Intelligent rehabilitation evaluation and data analysis method and system

Abstract

The invention relates to the technical field of medical informatization, in particular to an intelligent rehabilitation assessment and data analysis method and system.The method comprises the steps that multi-source heterogeneous rehabilitation assessment data collected by a wearable sensor when a patient executes a preset rehabilitation action is obtained in real time; based on a preset time window, processing the multi-source heterogeneous rehabilitation evaluation data into a series of time sequence dynamic graphs; inputting the series of time sequence dynamic graphs into a pre-trained time sequence dynamic graph neural network model, and generating a quantized rehabilitation evaluation result based on the output of the time sequence dynamic graph neural network model. According to the method, deep fusion and time sequence calibration of the multi-source heterogeneous rehabilitation data are realized, the precision of an evaluation model and the recognition capability of a complex motion mode are remarkably improved, the robustness of the model to noise and data artifacts is enhanced through adaptive weighting of the data quality, and the accuracy of the evaluation model is improved. And the reliability and the stability of a rehabilitation evaluation result in a real clinical environment are ensured.

Description

Technical Field

[0001] This application relates to the field of medical information technology, specifically to an intelligent rehabilitation assessment and data analysis method and system. Background Technology

[0002] In scenarios such as stroke, post-orthopedic surgery, or neurodegenerative diseases, the quantitative and objective assessment of patients' motor function is crucial for developing personalized rehabilitation plans, monitoring rehabilitation progress, and predicting rehabilitation outcomes. Traditional rehabilitation assessments largely rely on visual observation and scale evaluation by rehabilitation therapists, which have limitations such as strong subjectivity, limited accuracy, and difficulty in capturing subtle dynamic changes. With the development of sensor technology and artificial intelligence, using wearable devices to collect biomechanical data such as electromyography and inertial measurement units, and then conducting intelligent analysis through algorithmic models, has become an important development direction for improving the objectivity and accuracy of rehabilitation assessments.

[0003] In the existing technology, there have been some attempts to apply neural networks to rehabilitation data analysis. However, after in-depth analysis, the existing technology still has the following technical problems that have not been effectively solved in practical applications.

[0004] Existing methods typically employ simple feature splicing or independent processing, making it difficult to deeply mine and utilize the complex spatiotemporal dependencies between different data modalities, such as the intrinsic coupling between muscle activation patterns and limb kinematics, thus limiting the accuracy and robustness of the evaluation model.

[0005] In actual data acquisition, sensor data is inevitably affected by factors such as motion artifacts and poor electrode contact, resulting in dynamic changes in data quality. Existing technologies usually do not consider real-time and dynamic weighted adjustment of the contribution of different data sources during model analysis, making the evaluation results susceptible to the influence of low-quality or noisy data.

[0006] Existing models are often difficult to adapt quickly to the individualized pathological characteristics of new patients, and cannot be continuously optimized using multi-center data while protecting patient data privacy. In addition, their "black box" nature makes the assessment results difficult to interpret, and cannot provide therapists or patients with clear and quantifiable intervention guidance on "why the score is low" and "how to improve", which greatly limits their application value in clinical practice. Summary of the Invention

[0007] The objective of this application is to provide an intelligent rehabilitation assessment and data analysis method, comprising: real-time acquisition of multi-source heterogeneous rehabilitation assessment data collected by wearable sensors while a patient performs predetermined rehabilitation movements, wherein the multi-source heterogeneous rehabilitation assessment data includes at least time-series data of electromyographic signals from an electromyography sensor and time-series data of attitude angles from an inertial measurement unit; processing the multi-source heterogeneous rehabilitation assessment data into a series of time-series dynamic graphs based on a preset time window, wherein each time-series dynamic graph is a graph snapshot, and its heterogeneous data nodes are composed of time-series data from different sources, and the edges between nodes represent the biomechanical or temporal correlation of the time-series data; inputting the series of time-series dynamic graphs into a pre-trained time-series dynamic graph neural network model, which then performs the analysis. The following operations are performed: The spatiotemporal dependency between the muscle activation patterns reflected by the electromyographic signals and the kinematic features reflected by the posture angles is captured and fused through the message passing and aggregation mechanism of the graph neural network, thereby achieving deep fusion and temporal calibration of the multi-source heterogeneous rehabilitation assessment data; Furthermore, the data quality of the heterogeneous data nodes is dynamically evaluated based on the signal-to-noise ratio or information entropy in the temporal evolution of the attention mechanism integrated into the temporal dynamic graph neural network model, and a dynamic weight coefficient is assigned to each heterogeneous data node to adaptively adjust the contribution of different data sources during the information aggregation process; Finally, a quantitative rehabilitation assessment result is generated based on the output of the temporal dynamic graph neural network model, which characterizes the accurate identification of the temporal pattern of the patient's rehabilitation status.

[0008] By adopting the above scheme and utilizing the message passing and aggregation mechanism of the temporal dynamic graph neural network, the information of different data nodes can be deeply fused along the preset edges in the graph structure, thereby effectively capturing and modeling this inherent spatiotemporal dependency. At the same time, an attention mechanism is integrated into the model to address the technical deficiency of existing technologies in handling dynamic changes in sensor data quality. This attention mechanism can dynamically evaluate the data quality of heterogeneous data nodes and assign weights based on the signal-to-noise ratio or information entropy of the heterogeneous data nodes in the temporal evolution.

