlinear time warping algorithm to generate a quantitative risk level. The generation module is used to generate a dynamic heat map of the anatomical structure based on the quantified risk level, overlay biomechanical load streamline animation on the damaged tissue area, and output an interactive 3D report containing the location of key rehabilitation bottlenecks.

[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 self-assessment method for orthopedic patient rehabilitation. It collects multimodal biomechanical data from patients performing standardized rehabilitation movements, generating a spatiotemporally synchronized action-mechanical coupling dataset. This dataset is then input into a variational autoencoder to form a three-dimensional biomechanical feature matrix. This matrix is ​​further input into a dynamic graph neural network, outputting dual-channel prediction results for joint stability risk probability and tissue healing abnormality probability. The dual-channel prediction results are spatiotemporally aligned with a personalized knowledge graph constructed from the patient's historical rehabilitation data to generate a quantified risk level. Based on the quantified risk level, a dynamic heatmap of the anatomical structure is generated, and a biomechanical load streamline animation is overlaid on the damaged tissue area. An interactive three-dimensional report containing the location of key rehabilitation bottlenecks is then output. This enables three-dimensional, visually interactive self-assessment of orthopedic patient rehabilitation, improving the objectivity, accuracy, and clinical applicability of rehabilitation assessment. Attached Figure Description

[0014] Figure 1 Hardware structure block diagram of a computer terminal for a self-examination method for orthopedic patient rehabilitation provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating a self-examination method for orthopedic patient rehabilitation provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of an orthopedic patient rehabilitation self-testing 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 self-examination method for orthopedic patients' rehabilitation, 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 self-examination method for orthopedic patient rehabilitation provided in an embodiment of the present invention. Figure 1 As 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] Non-volatile storage media can store operating systems and computer programs. These computer programs include program instructions that, when executed, cause the processor to perform any orthopedic patient rehabilitation self-test 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 executed by a processor, the computer program can enable the processor to perform any self-examination method for orthopedic patient rehabilitation.

[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 2The present invention provides a self-examination method for orthopedic patients' rehabilitation, which may include the following steps: S201: Collect multimodal biomechanical data of patients performing standardized rehabilitation movements to generate a spatiotemporally synchronized motion-mechanical coupling dataset; Specifically, the patient's joint angular displacement sequence can be obtained through the optical motion capture system, and the rigid body transformation of the skeleton can be calculated to obtain high-precision motion trajectory data. Deployment and data acquisition of optical motion capture systems: Equipment configuration: Eight infrared optical cameras (such as the Vicon Vero series) are used, arranged around the patient at a sampling rate of 200 frames per second (200 FPS), with the camera lenses facing the center of the rehabilitation training area. Reflective markers (6 mm in diameter) are affixed to key anatomical points on the patient (including the tip of the lateral malleolus, the base of the fifth metatarsal, the calcaneal tuberosity, etc.), and the layout of the markers follows international standards for human bony landmarks (such as the ISB standard).

[0024] Dynamic capture process: The patient performs standardized rehabilitation movements (such as seated ankle pump exercises), the camera emits infrared light and receives the reflected signals from marker points, generating raw point cloud data in three-dimensional coordinates. Each motion capture lasts 10 seconds, generating approximately 2000 frames of raw position data (10 seconds × 200 FPS).

[0025] Noise filtering mechanism: Moving average filtering (a smoothing algorithm based on time window) is used to process the original coordinate jitter: the time window width is set to 5 frames (i.e. 25 milliseconds), the average position of the marker point within the window is calculated, and the ±2 mm position drift caused by skin slippage is eliminated.

[0026] Core technologies for calculating rigid body transformations of skeletons: Rigid body definition rules: Divide the foot into 3 rigid bodies (i.e., object models with negligible deformation): Rigid body of tibia and fibula: defined by the tibial plateau and the medial and lateral malleolar landmarks; Rigid body of talus: defined by the talus trochlea and the talus neck landmarks; Rigid body of foot: defined by the calcaneus and the first and fifth metatarsal heads landmarks.

[0027] Coordinate system transformation: Establish a global coordinate system: with the laboratory floor as the XY plane and the vertical direction as the Z axis.

[0028] Calculate the local coordinate system of each rigid body: for example, the rigid body of the foot takes the calcaneal marker as the origin and the direction pointing to the head of the first metatarsal head as the X-axis.

[0029] Joint angle displacement calculation: Ankle flexion / extension angle (Plantarflexion / Dorsiflexion): Calculate the angle between the tibial axis and the plantar plane. For example, it is 0° in the neutral position and increases to 20° in dorsiflexion.

[0030] Inversion / Eversion: The angle between the plantar plane and the horizontal plane is measured. When everted, the angle is +5°.

[0031] Output data: timestamp × joint angle matrix (2000 frames × 3 degrees of freedom angle values).

[0032] Accuracy calibration measures: Static Calibration: The baseline coordinates of the marker points are recorded when the patient is standing in a neutral position to compensate for individual differences in bone size.

[0033] Dynamic error compensation: When the visibility of the marker point is lost for more than 3 frames (such as due to clothing obstruction), the coordinate prediction algorithm based on rigid body constraints is activated (Predictive Algorithm, which uses rigid body kinematics to estimate the position of the lost point).

[0034] Based on the raw electromyographic signals collected by the surface electromyography sensor array, power frequency interference was removed by wavelet packet decomposition to obtain denoised muscle activation time series data. Deployment and preprocessing of the electromyography signal acquisition system: Electrode placement scheme: Bipolar surface electrodes (Ag / AgCl electrodes, electrode spacing 20 mm) are attached to the highest point of the muscle belly of the target muscles (lateral head of gastrocnemius, tibialis anterior, peroneus longus and brevis). The reference electrode is placed on the fibular head. The sampling rate is set to 2000 Hz, and the bandwidth is 10-500 Hz.

[0035] Raw signal characteristics: Electromyographic signal amplitude range ±5 millivolts (mV), containing three typical types of noise: Power line interference: 50 Hz sine wave (domestic power grid frequency) and its harmonics (100 Hz, 150 Hz).

[0036] Motion artifact: low-frequency large fluctuations (<5 Hz) caused by the relative displacement between the electrode and the skin.

[0037] Crosstalk: Signals from adjacent muscles are mixed together (e.g., signals from the tibialis anterior muscle seep into the peroneus muscle).

[0038] Baseline calibration: Acquire a 5-second signal while the patient is at rest, calculate the DC offset for each channel, and subtract this offset from subsequent data.

[0039] Denoising process using wavelet packet decomposition (an algorithm that divides a signal into frequency-based tree structures).

[0040] Decomposition parameter settings: Select Daubechies 4 wavelet basis (db4, a commonly used wavelet function); perform 4-level decomposition (Decomposition Level=4) to generate 16 frequency band sub-signals.

[0041] Power frequency noise cancellation: Locate the dominant power frequency interference band: 50 Hz, located in the 2nd sub-band of the 3rd layer (bandwidth range 37.5-62.5 Hz).

[0042] The sub-band coefficients are subjected to soft thresholding: the threshold λ = 0.05 mV, coefficients less than λ are set to zero, and the remaining coefficients are contracted according to their signs.

[0043] Motion artifact suppression: The low-frequency subband (0-12.5 Hz) uses a high-pass filter with a cutoff frequency of 5 Hz.

[0044] Signal reconstruction and effect verification: retain the denoised subband coefficients and perform wavelet packet reconstruction.

[0045] Calculate the signal-to-noise ratio (SNR) improvement: typical SNR after noise suppression > 25 dB (original signal is about 10 dB).

