A personalized orthopedic rehabilitation plan recommendation method and system

Abstract

The application discloses a kind of personalized orthopedics rehabilitation plan recommendation method and system, method includes: through wearable sensor, patient gait, joint angle and electromyogram are acquired, and fusion biomechanics feature matrix is generated;Fusion biomechanics feature matrix is registered with the bone density distribution map of patient CT / MRI image, and output quantization injury-compensation relationship diagram;Quantization injury-compensation relationship diagram is input adaptive action generator, and candidate rehabilitation action sequence meeting biomechanics constraint is generated based on joint kinematic chain inverse solution algorithm;Real-time physiological feedback test is carried out to candidate rehabilitation action sequence, and output safe action scheme set;Safe action scheme set is input metabolic efficiency optimization module, and final personalized rehabilitation plan table is generated with metabolic equivalent threshold as constraint condition.Utilize the embodiment of the application, the precision and safety of orthopedics rehabilitation can be improved.

Description

Technical Field

[0001] This invention belongs to the field of medical technology, specifically a method and system for recommending personalized orthopedic rehabilitation plans. Background Technology

[0002] In the field of orthopedic rehabilitation, traditional rehabilitation planning relies primarily on physician experience and static assessment data, making it difficult to comprehensively reflect the biomechanical changes and individual differences during the dynamic rehabilitation process. Existing technologies typically employ single-modal assessment methods, such as gait analysis or joint range of motion measurements alone, lacking comprehensive consideration and resulting in insufficiently targeted rehabilitation programs. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for recommending personalized orthopedic rehabilitation plans, in order to overcome the shortcomings of the prior art and improve the accuracy and safety of orthopedic rehabilitation.

[0004] One embodiment of this application provides a method for recommending personalized orthopedic rehabilitation plans, the method comprising: The patient's gait, joint angles and electromyography signals are acquired by wearable sensors. The original signals are spatiotemporally synchronized using a multimodal time series alignment algorithm to generate a fused biomechanical feature matrix. The fused biomechanical feature matrix is ​​registered with the bone density distribution map of the patient's CT / MRI images. The abnormal stress area and muscle compensation pattern are identified through the compensation effect detection model, and a quantitative damage-compensation relationship map is output. The quantified damage-compensation relationship diagram is input into the adaptive motion generator, and a candidate rehabilitation motion sequence that conforms to biomechanical constraints is generated based on the joint kinetic chain inverse solution algorithm. Each motion includes intensity, angle, and duration parameters. Real-time physiological feedback testing is performed on candidate rehabilitation movement sequences. Based on the coefficient of variation of electromyography signals and joint pressure sensor data, movement schemes that cause abnormal muscle activation or joint overload are dynamically screened out, and a set of safe movement schemes is output. The set of safe action plans is input into the metabolic efficiency optimization module. Based on the cardiopulmonary function monitoring data and the exercise oxygen consumption model, the metabolic load of different plans is predicted. The final personalized rehabilitation plan is generated with the metabolic equivalent threshold as a constraint.

[0005] Optionally, the step of acquiring patient gait, joint angles, and electromyographic signals through wearable sensors, and performing spatiotemporal synchronization processing on the original signals according to a multimodal time series alignment algorithm to generate a fused biomechanical feature matrix includes: Based on the joint six-degree-of-freedom motion trajectory and surface electromyography signal output by the IMU sensor, motion artifacts and physiological signals are separated by bandpass filtering to obtain the denoised original motion-electromyography time series data; Based on motion-electromyography time-series data, a dynamic time warping (DTW) algorithm is used to compensate for the differences in sampling rates of each sensor and generate a time-aligned multimodal signal stream. The multimodal signal stream is input into the biomechanical coordinate system transformation module, which maps the sensor's local coordinate system to the human body's global coordinate system through an inverse kinematics model, and outputs a spatially synchronized joint torque-muscle activation matrix. Tensor decomposition is performed on the joint torque-muscle activation matrix, and feature modes whose principal component energy ratio exceeds a preset ratio threshold are extracted and combined into a dimension-compressed fusion biomechanical feature matrix.

[0006] Optionally, the step of registering the fused biomechanical feature matrix with the bone mineral density distribution map of the patient's CT / MRI images, identifying abnormal stress areas and muscle compensation patterns through a compensation effect detection model, and outputting a quantitative damage-compensation relationship map includes: Based on the Hounsfield values ​​of the bone density distribution map, the spatial coordinates of the biomechanical feature matrix are mapped to the image anatomical coordinate system using a non-rigid B-spline registration algorithm to generate a bone-muscle coupling topology map. Based on the bone-muscle coupling topology, the finite element analysis method is used to calculate the deviation between the stress concentration area and the theoretical load-bearing threshold, and output the thermal map of the abnormal stress area. Muscle co-activation analysis was triggered by thermal mapping of abnormal stress areas. Compensatory and non-compensatory electromyographic signals were separated by independent component decomposition to generate a muscle compensation pattern coding table. By integrating the thermal map of the abnormal stress area with the coding table of muscle compensation patterns, a quantitative damage-compensation relationship diagram of damage force value and electromyographic compensation intensity in Newtons / square millimeter is constructed.

[0007] Optionally, the quantitative damage-compensation relationship diagram is input into an adaptive motion generator to generate a candidate rehabilitation motion sequence that conforms to biomechanical constraints based on a joint kinetic chain inverse solution algorithm. Each motion includes parameters of intensity, angle, and duration, including: Based on the distribution of damage force values ​​in the quantified damage-compensation relationship diagram, a mask of forbidden motion regions of the joint kinematic chain is generated by a convolutional neural network, and the biomechanical constraint boundary is output. Based on biomechanical constraints, the feasible region of joint angles is solved by inverse Jacobian matrix to generate a set of candidate joint motion trajectories that satisfy torque balance. Muscle length-tension optimization is performed on the candidate joint motion trajectory set. The Hill muscle model is used to dynamically adjust the motion amplitude and output the motion angle parameters with the optimal muscle activation efficiency. By combining cardiopulmonary function monitoring data, the energy consumption gradient of different action combinations is calculated through a metabolic equivalent prediction model, and action energy consumption labels containing intensity and duration are generated. By integrating motion angle parameters and motion energy consumption labels, Monte Carlo sampling is used to generate candidate rehabilitation motion sequences that meet Pareto optimality.

[0008] Optionally, the real-time physiological feedback test of the candidate rehabilitation movement sequence, dynamically filtering out movement schemes that cause abnormal muscle activation or joint overload based on the electromyographic signal variation coefficient and joint pressure sensor data, outputs a set of safe movement schemes, including: When executing candidate rehabilitation action sequences, muscle activation stability is detected by the coefficient of variation of electromyography signals, and a real-time alarm vector for abnormal muscle activation is generated. Based on joint pressure sensor data, a support vector regression model is used to predict the peak joint contact stress and output the joint overload risk index. By integrating abnormal muscle activation alarm vectors and joint overload risk indices, a fuzzy logic decision-maker is used to dynamically calculate action hazard scores. When the hazard score of an action exceeds the preset safety threshold, a real-time screening mechanism is triggered and the action sequence is updated, outputting a set of safe action plans.

[0009] Optionally, the step of inputting the set of safe action plans into the metabolic efficiency optimization module, predicting the metabolic load of different plans based on cardiopulmonary function monitoring data and exercise oxygen consumption models, and generating a final personalized rehabilitation plan using a metabolic equivalent threshold as a constraint includes: Based on the energy consumption labels of the safe action program set, the exercise oxygen consumption model is calibrated using cardiopulmonary function monitoring data, and the predicted value of metabolic load is output. Based on the predicted metabolic load, a non-dominated sorting genetic algorithm is used to perform multi-objective optimization with the goal of minimizing metabolic load and maximizing rehabilitation benefits. Set a preset multiple of the patient's resting metabolic equivalent as a threshold constraint to screen out action combinations that exceed the tolerance range; By integrating and optimizing the movement sequence and metabolic parameters, a final personalized rehabilitation plan is generated, which includes intensity, angle, duration, and energy consumption labels.

[0010] Another embodiment of this application provides a personalized orthopedic rehabilitation plan recommendation system, the system comprising: The processing module is used to acquire the patient's gait, joint angles and electromyography signals through wearable sensors, and to perform spatiotemporal synchronization processing on the raw signals according to a multimodal time series alignment algorithm to generate a fused biomechanical feature matrix. The registration module is used to register the fused biomechanical feature matrix with the bone density distribution map of the patient's CT / MRI images, identify abnormal stress areas and muscle compensation patterns through the compensation effect detection model, and output a quantitative damage-compensation relationship map. The generation module is used to input the quantified damage-compensation relationship diagram into the adaptive motion generator and generate a candidate rehabilitation motion sequence that conforms to biomechanical constraints based on the joint kinetic chain inverse solution algorithm. Each motion includes intensity, angle, and duration parameters. The screening module is used to perform real-time physiological feedback testing on candidate rehabilitation movement sequences. Based on the coefficient of variation of electromyography signals and joint pressure sensor data, it dynamically screens out movement schemes that cause abnormal muscle activation or joint overload, and outputs a set of safe movement schemes. The prediction module is used to input the set of safe action plans into the metabolic efficiency optimization module, predict the metabolic load of different plans based on cardiopulmonary function monitoring data and exercise oxygen consumption model, and generate the final personalized rehabilitation plan table with metabolic equivalent threshold as a constraint.