[0009] Optionally, the method further includes personalized adaptation and continuous optimization of the temporal dynamic graph neural network model through a federated meta-learning mechanism. Specific steps include: on a central server, using historical rehabilitation assessment data from multiple rehabilitation centers, a global temporal dynamic graph neural network model is pre-trained as the base model for meta-learning; when a new patient is admitted, on their local computing device, using a small amount of personalized rehabilitation assessment data collected from that patient, the base model is rapidly fine-tuned to generate personalized model parameters adapted to the patient's physiological and pathological characteristics. This process is completed while protecting the patient's original data from leaving the local device; and during the patient's rehabilitation cycle, the personalized model parameters are continuously updated using newly collected data, and the non-sensitive parameter gradient information generated by the update is encrypted and uploaded to the central server for aggregation and updating of the global base model on the central server, thereby achieving continuous model optimization.

[0010] By adopting the above approach, rapid model personalization is achieved, solving the cold start problem. Only a small number of samples are needed for the model to adapt to the unique physiological and pathological characteristics of new patients. The privacy and security of patient data are guaranteed, with the original data always stored locally and only encrypted, non-sensitive parameter gradients are uploaded. Distributed and continuous learning of the model is realized, with the central server able to aggregate knowledge from various sources, continuously optimize the global model, and form a virtuous cycle ecosystem that becomes smarter with use.

[0011] Optionally, before processing the multi-source heterogeneous rehabilitation assessment data into a series of time-series dynamic graphs, a preprocessing step is further included for the multi-source heterogeneous rehabilitation assessment data. The preprocessing includes: performing bandpass filtering and notch filtering on the electromyographic signal time-series data to filter out power frequency interference and motion artifacts, thereby improving the signal-to-noise ratio; and performing attitude calculation on the raw acceleration and angular velocity data collected by the inertial measurement unit to obtain attitude angle time-series data characterizing the orientation of limb segments in three-dimensional space.

[0012] By adopting the above scheme, the technical problem that the original sensor signals contain a lot of noise and interference, which would affect the accuracy of subsequent model analysis if used directly, is solved. The signal-to-noise ratio of the input signal is significantly improved, and more physically meaningful attitude angle time series data is extracted, thereby further improving the accuracy of the overall evaluation method.

[0013] Optionally, the step of processing the multi-source heterogeneous rehabilitation assessment data into a series of time-series dynamic graphs specifically involves: defining each channel of the preprocessed electromyographic signal time-series data and each dimension of the posture angle time-series data as a heterogeneous data node; within the same time slice, constructing edges representing biomechanical coupling between the corresponding heterogeneous data nodes based on the linkage relationship between muscles and joints in human anatomy; and constructing edges representing their temporal autocorrelation between heterogeneous data nodes representing the same physical quantity, thereby forming a graph snapshot.

[0014] By adopting the above scheme, prior knowledge is explicitly encoded in the graph structure, so that the temporal dynamic graph neural network model does not have to learn these basic physical relationships from scratch, thereby reducing the difficulty of model training, accelerating the convergence speed, and learning more interpretable and generalizable feature representations.

[0015] Optionally, the step of assigning a dynamic weight coefficient to each of the heterogeneous data nodes specifically includes: during the information aggregation process of each layer of the temporal dynamic graph neural network model, the attention mechanism calculates the similarity score between the feature vector of each heterogeneous data node and the feature vector of its neighboring nodes; based on the similarity score, combined with a measure of the stability of the heterogeneous data node's own features, a normalized attention weight is assigned to the information transmission path from each neighboring node to the central node, so that nodes with stable features and rich information content receive higher weights when aggregating neighbor information.

[0016] By adopting the above scheme, an adaptive weighting mechanism based on the intrinsic characteristics of data (rather than external labels) was established, which enables the model to intelligently determine which information sources are more reliable or richer in information at each step of information aggregation. This allows the model to maximize the use of high-quality information and suppress noise interference when aggregating neighbor information, thereby further enhancing the robustness and accuracy of the evaluation results.

[0017] Optionally, after generating the rehabilitation assessment results, an attribution-counterfactual dual-track interpretability engine is activated to provide decision-making explanations and intervention guidance. The operation steps of the attribution-counterfactual dual-track interpretability engine include: identifying and highlighting key biomechanical features that have a decisive impact on the current rehabilitation assessment results through an attribution analysis module; and constructing and presenting one or more feasible minimal virtual movement adjustment schemes through a counterfactual interpretation generation module. These virtual movement adjustment schemes can guide the patient's rehabilitation assessment results towards the expected better state and provide actionable rehabilitation intervention guidance to rehabilitation therapists or patients in a preset form.

[0018] By adopting the above approach, not only is "what" (assessment score) provided, but "why" (key influencing factors) is also clearly explained, and further "how" (quantifiable intervention guidance) is provided. This greatly enhances the system's transparency, credibility, and clinical applicability, transforming it from a simple assessment tool into an intelligent rehabilitation decision support system that integrates assessment, diagnosis, and guidance.