[0046] Output data: Denoising electromyography timing (2000 Hz × 10 seconds = 20,000 sampling points / channel).

[0047] Muscle activation time sequence extraction technology: Envelope generation: After full-wave rectification of the denoised signal, a 4th-order Butterworth low-pass filter (a smoothing filter) is used to process it with a cutoff frequency of 5 Hz to extract the active envelope.

[0048] Activation range identification: Set a dynamic threshold: Threshold = Root mean square of resting period (RMS) × 3 (e.g., resting RMS = 0.02 mV → threshold 0.06 mV).

[0049] Activation period marked: continuous periods when the envelope exceeds the threshold (e.g., peroneal muscle activation lasts for 0.8 seconds).

[0050] Synergistic contraction analysis: Calculate the percentage of overlapping activation time of antagonistic muscle groups (such as tibialis anterior and gastrocnemius muscles). In normal gait, this should be <30%.

[0051] Based on motion trajectory data and muscle activation timing data, a dynamic time warp algorithm is used to align timestamps and generate a spatiotemporally synchronized skeletal-muscle joint timing stream. Unifying the time base of multi-source data: Hardware clock synchronization: The motion capture system (Vicon) and the electromyography acquisition system (Noraxon) receive the same 10 MHz clock source signal through the BNC interface, ensuring hardware-level time synchronization (error < 1 millisecond).

[0052] Software timestamp alignment: Define the action start event (Event Marker): the instant the patient's heel contacts the pressure plate (pressure > 20 Newtons) as the time zero point (T0).

[0053] The motion capture system records the first frame at T0+5 milliseconds, and the electromyography system triggers acquisition at T0+3 milliseconds.

[0054] The core process of the Dynamic Time Warping (DTW) algorithm (an algorithm for aligning time sequences at different rates).

[0055] Input data preprocessing: Motion trajectory data: downsampled to 100 Hz (consistent with the frequency of electromyography envelope), for example, dorsiflexion angle sequence (1000 points in length).

[0056] Electromyographic activation data: Peroneus muscle envelope sequence (1000 points in length).

[0057] Similarity matrix construction: Calculate the Euclidean distance matrix: Matrix element D(i,j) = |point i in the angular sequence -point j in the electromyographic sequence|.

[0058] Example: When the dorsiflexion peak (i=500) aligns with the electromyography peak (j=520), D(500,520)=0.1 (ideal alignment distance is close to 0).

[0059] Optimal path search: Dynamic programming is used to search for the path with the minimum cumulative distance from the top left corner (1,1) to the bottom right corner (1000,1000) of the matrix.

[0060] Constraints: The path slope is limited to between 0.5 and 2 (to prevent excessive distortion).

[0061] Time axis remapping: Adjusting EMG timestamps according to the optimal path: for example, mapping point 520 of the original EMG sequence to point 500 of the motion sequence.

[0062] Output synchronized data: timestamp-aligned pairs of angle values ​​and electromyographic values ​​(100 sets of data per second).

[0063] Time-series data structures and quality control: The combined time-series stream format is shown in Table 1: Table 1 ; Synchronization accuracy verification: Calculate the reasonable range of physiological delay: the neuromuscular conduction delay should be between 20-50 milliseconds (e.g., the electromyography peak lags the angle peak by 30 milliseconds, which is normal).

[0064] Anomaly handling: If a negative delay is detected (EMG precedes movement by more than 10 milliseconds), it is determined as a synchronization failure and re-alignment is performed.

[0065] Based on the joint temporal flow of the skeletal-muscle system, the three-dimensional force vectors measured by the plantar pressure plate are fused, and the joint torque vectors are calculated through the inverse dynamics model to output the motion-mechanics coupling dataset.

[0066] Plantar pressure data fusion technology: Pressure plate system configuration: Employs a piezoresistive pressure plate (such as Bertec FP4060), with a sampling rate of 1000 Hz and a spatial resolution of 1.5 sensors / cm². Outputs a three-dimensional force vector (Fx, Fy, Fz) and pressure center coordinates (COPx, COPy).

[0067] Spatiotemporal alignment operation: The pressure data was downsampled to 100 Hz (the same frequency as the skeletal-muscle flow).

[0068] Align the time axis with the heel strike event as the reference (error < 2 milliseconds).

[0069] Calculation of dynamic parameters: Ground Reaction Force: Directly read the Fz component (vertical force), for example, the peak force during the push-off phase is 800 Newtons.

[0070] Friction force vector: composed of Fx (front-back direction) and Fy (inward and outward direction) components, for example, the maximum lateral friction force is 50 Newtons.

[0071] Solution process for inverse dynamics (a method of calculating joint forces based on motion and external forces).

[0072] Modeling of a multi-rigid-body system for the foot: Define three rigid bodies for the foot (see step 1), and set the mass properties of each rigid body according to the CT data (e.g., the total mass of the foot is 1.2 kg, and the center of mass is 60 mm away from the calcaneus).

[0073] Joint constraints: The ankle joint is defined as a ball-and-socket joint (3 degrees of freedom), and the subtalar joint is defined as a hinge joint (1 degree of freedom).

[0074] Iterative solution of the Newton-Euler equations (fundamental laws of mechanics): Progressing from the far end to the near end: Step 1: Calculate the rigid body inertial force of the foot: The acceleration is obtained by the difference of the motion trajectory (e.g., the acceleration of the foot's center of mass is 3 m / s²). 2 →Inertial force = mass × acceleration = 1.2 kg × 3 m / s² 2 =3.6 Newtons).

[0075] Step 2: Calculate the joint reaction force: Ground reaction force + inertial force = joint force (e.g., ankle joint force = 800 Newtons + 3.6 Newtons ≈ 804 Newtons).

[0076] Joint torque calculation: Torque = Force × Lever arm (three-dimensional vector cross product).

[0077] Example: Ankle dorsiflexion torque = Achilles tendon force × Achilles tendon lever arm (lever arm length 0.05 meters) × mechanical efficiency coefficient.

[0078] Output data: Ankle joint triaxial moment vector (dorsiflexion-plantarflexion moment, inversion-eversion moment, internal rotation-external rotation moment).

[0079] Action-mechanics coupled dataset integration: The data structure design is shown in Table 2: Table 2 ; Outlier handling mechanism: Momentum mutation detection: If the torque difference between adjacent frames is greater than 30 N·m (e.g. due to device jitter), cubic spline interpolation (a smooth interpolation algorithm) is initiated for repair.

[0080] Clinical application example: Peroneal tendinitis patients: peak eversion torque decreased by 40% (normal value 15 N·m → patient 9 N·m), and electromyographic activation delay was 100 milliseconds.

[0081] S202, the motion-mechanics coupled dataset is input into the variational autoencoder, and the joint range vector, muscle load vector and neural control coordination vector are separated through deentanglement representation learning to form a three-dimensional biomechanical feature matrix; Specifically, spatiotemporal features can be extracted from the action-mechanics coupled dataset using a convolutional encoder to output the initial encoding of the latent space; Preprocessing workflow for action-mechanics coupled datasets.

[0082] Data standardization processing: Receive the motion-mechanical coupling dataset (100 frames per second × 20 biomechanical parameters) from claim 2, and normalize the raw data. Joint angle parameters are mapped to the [0,1] interval (e.g., dorsiflexion angle 0°-30° linearly corresponds to 0-1), the percentage of electromyographic activation intensity is directly divided by 100 (85%→0.85), and joint torque parameters are normalized by dividing by the patient's weight (e.g., for a 70 kg patient, 42.5 N·m torque → 0.61 N·m / kg).