[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 personalized orthopedic rehabilitation plan recommendation method. It acquires patient gait, joint angles, and electromyographic signals using wearable sensors to generate a fused biomechanical feature matrix. This matrix is ​​then registered with the bone density distribution map of the patient's CT / MRI images to output a quantified damage-compensation relationship map. This map is input into an adaptive motion generator, which generates candidate rehabilitation motion sequences that conform to biomechanical constraints based on a joint kinetic chain inverse solving algorithm. Real-time physiological feedback testing is performed on the candidate rehabilitation motion sequences to output a safe motion scheme set. Finally, this safe motion scheme set is input into a metabolic efficiency optimization module, using a metabolic equivalent threshold as a constraint to generate a final personalized rehabilitation plan, thereby improving the accuracy and safety of orthopedic rehabilitation. Attached Figure Description

[0014] Figure 1 Hardware structure block diagram of a computer terminal for a personalized orthopedic rehabilitation plan recommendation method provided in an embodiment of the present invention; Figure 2A flowchart illustrating a personalized orthopedic rehabilitation plan recommendation method provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a personalized orthopedic rehabilitation plan recommendation 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 method for recommending personalized orthopedic rehabilitation plans. This method 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 personalized orthopedic rehabilitation plan recommendation method 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 recommended method for a personalized orthopedic rehabilitation plan.

[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 on non-volatile storage media, which, when executed by a processor, enable the processor to perform any personalized orthopedic rehabilitation plan recommendation.

[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 a method for recommending personalized orthopedic rehabilitation plans, which may include the following steps: S201 acquires the patient's gait, joint angles, and electromyographic signals through wearable sensors, and performs spatiotemporal synchronization processing on the original signals according to a multimodal time series alignment algorithm to generate a fused biomechanical feature matrix; Specifically, based on the joint six-degree-of-freedom motion trajectory and surface electromyography signal output by the IMU sensor, motion artifacts and physiological signals can be separated by bandpass filtering to obtain the denoised original motion-electromyography time series data; Multi-source sensor data acquisition and noise identification: IMU (Inertial Measurement Unit) data analysis: Two nine-axis IMU sensors (including a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer) are fixed to the patient's target joint (such as the knee joint) to collect six degrees of freedom motion data of the joint in real time at a sampling rate of 200 Hz (a unit of frequency, representing 200 samples per second). (Degrees of freedom refer to the number of directions of motion, including translation (forward / backward / left / right / up / down) and translation (roll / pitch / yaw / rotation). For example, the femoral end IMU records thigh movements, and the tibial end IMU records lower leg movements. The fusion of the two data can accurately calculate the knee joint flexion and extension angles.

[0024] Motion artifact (noise caused by equipment shaking) feature identification: When the accelerometer reading changes abruptly by more than 5G (G represents gravitational acceleration, 1G = 9.8 m / s²), the artifact is identified. 2 When a segment does not match the physiological movement of the joint, it is marked as a noise segment.

[0025] Optimization of surface electromyography (sEMG) signal acquisition: Bipolar electrodes are attached to the target muscle group (such as the quadriceps femoris) and muscle electrical signals are acquired at a sampling rate of 1000 Hz.

[0026] Distinguishing between physiological signals and noise: The effective frequency range of electromyography is 20-450 Hz. Components below 20 Hz are mostly skin friction artifacts, while components above 450 Hz are mostly environmental electromagnetic interference.

[0027] Band-pass filtering (a noise reduction technique that retains only signals at specific frequencies) is performed dynamically.

[0028] Dynamic configuration of filter parameters: For IMU data: A fourth-order Butterworth filter (a filter design for smoothing frequency response) is used with a passband range of 0.1-10 Hz to filter out high-frequency jitter (>10Hz) caused by clothing friction during walking and low-frequency noise (<0.1Hz) caused by equipment drift.

[0029] For electromyography data: Set the passband to 20-450 Hz, and add a notch filter (specifically for eliminating mains interference) at 60 Hz to eliminate power supply frequency interference.

[0030] Clinical noise reduction example: When a patient walks, a high-frequency jitter of 12 Hz (caused by clothing friction) is detected in the raw IMU data. After filtering, the amplitude is reduced by 90%; the mains interference of 50 Hz in the electromyography signal is completely suppressed, and the fidelity of the physiological signal is improved to 95%.

[0031] Output verification mechanism: The signal-to-noise ratio (SNR) of the filtered data must be greater than 30 dB (logarithmic ratio unit). If the standard is not met, the passband boundary will be automatically widened and the data will be re-filtered.

[0032] Based on motion-electromyography time-series data, a dynamic time warping (DTW) algorithm is used to compensate for the differences in sampling rates of each sensor and generate a time-aligned multimodal signal stream. Multimodal signal timing deviation analysis: Sampling rate discrepancy issue: The IMU sampling rate of 200 Hz (200 points per second) and the EMG sampling rate of 1000 Hz (1000 points per second) result in an asymmetry in the number of data points (for example, within the same second, the IMU outputs 200 data points and the EMG outputs 1000 points).

[0033] Motion event binding failure: For example, at the peak of knee flexion, the IMU marks it at 5 milliseconds (ms), but the electromyographic signal needs to be interpolated and matched at the corresponding time.

[0034] Critical event marking strategy: Extracting motion feature events: using the point where the joint angular velocity crosses zero (such as the moment of transition from knee flexion to knee extension) detected by the IMU as the time anchor point.

[0035] Synchronization of electromyographic signal events: the electromyographic burst period (such as the quadriceps activation peak) within 50 milliseconds (ms) before and after locking the anchor point.

[0036] Execution flow of Dynamic Time Warping (DTW, a non-linear time alignment method).

[0037] Core operations of the algorithm: Construct the cost matrix: calculate the difference between the IMU and the electromyographic signal at each time point (e.g., Euclidean distance).

[0038] Find the optimal warping path: Search the matrix for the path with the minimum cumulative cost from the starting point (1,1) to the ending point (m,n) (the slope of the path is limited to between 1:2 and 2:1 to avoid excessive time distortion).

[0039] Clinical alignment case: During a 3-second walking cycle, the patient's electromyographic signals need to be compressed to 600 key points (originally 3000) to match the IMU's 600 points. DTW identified: The moment of heel strike (IMU point 120) corresponds to the gastrocnemius muscle electromyography burst (original electromyography point 590); the moment of toe lift-off (IMU point 480) corresponds to the tibialis anterior muscle activation (original electromyography point 2450).

[0040] Output standardization: Generate a uniform 500 Hz resampled signal stream (one data point every 2 milliseconds) to ensure strict time synchronization between the IMU and EMG data (error < 1 millisecond).

[0041] The multimodal signal stream is input into the biomechanical coordinate system transformation module, which maps the sensor's local coordinate system to the human body's global coordinate system through an inverse kinematics model, and outputs a spatially synchronized joint torque-muscle activation matrix. Anatomical basis of coordinate system transformation: Limitations of local coordinate systems: The coordinate system of the femoral end IMU is based on the sensor itself (with the X-axis pointing to the long side of the device), which cannot directly reflect the knee flexion angle (it needs to be converted to the rotation of the lower leg relative to the thigh).

[0042] Directional influence of electromyographic electrodes: Deviation in electrode attachment angle can lead to signal amplitude error (e.g., when the electrode deviates from the direction of the muscle fiber by 30°, the amplitude decreases by 15%).

[0043] Global coordinate system definition: Based on the anatomical planes of the human body: the sagittal plane (the plane that divides the anteroposterior direction) is used to calculate the flexion and extension angles; the coronal plane (the plane that divides the left and right directions) is used to calculate adduction and abduction; the origin is set at the center of joint rotation (such as the midpoint of the intercondylar fossa of the femur in the knee joint).

[0044] Spatial registration of the inverse kinematics model (an algorithm that infers joint angles from end-effector motion).

[0045] Spatial mapping from sensors to bones: Calibration phase: Record the initial posture of the IMU when the patient is upright and align it with the skeletal coordinate system obtained from the CT scan (error < 2 mm).

[0046] Dynamic conversion: IMU data → converted into Euler angles (a three-dimensional rotation representation) of the femur relative to the pelvis via a quaternion rotation matrix. Electromyographic signal → Based on the angle between the electrode and the muscle direction → Corrected to activation intensity along the muscle fiber direction.

[0047] Calculation of joint torque (the force required to rotate a joint): Input: Angular acceleration of the IMU (unit: radians / second) 2 ) and human segment parameters (such as lower leg weight 1.5 kg, length 40 cm); Equivalent description of the formula: Torque = Moment of inertia × Angular acceleration (Moment of inertia is provided by the limb mass distribution model).

[0048] Example: Lower leg angular acceleration 30 radians / second 2 → Knee extension torque is 35 Newton-meters (Nm).

[0049] Muscle activation matrix generation: Time-aligned electromyographic signals are normalized to the percentage of maximum voluntary contraction (%MVC), such as a quadriceps activation level of 45%MVC.

[0050] By integrating joint torque and muscle activation data, a matrix structure of rows (time points) × columns (torque values ​​+ activation intensity of each muscle) is formed.

[0051] Tensor decomposition is performed on the joint torque-muscle activation matrix, and feature modes whose principal component energy ratio exceeds a preset ratio threshold are extracted and combined into a dimension-compressed fusion biomechanical feature matrix.

[0052] The clinical necessity of dimensionality reduction for high-dimensional data: Original data redundancy problem: A single gait cycle generates 2000 time points × (1 joint torque + 8 sets of electromyographic signals) = 18,000 dimensions of data, of which 90% is noise or redundancy.

[0053] Key features are hidden: for example, the compensatory activation pattern of the gluteus medius in the early stage of the gait support phase may be drowned out by noise.