[0019] Optionally, the step of identifying and highlighting key biomechanical features that have a decisive impact on the current rehabilitation assessment result through the attribution analysis module includes: using a gradient-based attribution algorithm to calculate the partial derivative of the rehabilitation assessment result with respect to the input feature value of each heterogeneous data node in the time-series dynamic graph sequence at the key motion time point; and identifying the biomechanical events corresponding to the input feature values ​​with higher partial derivative values ​​as key biomechanical features affecting the assessment result.

[0020] By adopting the above approach, it is possible to accurately and quantitatively pinpoint the specific biomechanical events that affect the final score, such as an abnormal activation peak of a certain muscle at a certain time point, providing rehabilitation therapists with strong data support for root cause analysis.

[0021] Optionally, the step of constructing and presenting one or more achievable minimization virtual motion adjustment schemes through the counterfactual interpretation generation module includes: performing an optimization search based on a pre-trained generative adversarial network or variational autoencoder model, with the optimization objective of minimizing the perturbation distance from the patient's original multi-source heterogeneous rehabilitation assessment data, while constraining the rehabilitation assessment results to reach a preset higher score; decoding and generating one or more virtual counterfactual data samples from the search results; and generating the quantifiable rehabilitation intervention guidance by comparing the differences between the counterfactual data samples and the original data in key biomechanical characteristics.

[0022] By adopting the above approach, the generated intervention guidance is quantifiable, such as "increase elbow flexion by 5 degrees," rather than vague qualitative suggestions. At the same time, since the optimization goal is to "minimize perturbation," it ensures that the guidance is a realistic, feasible, and most easily achievable improvement path for patients, thereby significantly improving the efficiency and compliance of rehabilitation training.

[0023] Optionally, the step of aggregating and updating the global base model specifically includes: after receiving encrypted parameter gradients uploaded by local devices from multiple rehabilitation centers, the central server uses a federated averaging algorithm to perform a weighted average of the parameter gradients to calculate the update amount of the global model; the update amount of the global model is used to update the global base model, and the updated global base model parameters are distributed to each local computing device as the starting point for the next round of personalized fine-tuning.

[0024] By adopting the above scheme, it is ensured that the global model can benefit stably and effectively from the local learning of all participants, and a globally intelligent system with continuously improving performance is built under the premise of protecting the data sovereignty and privacy of all parties. This makes the evaluation capability of the entire system increase with the expansion of the user group.

[0025] The second objective of this application is to provide an intelligent rehabilitation assessment and data analysis system, including a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed by the aforementioned intelligent rehabilitation assessment and data analysis method. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the intelligent rehabilitation assessment and data analysis method of this application. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0028] like Figure 1 As shown in the figure, this application provides an intelligent rehabilitation assessment and data analysis method, which includes the following steps.

[0029] S01: Real-time acquisition of multi-source heterogeneous rehabilitation assessment data collected by wearable sensors when the patient performs predetermined rehabilitation movements. The multi-source heterogeneous rehabilitation assessment data includes at least electromyographic signal timing data from electromyography sensors and attitude angle timing data from inertial measurement units.

[0030] Specifically, in this embodiment, the electromyography (EMG) sensor used in the data acquisition stage can be the FREEEMG 1000 wireless surface EMG acquisition system manufactured by BTS-Bioengineering in Italy. This system is equipped with eight independent wireless EMG probes, with a sampling frequency set at 2048 Hz. Each probe has a built-in bipolar silver / silver chloride electrode with an electrode spacing of 20 mm. Before acquisition, the skin of the muscle to be tested is cleaned with alcohol swabs and lightly abraded to reduce the skin impedance to below 5 kΩ, thereby ensuring signal quality. These eight EMG probes are strategically attached to the surface of key muscle groups on the affected side of the patient's upper limb, specifically distributed as follows: anterior deltoid, middle deltoid, long head of biceps brachii, long head of triceps brachii, flexor carpi radialis, extensor carpi ulnaris, flexor digitorum superficialis, and extensor digitorum commonis. This layout can comprehensively cover the main dynamic muscle groups of the shoulder, elbow, wrist, and finger joints, thereby completely capturing the muscle activation patterns of the upper limb kinetic chain. Meanwhile, the inertial measurement unit can use the MTw Awinda wireless inertial sensor kit from the Dutch company Xsens. This kit contains three independent inertial measurement units, each of which integrates a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer. Their sampling frequency is uniformly set to 100 Hz. These three inertial measurement units are fixed to the middle of the patient's upper arm, the middle of the forearm, and the back of the hand on the affected side by elastic straps, respectively, to accurately measure the kinematic state of these three key segments in three-dimensional space.