[0083] Spatiotemporal data reconstruction: Convert time series data into a three-dimensional structure: Time dimension: 1000 time points (10-second data); Spatial dimension: 4 anatomical units (ankle joint, peroneal muscle group, tibialis muscle group, plantar pressure zone); Feature dimension: 5 parameters per group (e.g., ankle joint includes dorsiflexion angle, eversion angle, internal rotation angle, dorsiflexion moment, eversion moment).

[0084] Form a 20×1000×4 data cube (features×time×space).

[0085] Architecture design of a convolutional encoder.

[0086] Spatiotemporal convolution kernel configuration: Design two-level convolutional layers: Temporal convolutional layer: One-dimensional convolutional kernels sliding along the time axis (Kernel Size=5 frames) extract motion continuity features (such as the acceleration phase of backflexion); each spatial unit is processed independently, and the output temporal dimension is compressed to 200 frames (Stride=5).

[0087] Spatial convolutional layer: A three-dimensional convolutional kernel spanning anatomical units (Kernel Size = 4×1×1), fusing ankle joint and muscle group association features (such as peroneal muscle activation and ankle eversion torque synchronization). The output spatial dimension is compressed into a single integrated node.

[0088] Feature dimensionality reduction mechanism: After convolution, a max pooling layer is applied to downsample the temporal dimension by a factor of 5 (200 frames → 40 frames), preserving key peak features of key actions (such as the maximum dorsiflexion angle at the moment of heel lift-off). Finally, a fully connected layer compresses the 20-dimensional features into an 8-dimensional latent vector.

[0089] Example of initial encoding output: A patient's single action outputs an 8-dimensional vector: [0.32, -0.15, 1.08, 0.74, -0.03, 0.91, 0.12, -0.45]. Positive values ​​represent high activity characteristics, and negative values ​​represent electromyographic delay characteristics.

[0090] Based on the initial encoding of the latent space, the entangled subspace is separated by adversarial regularization constraints to obtain three types of latent variables: orthogonalized joint mobility, muscle load, and neural coordination. The implementation mechanism of Adversarial Regularization.

[0091] Principle of entanglement decoupling: Three discriminators (neural network modules used to distinguish feature types) are introduced: Joint range of motion discriminator: forces independent aggregation of features related to range of motion (such as rate of change of angle) in the latent encoding; Muscle load discriminator: separates features related to electromyographic intensity and torque (such as peak activation intensity); Neural coordination discriminator: focuses on the difference in activation timing of antagonistic muscle groups (such as the delay time between the tibialis anterior and gastrocnemius muscles).

[0092] Combat training process: The encoder generates an initial latent code; each discriminator predicts the category to which the code belongs (e.g., joint mobility class); the encoder performs inverse optimization: adjusting parameters so that the discriminator cannot distinguish feature types (i.e., confusing the discriminator) while maintaining reconstruction accuracy.

[0093] Training termination condition: The discriminator classification accuracy is less than 55% (close to random guessing).

[0094] Feature orthogonalization technique.

[0095] Subspace separation operation: The Gram-Schmidt process (a vector orthogonalization method) forces three types of feature vectors to be perpendicular: the direction of the joint range of motion vector is defined as the reference axis; the muscle load vector is subtracted from its projection component onto the range of motion vector; and the neural coordination vector is simultaneously orthogonal to both of the above.

[0096] Mathematical verification: Calculate the vector dot product to ensure that |cosθ| < 0.05 (i.e., the included angle > 87°).

[0097] Latent variable dimension allocation: Joint range of motion vector: 3D (corresponding to flexion / extension, inversion / valgus, and rotation degrees of freedom); Muscle load vector: 2D (load intensity of peroneal muscles and tibialis posterior muscles); Neural coordination vector: 3D (intermuscular delay, activation sequence index, and synergistic contraction rate); Example output: Joint range of motion vector = [0.82, 0.15, -0.03]; Muscle load vector = [0.12, 0.91]; Neural coordination vector = [0.04, -0.33, 0.57].

[0098] Based on the three types of latent variables, the joint kinematic parameters, electromyographic envelope signals and motor coordination indices are reconstructed through a multi-branch decoder, and the unentangled feature vector set is output. Architecture design of a multi-branch decoder.

[0099] Branch-specific reconstruction: Joint kinematics decoder: Input: Joint motion vector (3D); Structure: 3-layer deconvolutional network; Output: Reconstructed 10-second joint angle time sequence (100 frames × 3 degrees of freedom), with an accuracy error of <1.2°.

[0100] Myoelectronic decoder: Input: Muscle load vector (2D); Structure: Long Short-Term Memory (LSTM) network; Output: Activation envelope of the peroneus and tibialis posterior muscles, with a correlation coefficient r > 0.89.

[0101] Coordination Indicator Sub-Decoder: Input: Neural coordination vector (3D); Structure: Fully connected layer + Gaussian Process Regression (GPR); Output: 5 indicators including synergistic contraction rate and peak delay time.

[0102] Reconstructing the loss function strategy.

[0103] Independent branch optimization: Each sub-decoder uses its own loss function: Kinematics branch: Minimize root mean square error (RMSE) of joint angles (target <1.5°); Electromyography branch: Maximize envelope shape similarity (Dynamic Time Warping Distance <0.3); Coordination branch: Minimize time deviation of critical events (e.g., activation peak time error <15 milliseconds).

[0104] Clinical interpretability constraint: The results of forced reconstruction conformed to biomechanical principles: the peroneus muscle was activated 20-50 milliseconds earlier than the tibialis posterior muscle; the time difference between the peak dorsiflexion angle and the peak electromyography was <80 milliseconds.

[0105] Based on the unentangled feature vector set, it is stacked into a tensor along the time axis, and a three-dimensional biomechanical feature matrix is ​​generated by normalizing the feature dimension.

[0106] Timeline stacking operation: Multi-movement cycle integration: The patient performs 5 consecutive standardized ankle pump movements. For each movement, three types of feature vectors are extracted (3D joint range of motion, 2D muscle load, and 3D neural coordination) to form a time series: 100 frames × 8-dimensional vector for a single movement (3+2+3); a total of 500 frames × 8-dimensional vector for 5 movements.

[0107] Tensor data structures: Construct a three-dimensional matrix: Dimension 1: Time frame (500 frames); Dimension 2: Feature category (3 categories); Dimension 3: Feature vector (maximum dimension 3).

[0108] Filling rule: The 2D vector of muscle load is padded with zeros in the third dimension to make it 3D.

[0109] Dynamic normalization technique: Patient adaptive normalization: Joint range of motion characteristics: divided by the individual patient's physiological limit (e.g., maximum dorsiflexion angle 25° → characteristic value 0.82 normalized to 0.82 / 25=0.0328); Muscle load characteristics: standardized based on maximum voluntary contraction force (MVC, Maximum Voluntary Contraction) (e.g., peroneus muscle MVC=110 Newtons → load value 0.91 → 0.91 / 110=0.0083); Neural coordination characteristics: standardized using Z-score (subtract the mean of healthy individuals and divide by the standard deviation).

[0110] Outlier handling: Set the effective range of feature values ​​(e.g., joint range of motion [-0.5, 1.5]), and cut off values ​​exceeding the limit to the boundary and mark them with an anomaly flag.