[0054] Preset energy threshold setting: The threshold for the cumulative energy percentage of principal components is set to 95% (meaning that the features after dimensionality reduction must retain 95% of the information content of the original data).

[0055] Adaptive adjustment: When the patient's movement pattern is complex (e.g., gait asymmetry > 15%), the threshold is increased to 98%.

[0056] Tensor Decomposition (a high-dimensional data feature extraction technique) execution flow.

[0057] Implementation of CP decomposition (CANDECOMP / PARAFAC Decomposition, a tensor decomposition model): The three-dimensional tensor (time × biomechanical variable × action cycle) is decomposed into three factor matrices: Time factor matrix: characterizes the features of the movement phase (e.g., support phase / swing phase); Variable factor matrix: reveals the torque-electromyographic coupling pattern (e.g., peak torque is synchronized with gastrocnemius muscle activation); Periodic factor matrix: reflects the consistency of movement repetition.

[0058] Decomposition rank (feature complexity) selection: Calculate using alternating least squares (ALS, an iterative optimization algorithm) until the residual (amount of unexplained data) is <5%.

[0059] Feature modality filtering example: The decomposition yields 20 component modes, sorted by energy: Modus 1: Knee flexion and extension torque dominates (energy percentage 48%); Modus 2: Vas lateralis-gastrocnemius synergistic activation (22%); Modus 3: Tibialis anterior burst mode during swing phase (15%); other modalities total <15% → only the first 3 modalities are retained (cumulative percentage 85%, more modalities need to be added if the 95% threshold is not reached); finally, the first 5 modalities are retained (cumulative energy 97%), and the original 18,000-dimensional data is compressed into a 5-dimensional feature vector.

[0060] The fusion matrix output structure is as follows: each row represents an action cycle (such as walking in one step), and each column represents the feature modality intensity value (such as [-0.35, 1.2, 0.8, -0.6, 0.4]).

[0061] Clinical interpretability mapping: negative values ​​represent compensatory patterns (e.g., a negative value in mode 1 indicates insufficient knee flexion torque).

[0062] S202, the fused biomechanical feature matrix is ​​registered with the bone density distribution map of the patient's CT / MRI images, and the abnormal stress area and muscle compensation pattern are identified through the compensation effect detection model, and a quantitative damage-compensation relationship map is output. Specifically, based on the Hounsfield value of the bone density distribution map, the spatial coordinates of the biomechanical feature matrix can be mapped to the imaging anatomical coordinate system using a non-rigid B-spline registration algorithm to generate a bone-muscle coupling topology map. The spatial correlation basis between bone mineral density data and biomechanical characteristics: Clinical interpretation of Hounsfield value (CT image grayscale value, reflecting bone mineral density): Bone mineral density (BMD) maps are extracted from patient CT images, with each pixel's Hounsfield value representing the degree of local bone mineralization. For example, healthy cortical bone has a value >1000 (HU), while osteoporotic areas have a value <300 HU (HU is Hounsfield unit). Patients with femoral neck fractures may experience a sharp drop in BMD to 200 HU, requiring special attention.

[0063] Spatial coordinate alignment challenge: Joint pressure data (such as tibial plateau pressure in the knee joint) in the biomechanical feature matrix need to be precisely matched with the anatomical position in the CT image. However, there is a rotational and translational deviation between the sensor coordinate system (with the IMU as the origin) and the image coordinate system (with the anterior superior iliac spine as the origin).

[0064] The necessity of non-rigid registration (an image alignment technique that allows for deformation): Individual patient differences can cause errors of up to 10 millimeters (mm) in standard rigid registration (translation and rotation only), which cannot meet the requirements for precise positioning of the force-bearing area.

[0065] Dynamic deformation compensation mechanism: Soft tissue deformation during gait can cause changes in the relative position of bones (e.g., the patella moves 30 mm when the knee is flexed), so a deformable registration algorithm is required.

[0066] Execution flow of the non-rigid B-spline based registration algorithm.

[0067] Registration control point mesh construction: Three-dimensional mesh control points (10 mm spacing) are set on CT images, and each point drives local image deformation. For example, the mesh is refined (5 mm spacing) in the distal femur to accommodate complex articular surfaces.

[0068] Biomechanical feature mapping: The pressure point (such as the pressure peak point on the anterior side of the tibial plateau) in the sensor coordinate system is assigned initial coordinates, and the point is moved to the corresponding anatomical position on the CT image (10 mm lateral to the tibial tuberosity) through iterative optimization.

[0069] Multimodal similarity measurement and optimization: Similarity criterion: Using mutual information as an indicator, maximize the statistical correlation between bone mineral density gradient and biomechanical pressure gradient (e.g., the pressure peak area should correspond to the high bone mineral density area).

[0070] Gradient descent optimization: The algorithm automatically adjusts the displacement of 800-1200 control points until the registration error is <2 mm. Clinical validation shows that the average error after femoral condyle registration is 1.8 mm.

[0071] Generation of the Bone-Muscle Coupling Topology Map: Output three-dimensional fusion atlas: red marks osteoporotic areas of CT images (e.g., <300 HU), blue overlays high biomechanical pressure areas (e.g., >0.5 N / mm² pressure), and yellow areas represent muscle attachment points (e.g., patellar tendon insertion).

[0072] Dynamic association rule: When the knee is flexed at 45°, the pressure distribution on the patellar articular surface and the direction of the quadriceps muscle tension line are displayed in the topology diagram as linked vector arrows.

[0073] Based on the bone-muscle coupling topology, the finite element analysis method is used to calculate the deviation between the stress concentration area and the theoretical load-bearing threshold, and output the thermal map of the abnormal stress area. Finite Element Analysis (FEM) Model Building and Personalized Adaptation: Skeletal geometry reconstruction and material property assignment: A 3D skeletal model (0.5 mm precision) is extracted from the registered CT images and divided into 100,000 to 500,000 tetrahedral elements. For example, the tibia model is divided into 120,000 elements, with the element size refined to 0.8 mm near the articular surfaces.

[0074] Material property definition: Bone mineral density → elastic modulus mapping: The elastic modulus of the osteoporotic area (300 HU) is set to 1 GPa, and that of healthy bone (1000 HU) is 15 GPa (GPa is a unit of material stiffness).

[0075] Cartilage layer addition: A 2 mm thick cartilage layer (elastic modulus 5 MPa) is applied to the surface of the tibial plateau.

[0076] Physiological load boundary condition settings: Ligament constraint: The medial collateral ligament (MCL) is simulated as a nonlinear spring (stiffness 200 N / mm) to limit excessive tibial displacement.

[0077] Muscle force loading: Based on the intensity of electromyographic activation (e.g., 45% MVC activation in the quadriceps), it is converted into actual muscle force (approximately 350 Newtons) and applied to the bone through the tendon attachment point.

[0078] Stress concentration identification and load-bearing threshold deviation calculation: The principle behind the generation of the theoretical load threshold: Health benchmark model: Based on data from 100 healthy individuals, establish upper limits for weight-bearing capacity in different bone density regions. For example, the threshold for the femoral neck cortical bone region is 15 MPa, and the threshold for the cancellous bone region is 5 MPa.

[0079] Individualized downsizing for patients: The osteoporosis threshold (<300 HU) is reduced to 60% of the baseline value (e.g., from 5 MPa to 3 MPa).

[0080] Stress Deviation Quantification Methods: Finite element solution: Calculate the maximum principal stress for each element and identify regions that exceed the local threshold (e.g., the stress of an element is 4.2 MPa, which is greater than the threshold of 3 MPa).

[0081] Deviation grading: Level 1 deviation (yellow zone): exceeding the threshold by less than 10% (3.3 MPa); Level 2 deviation (orange zone): exceeding the threshold by 10-30% (3.3-3.9 MPa); Level 3 deviation (red zone): exceeding the threshold by more than 30% (>3.9 MPa).

[0082] Heatmap visualization output: Spatial mapping: The stress values ​​of the finite element elements are mapped back to the CT image space to generate a two-dimensional coronal / sagittal thermal map. The red area represents the stress concentration on the medial tibial plateau (peak value 4.8 MPa, exceeding the threshold by 60%).

[0083] Dynamic marking: During the mid-gait support phase, the patellofemoral joint stress suddenly increases from 1.2 MPa to 5.1 MPa (exceeding the threshold by 70%), and the thermal map flashes an alarm in real time.

[0084] Muscle co-activation analysis was triggered by thermal mapping of abnormal stress areas. Compensatory and non-compensatory electromyographic signals were separated by independent component decomposition to generate a muscle compensation pattern coding table. Causal association between muscle synergy and abnormal stress: Heatmap event triggering mechanism: When the area of ​​the red zone on the heat map is greater than 50 square millimeters or the peak stress lasts for more than 0.5 seconds, the electromyography compensation analysis module is automatically activated.

[0085] Spatiotemporal correlation locking: Electromyographic signals within 200 milliseconds (ms) before and after the stress peak are marked as key analysis segments. For example, when the tibial plateau stress is 4.8 MPa, the activation sequence of the quadriceps and hamstring muscles is analyzed simultaneously.

[0086] Physiological characteristics of the compensation pattern: Normal synergy: The activation ratio of the vastus lateralis (VL) to the vastus medialis (VM) is 1:1 (time difference < 30 milliseconds).

[0087] Typical compensation: When stress exceeds the limit, the gastrocnemius muscle (GAS) is abnormally activated prematurely (150 milliseconds before heel strike), thus sharing the load on the knee.

[0088] Signal separation techniques using Independent Component Analysis (ICA).

[0089] Blind source separation of multichannel electromyography signals: Input signal: 8-channel electromyography timing data (rectus femoris / vastus lateralis / vastus medialis / semitendinosus / gastrocnemius, etc.), sampling rate 1000 Hz.