[0031] Understandably, the planned rehabilitation exercise can be a "forward extension-grasping-placement" task. The specific procedure is as follows: Starting in a seated position with arms hanging naturally, upon hearing the instruction, the patient extends their upper limb forward to reach and grasp a cylindrical wooden block placed in front of them at a horizontal distance of 80% of their arm length and a vertical height equal to their shoulder height. The cylindrical block is 5 cm in diameter and 10 cm high. The patient then moves the block to a target area on the same side of their body and places it down, for example, 30 cm to the side of the initial position. Finally, the arm returns to the initial position. The entire task must be completed within 5 seconds, and the entire data acquisition process must be conducted in a shielded laboratory with a constant temperature of 23±1°C, relative humidity of 50±5%, and free from strong electromagnetic interference to eliminate the potential impact of environmental factors on data quality. The final output of this step is eight channels of raw electromyographic voltage timing signals and nine-axis raw data streams (including three-axis acceleration, three-axis angular velocity, and three-axis magnetic field strength) output by each of the three inertial measurement units.

[0032] S02: Based on a preset time window, the multi-source heterogeneous rehabilitation assessment data is processed into a series of time-series dynamic graphs, wherein each time-series dynamic graph is a graph snapshot, and its heterogeneous data nodes are composed of time-series data from different sources, and the edges between nodes represent the biomechanical or temporal correlation of the time-series data.

[0033] Specifically, the preset time window length can be set to 250 milliseconds, and the sliding step size is 125 milliseconds, meaning there is a 50% overlap between windows. The choice of a 250-millisecond window length is based on a comprehensive consideration of the recruitment frequency and kinematic smoothness of human motor units. This duration is sufficient to capture the complete waveform of a rapid muscle contraction-relaxation event, while avoiding blurring of motion details due to an excessively long window. The 50% overlap ensures the continuity and smooth transition of timing information. Within each time window, all data is used to construct an independent graph snapshot, a graph structure consisting of a set of nodes and a set of edges. The node set in this snapshot comprises all data channels, specifically including electromyography (EMG) activation values ​​obtained after processing eight EMG channels, and nine attitude angle values ​​calculated by three inertial measurement units (IMUs). Each IMU outputs three Euler angles: roll, pitch, and yaw. Therefore, each graph snapshot contains 17 heterogeneous data nodes. The characteristic of each node at a given moment is its mean or root mean square value within that time window. The set of edges between nodes represents the intrinsic relationships between these data channels. By constructing prior domain knowledge, for example, there is a direct biomechanical driving relationship between the activation of the biceps brachii node and the change in the pitch angle of the forearm (corresponding to elbow flexion), a "biomechanical coupling edge" is established between these two nodes. In this way, multiple time series data that were originally independent are reconstructed into a structured data representation that can reflect their inherent physical connection and functional coupling, namely a graph sequence. The final data structure of this series of time series dynamic graphs can be represented as a four-dimensional tensor, whose dimensions represent the batch size, the number of graph snapshots, the number of nodes, and the node feature dimension, respectively. This tensor serves as the direct input to the subsequent neural network model.

[0034] S03: Input a series of the aforementioned time-series dynamic graphs into a pre-trained time-series dynamic graph neural network model, and have the time-series dynamic graph neural network model perform the following operations: capture and fuse the spatiotemporal dependency between the muscle activation pattern reflected by the electromyographic signals and the kinematic features reflected by the posture angles through the message passing and aggregation mechanism of the graph neural network, so as to achieve deep fusion and temporal calibration of the multi-source heterogeneous rehabilitation assessment data; and dynamically evaluate the data quality of the heterogeneous data nodes based on the signal-to-noise ratio or information entropy in the temporal evolution through the attention mechanism integrated in the time-series dynamic graph neural network model, and assign a dynamic weight coefficient to each of the heterogeneous data nodes to adaptively adjust the contribution of different data sources during the information aggregation process.

[0035] In this embodiment, the temporal dynamic graph neural network model is specifically structured as a temporal graph convolutional network. This model consists of three stacked temporal graph convolutional layers, each followed by a batch normalization layer and a modified linear unit activation function. Within each temporal graph convolutional layer, the core operation comprises two parts: First, a graph convolutional network module aggregates neighbor node information in the spatial (i.e., graph structure). Specifically, for any node in the graph, its feature vector in the next layer is calculated by aggregating the feature vectors of all its neighbor nodes in the current layer and combining them with its own features. Through this operation, muscle activation information and limb kinematics information can be effectively transmitted and fused along the preset biomechanical coupling edges in the graph, enabling the temporal dynamic graph neural network model to learn deep spatiotemporal coupling patterns. Next, the output of the graph convolutional network module is fed into a gated recurrent unit module to capture... By capturing temporal dependencies, the gated recurrent unit effectively models the dynamic evolution of the graph structure through its internal update and reset gate mechanisms. An integrated attention mechanism is activated during the message passing phase of the graph convolutional network to dynamically evaluate the importance of information transmitted by different neighboring nodes. Specifically, the temporal dynamic graph neural network model calculates an approximate value of the information entropy or signal-to-noise ratio for each data node based on its feature vector sequence over the most recent few time steps. If a node's signal exhibits extremely low variance (potentially indicating sensor disconnection or inactivity) or extremely high and irregular variance (potentially indicating noise interference), its assigned attention weight is reduced accordingly. This weight coefficient (a scalar in the range of 0 to 1) is multiplied by the message transmitted by that neighboring node, thus adaptively suppressing the contribution of low-quality or noisy data sources during the aggregation process, enhancing the model's robustness to common data quality problems in real-world clinical environments. After processing by a three-layer temporal graph convolutional network, the temporal dynamic graph neural network model ultimately outputs a fixed-dimensional graph-level representation vector, for example, 128 dimensions.