[0111] Final output example: The 3D feature matrix size is 500×3×3. Example data from frame 120: Joint range of motion: [0.028, 0.006, -0.001]; Muscle load: [0.008, 0.012, 0.000] (padding with zeros); Neural coordination: [-0.32, 0.18, 0.07]. S203, the three-dimensional biomechanical feature matrix is ​​input into a dynamic graph neural network, the connection weights of the graph nodes are dynamically adjusted according to the individual patient's rehabilitation stage, and the dual-channel prediction results of joint stability risk probability and tissue healing abnormality probability are output. Specifically, based on the three-dimensional biomechanical feature matrix, an initial graph structure with joints as nodes and biomechanical associations as edges can be constructed, and a static topological graph can be output. Rules for defining nodes and edges: Joint node attribute configuration: The three-dimensional biomechanical feature matrix contains three core features (joint range of motion, muscle load, and neural coordination), divided into six key nodes according to anatomical structure: tibiotalar joint (core of the ankle joint), subtalar joint, peroneus longus and brevis muscle attachment points, tibialis posterior muscle attachment points, calcaneus, and metatarsal base. Each node carries a 15-dimensional feature vector (5-second movement cycle × 3 types of features). For example, the tibiotalar joint node includes: mean dorsiflexion angle, standard deviation of eversion angle, peak peroneus muscle load, etc.

[0112] Weight assignment for biomechanical edges: Mechanical transmission edge: If there is a direct mechanical interaction between two nodes (such as the tibiotalar joint and the subtalar joint), the weight is equal to the characteristic correlation coefficient (such as the Pearson correlation coefficient between range of motion and load). Calculation rule: Take the average correlation coefficient of 20 movements. For example, if the correlation between joint range of motion and electromyographic activation is 0.78, then the edge weight is 0.78.

[0113] Neural control edge: Coordination relationship between muscle nodes and joint nodes (e.g., peroneus muscle node and tibiotalar joint node), weight = mutual information of neural coordination vector (quantification of signal dependence), normal value > 0.35 bits.

[0114] Anatomical constraint edge: Fixed bone nodes (such as calcaneus-metatarsal base), with a weight of 1.0 (non-adjustable rigid connection).

[0115] Static topology graph construction process: Graph data structure generation: Node set: 6 nodes × 15-dimensional feature vectors; Edge set: 9 directed edges (e.g., tibiotalar joint → peroneus muscle node); Weight matrix: 9×9 symmetric matrix (diagonal elements are node self-weights, fixed at 0).

[0116] Pathological specialization adjustment: For patients with ligament injuries: Add a virtual border for the "tibiotalar joint-calcaneus" (to compensate for ligament dysfunction), with an initial weight of 0.5 (this border is not present in healthy individuals).

[0117] After joint replacement surgery: reduce the weight of the "tibial-talar joint-subtalar joint" side to 0.6 (normal 0.9) to reflect the kinematic differences of artificial joints.

[0118] Based on the patient's recovery stage label, the edge weight update matrix of the static topology graph is generated through the meta-learning controller, and the updated dynamic adaptive graph is output. Rehabilitation phase labeling and coding mechanism: Stage Grading Standards: The rehabilitation stage label is divided into 5 levels (1=postoperative acute phase, 2=fibrosis phase, 3=remodeling phase, 4=functional recovery phase, 5=return to sports phase), which are determined by clinicians based on imaging (such as MRI cartilage healing score) and functional assessment (such as ankle joint scoring scale).

[0119] Label vectorization: Stage labels are converted into 5-dimensional one-hot encodings, for example, stage 2 is encoded as [0,1,0,0,0]. This vector is concatenated with the patient ID and the number of days of illness (e.g., day 28 post-surgery) to form an 8-dimensional control vector.

[0120] How the Meta-Learning Controller works.

[0121] Neural network architecture: The controller uses a 3-layer fully connected network (128-64-9 neurons), with an 8-dimensional control vector as input and a 9-dimensional edge weight update (Delta Weight) as output. The update range is [-0.3, +0.3] to prevent sudden weight changes.

[0122] Training data strategy: Pre-training phase: Using data from 300 historical patients, a static topology graph was fixed, and network parameters were optimized based on the recovery phase.

[0123] Online Adaptive: When a patient enters a new rehabilitation stage (such as upgrading from stage 2 to stage 3), weight updates are automatically triggered.

[0124] Example: In stage 3 patients, the weight of the "tibiotalar joint-peroneus muscle" side increases from 0.75 to 0.82 (increased demand for muscle compensation).

[0125] Output update matrix: Static graph weight matrix + update matrix → dynamic adaptive weight matrix. For example: Initial weight matrix: [0.0, 0.78, 0.0; 0.0, 0.0, 0.65; ...]; Update matrix: [0.0, +0.04, 0.0; 0.0, 0.0, -0.02; ...]; Dynamic weight matrix: [0.0, 0.82, 0.0; 0.0, 0.0, 0.63; ...]. Based on the dynamic adaptive graph, a gated graph convolutional network is used to aggregate the features of neighboring nodes and output a spatiotemporal embedding vector. The working mechanism of Gated Graph Convolutional Network.

[0126] Neighbor feature aggregation strategy: Each node aggregates features from its first-order neighbors (directly connected nodes), weighted as follows: Node's own features × self-weight coefficient (fixed at 0.7); neighbor node features × dynamic edge weight (e.g., the peroneus muscle node contributes a weight of 0.82 to the tibiotalar joint); aggregation result = Σ(neighbor features × edge weight) + own features × 0.7.

[0127] Core functions of the gating mechanism: Forget Gate: Determines how much historical features to retain (weight matrix W_f, 128×128 dimensions, W_f meaning: weight matrix). For example, when joint stability is poor, the forget gate reduces the weight of historical data (from 0.8 to 0.5) to avoid error accumulation.

[0128] Update Gate: Controls the proportion of new features written (weight matrix W_u, 128×128 dimensions). When tissue healing signals are detected (such as increased electromyographic coordination), the weight of the new feature is increased to 0.9.

[0129] Spatiotemporal embedding vector generation process: Iteration in the time dimension: Progressing in 100-millisecond time steps, each step executes: Update node features (e.g., add dorsiflexion angle value for the tibiotalar joint); aggregate neighbor data using gated convolution; output the node state vector at the current time step.

[0130] Spatial dimension compression: 6-node × 128-dimensional state vector → Global Average Pooling → Generate a single 256-dimensional spatio-temporal embedding vector.

[0131] Example vector fragment: [0.32, -1.08, ..., 0.75] (positive values ​​represent stability, negative values ​​warn of risk).

[0132] Based on the spatiotemporal embedding vector, the joint stability risk score and tissue healing abnormality score are calculated through a dual-channel fully connected layer, and the original probability distribution is output. Dual-channel predictor design: Joint stability channel: Input 256-dimensional spatiotemporal embedding vector → 3-layer fully connected network (128-64-1 neurons) → Output risk score (range [-5.0, +5.0]). A score > 2.0 triggers a high-risk warning (e.g., a score of 3.8 for patients with ankle laxity).

[0133] Key features to focus on: This channel focuses on mutations in neural coordination vectors (e.g., a 30% decrease in the rate of coordinated contraction contributes +1.5 points) and abnormalities in activity (e.g., an excessive eversion angle contributes +0.9 points).

[0134] Tissue healing pathway: Input the same embedding vector → Independent 3-layer network (128-64-1) → Output anomaly score (range [-5.0, +5.0]). A score > 3.0 indicates delayed healing (e.g., a patient 8 weeks after Achilles tendon repair surgery scores 3.5).

[0135] The core judgment criteria are: muscle load vector fluctuation (standard deviation > 0.12 contributes +2.1 points) and mechanical transmission efficiency (edge ​​weight < 0.6 contributes +1.2 points).