[0090] Decomposition objective: to separate mutually independent activation pattern components (such as native electromyographic components vs. compensatory cross-interference components).

[0091] Compensatory signal identification criteria: Non-physiological frequency band characteristics: compensatory signals often contain discrete high-frequency bursts (>200 Hz), while the main energy of normal electromyography is 50-150 Hz.

[0092] Correlation with peak stress time: The correlation coefficient between the compensatory component and the time when the red zone appears on the heat map is >0.7 (normal component <0.3).

[0093] Generation of the Muscle Compensation Coding Table: Quantification of compensation levels: Grade 0: No non-compensatory components (normal); Grade 1: Single muscle compensation (e.g., gastrocnemius muscle contribution > 40%); Grade 2: Multi-muscle group chain compensation (e.g., gastrocnemius muscle + soleus muscle synergistic compensation).

[0094] Clinical example outputs are shown in Table 1: Table 1 ; By integrating the thermal map of the abnormal stress area with the coding table of muscle compensation patterns, a quantitative damage-compensation relationship diagram of damage force value and electromyographic compensation intensity in Newtons / square millimeter is constructed.

[0095] Spatial alignment and weight allocation in multimodal data fusion: Heatmap - Coding Table Coordinates Unified: Map the anatomical coordinates of the heat map (e.g., medial tibial plateau X=32mm, Y=15mm) to the muscle attachment area (e.g., semimembranosus insertion X=30mm, Y=18mm).

[0096] Spatial association rule: Based on Euclidean distance, bone areas with a distance of less than 5 mm are considered to be directly associated with muscles (such as the patellar stress area and the quadriceps tendon insertion point).

[0097] Damage Force Value Definition: Calculate the area-weighted value of the stress exceeding the threshold: Damage force per unit area in the red zone = element stress × over-limit ratio (e.g., 4 MPa over-threshold 33% → 1.32 N / mm²); Total damage force = Σ (red zone area × unit damage force).

[0098] Example: Tibial plateau red zone area 80 square millimeters, average damage force 1.5 Newtons / square millimeter → total damage 120 Newtons.

[0099] Quantitative modeling of EMG compensation intensity.

[0100] Intensity calculation formula: Single muscle compensatory intensity = Root mean square of compensatory signal (RMS) ÷ RMS of normal signal × 100% (e.g., RMS 0.25 mV vs 0.4 mV → intensity 62.5%).

[0101] Regional integration: Anterior knee region compensation strength = (mean strength of rectus femoris + vastus lateralis + vastus medialis) × compensation level coefficient (level 1 × 1.0, level 2 × 1.3).

[0102] Modeling the damage-compensation coupling relationship: Establish a two-dimensional relationship matrix: the horizontal axis represents the damage force value (0-10 N / mm²), and the vertical axis represents the compensation strength (0-100%).

[0103] Clinical distribution pattern: High-risk zone (red): Damage > 5 N / mm² + compensation > 70% → immediate intervention required; Warning zone (yellow): Damage 2-5 N / mm² + compensation 40-70% → warning.

[0104] Clinical applications of the Quantified Damage-Compensation Relationship Map.

[0105] Visual output format: 3D skeletal model overlay: Bone surface coloring: Damage force values ​​are scaled in Newtons per square millimeter (gradient from blue to yellow to red); Muscle vector arrows: Compensation strength is represented by arrow thickness (thin arrows 30%, thick arrows 80%); Dynamic correlation curves: Plotting the curves of damage force and compensation strength changes during the gait cycle (e.g., when the heel strikes, the peak damage force is 3.8 Newtons per square millimeter, and the compensation strength rises to 75% simultaneously).

[0106] Decision support functions: Automatic identification of compensation source: When the hamstring compensation strength is >50% and the distance from the posterior femoral condyle stress zone is <10 mm, it is determined as "hamstring overcompensation leading to posterior femoral condyle overload".

[0107] Treatment plan correlation: High injury-high compensation areas are directly associated with the rehabilitation movement contraindication library (e.g., the red zone corresponds to the prohibition of squatting >60°).

[0108] S203, the quantified damage-compensation relationship diagram is input into the adaptive motion generator, and a candidate rehabilitation motion sequence that conforms to biomechanical constraints is generated based on the joint kinetic chain inverse solution algorithm. Each motion includes intensity, angle and duration parameters. Specifically, based on the distribution of damage force values ​​in the quantified damage-compensation relationship diagram, a mask of forbidden motion regions of the joint kinematic chain can be generated through a convolutional neural network, and biomechanical constraint boundaries can be output. Clinical mapping and feature extraction of injury force distribution: The core input for quantifying the damage-compensation relationship map is the bone surface area measured in Newtons per square millimeter (N / mm²). 2 The data represents the distribution of injury force values ​​in units of 0.5. For example, the medial region of the tibial plateau is marked as a high-injury area (peak value 3.8 N / mm). 2 The outer region is a low-damage area (peak value 0.9 N / mm²). 2 These values ​​are mapped to the joint anatomy through a three-dimensional mesh, with each mesh cell (1×1 mm in size) containing an independent damage force value.

[0109] Architecture design of Convolutional Neural Networks (CNNs, a type of deep learning model adept at processing spatial data): The input layer receives a 512×512 pixel heatmap of damage force values ​​(pixel value range 0-10 N / mm). 2 Features are extracted using three sets of convolutional-pooling modules: The first group consists of 64 convolutional kernels (feature detectors) of size 3×3 with a stride of 1 (used to scan heatmaps to extract local injury patterns); the second group consists of 128 5×5 convolutional kernels to capture cross-regional injury correlations (such as the force transmission path between the femoral condyle and the tibial plateau); the third group consists of 256 7×7 convolutional kernels to identify large-scale taboo patterns (such as the cumulative damage across the entire knee flexion chain); ReLU is used as the activation function to avoid gradient vanishing.

[0110] Forbidden region mask (binary forbidden region marker) generation logic: The output layer uses the Sigmoid function (which compresses values ​​to the range of 0-1) to predict the taboo probability of each pixel. Regions with a probability > 0.65 are marked as taboo regions (value 1), and the rest are safe regions (value 0).

[0111] Clinical example: Patient with posterior femoral condyle injury force > 4.2 N / mm 2 The area is marked as a red forbidden zone, corresponding to the prohibited range of movements where the knee flexion is greater than 60°.

[0112] Boundary optimization mechanism: Morphological closing (an image processing technique) is used to smooth mask edges and eliminate isolated taboo points.

[0113] Global constraints on output joint kinetic chains (such as the knee-ankle linkage chain): Prohibit combined movements involving knee flexion of 60°-90° and ankle dorsiflexion of >15°.

[0114] Based on biomechanical constraints, the feasible region of joint angles is solved by inverse Jacobian matrix to generate a set of candidate joint motion trajectories that satisfy torque balance. Transformation from constraint boundary to kinematic model: Biomechanical constraints are translated into degrees of freedom (DOF) restrictions on joint kinematic chains. For example, the knee joint's flexion and extension degrees of freedom are limited to 0°-50° (avoiding the 60°-90° contraindication zone), and the hip joint's adduction and abduction degrees of freedom are set to -10° to +15° (to prevent pelvic compensation).

[0115] The principle behind solving the inverse Jacobian matrix (a mathematical tool for inferring joint angles from end-effector motion): Define the end-effector of rehabilitation movements (such as the target position of the foot or hand): for example, requiring the heel movement trajectory to be a straight path of 30 cm (the heel-to-toe lift-off section in gait training).

[0116] The Jacobian matrix establishes the relationship between joint angular velocity (θ') and end-effector linear velocity (v): v = J(θ) × θ'. The reverse solution involves deriving the joint angular velocity combination from the end-effector target velocity.

[0117] Dynamic calculation of the feasible region: Taking the gait support phase as an example: Input: The heel must move 20 cm within 0.5 seconds (end velocity 40 cm / s); Constraints: Knee angle ≤ 50°, hip torque ≤ 45 Nm; Output: Filter out the combined movement range of knee flexion 15°-45° and hip flexion 10°-25° (excluding schemes with knee > 50°); Tens of thousands of candidate trajectories are generated: each trajectory contains a combination of joint angles at 50 time points (time resolution 10 milliseconds).

[0118] Torque Balance Verification Mechanism: Each candidate trajectory must be validated using Newton-Euler dynamics equations. Calculate joint torques based on limb mass distribution (e.g., lower leg mass 1.8 kg, length 0.4 m) and measured accelerations (e.g., knee extension angle acceleration 50 radians / second). 2 ); Balance principle: Active muscle torque (such as quadriceps exertion) must be greater than 110% of the resistance torque (gravity + external load) to avoid compensation.

[0119] Trajectory set screening criteria: Remove schemes that cause joint torque to exceed the limit (e.g., knee flexion torque > 80 Nm) or muscle synergy imbalance (e.g., hamstring / quadriceps activation ratio > 1.5).

[0120] Muscle length-tension optimization is performed on the candidate joint motion trajectory set. The Hill muscle model is used to dynamically adjust the motion amplitude and output the motion angle parameters with the optimal muscle activation efficiency. Modeling of muscle biomechanical properties: The core components of the Hill-type Muscle Model: Contractile Element (CE): Simulates the active contractile force of muscle fibers; the force value depends on the activation level (0-100%) and the contraction speed. Parallel Elastic Element (PE): Represents the passive tension of connective tissue, which increases non-linearly with muscle stretching; Series Elastic Element (SE): Simulates tendon elasticity, with a force transmission delay of approximately 30 milliseconds.