[0036] S04: Based on the output of the temporal dynamic graph neural network model, a quantitative rehabilitation assessment result is generated, which characterizes the accurate identification of the temporal pattern of the patient's rehabilitation status.

[0037] Specifically, this 128-dimensional graph-level representation vector is fed into a two-layer multilayer perceptron. Depending on the nature of the assessment task, the output layer can employ different activation functions. In this embodiment, to achieve more refined quantitative assessment, a regression task is used, where the output layer directly outputs a continuous scalar value, which is normalized to a score between 0 and 100, serving as the final quantitative rehabilitation assessment result. To train the model, this embodiment uses a dataset containing 500 patients at different rehabilitation stages and 100 healthy individuals. Each subject's data is scored by multiple senior rehabilitation therapists using the internationally recognized Fogg-Meyer Upper Limb Rating Scale, and this expert score serves as the standard label for model training. The objective function of model training is to minimize the mean squared error between the model's predicted score and the expert score of the Fogg-Meyer Upper Limb Rating Scale. Through this end-to-end training, the model learns how to automatically extract key features from raw sensor data and map them onto clinically significant rehabilitation scores, achieving accurate identification and quantification of the temporal patterns of patient rehabilitation status.

[0038] It is understood that before processing the multi-source heterogeneous rehabilitation assessment data into a series of time-series dynamic graphs, a preprocessing step is included for the multi-source heterogeneous rehabilitation assessment data. The preprocessing includes: performing bandpass filtering and notch filtering on the electromyographic signal time-series data to filter out power frequency interference and motion artifacts; and performing attitude calculation on the raw acceleration and angular velocity data collected by the inertial measurement unit to obtain attitude angle time-series data characterizing the orientation of limb segments in three-dimensional space.

[0039] Specifically, the preprocessing step for the electromyography (EMG) signal time-series data can take the input of raw EMG voltage signals sampled at 2048 Hz from an 8-channel FREEEMG1000 system. The processing mechanism involves applying a series of digital signal processing filters. First, to filter out 50 Hz power frequency interference and its harmonics, a second-order infinite impulse response digital notch filter is applied to the raw signal of each channel. Then, to eliminate motion artifacts and high-frequency noise, the signal is fed into a fourth-order Butterworth digital bandpass filter, with its passband range set... The frequency ranges from 20 Hz to 450 Hz. After filtering, the signal is rectified and its envelope is calculated using the moving root mean square algorithm to obtain an activation index characterizing the intensity of muscle contraction. The calculation window length of the root mean square is set to 150 milliseconds. Finally, in order to make the electromyographic activation levels comparable between different channels or different subjects, the maximum voluntary contraction method is used for normalization to obtain dimensionless normalized electromyographic activation levels (in percentage form). The time series data of the normalized electromyographic activation levels of these 8 channels are the electromyographic signal part that is finally input into the graph construction module.

[0040] In this embodiment, we can use an extended Kalman filter as the core attitude calculation algorithm. This algorithm can optimally fuse data from the gyroscope, accelerometer, and magnetometer. The state vector of the algorithm is defined as an attitude representation in quaternion form to avoid gimbal lock-up. The calculated attitude quaternions are then converted into more intuitive Euler angles, including roll, pitch, and yaw angles, for subsequent biomechanical analysis and graph node construction. Thus, we obtain three attitude angles for each of the three segments of the upper arm, forearm, and hand, for a total of nine dimensions of kinematic time-series data.

[0041] In this embodiment of the application, the step of processing the multi-source heterogeneous rehabilitation assessment data into a series of time-series dynamic graphs specifically involves: defining each channel of the preprocessed electromyographic signal time-series data and each dimension of the posture angle time-series data as a heterogeneous data node; within the same time slice, constructing edges representing biomechanical coupling between the corresponding heterogeneous data nodes based on the linkage relationship between muscles and joints in human anatomy; and constructing edges representing their temporal autocorrelation between heterogeneous data nodes representing the same physical quantity, thereby forming a graph snapshot.

[0042] Specifically, firstly, the 17 data channels are mapped one by one to the 17 nodes of the graph. Secondly, biomechanical coupling edges are constructed. These edges are fixed and reflect the inherent anatomical structure and motor function relationship of the human body. Specific connection examples are as follows: (1) An edge is established between the anterior deltoid node and the upper arm pitch angle node (representing shoulder flexion / extension); (2) An edge is established between the biceps brachii node and the forearm pitch angle node (representing elbow flexion / extension); (3) An edge is also established between the triceps brachii node and the forearm pitch angle node, because the former is the antagonist of the latter; (4) An edge is established between the flexor carpi radialis node and the hand pitch angle node (representing wrist flexion / extension). In this way, we directly link muscle activation with the joint movement it drives on the graph structure. Furthermore, edges are established between functionally synergistic or antagonistic muscles. Finally, the edges representing their temporal autocorrelation are typically implemented in specific graph convolutional network implementations by adding an identity matrix to the adjacency matrix. This operation is equivalent to adding an edge to each node pointing to itself.