[0136] Transformation of the original probability distribution: Fraction-probability mapping function: Using the sigmoid curve transformation: Stability risk probability = 1 / [1 + e^(-0.8 × risk score)]. For example, risk score 2.0 → probability P = 1 / (1 + e^{-1.6}) ≈ 83.2%.

[0137] Dual-channel output format: Joint stability risk probability: 0.832 (83.2%); Tissue healing abnormality probability: 0.654 (65.4%).

[0138] Based on the original probability distribution, the bias is adjusted using a Bayesian uncertainty calibration module, and the calibrated dual-channel prediction results are output.

[0139] The necessity of uncertainty calibration: Sources of systematic bias: Data sparsity: Few patient movement samples in the acute postoperative period (only 3 ankle pumps can be completed); Individual differences: Connective tissue degeneration in elderly patients leads to abnormal mechanical conduction; Sensor noise: Electromyographic signals fluctuate randomly by ±5%.

[0140] Deviation manifests as follows: the original probability often overestimates the risk (e.g., outputting 82% when the actual risk is 60%), and needs to be calibrated downwards.

[0141] Operation flow of the Bayesian Calibration Module.

[0142] Prior distribution settings: Prior probabilities are established based on historical databases: stability risk prior: Beta distribution (α=8, β=2) (α means: count of historical high-risk events, β means: count of low-risk events); tissue healing prior: Gamma distribution (k=6, θ=0.4) (k means: morphological parameter, θ means: scale parameter).

[0143] Posterior probability calculation: Collect the most recent 10 prediction results of the current patient as the likelihood sample.

[0144] By fusing prior and likelihood using Bayes' theorem: Calibrated probability = (Prior count × Prior probability + Likelihood count × Original probability) / Total count.

[0145] Example: Original stability risk probability = 0.832; prior probability = 0.75 (calculated from α / (α+β) = 8 / 10); calibrated probability = (8×0.75 + 10×0.832) / (8+10) = 14.32 / 18 ≈ 0.796 (79.6%).

[0146] Dynamic confidence interval: The output results are appended with a 95% confidence interval, such as "Stability risk 79.6% [74.2%, 84.3%]".

[0147] S204, the dual-channel prediction results are spatiotemporally aligned with the personalized knowledge graph constructed from the patient's historical rehabilitation data, and the deviation between the current rehabilitation status and the ideal rehabilitation trajectory is calculated using a nonlinear time warping algorithm to generate a quantitative risk level; Specifically, based on the patient's historical rehabilitation data, a graph embedding algorithm can be used to extract the joint-tissue association topology and output a personalized knowledge graph embedding vector. Structured processing of historical rehabilitation data: Collect patient data from at least three historical assessments (e.g., at weeks 2, 4, and 6 post-surgery). Each assessment should include a three-dimensional biomechanical feature matrix (500 frames × 3 × 3), clinical assessment scales (e.g., AOFAS ankle joint scoring system scores), and imaging features (e.g., fascia thickness measured in millimeters by ultrasound and T2 relaxation time in milliseconds from MRI images). Construct a four-dimensional data cube: time axis × anatomical structure × feature dimension × assessment type.

[0148] Association topology construction rules: Node definition: 6 core anatomical nodes (tibiotalar joint cartilage, peroneal tendon, calcaneal callus, etc.) and 4 functional nodes (joint stability, muscle strength recovery, pain index, and activity tolerance).

[0149] Edge connection logic: Physical connection edge (e.g., tibiotalar joint cartilage → peroneus tendon) weight = 1.0; Functional dependence edge (e.g., muscle strength recovery → joint stability) weight = Pearson correlation coefficient (historical data calculation value, e.g., 0.78); Pathological influence edge (e.g., pain index → ​​activity tolerance) weight = -0.62 (negative value indicates inhibitory effect).

[0150] Implementation process of the graph embedding algorithm: The GraphSAGE algorithm (SAmple and aggreGatE, a graph embedding technique based on neighbor sampling) was selected: Neighbor sampling: Each node randomly selects 3 first-order neighbors (e.g., the "peroneal tendon" node is associated with "muscle strength recovery", "pain index", and "calcaneal callus").

[0151] Feature aggregation: A mean aggregator is used to average the features of neighboring nodes and then concatenate them with the features of the node itself. For example, tendon node features = [0.32, 0.15] + neighbor mean [0.18, 0.07] → output [0.32, 0.15, 0.18, 0.07].

[0152] Deep iterative mechanism: Perform two-layer aggregation (Layer-1 aggregates direct neighbors, Layer-2 aggregates neighbors' neighbors) to generate a globally aware 128-dimensional embedding vector. For patients with postoperative adhesions, the "calcaneal callus" node is associated with the "activity endurance" node in the second layer, capturing cross-layer dependencies.

[0153] Example of output embedding vector: [0.85, -0.12, 0.73, ...] (positive components indicate promotion of healing, negative components warn of risk).

[0154] Based on the dual-channel prediction results and the knowledge graph embedding vector, an attention mechanism is used to project them into a unified semantic space, and cross-modal alignment features are output. Dual-modal data preprocessing: Vectorization of dual-channel prediction results: The probability of joint stability risk (e.g., 79.6%) and the probability of abnormal tissue healing (e.g., 38.6%) are converted into a 2-dimensional vector: [0.796, 0.386].

[0155] Knowledge graph vector dimensionality reduction: The 128-dimensional knowledge graph embedding vector is compressed to 16 dimensions through Principal Component Analysis (PCA), retaining 95% of the information.

[0156] The alignment process of the attention mechanism.

[0157] Semantic Space Construction: Define a 32-dimensional unified semantic space, with each dimension corresponding to a specific clinical meaning (e.g., dimension 1 = mechanical stability, dimension 5 = degree of inflammation).

[0158] Cross-modal attention weight allocation: Calculate the similarity score between the dual-channel vector and the graph vector: Similarity = Dual-channel vector · Graph vector (dot product) / Vector magnitude product.

[0159] Example: The stability probability of 0.796 is similar to the "Joint Ligament Tension" dimension in the atlas with a similarity score of 0.91.

[0160] Generate attention distribution: assign weights ≥0.7 to high similarity dimensions (score>0.8); assign weights ≤0.1 to low similarity dimensions (score<0.3).

[0161] Feature projection operation: Simplified description of weighted fusion formula: Dual-channel vector × 60% weight + spectral vector × 40% weight → Output 32-dimensional aligned features.

[0162] Example output snippet: [0.83, 0.12, -0.05, ...] (Dimension 1 value of 0.83 indicates that the current mechanical stability is close to the historical best state).

[0163] Based on cross-modal alignment features, a nonlinear time warping algorithm is used to match the temporal path between the current state and the ideal rehabilitation trajectory, and output the dynamic time warping distance. Construction of a rehabilitation trajectory database: Ideal trajectory generation: Based on data from 100 patients with similar successful rehabilitation (e.g., grade II ankle ligament injury), segmented modeling is performed according to rehabilitation stages. Phase 1 (0-2 weeks): Joint range of motion recovers to 35%; Phase 2 (3-5 weeks): Muscle strength reaches 60% of normal value; Phase 3 (6-8 weeks): Pain index <2 (out of 10).

[0164] Current state serialization: The cross-modal aligned features are divided into time windows (each segment is 5 days) to generate a 6-dimensional time series sequence (e.g., [0.83, 0.71, 0.62, 0.59, 0.66, 0.70]).