[0121] Length-Tension Relationship Optimization Objective: Muscles exert the most force at a specific length (e.g., the quadriceps muscle has the optimal muscle fiber length when the knee is flexed at 30°).

[0122] The optimization algorithm traverses the joint angle sequence of each trajectory and calculates the corresponding muscle length (e.g., the rectus femoris muscle length is 25 cm when the knee is flexed at 40°).

[0123] Dynamic amplitude adjustment and efficiency evaluation: Adjustment strategy: If the muscle length deviates from the optimal value by more than 10% (e.g., the length of the vastus lateralis is 23 cm < the optimal value of 25 cm), then adjust the range of motion: reduce knee flexion by 5° to 35°; if the tendon strain (extension rate) is greater than 5% (e.g., the Achilles tendon is extended by 8 mm, exceeding the safety limit), then limit the ankle dorsiflexion angle by 3°.

[0124] Activation Efficiency Quantification Methods: Calculate the ratio of actual muscle force to maximum isometric contraction force (MVC), and after optimization, the efficiency of the target muscle group (such as the quadriceps) should be >65%.

[0125] Example of output parameters: Optimizing the angle of the wall squat from 50° to 42° increases the activation efficiency of the vastus medialis muscle from 58% to 73%.

[0126] By combining cardiopulmonary function monitoring data, the energy consumption gradient of different action combinations is calculated through a metabolic equivalent prediction model, and action energy consumption labels containing intensity and duration are generated. Multi-source physiological data fusion and energy consumption modeling: Real-time acquisition of cardiopulmonary function monitoring data: Oxygen uptake (VO2, in ml / kg / min): Measured by a portable metabolic analyzer (e.g., the patient's resting VO2 is 3.5 ml / kg / min).

[0127] Heart rate (HR, beats per minute): The chest strap sensor records changes in heart rate during exercise (e.g., heart rate increases from 72 to 105 beats per minute during the exercise).

[0128] Construction of a Metabolic Equivalent (MET) prediction model: Formula concept description: 1 MET = resting energy consumption (approximately 3.5 ml / kg / min), exercise MET value = exercise oxygen uptake ÷ resting oxygen uptake.

[0129] Input motion parameters (intensity / angle / duration) and physiological data (HR+VO2), and output the MET prediction value through random forest regression (e.g., wall squat MET=2.8).

[0130] Energy Cost Gradient and Label Generation: Gradient calculation: Energy consumption per action: MET value × duration (minutes) × weight (kg) × 0.0175 (conversion factor). For example, the energy consumption of a 70 kg patient performing a 10-minute wall squat is 2.8 × 10 × 70 × 0.0175 = 343 kcal.

[0131] Total energy consumption of the combined scheme: sum up the energy consumption of all movements (e.g., static squat 343 kcal + seated knee extension 210 kcal = 553 kcal).

[0132] Energy Cost Label Definition: Low energy consumption scheme: MET peak < 3.0, total energy consumption < 400 kcal; Medium energy consumption scheme: MET peak 3.0-5.0, total energy consumption 400-600 kcal; High energy consumption scheme: MET peak > 5.0, total energy consumption > 600 kcal; A tag is attached to each action sequence (e.g., "Scheme A: Medium energy consumption, total energy consumption 520 kcal").

[0133] By integrating motion angle parameters and motion energy consumption labels, Monte Carlo sampling is used to generate candidate rehabilitation motion sequences that meet Pareto optimality.

[0134] Definition and sampling strategy of multi-objective optimization problem: Optimize conflicting objectives: maximize muscle activation efficiency (requires large range of motion); minimize metabolic load (requires low intensity); minimize joint injury risk (requires avoiding prohibited angles).

[0135] Monte Carlo Sampling (a search algorithm based on random sampling) execution flow: Randomly generate 500,000 combinations of action parameters within the feasible region: Knee flexion angle: uniformly sampled from 0° to 50°; ankle dorsiflexion angle: randomly selected from -5° to +20°; intensity: elastic band resistance 5-30 Newtons (N) randomly selected; duration: 10-60 seconds per set of movements randomly set; three target values ​​(efficiency value / energy consumption value / risk value) are calculated for each set of parameters to form a three-dimensional solution space.

[0136] Pareto Optimality (a set of high-quality solutions for multi-objective optimization) selection: Non-dominated solution identification criteria: The condition for solution A to dominate solution B is: A's efficiency ≥ B, energy consumption ≤ B, risk ≤ B, and at least one of these conditions must be strictly better than B.

[0137] Pareto Front Construction: Retain all solutions that are not dominated by other solutions (e.g., Solution X: efficiency 75% / energy consumption 450 kcal / risk 0.1; Solution Y: efficiency 68% / energy consumption 380 kcal / risk 0.05).

[0138] Enhanced clinical suitability: Eliminate regimens with muscle efficiency <60% or risk >0.3 (the risk value is a normalized index of 0-1, and >0.3 indicates a high probability of exceeding the red zone limit).

[0139] The final output is 200-300 Pareto optimal action sequences, for example: Sequence 1: Wall squat (angle 42°, resistance 15N, 3 sets x 30 seconds) + seated knee extension (angle 0°-30°, resistance 10N, 3 sets x 45 seconds); Sequence 2: Straight leg raise (height 20 cm, 5 sets x 20 seconds) + ankle pump (amplitude -5° to +15°, 4 sets x 40 seconds).

[0140] S204 performs real-time physiological feedback testing on candidate rehabilitation movement sequences, dynamically filters out movement schemes that cause abnormal muscle activation or joint overload based on electromyographic signal variation coefficient and joint pressure sensor data, and outputs a set of safe movement schemes. Specifically, when executing candidate rehabilitation action sequences, the stability of muscle activation can be detected by the coefficient of variation of electromyography signals, and a real-time alarm vector for abnormal muscle activation can be generated. Real-time calculation mechanism of the coefficient of variation for electromyography (CV-EMG): The coefficient of variation (CV) is defined as the standard deviation of the electromyographic (EMG) signal amplitude divided by the mean and then multiplied by 100% (the result is a percentage). It is used to quantify the degree of fluctuation in muscle activation. For example, the mean EMG amplitude of the rectus femoris muscle during a static squat is 0.4 mV, and the standard deviation is 0.12 mV. Therefore, the CV = (0.12 / 0.4) × 100% = 30%. Clinically, a CV > 35% is considered abnormal fluctuation (indicating impaired muscle coordination).

[0141] Sliding window real-time analysis: The system uses a 500-millisecond (ms) time window and updates the calculation every 100 milliseconds. Each window contains 500 electromyographic data points (sampling rate 1000 Hz, i.e., 1000 points per second). When the coefficient of variation of a certain channel of the quadriceps femoris is detected to be >35% for 3 consecutive times (i.e., for 300 milliseconds), a primary alarm is triggered.

[0142] Multi-muscle synergy monitoring strategy: Simultaneously analyze the coefficient of variation ratio of the target muscle group (such as vastus medialis and vastus lateralis) to the antagonist muscle (such as hamstrings). Under normal synergy, the coefficient of variation ratio of quadriceps femoris / hamstrings is 1.0-1.8. If it is >2.5 (e.g., quadriceps femoris 40%, hamstrings 15%), it is judged as compensatory overactivation of quadriceps femoris.

[0143] Abnormal activation pattern recognition and alarm generation logic: Alert Vector Data Structure: Construct an 8-dimensional vector (corresponding to 8 monitored muscles), with each dimension containing three parameters: Status codes: 0 (normal), 1 (mild anomaly), 2 (severe anomaly); Anomaly type: Code 1 = intermittent activation, 2 = continuous high activation, 3 = delayed activation; Confidence level: Probability value (0-100%) based on historical data matching degree.

[0144] Typical Case Analysis: When patients performed single-leg standing after anterior cruciate ligament surgery: the coefficient of variation of the vastus lateralis muscle suddenly increased to 52% (status code 2), and the abnormality type was 2 (persistent high activation); the coefficient of variation of the gluteus medius muscle decreased to 18% (status code 1), and the abnormality type was 3 (delayed activation); the output alarm vector was: [(2,2,85%), (1,3,90%), 0, 0, 0, 0, 0, 0].

[0145] Dynamic threshold adjustment mechanism: The judgment criteria are automatically adjusted according to the patient's recovery stage. In the early postoperative stage, the coefficient of variation is allowed to be relaxed to 50% (due to the lack of recovery of muscle control), while in the later stage of recovery, it is strictly limited to 30%.

[0146] Based on joint pressure sensor data, a support vector regression model is used to predict the peak joint contact stress and output the joint overload risk index. Construction and data fusion of multi-source pressure sensing networks: Sensor deployment solutions: The knee joint cavity is equipped with a miniature piezoresistive film sensor (0.2 mm thick), with 6 measuring points covering the anterior / middle / posterior regions of the tibial plateau (e.g., measuring point P1 is located on the anterior medial side of the tibia, and P6 is located on the posterolateral side).

[0147] The hip joint uses a surface array sensor (8×8 matrix, 5 mm spacing) to fit the projection area of ​​the greater trochanter and femoral neck.

[0148] Physical signal preprocessing: Temperature compensation: When the sensor temperature drift is greater than ±2℃, real-time correction based on thermocouple is initiated (e.g., the pressure reading calibration coefficient is 0.98 at 30℃).

[0149] Dynamic range switching: 0-5 N / cm² in resting state. 2 The measuring range automatically switches to 0-20 N / cm during motion. 2 (To prevent signal saturation).