[0043] It is understood that the step of assigning a dynamic weight coefficient to each of the heterogeneous data nodes specifically involves: during the information aggregation process of each layer of the temporal dynamic graph neural network, the attention mechanism calculates the similarity score between the feature vector of each heterogeneous data node and the feature vector of its neighboring nodes; based on the similarity score, combined with a measure of the stability of the heterogeneous data node's own features, a normalized attention weight is assigned to the information transmission path from each neighboring node to the central node.

[0044] Specifically, this attention mechanism is implemented through a multi-head graph attention network layer. For a central node in the graph, the graph attention network first calculates an attention coefficient for each edge from neighboring nodes to the central node. This coefficient indicates the importance of the features of neighboring nodes to the central node. To further incorporate considerations of data quality, namely "a measure of the stability of the node's own features," a regulation term is introduced when calculating the attention coefficient. This regulation term is a function of the stability of the features of neighboring nodes. For example, this regulation term can be defined as the reciprocal of the rate of change of the node's feature vector over a preset time step. A node with stable features (small variance) will receive a larger regulation term value, and vice versa. The modified attention coefficient calculation process is as follows: First, the feature vectors of the central node and neighboring nodes are linearly transformed and then concatenated. Next, the concatenated vector is inner-producted with a learnable attention parameter vector and processed through a linear unit activation function with leakage correction. Finally, the result is multiplied by the feature stability regulation term of the neighboring nodes. Through this multiplication term, we effectively inject prior information about data quality into the calculation of attention weights. This allows neighboring nodes with stable signal characteristics and rich information content to naturally receive higher weights during aggregation, while the contributions of nodes with unstable signals or potential noise interference are dynamically suppressed. Finally, the updated feature vector of the center node is obtained by weighting and summing the feature vectors of all its neighboring nodes after linear transformation, according to the final attention weights obtained by normalizing the attention coefficients using a flexible maximum function, and then processing this sum using a non-linear activation function.

[0045] The embodiments of this application can further introduce a federated meta-learning mechanism and an attribution-counterfactual dual-track interpretability engine, which solves the three major technical challenges in clinical applications: personalized model adaptation, data privacy protection, and interpretability of evaluation results.

[0046] Specifically, this also includes personalized adaptation and continuous optimization of the temporal dynamic graph neural network model through a federated meta-learning mechanism. The specific steps are as follows: On a central server, a global temporal dynamic graph neural network model is pre-trained as the base model for meta-learning using historical rehabilitation assessment data from multiple rehabilitation centers; When a new patient is admitted, the base model is rapidly fine-tuned on their local computing device using a small amount of personalized rehabilitation assessment data collected from that patient, generating personalized model parameters adapted to the patient's physiological and pathological characteristics; During the patient's rehabilitation cycle, the personalized model parameters are continuously updated using newly collected data, and the non-sensitive parameter gradient information generated by the update is encrypted and uploaded to the central server for aggregation and updating of the global base model.

[0047] Specifically, the implementation details of this federated meta-learning mechanism are as follows: The step of pre-training the global base model takes as input a historical patient dataset from multiple rehabilitation centers and processes it by using an algorithm called model-independent meta-learning to train a temporal dynamic graph neural network model. The output is an initial global base model parameter with good generalization and fast learning ability. The model-independent meta-learning algorithm trains on a large amount of highly heterogeneous patient data to "learn how to learn," and the final initial global base model parameter is in an optimal parameter space position. Starting from this position, it can quickly converge to a personalized model with excellent performance for new patients with very few samples and update steps.

[0048] The step of fine-tuning the personalized model for new patients involves taking a small amount of data collected when the new patient first uses the system as input. The processing mechanism involves loading and fine-tuning the global base model parameters on the patient's local computing device, and outputting a set of personalized model parameters specific to that patient. When a new patient accesses the system, their local device downloads the latest initial global base model parameters from the central server. Then, using a small amount of locally collected task data as a training set, the model is trained using standard backpropagation. Since the initial global base model parameters already possess rapid learning capabilities, this small amount of local data is sufficient to guide the model parameters to a new position that better captures the unique physiological and pathological characteristics of the new patient, forming personalized model parameters. The entire fine-tuning process is completed in a closed loop on the local device, without uploading the original data, thus technically ensuring the patient's data sovereignty and privacy.

[0049] The continuous optimization and aggregation update steps take into account the new data continuously generated by the patient throughout the recovery period and encrypted parameter gradients from multiple local devices. The processing mechanism involves continuous updates to the local model and federated averaging aggregation from the central server. After one or more updates to the local model, the local device calculates the change in model parameters, i.e., the parameter gradient; this gradient information is processed using homomorphic encryption or differential privacy techniques before uploading. After collecting parameter gradients from multiple different local devices, the central server performs an aggregation update operation. The updated global base model is then redistributed to all local devices as a better starting point for their next round of personalized fine-tuning.

[0050] The specific steps for aggregating and updating the global base model are as follows: after receiving encrypted parameter gradients uploaded from local devices of multiple rehabilitation centers, the central server uses a federated averaging algorithm to perform a weighted average of the parameter gradients to calculate the update amount of the global model; and uses the update amount of the global model to update the global base model.