[0165] Execution flow of the Nonlinear Time Warping (NTW) algorithm Similarity matrix calculation: Construct a 6×3 matrix (current 6 time periods vs. ideal 3 stages); matrix element value = Euclidean distance (mean of current segment features - target value of ideal stage).

[0166] Example: Current average value of segment 4 is 0.59 vs. target value of 0.60 in stage 2 → distance 0.01.

[0167] Optimal path search strategy: Allows multiple time periods to be mapped to the same ideal stage (e.g., when postoperative recovery is slow, stage 1 lasts for 4 weeks); Path constraints: Prohibits jumping across stages (e.g., stage 1 cannot directly reach stage 3).

[0168] Dynamic Time Warping Distance (DTWD) output: Minimum cumulative path distance (unit: standard offset of feature space); Example: DTWD = 1.83 4 weeks postoperatively for ligament injury patients (<2.0 is within the allowable range).

[0169] Based on the dynamic time curvature distance, it is mapped to a preset risk threshold range and a four-level quantitative risk level is generated, which includes low, medium, high and critical.

[0170] Risk threshold range division rules: Clinical validation thresholds: Based on retrospective analysis of patient recovery outcomes, as shown in Table 1: Table 1 ; Hierarchy generation and clinical decision mapping: Example of grading logic: Patient A's DTWD=2.73 → falls into [2.5,3.5) → high risk level; corresponding clinical measures: increase the frequency of ultrasound examinations (twice a week), and adjust the weight training plan.

[0171] Composite risk enhancement mechanism: When the probability of joint stability in dual-channel prediction is >80% and DTWD is >2.5, it is automatically upgraded to the critical level.

[0172] Example: Patient B DTWD=2.43 (originally medium risk), upgraded to high risk due to stability probability of 85.2%.

[0173] S205, Based on the quantified risk level, generate a dynamic heat map of the anatomical structure, overlay biomechanical load streamline animation on the damaged tissue area, and output an interactive 3D report containing the location of key rehabilitation bottlenecks.

[0174] Specifically, based on the quantified risk level and anatomical atlas, a multi-scale rendering engine can be used to generate a color gradient heatmap of the joint region and output a risk visualization base map. Spatial mapping between anatomical structures and risk levels: Anatomical atlas partitioning rules: The target joint (such as the ankle joint) is divided into 8 key anatomical blocks: distal tibial articular surface (block 1), talus top (block 2), distal fibula (block 3), calcaneal anterior process (block 4), subtalar articular surface (block 5), peroneus tendon groove (block 6), Achilles tendon insertion (block 7), and plantar fascia origin (block 8). Each block is bound to a preset coordinate range (e.g., the coordinate range of block 1 is X[120-150], Y[80-100]).

[0175] Risk-Color Mapping Strategy: HSV color space parameters are set based on a four-level quantified risk rating (low, medium, high, critical). Low risk (Level 0): Hue H=120° (green), Saturation S=30%, Brightness V=90%; Medium risk (Level 1): H=60° (yellow), S=60%, V=85%; High risk (Level 2): ​​H=30° (orange), S=80%, V=80%; Critical (Level 3): H=0° (red), S=100%, V=100%.

[0176] Gradient generation rule: Within a single anatomical block, for every 0.1 level increase in risk value, the hue shifts 3° towards red (e.g., high risk 0.2 level → H=24°).

[0177] Multi-scale Rendering Engine (MLE) Operation Process.

[0178] Macro-scale rendering (overall joint): Based on the joint capsule boundary, generate a 512×512 pixel base map; global risk value = average risk level of each block (e.g., average value 2.3 → overall orange tone).

[0179] Microscale rendering (organizational level): Activate the Level of Detail (LOD) mechanism; when the user's viewing distance is <30 cm virtual unit: render subchondral bone microstructure (accuracy 0.1 mm); display trabecular fracture patterns in high-risk areas (≥2 level) (pattern density is proportional to risk level).

[0180] Dynamic update mechanism: Receive risk level data stream in real time (10 frames per second update), and trigger a color gradient transition animation (duration 0.3 seconds) when the risk change of adjacent frames is greater than 0.5.

[0181] Based on biomechanical load streamline animation, the direction of muscle force transmission is simulated through a particle system, and a dynamic streamline overlay layer is output. Particle System Parameter Configuration: Particle generation rules: Set emission sources at the origin and insertion points of muscles (e.g., the origin of the medial head of the gastrocnemius muscle at the medial epicondyle of the femur); Particle quantity = electromyographic activation intensity × 100 (e.g., 80% activation → 80 particles / second emission); Particle lifespan = muscle contraction duration (e.g., isometric contraction for 2 seconds → particle survival for 2 seconds).

[0182] Mechanics-driven logic: Particle motion direction: along the direction of muscle fiber vector (vector direction is calculated from motion capture data); Particle velocity = joint torque value × scaling factor (e.g., torque 25 Nm → velocity 2.5 units / second).

[0183] Particle color-coded payload type: Tensile load: blue particles (H=240°); compressive load: purple particles (H=270°); shear load: magenta particles (H=300°).

[0184] Streamline animation compositing technology: Path smoothing: The Catmull-Rom spline curve algorithm (a smooth path generation algorithm) is used to connect particle trajectory points to eliminate motion jitter. Trajectory points are sampled every 0.1 seconds to generate continuous streamlines.

[0185] Visualization of pathological conditions: Patients with tendinitis: The flow lines appear vortex-like disorder in the inflamed area (such as the Achilles tendon insertion point); neurological control disorder: the particle flow is bidirectionally opposed (such as the particle flow of the peroneal muscle group and the posterior tibial muscle group colliding).

[0186] Output overlay format: semi-transparent RGBA layer (alpha=0.7), blended with the heatmap base map using the Multiply blending mode.

[0187] Based on the risk visualization base map and dynamic streamline overlay, a keyframe interpolation algorithm is used to locate the rehabilitation bottleneck area and output the coordinate set of the abnormal area. Rehabilitation bottleneck feature identification logic: Thermodynamic-streamline coupling analysis: Define the conditions for abnormal areas: Condition A: The risk level of the heat map is ≥2 (orange / red); Condition B: The streamline animation shows eddies or flow interruptions; Areas that meet both conditions are marked as candidate bottleneck areas.

[0188] Typical scenario example: Patient with tibial plateau fracture: The heat map shows that block 1 has a risk level of 2.8 (red), and the flow line is interrupted at the fracture line → confirmed as a bottleneck area.

[0189] The keyframe interpolation algorithm provides precise localization.

[0190] Keyframe definition: Select three biomechanical feature points as reference frames within the motion cycle: Start frame (T0): the moment the heel contacts the ground; peak frame (T1): the moment of maximum dorsiflexion angle; end frame (T2): the moment the toes leave the ground.

[0191] Interpolation rules: Insert 5 intermediate frames (0.2 seconds apart) between T0 and T1. Anomaly area coordinates = (previous frame coordinates × 0.4 + next frame coordinates × 0.6) + position correction amount; The correction amount is calculated based on the rate of change of streamline direction (if the change in direction is greater than 30° / second, the offset is increased).

[0192] Example of coordinate set output format: Abnormal area 1: Tibiotalar joint cartilage Coordinate sequence: T0: (120.5, 80.3) T0.2: (122.1, 81.2) ... T1: (135.8, 85.6) Abnormal area 2: Peroneus tendon groove ... Based on the coordinate set of abnormal areas, the system integrates the patient's three-dimensional skeletal model with biomechanical data to generate a rotatable, scalable, and interactive three-dimensional rehabilitation examination report.