[0150] The prediction process of the Support Vector Regression (SVR) model: Input feature engineering: Spatiotemporal characteristics: spatial distribution of pressure measurement points (e.g., pressure gradient in the front / back zone), pressure rise rate (e.g., an increase of 4 N / cm within 200 milliseconds). 2 ); Movement-related features: current joint angle (e.g., knee flexion 35°), angular velocity (e.g., extension speed 20 degrees / second); Model Training and Prediction: The model was trained based on a database of 300 orthopedic patients. The input features were 32-dimensional (including pressure, angle, velocity, etc.), and the output was the peak stress (unit: Newtons / square millimeter, N / mm). 2 ).

[0151] Real-time prediction example: When a patient performs a lunge, input the feature values ​​[knee flexion 42°, pressure rise rate 1.8 N / cm]. 2 / 100ms, front / back gradient ratio 2.3] → SVR output predicted peak stress 4.5 N / mm 2 .

[0152] Risk Index (RI) tiered output: Index = Predicted peak stress ÷ Individualized load-bearing threshold × 100%.

[0153] Grading criteria: Low risk (green zone): <80% (e.g., predicted value 3.0 N / mm) 2 vs threshold 3.8 N / mm 2 Medium risk (yellow zone): 80%-120%; High risk (red zone): >120% (warning value).

[0154] By integrating abnormal muscle activation alarm vectors and joint overload risk indices, a fuzzy logic decision-maker is used to dynamically calculate action hazard scores. Fuzzification of multimodal input parameters: Fuzzy transformation of alarm vectors: The status codes (0 / 1 / 2) are converted into fuzzy membership degrees: "Normal" membership degree: 1.0 for status code 0, 0.3 for status code 1, and 0 for status code 2; "Abnormal" membership degree: 0 for status code 0, 0.7 for status code 1, and 1.0 for status code 2; Confidence mapping: 90% confidence is directly used as the rule weight.

[0155] Fuzzy classification of risk index: Define fuzzy sets as: low risk (center of the triangular function 0%), medium risk (center 100%), and high risk (center 200%). Example: Risk index 115% → Medium risk membership degree 0.8, high risk membership degree 0.2.

[0156] Clinical logic design of a fuzzy rule base: Examples of core rules: Rule R1: IF Quadriceps abnormality (membership > 0.7) AND knee risk > 100% THEN Risk score increase (weight = mean confidence level); Rule R2: IF antagonist muscle activation delay (status code = 1) AND rate of stress rise > 1.5 N / cm 2 / 100msTHEN Risk score increases moderately; Rule R3: IF Multi-muscle group synergistic imbalance (≥3 muscle status codes ≥1) OR Risk index > 150% THEN High risk is triggered directly.

[0157] Dynamic weight allocation: Early rehabilitation patients: joint risk weight accounts for 70%; later functional training: muscle abnormality weight accounts for 60%; defuzzification outputs risk score.

[0158] Center of gravity calculation: Converts the fuzzy output set (hazard level: low / medium / high) into precise values ​​from 0 to 100. For example: Low-risk membership score: 0.2 (range: 0-30 points); Medium-risk membership score: 0.6 (range: 31-70 points); High-risk membership score: 0.3 (range: 71-100 points). Risk score = (0.2×15+0.6×50+0.3×85) / (0.2+0.6+0.3)=58 points.

[0159] Real-time update frequency: Calculated every 200 milliseconds to ensure synchronization with action execution.

[0160] When the hazard score of an action exceeds the preset safety threshold, a real-time screening mechanism is triggered and the action sequence is updated, outputting a set of safe action plans.

[0161] Tiered safety thresholds and screening strategies: Threshold settings: Early postoperative patients: safety threshold 45 points (>45 points will be screened out); Mid-term muscle strength training: safety threshold 60 points; Late-term functional training: safety threshold 75 points.

[0162] Screening mechanism trigger scenarios: Instantaneous overrun: If the danger score exceeds the threshold for 3 consecutive times (i.e., lasts for 600 milliseconds) → immediately pause the current action; Trend warning: if the scoring slope is greater than 2 points / 100ms (e.g., the score rises from 40 to 50 within 50ms), the next action will be pre-screened.

[0163] Sequence dynamic update algorithm: Alternative solution matching principle: Downgrade the same type of exercise: Lunge training (original knee flexion 50°) → downgrade to shallow squat (knee flexion 30°); load adjustment: reduce the resistance of the elastic band from 30 Newtons (N) to 15 Newtons.

[0164] Linked parameter optimization: If the screening action is excluded due to abnormal activation of the quadriceps, the new protocol must meet the following requirements: quadriceps activation intensity ≤ 40% of maximum voluntary contraction (MVC); hamstring / quadriceps activation ratio ≥ 0.6.

[0165] Case description: When the patient performed step training (original plan step height 25 cm), the risk score reached 68 points (threshold 60 points); the system automatically switched: the step height was reduced to 15 cm → re-monitoring the score reduced to 52 points (safe).

[0166] Output specifications for Safe Action Set.

[0167] The data structure is shown in Table 2: Table 2 ; Solution set generation logic: retain all actions with scores < threshold (such as A03 and B12 above); automatically supplement 3-5 backup solutions (such as C07: straight leg raise, score 45 points); mark the source of risk when outputting (such as "Solution B12: vastus lateralis coefficient of variation is safe").

[0168] S205, the set of safe action plans is input into the metabolic efficiency optimization module, and the metabolic load of different plans is predicted based on cardiopulmonary function monitoring data and exercise oxygen consumption model. The final personalized rehabilitation plan is generated with metabolic equivalent threshold as a constraint.

[0169] Specifically, based on the energy consumption labels in the set of safe action plans, the exercise oxygen consumption model can be calibrated using cardiopulmonary function monitoring data to output predicted values ​​of metabolic load. Clinical analysis and data fusion of motion energy consumption labels: Each movement in the safety movement protocol set carries an Energy Cost Label, which includes the estimated metabolic equivalent (MET) for the intensity, angle, and duration parameters (e.g., a wall squat with MET=2.8 means that the energy consumption of this movement is 2.8 times that of the resting state). The system extracts the MET value (Metabolic Equivalent), duration (in minutes), and number of repetitions (e.g., 3 sets × 30 seconds) from the label, and combines this with the patient's real-time weight (e.g., 70 kg) to calculate the theoretical energy consumption of a single movement: Energy consumption (kcal) = MET × duration (minutes) × weight (kg) × 0.0175 (standard conversion factor); Example: The theoretical energy consumption of a 70 kg patient performing a 10-minute wall squat (MET=2.8) is 2.8 × 10 × 70 × 0.0175 = 343 kcal.

[0170] Dynamic calibration of cardiopulmonary function monitoring data: Real-time acquisition of oxygen uptake (VO2) (mL / kg / min) and heart rate (HR) (beats / min) via chest strap sensors. For example, if a patient's resting VO2 is 3.5 mL / kg / min, and VO2 rises to 11.2 mL / kg / min during exercise (measured value).

[0171] Calibration mechanism: When the deviation between the theoretical MET value (label value) and the measured MET value (measured VO2 ÷ resting VO2) is greater than 15%, linear correction is initiated. If the label MET for a wall squat is 2.8 but the measured value is only 2.3 (deviation 18%), the MET label for this action will be updated to 2.3, and the energy consumption will be recalculated to 282 kcal.

[0172] Personalized adaptation of the Exercise Oxygen Consumption Model: Model input parameter expansion: Mechanics parameters: joint range of motion (e.g., energy consumption is 12% higher when the knee is flexed at 35° than at 0°), external load (e.g., an elastic band resistance of 15 Newtons (N) increases energy consumption by 18%). Physiological parameters: patient age correction factor (basal metabolic rate is reduced by 10% for those >65 years old), blood oxygen saturation SpO2 (energy consumption compensation factor of 1.15 when <95%).

[0173] Output Metabolic Load Prediction: The predicted value is the equivalent resting metabolic time (in minutes). For example, if a certain protocol has a total energy consumption of 550 kcal and the patient's resting metabolic rate is 1.2 kcal / min, then the metabolic load = 550 ÷ 1.2 = 458 minutes (that is, completing this protocol is equivalent to continuous resting energy consumption of 7.6 hours).

[0174] Classification labeling: low load (<300 minutes), medium load (300-600 minutes), high load (>600 minutes).

[0175] Based on the predicted metabolic load, a non-dominated sorting genetic algorithm is used to perform multi-objective optimization with the goal of minimizing metabolic load and maximizing rehabilitation benefits. Quantitative Model for Rehabilitation Benefits and Definition of Optimization Objectives: Multidimensional assessment of Rehabilitation Benefit: Muscle function gain: percentage increase in the activation efficiency of the target muscle group (e.g., an increase in quadriceps activation from 60% to 75% is recorded as a 15% gain). Reduced joint risk: The rate of decrease in peak stress predicted by finite element analysis (e.g., a 28.6% improvement in tibial plateau stress from 4.2 N / mm² to 3.0 N / mm²). Metabolic efficiency ratio: the increase in muscle strength per unit of energy consumption (e.g., 5 Newtons of quadriceps muscle strength per 100 kcal of energy consumption).

[0176] Analysis of the conflict between dual objectives: Objective 1: Minimize metabolic load (requires low-intensity, short-duration movements); Objective 2: Maximize rehabilitation benefits (requires medium- to high-intensity movement stimulation).

[0177] Typical paradox: A wall squat with knee flexion of 40° (MET=3.1) consumes 29% more energy than a squat with knee flexion of 30° (MET=2.4), but the quadriceps activation efficiency is 22% higher.

[0178] Execution flow of the Non-dominated Sorting Genetic Algorithm (NSGA-II).