[0051] Specifically, the core idea of ​​the federated averaging algorithm is to determine the influence of each local device in the global model update based on the amount of data contributed by each device. Devices with larger data volumes are generally considered to have more representative gradients and should therefore be given higher weights. The central server performs weighted aggregation operations on encrypted gradients in a secure multi-party computation environment or by utilizing the characteristics of homomorphic encryption without decryption. The calculated global update is used to update the global base model, that is, the new global model parameters are obtained by adding the current global model parameters to the global update after scaling by the learning rate. After the update is completed, the new global model parameters are broadcast to the local devices in the system as the starting point for a new round of personalized fine-tuning.

[0052] The method further includes, after generating the rehabilitation assessment results, launching an attribution-counterfactual dual-track interpretability engine to provide decision-making explanations and intervention guidance. The engine's operation steps include: identifying and highlighting key biomechanical features that have a decisive impact on the current rehabilitation assessment results through an attribution analysis module; and constructing and presenting one or more achievable, minimal virtual movement adjustment schemes through a counterfactual interpretation generation module, which can guide the patient's assessment results toward the expected better state.

[0053] The step of identifying and highlighting key biomechanical features that have a decisive impact on the current rehabilitation assessment result through the attribution analysis module is as follows: using a gradient-based attribution algorithm, such as integral gradient, to calculate the partial derivative of the quantified rehabilitation assessment result with respect to the input feature value of each heterogeneous data node in the time-series dynamic graph sequence at the key motion time point; and identifying the biomechanical events corresponding to the input feature values ​​with higher partial derivative values ​​as key biomechanical features affecting the assessment result.

[0054] Specifically, the attribution analysis module uses an "integral gradient" algorithm to calculate an attribution score for each data point in the input data sequence. This score intuitively represents the contribution of that data point to the final evaluation score. We identify the feature points with the highest absolute values ​​as "key biomechanical features." For example, the system may find that for a patient with a low score, the feature of "triceps activation during elbow extension" has a very large negative attribution score. The system can then generate the following explanation: "The low evaluation score is mainly due to the abnormal and excessive co-activation of the triceps brachii, an antagonist muscle, during elbow extension, which hinders the smoothness of the movement (key negative factor)."

[0055] The step of constructing and presenting one or more achievable minimal virtual action adjustment schemes through the counterfactual interpretation generation module is as follows: Based on a pre-trained generative adversarial network or variational autoencoder model, with the optimization objective of minimizing the perturbation distance from the patient's original multi-source heterogeneous rehabilitation assessment data, and with the constraint of achieving a preset higher score for the rehabilitation assessment results, an optimization search is performed; one or more virtual counterfactual data samples are decoded and generated from the search results; by comparing the differences between the counterfactual data samples and the original data in key biomechanical characteristics, the quantifiable rehabilitation intervention guidance is generated.

[0056] Specifically, the core of the counterfactual explanation generation module is a pre-trained conditional variational autoencoder. When a counterfactual explanation needs to be generated, the system sets a "target score" and searches in the model's latent space for a new representation point that is closest to the original data representation and can achieve the target score after decoding. After finding this new representation point, the decoder generates a better sequence of action data, i.e., a "counterfactual data sample". Finally, the system compares the differences between the original data and the counterfactual data sample and converts significant differences into natural language instructions, such as: "If your elbow can extend about 5 degrees further when it is fully extended, while consciously relaxing the muscles on the back of your upper arm (triceps) and reducing their activation level by about 10%, then your evaluation score is expected to improve by 5 points."

[0057] This application also discloses an intelligent rehabilitation assessment and data analysis system, including a memory and a processor. The memory stores a computer program that can be loaded by the processor and executed by the intelligent rehabilitation assessment and data analysis method described above.

[0058] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0059] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0060] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0061] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0062] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0063] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. An intelligent rehabilitation assessment and data analysis method, characterized in that, include: The system acquires multi-source heterogeneous rehabilitation assessment data collected by wearable sensors in real time when the patient performs predetermined rehabilitation movements. The multi-source heterogeneous rehabilitation assessment data includes at least electromyographic signal timing data from electromyography sensors and attitude angle timing data from inertial measurement units. Based on a preset time window, the multi-source heterogeneous rehabilitation assessment data is processed into a series of time-series dynamic graphs. Each time-series dynamic graph is a graph snapshot, and its heterogeneous data nodes are composed of time-series data from different sources. The edges between nodes represent the biomechanical or temporal correlation of the time-series data. A series of the aforementioned temporal dynamic graphs are input into a pre-trained temporal dynamic graph neural network model, which then performs the following operations: By employing the message passing and aggregation mechanism of graph neural networks, the spatiotemporal dependency between the muscle activation patterns reflected by the electromyographic signals and the kinematic characteristics reflected by the posture angles is captured and fused to achieve deep fusion and temporal calibration of the multi-source heterogeneous rehabilitation assessment data; and By integrating an attention mechanism into the temporal dynamic graph neural network model, the data quality of the heterogeneous data nodes is dynamically evaluated based on their signal-to-noise ratio or information entropy during temporal evolution. A dynamic weight coefficient is assigned to each heterogeneous data node to adaptively adjust the contribution of different data sources during information aggregation. Based on the output of the temporal dynamic graph neural network model, a quantitative rehabilitation assessment result is generated, which characterizes the accurate identification of the temporal pattern of the patient's rehabilitation status.