[0193] Multi-source data fusion mechanism: 3D skeletal model reconstruction: A triangular patch model (approximately 50,000 patches) was constructed based on CT scan data (0.6 mm slice thickness). Implant biomechanical markers: attach luminescent markers (3 mm diameter spheres) at abnormal coordinate locations.

[0194] Biomechanical data binding: joint range of motion data → bone rotation angle drive (e.g., 25° dorsiflexion → 15° rotation of the talus around the X-axis); muscle load data → tendon translucency mapping (the greater the load, the less transparent).

[0195] Interactive reporting functionality implementation: Core interactive controls: Rotation control: Free rotation along three axes (X / Y / Z axis sensitivity 0.5° / pixel displacement); Profile cutting: Dynamic profile cutting along the coronal / sagittal plane (profile thickness adjustable from 0.1 to 5 mm); Data probe: Clicking on an abnormal area will pop up a details card (displaying risk value, historical comparison, and clinical recommendations).

[0196] Automatic annotation function: When the viewing angle focuses on the abnormal area for more than 3 seconds: automatically zoom in to 2 times the field of view; overlay text annotation (such as "abnormal load on the peroneal tendon: peak value exceeds limit by 32%).

[0197] Report output format: WebGL standard format (loaded directly on web pages), supports immersive viewing with VR glasses.

[0198] Typical clinical case example (patient 10 weeks after Achilles tendon repair surgery): Risk visualization base map: Posterior calcaneal region (block 7) critical level red (risk level 3.4); Peroneus tendon groove (block 6) high risk orange (level 2.6).

[0199] Streamline animation anomaly: The blue particle flow at the Achilles tendon insertion point is interrupted (load transmission failure); vortices appear in the purple particle flow of the peroneal muscle group (abnormal shear force).

[0200] Rehabilitation bottleneck location: Keyframe T1 (heel lift phase) location coordinates (128.7, 92.3); interpolation revealed that the abnormal area's movement trajectory deviated from the normal path by 3.2 mm.

[0201] Interactive reporting: The rotating model shows calcifications in the Achilles tendon (2.1 mm in diameter); the data probe indicates: "Achilles tendon load capacity has only recovered to 62%, jump training is recommended to be postponed."

[0202] As can be seen, by collecting multimodal biomechanical data from patients performing standardized rehabilitation movements, a spatiotemporally synchronized motion-mechanical coupling dataset is generated. This dataset is then input into a variational autoencoder to form a three-dimensional biomechanical feature matrix. This matrix is ​​further input into a dynamic graph neural network, which outputs dual-channel predictions of joint stability risk probability and tissue healing abnormality probability. These dual-channel predictions are then spatiotemporally aligned with a personalized knowledge graph constructed from the patient's historical rehabilitation data to generate a quantified risk level. Based on this quantified risk level, a dynamic heatmap of the anatomical structure is generated, and a biomechanical load streamline animation is overlaid on the damaged tissue area. An interactive three-dimensional report containing the location of key rehabilitation bottlenecks is then output. This enables three-dimensional, visually interactive self-assessment of orthopedic patient rehabilitation, improving the objectivity, accuracy, and clinical applicability of rehabilitation assessments.

[0203] Another embodiment of the present invention provides a self-testing system for orthopedic patient rehabilitation, see [link to documentation]. Figure 3 The system may include: The acquisition module 301 is used to acquire multimodal biomechanical data of patients when performing standardized rehabilitation movements, and generate a spatiotemporally synchronized motion-mechanical coupling dataset. The separation module 302 is used to input the motion-mechanics coupled dataset into the variational autoencoder, and learn to separate the joint range vector, muscle load vector and neural control coordination vector through deentanglement representation learning to form a three-dimensional biomechanical feature matrix. The adjustment module 303 is used to input the three-dimensional biomechanical feature matrix into the dynamic graph neural network, dynamically adjust the connection weights of the graph nodes according to the individual patient's rehabilitation stage, and output the dual-channel prediction results of joint stability risk probability and tissue healing abnormality probability. Alignment module 304 is used to spatiotemporally align the dual-channel prediction results with the personalized knowledge graph constructed from the patient's historical rehabilitation data, calculate the deviation between the current rehabilitation status and the ideal rehabilitation trajectory through a nonlinear time warping algorithm, and generate a quantitative risk level. The generation module 305 is used to generate a dynamic heat map of the anatomical structure based on the quantified risk level, overlay biomechanical load streamline animation on the damaged tissue area, and output an interactive three-dimensional report containing the location of key rehabilitation bottlenecks.

[0204] 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.

[0205] Specifically, in this embodiment, the storage medium can be configured to store a computer program for performing the following steps: S201: Collect multimodal biomechanical data of patients performing standardized rehabilitation movements to generate a spatiotemporally synchronized motion-mechanical coupling dataset; S202, the motion-mechanics coupled dataset is input into the variational autoencoder, and the joint range vector, muscle load vector and neural control coordination vector are separated through deentanglement representation learning to form a three-dimensional biomechanical feature matrix; S203, the three-dimensional biomechanical feature matrix is ​​input into a dynamic graph neural network, the connection weights of the graph nodes are dynamically adjusted according to the individual patient's rehabilitation stage, and the dual-channel prediction results of joint stability risk probability and tissue healing abnormality probability are output. S204, the dual-channel prediction results are spatiotemporally aligned with the personalized knowledge graph constructed from the patient's historical rehabilitation data, and the deviation between the current rehabilitation status and the ideal rehabilitation trajectory is calculated using a nonlinear time warping algorithm to generate a quantitative risk level; S205, Based on the quantified risk level, generate a dynamic heat map of the anatomical structure, overlay biomechanical load streamline animation on the damaged tissue area, and output an interactive 3D report containing the location of key rehabilitation bottlenecks.

[0206] 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.

[0207] 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.

[0208] Specifically, in this embodiment, the processor can be configured to perform the following steps via a computer program: S201: Collect multimodal biomechanical data of patients performing standardized rehabilitation movements to generate a spatiotemporally synchronized motion-mechanical coupling dataset; S202, the motion-mechanics coupled dataset is input into the variational autoencoder, and the joint range vector, muscle load vector and neural control coordination vector are separated through deentanglement representation learning to form a three-dimensional biomechanical feature matrix; S203, the three-dimensional biomechanical feature matrix is ​​input into a dynamic graph neural network, the connection weights of the graph nodes are dynamically adjusted according to the individual patient's rehabilitation stage, and the dual-channel prediction results of joint stability risk probability and tissue healing abnormality probability are output. S204, the dual-channel prediction results are spatiotemporally aligned with the personalized knowledge graph constructed from the patient's historical rehabilitation data, and the deviation between the current rehabilitation status and the ideal rehabilitation trajectory is calculated using a nonlinear time warping algorithm to generate a quantitative risk level; S205, Based on the quantified risk level, generate a dynamic heat map of the anatomical structure, overlay biomechanical load streamline animation on the damaged tissue area, and output an interactive 3D report containing the location of key rehabilitation bottlenecks.