[0179] Population initialization: 100 schemes are randomly selected from the set of safe action schemes as the initial population (e.g., scheme A: static squat + seated knee extension; scheme B: straight leg raise + ankle pump); each scheme is encoded as a gene sequence: action type-angle-intensity-duration (e.g., gene locus 1: static squat_40°_15N_30sec).

[0180] Fitness assessment and non-dominated ranking: Calculate the metabolic load (e.g., Plan A = 420 minutes) and rehabilitation benefit (e.g., muscle strength gain of 12%) for each plan; screen for non-dominated solutions (Pareto solutions): Plan X dominates Plan Y if the metabolic load of X is ≤ Y and the rehabilitation benefit is ≥ Y, and at least one of them is strictly superior. The first round of screening yielded 30 non-dominated solutions that constitute the Pareto Front.

[0181] Iterative optimization of genetic operations: Crossover: Randomly swap 50% of the movement genes between two schemes (e.g., replacing "static squat" in scheme A with "straight leg raise" in scheme B); Mutation: Adjusting single-action parameters with a 5% probability (e.g., randomly adjusting the knee flexion angle from 40°±5° to 35° or 45°). After 200 generations of evolution, the Pareto front converges to the optimal solution set (e.g., the front includes scheme P1: metabolic load 380 minutes / benefit 14%; scheme P2: load 450 minutes / benefit 18%).

[0182] Set a preset multiple of the patient's resting metabolic equivalent as a threshold constraint to screen out action combinations that exceed the tolerance range; Personalized metabolic equivalent threshold (MET) setting strategy: Resting metabolic equivalent (MET) baseline determination: The patient's resting oxygen uptake (e.g., VO2 = 3.3 ml / kg / min) was measured by indirect calorimetry while the patient was in bed upon waking in the morning. 1 MET = this measured value.

[0183] Clinical stage differentiation thresholds: Postoperative acute phase (0-2 weeks): upper limit 2.0 MET (to avoid exacerbating the inflammatory response); recovery phase (3-6 weeks): upper limit 3.5 MET (to promote muscle strength reconstruction); functional phase (>6 weeks): upper limit 5.0 MET (to simulate the intensity of daily activities).

[0184] Intensity assessment of movement combinations: The peak value of the MET for a single movement is taken (e.g., the peak MET value for lunge training is 4.2). Combined treatment MET threshold = resting MET × preset multiple (e.g., for patients in the recovery period, resting MET = 1.0, preset multiple 3.5 → threshold 3.5 MET); Screening Case: The program includes lunge training (MET peak 4.2 > 3.5) → triggers screening.

[0185] The dynamic expansion mechanism of the tolerance range: Fatigue accumulation model: Define the metabolic fatigue index (MFI) as total load of the program ÷ maximum weekly tolerance load (e.g., weekly upper limit of 2000 minutes); if MFI > 0.85 (e.g., 1700 minutes of load has been performed this week, new program of 300 minutes → MFI = 1.0), then the intensity of the exercise is forcibly reduced by 20%.

[0186] Real-time constraints on physiological indicators: Blood oxygen alert: If SpO2 remains <92% for more than 10 seconds during the action, the MET threshold will be automatically lowered by 0.5; Heart rate limit: 75% of the age-corrected maximum heart rate (220 - age) is used as a hard cutoff point (e.g., if a 60-year-old patient's heart rate is >120 beats / minute, the current action will be stopped).

[0187] By integrating and optimizing the movement sequence and metabolic parameters, a final personalized rehabilitation plan is generated, which includes intensity, angle, duration, and energy consumption labels.

[0188] Clinical adaptability integration of multi-dimensional parameters: The timing arrangement principles of action sequences: Fatigue management logic: High-intensity activities (such as MET > 3.0) should be performed in the first 1 / 3 of the session (while the muscles are not fatigued), and low-intensity activities (such as MET < 2.0) should be performed in the later part of the session; Joint protection strategies: Limit continuous knee flexion to no more than 10 minutes (to prevent cartilage wear), and incorporate ankle pumps or upper limb training.

[0189] Example arrangement: 0-10 minutes: wall squat (knee flexion 40°, 3 sets x 45 seconds, 60 seconds rest between sets); 11-20 minutes: seated knee extension (0°-45°, 3 sets x 30 seconds); 21-30 minutes: resistance band ankle dorsiflexion (resistance 10 Newtons, 4 sets x 20 seconds).

[0190] Visualization of metabolic parameters: Each action is marked with a real-time energy consumption bar chart (e.g., static squat: 45 kcal per set, red bar length accounts for 15% of the total); Total load of the whole plan prompt: "The total metabolic load of this plan is 480 minutes (about 8 hours of resting consumption), and the daily execution rate is 30%".

[0191] Specifications for generating the Final Rehabilitation Schedule.

[0192] The data structure and clinical delivery format are shown in Table 3: Table 3 ; Smart reminders and emergency rules: Real-time monitoring and linkage: When the MET value during execution exceeds 10% of the peak value marked on the plan (e.g., the measured MET for static squatting is 3.4, which is greater than the marked 3.1), a voice prompt will be triggered: "Please reduce the squatting depth." Emergency termination conditions: Two consecutive instances of exceeding the limit or a subjective pain score > 6 (out of 0-10), automatically switching to the backup plan.

[0193] Version iteration mechanism: The plan is updated weekly based on new biomechanical data (e.g., after a 5% increase in bone density, the knee flexion angle is allowed to increase from 35° to 40°); Patient compliance feedback: Daily completion rate is recorded by scanning a code (>90% maintain the original plan, <70% reduce the intensity by 10%).

[0194] As can be seen, by acquiring patient gait, joint angles, and electromyographic signals through wearable sensors, a fused biomechanical feature matrix is ​​generated. This matrix is ​​then registered with the bone density distribution map of the patient's CT / MRI images to output a quantitative damage-compensation relationship map. This map is then input into an adaptive motion generator, which generates candidate rehabilitation motion sequences that conform to biomechanical constraints based on a joint kinetic chain inverse solution algorithm. Real-time physiological feedback testing is performed on the candidate rehabilitation motion sequences, outputting a safe motion scheme set. Finally, this safe motion scheme set is input into a metabolic efficiency optimization module, using a metabolic equivalent threshold as a constraint to generate a final personalized rehabilitation plan, thereby improving the accuracy and safety of orthopedic rehabilitation.

[0195] Another embodiment of the present invention provides a personalized orthopedic rehabilitation plan recommendation system, see [link to relevant documentation]. Figure 3 The system may include: Processing module 301 is used to acquire patient gait, joint angle and electromyographic signals through wearable sensors, and perform spatiotemporal synchronization processing on the original signals according to a multimodal time series alignment algorithm to generate a fused biomechanical feature matrix. The registration module 302 is used to register the fused biomechanical feature matrix with the bone density distribution map of the patient's CT / MRI images, identify abnormal stress areas and muscle compensation patterns through the compensation effect detection model, and output a quantitative damage-compensation relationship map. The generation module 303 is used to input the quantified damage-compensation relationship diagram into the adaptive motion generator and generate a candidate rehabilitation motion sequence that conforms to biomechanical constraints based on the joint kinetic chain inverse solution algorithm. Each motion includes intensity, angle, and duration parameters. The screening module 304 is used to perform real-time physiological feedback testing on candidate rehabilitation movement sequences. Based on the coefficient of variation of electromyography signals and joint pressure sensor data, it dynamically screens out movement schemes that cause abnormal muscle activation or joint overload, and outputs a set of safe movement schemes. The prediction module 305 is used to input the set of safe action plans into the metabolic efficiency optimization module, predict the metabolic load of different plans based on cardiopulmonary function monitoring data and exercise oxygen consumption model, and generate the final personalized rehabilitation plan table with metabolic equivalent threshold as a constraint condition.

[0196] 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.

[0197] Specifically, in this embodiment, the storage medium can be configured to store a computer program for performing the following steps: S201 acquires the patient's gait, joint angles, and electromyographic signals through wearable sensors, and performs spatiotemporal synchronization processing on the original signals according to a multimodal time series alignment algorithm to generate a fused biomechanical feature matrix; S202, the fused biomechanical feature matrix is ​​registered with the bone density distribution map of the patient's CT / MRI images, and the abnormal stress area and muscle compensation pattern are identified through the compensation effect detection model, and a quantitative damage-compensation relationship map is output. S203, the quantified damage-compensation relationship diagram is input into the adaptive motion generator, and a candidate rehabilitation motion sequence that conforms to biomechanical constraints is generated based on the joint kinetic chain inverse solution algorithm. Each motion includes intensity, angle and duration parameters. S204 performs real-time physiological feedback testing on candidate rehabilitation movement sequences, dynamically filters out movement schemes that cause abnormal muscle activation or joint overload based on electromyographic signal variation coefficient and joint pressure sensor data, and outputs a set of safe movement schemes. S205, the set of safe action plans is input into the metabolic efficiency optimization module, and the metabolic load of different plans is predicted based on cardiopulmonary function monitoring data and exercise oxygen consumption model. The final personalized rehabilitation plan is generated with metabolic equivalent threshold as a constraint.

[0198] 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.

[0199] 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.