2. The method according to claim 1, characterized in that, The method further includes personalized adaptation and continuous optimization of the temporal dynamic graph neural network model through a federated meta-learning mechanism, specifically including the following steps: On the central server, a global temporal dynamic graph neural network model is pre-trained using historical rehabilitation assessment data from multiple rehabilitation centers as the base model for meta-learning; When a new patient is admitted, the basic model is rapidly fine-tuned on their local computing device using a small amount of personalized rehabilitation assessment data collected from that patient. This generates personalized model parameters adapted to the patient's physiological and pathological characteristics. This process is completed while protecting the patient's original data from leaving the local device. During the patient's recovery period, the parameters of the personalized model are continuously updated using newly collected data. The gradient information of the non-sensitive parameters generated by the update is encrypted and uploaded to the central server for aggregation and updating of the global basic model in the central server, thereby achieving continuous optimization of the model.

3. The method according to claim 1, characterized in that, Before processing the multi-source heterogeneous rehabilitation assessment data into a series of time-series dynamic graphs, a preprocessing step is included for the multi-source heterogeneous rehabilitation assessment data, the preprocessing including: The electromyography signal time-series data is subjected to bandpass filtering and notch filtering to remove power frequency interference and motion artifacts, thereby improving the signal-to-noise ratio. Furthermore, attitude calculation is performed on the raw acceleration and angular velocity data collected by the inertial measurement unit to obtain attitude angle time-series data characterizing the orientation of limb segments in three-dimensional space.

4. The method according to claim 3, characterized in that, The specific steps for processing the multi-source heterogeneous rehabilitation assessment data into a series of time-series dynamic graphs are as follows: Each channel of the preprocessed electromyographic signal time series data and each dimension of the attitude angle time series data are defined as a heterogeneous data node. Within the same time slice, based on the linkage relationship between muscles and joints in human anatomy, edges representing biomechanical coupling are constructed between corresponding heterogeneous data nodes. Furthermore, between heterogeneous data nodes representing the same physical quantity, edges representing their temporal autocorrelation are constructed to form a graph snapshot.

5. The method according to claim 1, characterized in that, The step of assigning a dynamic weight coefficient to each of the heterogeneous data nodes specifically includes: During each layer of information aggregation in the temporal dynamic graph neural network model, the attention mechanism calculates the similarity score between the feature vector of each heterogeneous data node and the feature vector of its neighboring nodes. Based on the similarity score and combined with a measure of the stability of the heterogeneous data nodes' own features, a normalized attention weight is assigned to each neighbor node's information transmission path to the central node, so that nodes with stable features and rich information content receive higher weights when aggregating neighbor information.

6. The method according to claim 1, characterized in that, After generating the rehabilitation assessment results, an attribution-counterfactual dual-track interpretability engine is activated to provide decision-making explanations and intervention guidance. The operation steps of the attribution-counterfactual dual-track interpretability engine include: The attribution analysis module identifies and highlights key biomechanical features that decisively influence the current rehabilitation assessment results; and Through the counterfactual interpretation generation module, one or more feasible minimal virtual motion adjustment schemes are constructed and presented. These virtual motion adjustment schemes can guide the patient's rehabilitation assessment results toward the expected better state and provide actionable rehabilitation intervention guidance to rehabilitation therapists or patients in a preset form.

7. The method according to claim 6, characterized in that, The step of identifying and highlighting key biomechanical features that have a decisive impact on the current rehabilitation assessment results through the attribution analysis module includes: A gradient-based attribution algorithm is used to calculate the partial derivative of the rehabilitation assessment result with respect to the input feature value of each heterogeneous data node in the time-series dynamic graph sequence at key motion time points; Biomechanical events corresponding to input feature values ​​with high partial derivative values ​​are identified as key biomechanical features that affect the evaluation results.

8. The method according to claim 6, characterized in that, The step of constructing and presenting one or more feasible minimal virtual action adjustment schemes through the counterfactual interpretation generation module includes: Based on a pre-trained generative adversarial network or variational autoencoder model, the optimization search is performed with the goal of minimizing the perturbation distance from the patient's original multi-source heterogeneous rehabilitation assessment data, while the constraint is to make the rehabilitation assessment results reach a preset higher score. Decode and generate one or more virtual counterfactual data samples from search results; and By comparing the differences in key biomechanical characteristics between the counterfactual data sample and the original data, the quantifiable rehabilitation intervention guidance is generated.

9. The method according to claim 2, characterized in that, The steps for aggregating and updating the global base model specifically include: After receiving encrypted parameter gradients uploaded from local devices in multiple rehabilitation centers, the central server uses a federated averaging algorithm to perform a weighted average of the parameter gradients in order to calculate the update amount of the global model. The global base model is updated using the update amount of the global model, and the updated global base model parameters are distributed to each local computing device as the starting point for the next round of personalized fine-tuning.

10. An intelligent rehabilitation assessment and data analysis system, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any one of claims 1 to 9.

Images