[0209] 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 self-examination method for orthopedic patients' rehabilitation, characterized in that, The method includes: Collect multimodal biomechanical data of patients performing standardized rehabilitation movements to generate a spatiotemporally synchronized motion-mechanics coupled dataset; The motion-mechanics coupled dataset is input into a variational autoencoder, and the joint range vector, muscle load vector, and neural control coordination vector are separated through deentanglement representation learning to form a three-dimensional biomechanical feature matrix. The three-dimensional biomechanical feature matrix is ​​input into a dynamic graph neural network, and the connection weights of the graph nodes are dynamically adjusted according to the individual patient's rehabilitation stage. The dual-channel prediction results of joint stability risk probability and tissue healing abnormality probability are output. The dual-channel prediction results are spatiotemporally aligned with a personalized knowledge graph constructed from the patient's historical rehabilitation data. A nonlinear time warping algorithm is used to calculate the deviation between the current rehabilitation status and the ideal rehabilitation trajectory, generating a quantitative risk level. Based on the quantified risk level, a dynamic heat map of the anatomical structure is generated, a biomechanical load streamline animation is superimposed on the damaged tissue area, and an interactive 3D report containing the location of key rehabilitation bottlenecks is output.

2. The method according to claim 1, characterized in that, The process involves collecting multimodal biomechanical data from patients performing standardized rehabilitation movements, generating a spatiotemporally synchronized motion-mechanical coupling dataset, including: Based on the patient's joint angular displacement sequence obtained by the optical motion capture system, the rigid body transformation of the skeleton is calculated to obtain high-precision motion trajectory data. Based on the raw electromyographic signals collected by the surface electromyography sensor array, power frequency interference was removed by wavelet packet decomposition to obtain denoised muscle activation time series data. Based on motion trajectory data and muscle activation timing data, a dynamic time warp algorithm is used to align timestamps and generate a spatiotemporally synchronized skeletal-muscle joint timing stream. Based on the joint temporal flow of the skeletal-muscle system, the three-dimensional force vectors measured by the plantar pressure plate are fused, and the joint torque vectors are calculated through the inverse dynamics model to output the motion-mechanics coupling dataset.

3. The method according to claim 2, characterized in that, The process involves inputting the motion-mechanics coupled dataset into a variational autoencoder, and learning to separate joint range vectors, muscle load vectors, and neural control coordination vectors through deentanglement representation learning, forming a three-dimensional biomechanical feature matrix, including: Based on the action-mechanics coupled dataset, spatiotemporal features are extracted using a convolutional encoder, and the initial encoding of the latent space is output. Based on the initial encoding of the latent space, the entangled subspace is separated by adversarial regularization constraints to obtain three types of latent variables: orthogonalized joint mobility, muscle load, and neural coordination. Based on the three types of latent variables, the joint kinematic parameters, electromyographic envelope signals and motor coordination indices are reconstructed through a multi-branch decoder, and the unentangled feature vector set is output. Based on the unentangled feature vector set, it is stacked into a tensor along the time axis, and a three-dimensional biomechanical feature matrix is ​​generated by normalizing the feature dimension.

4. The method according to claim 3, characterized in that, The process of inputting the three-dimensional biomechanical feature matrix into a dynamic graph neural network, dynamically adjusting the connection weights of graph nodes according to the individual patient's rehabilitation stage, and outputting a dual-channel prediction result of joint stability risk probability and tissue healing abnormality probability includes: Based on the three-dimensional biomechanical feature matrix, an initial graph structure is constructed with joints as nodes and biomechanical associations as edges, and a static topological graph is output. Based on the patient's recovery stage label, the edge weight update matrix of the static topology graph is generated through the meta-learning controller, and the updated dynamic adaptive graph is output. Based on the dynamic adaptive graph, a gated graph convolutional network is used to aggregate the features of neighboring nodes and output a spatiotemporal embedding vector. Based on the spatiotemporal embedding vector, the joint stability risk score and tissue healing abnormality score are calculated through a dual-channel fully connected layer, and the original probability distribution is output. Based on the original probability distribution, the bias is adjusted using a Bayesian uncertainty calibration module, and the calibrated dual-channel prediction results are output.

5. The method according to claim 4, characterized in that, The process involves spatiotemporally aligning the dual-channel prediction results with a personalized knowledge graph constructed from the patient's historical rehabilitation data, calculating the deviation between the current rehabilitation status and the ideal rehabilitation trajectory using a nonlinear time warping algorithm, and generating a quantitative risk level, including: Based on the patient's historical rehabilitation data, the joint-tissue association topology is extracted using a graph embedding algorithm, and a personalized knowledge graph embedding vector is output. Based on the dual-channel prediction results and the knowledge graph embedding vector, an attention mechanism is used to project them into a unified semantic space, and cross-modal alignment features are output. Based on cross-modal alignment features, a nonlinear time warping algorithm is used to match the temporal path between the current state and the ideal rehabilitation trajectory, and output the dynamic time warping distance. Based on the dynamic time curvature distance, it is mapped to a preset risk threshold range and a four-level quantitative risk level is generated, which includes low, medium, high and critical.

6. The method according to claim 5, characterized in that, The process involves generating a dynamic heatmap of the anatomical structure based on the quantified risk level, overlaying biomechanical load streamline animation on the damaged tissue area, and outputting an interactive 3D report containing the location of key rehabilitation bottlenecks, including: Based on the quantitative risk level and anatomical atlas, a color gradient heatmap of the joint region is generated through a multi-scale rendering engine, and a risk visualization base map is output. Based on biomechanical load streamline animation, the direction of muscle force transmission is simulated through a particle system, and a dynamic streamline overlay layer is output. Based on the risk visualization base map and dynamic streamline overlay, the keyframe interpolation algorithm is used to locate the rehabilitation bottleneck area and output the coordinate set of the abnormal area; Based on the coordinate set of abnormal areas, the system integrates the patient's three-dimensional skeletal model with biomechanical data to generate a rotatable, scalable, and interactive three-dimensional rehabilitation examination report.

7. A self-assessment system for orthopedic patient rehabilitation, characterized in that, The system includes: The acquisition module is used to collect multimodal biomechanical data of patients when performing standardized rehabilitation movements, and generate a spatiotemporally synchronized motion-mechanical coupling dataset. The separation module is used to input the motion-mechanics coupled dataset into the variational autoencoder, and learn to separate the joint range vector, muscle load vector and neural control coordination vector through deentanglement representation learning to form a three-dimensional biomechanical feature matrix. The adjustment module is used to input the three-dimensional biomechanical feature matrix into the dynamic graph neural network, dynamically adjust the connection weights of the graph nodes according to the individual patient's rehabilitation stage, and output a dual-channel prediction result of joint stability risk probability and tissue healing abnormality probability. The alignment module is used to align the dual-channel prediction results with a personalized knowledge graph constructed from the patient's historical rehabilitation data in a spatiotemporal manner. It calculates the deviation between the current rehabilitation status and the ideal rehabilitation trajectory using a nonlinear time warping algorithm to generate a quantitative risk level. The generation module is used to generate a dynamic heat map of the anatomical structure based on the quantified risk level, overlay biomechanical load streamline animation on the damaged tissue area, and output an interactive 3D report containing the location of key rehabilitation bottlenecks.

8. The system according to claim 7, characterized in that, The acquisition module is specifically used for: Based on the patient's joint angular displacement sequence obtained by the optical motion capture system, the rigid body transformation of the skeleton is calculated to obtain high-precision motion trajectory data. Based on the raw electromyographic signals collected by the surface electromyography sensor array, power frequency interference was removed by wavelet packet decomposition to obtain denoised muscle activation time series data. Based on motion trajectory data and muscle activation timing data, a dynamic time warp algorithm is used to align timestamps and generate a spatiotemporally synchronized skeletal-muscle joint timing stream. Based on the joint temporal flow of the skeletal-muscle system, the three-dimensional force vectors measured by the plantar pressure plate are fused, and the joint torque vectors are calculated through the inverse dynamics model to output the motion-mechanics coupling dataset.

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