[0200] Specifically, in this embodiment, the processor can be configured to perform the following steps via a computer program: S201 acquires the patient's gait, joint angles, and electromyographic signals through wearable sensors, and performs spatiotemporal synchronization processing on the original signals according to a multimodal time series alignment algorithm to generate a fused biomechanical feature matrix; S202, the fused biomechanical feature matrix is ​​registered with the bone density distribution map of the patient's CT / MRI images, and the abnormal stress area and muscle compensation pattern are identified through the compensation effect detection model, and a quantitative damage-compensation relationship map is output. S203, the quantified damage-compensation relationship diagram is input into the adaptive motion generator, and a candidate rehabilitation motion sequence that conforms to biomechanical constraints is generated based on the joint kinetic chain inverse solution algorithm. Each motion includes intensity, angle and duration parameters. S204 performs real-time physiological feedback testing on candidate rehabilitation movement sequences, dynamically filters out movement schemes that cause abnormal muscle activation or joint overload based on electromyographic signal variation coefficient and joint pressure sensor data, and outputs a set of safe movement schemes. S205, the set of safe action plans is input into the metabolic efficiency optimization module, and the metabolic load of different plans is predicted based on cardiopulmonary function monitoring data and exercise oxygen consumption model. The final personalized rehabilitation plan is generated with metabolic equivalent threshold as a constraint.

[0201] 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 method for recommending personalized orthopedic rehabilitation plans, characterized in that, The method includes: The patient's gait, joint angles and electromyography signals are acquired by wearable sensors. The original signals are spatiotemporally synchronized using a multimodal time series alignment algorithm to generate a fused biomechanical feature matrix. The fused biomechanical feature matrix is ​​registered with the bone density distribution map of the patient's CT / MRI images. The abnormal stress area and muscle compensation pattern are identified through the compensation effect detection model, and a quantitative damage-compensation relationship map is output. The quantified damage-compensation relationship diagram is input into the adaptive motion generator, and a candidate rehabilitation motion sequence that conforms to biomechanical constraints is generated based on the joint kinetic chain inverse solution algorithm. Each motion includes intensity, angle, and duration parameters. Real-time physiological feedback testing is performed on candidate rehabilitation movement sequences. Based on the coefficient of variation of electromyography signals and joint pressure sensor data, movement schemes that cause abnormal muscle activation or joint overload are dynamically screened out, and a set of safe movement schemes is output. The set of safe action plans is input into the metabolic efficiency optimization module. Based on the cardiopulmonary function monitoring data and the exercise oxygen consumption model, the metabolic load of different plans is predicted. The final personalized rehabilitation plan is generated with the metabolic equivalent threshold as a constraint.

2. The method according to claim 1, characterized in that, The process involves acquiring patient gait, joint angles, and electromyographic signals via wearable sensors, and performing spatiotemporal synchronization processing on the original signals using a multimodal time series alignment algorithm to generate a fused biomechanical feature matrix, including: Based on the joint six-degree-of-freedom motion trajectory and surface electromyography signal output by the IMU sensor, motion artifacts and physiological signals are separated by bandpass filtering to obtain the denoised original motion-electromyography time series data; Based on motion-electromyography time-series data, a dynamic time warping (DTW) algorithm is used to compensate for the differences in sampling rates of each sensor and generate a time-aligned multimodal signal stream. The multimodal signal stream is input into the biomechanical coordinate system transformation module, which maps the sensor's local coordinate system to the human body's global coordinate system through an inverse kinematics model, and outputs a spatially synchronized joint torque-muscle activation matrix. Tensor decomposition is performed on the joint torque-muscle activation matrix, and feature modes whose principal component energy ratio exceeds a preset ratio threshold are extracted and combined into a dimension-compressed fusion biomechanical feature matrix.

3. The method according to claim 2, characterized in that, The process involves registering the fused biomechanical feature matrix with the bone density distribution map of the patient's CT / MRI images, identifying abnormal stress areas and muscle compensation patterns using a compensation effect detection model, and outputting a quantitative damage-compensation relationship map, including: Based on the Hounsfield values ​​of the bone density distribution map, the spatial coordinates of the biomechanical feature matrix are mapped to the image anatomical coordinate system using a non-rigid B-spline registration algorithm to generate a bone-muscle coupling topology map. Based on the bone-muscle coupling topology, the finite element analysis method is used to calculate the deviation between the stress concentration area and the theoretical load-bearing threshold, and output the thermal map of the abnormal stress area. Muscle co-activation analysis was triggered by thermal mapping of abnormal stress areas. Compensatory and non-compensatory electromyographic signals were separated by independent component decomposition to generate a muscle compensation pattern coding table. By integrating the thermal map of the abnormal stress area with the coding table of muscle compensation patterns, a quantitative damage-compensation relationship diagram of damage force value and electromyographic compensation intensity in Newtons / square millimeter is constructed.

4. The method according to claim 3, characterized in that, The quantitative damage-compensation relationship diagram is input into an adaptive motion generator, which generates a candidate rehabilitation motion sequence that conforms to biomechanical constraints based on a joint kinetic chain inverse solution algorithm. Each motion includes intensity, angle, and duration parameters, including: Based on the distribution of damage force values ​​in the quantified damage-compensation relationship diagram, a mask of forbidden motion regions of the joint kinematic chain is generated by a convolutional neural network, and the biomechanical constraint boundary is output. Based on biomechanical constraints, the feasible region of joint angles is solved by inverse Jacobian matrix to generate a set of candidate joint motion trajectories that satisfy torque balance. Muscle length-tension optimization is performed on the candidate joint motion trajectory set. The Hill muscle model is used to dynamically adjust the motion amplitude and output the motion angle parameters with the optimal muscle activation efficiency. By combining cardiopulmonary function monitoring data, the energy consumption gradient of different action combinations is calculated through a metabolic equivalent prediction model, and action energy consumption labels containing intensity and duration are generated. By integrating motion angle parameters and motion energy consumption labels, Monte Carlo sampling is used to generate candidate rehabilitation motion sequences that meet Pareto optimality.

5. The method according to claim 4, characterized in that, The process involves real-time physiological feedback testing of candidate rehabilitation movement sequences, dynamically filtering out movement schemes that cause abnormal muscle activation or joint overload based on electromyographic signal variation coefficients and joint pressure sensor data, and outputting a set of safe movement schemes, including: When executing candidate rehabilitation action sequences, muscle activation stability is detected by the coefficient of variation of electromyography signals, and a real-time alarm vector for abnormal muscle activation is generated. Based on joint pressure sensor data, a support vector regression model is used to predict the peak joint contact stress and output the joint overload risk index. By integrating abnormal muscle activation alarm vectors and joint overload risk indices, a fuzzy logic decision-maker is used to dynamically calculate action hazard scores. When the hazard score of an action exceeds the preset safety threshold, a real-time screening mechanism is triggered and the action sequence is updated, outputting a set of safe action plans.

6. The method according to claim 5, characterized in that, The step involves inputting the set of safe action plans into the metabolic efficiency optimization module, predicting the metabolic load of different plans based on cardiopulmonary function monitoring data and an exercise oxygen consumption model, and generating a final personalized rehabilitation plan using a metabolic equivalent threshold as a constraint. This includes: Based on the energy consumption labels of the safe action program set, the exercise oxygen consumption model is calibrated using cardiopulmonary function monitoring data, and the predicted value of metabolic load is output. Based on the predicted metabolic load, a non-dominated sorting genetic algorithm is used to perform multi-objective optimization with the goal of minimizing metabolic load and maximizing rehabilitation benefits. Set a preset multiple of the patient's resting metabolic equivalent as a threshold constraint to screen out action combinations that exceed the tolerance range; By integrating and optimizing the movement sequence and metabolic parameters, a final personalized rehabilitation plan is generated, which includes intensity, angle, duration, and energy consumption labels.

7. A personalized orthopedic rehabilitation plan recommendation system, characterized in that, The system includes: The processing module is used to acquire the patient's gait, joint angles and electromyography signals through wearable sensors, and to perform spatiotemporal synchronization processing on the raw signals according to a multimodal time series alignment algorithm to generate a fused biomechanical feature matrix. The registration module is used to register the fused biomechanical feature matrix with the bone density distribution map of the patient's CT / MRI images, identify abnormal stress areas and muscle compensation patterns through the compensation effect detection model, and output a quantitative damage-compensation relationship map. The generation module is used to input the quantified damage-compensation relationship diagram into the adaptive motion generator and generate a candidate rehabilitation motion sequence that conforms to biomechanical constraints based on the joint kinetic chain inverse solution algorithm. Each motion includes intensity, angle, and duration parameters. The screening module is used to perform real-time physiological feedback testing on candidate rehabilitation movement sequences. Based on the coefficient of variation of electromyography signals and joint pressure sensor data, it dynamically screens out movement schemes that cause abnormal muscle activation or joint overload, and outputs a set of safe movement schemes. The prediction module is used to input the set of safe action plans into the metabolic efficiency optimization module, predict the metabolic load of different plans based on cardiopulmonary function monitoring data and exercise oxygen consumption model, and generate the final personalized rehabilitation plan table with metabolic equivalent threshold as a constraint.

8. The system according to claim 7, characterized in that, The processing module is specifically used for: Based on the joint six-degree-of-freedom motion trajectory and surface electromyography signal output by the IMU sensor, motion artifacts and physiological signals are separated by bandpass filtering to obtain the denoised original motion-electromyography time series data; Based on motion-electromyography time-series data, a dynamic time warping (DTW) algorithm is used to compensate for the differences in sampling rates of each sensor and generate a time-aligned multimodal signal stream. The multimodal signal stream is input into the biomechanical coordinate system transformation module, which maps the sensor's local coordinate system to the human body's global coordinate system through an inverse kinematics model, and outputs a spatially synchronized joint torque-muscle activation matrix. Tensor decomposition is performed on the joint torque-muscle activation matrix, and feature modes whose principal component energy ratio exceeds a preset ratio threshold are extracted and combined into a dimension-compressed fusion biomechanical feature matrix.

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