An intelligent evaluation method for upper limb motor function based on kinematics and deep learning for stroke rehabilitation assessment

Abstract

The application relates to the technical field of stroke rehabilitation evaluation, and discloses a kinematics and deep learning-based intelligent upper limb motor function evaluation method for stroke rehabilitation evaluation, which comprises the following steps: a wearable inertial sensing system is used to collect multi-dimensional motion data of rehabilitation actions of a patient; after length normalization, abnormal jump removal and numerical normalization preprocessing, joint angles are calculated based on upper limb kinematics modeling; two types of data are spliced and divided into segments according to a preset specification; features are extracted through a deep learning network with a fusion attention mechanism; a prediction model is trained in combination with an optimization algorithm; new patient data is input into the model after the same processing, and an upper limb motor function evaluation result is generated. The application combines wearable sensing, kinematics modeling and deep learning technology, realizes accurate, objective and intelligent evaluation of the upper limb motor function of a stroke patient, improves evaluation efficiency and accuracy, and provides a reliable basis for rehabilitation diagnosis and treatment.

Description

Technical Field

[0001] This invention relates to the field of stroke rehabilitation assessment technology, and more specifically, to an intelligent assessment method for upper limb motor function based on kinematics and deep learning for stroke rehabilitation assessment. Background Technology

[0002] Stroke patients often experience upper limb motor dysfunction, and upper limb motor function assessment is a core component in developing rehabilitation plans and evaluating rehabilitation effectiveness. Traditional upper limb motor function assessments largely rely on scale methods (such as the Fugl-Meyer scale), which are completed by assessors observing the patient's movements and assigning scores, providing a basis for clinical rehabilitation diagnosis and treatment.

[0003] However, the accuracy and objectivity of traditional assessment methods are difficult to guarantee. Assessment results are highly dependent on the assessor's clinical experience, and different assessors may score the same patient differently, leading to poor consistency in assessment results. The assessment process is mainly qualitative and descriptive, making it difficult to accurately quantify subtle motor characteristics such as joint range of motion and coordination, and thus failing to fully reflect the patient's true functional level. The assessment process is time-consuming and inefficient, and lacks a real-time feedback mechanism, making it impossible to dynamically capture functional changes in the patient's rehabilitation process, resulting in delayed adjustments to the rehabilitation plan and further affecting the rehabilitation outcome.

[0004] Therefore, it is necessary to design an intelligent assessment method for upper limb motor function based on kinematics and deep learning to solve the problems of poor consistency of results, low quantification, low assessment efficiency and lack of real-time feedback in traditional assessment methods. Summary of the Invention

[0005] In view of this, the present invention proposes an intelligent assessment method for upper limb motor function based on kinematics and deep learning for stroke rehabilitation assessment, aiming to solve the problems of poor consistency of results, low degree of quantification, low assessment efficiency and lack of real-time feedback in traditional assessment methods.

[0006] This invention proposes an intelligent assessment method for upper limb motor function based on kinematics and deep learning for stroke rehabilitation evaluation, comprising: Construct a wearable inertial sensing system to collect multi-dimensional motion data of patients when performing rehabilitation movements; The collected motion data is preprocessed, including length normalization, removal of anomalous jumps, and numerical normalization, to obtain standardized data. Based on human upper limb kinematic modeling, the joint movement angles are calculated from standardized data; Multidimensional motion data and joint range of motion data are stitched together, and the data segments are divided according to preset window specifications and stride length; A deep learning network with an attention fusion mechanism is used to extract features from data segments. The model parameters are then iteratively optimized using an optimization algorithm to train a motor function level prediction model based on a rehabilitation assessment scale. Preprocessing, joint angle calculation, and data segmentation are performed on the multidimensional motion data of newly assessed patients. The processed data is then input into the trained prediction model to generate upper limb motor function assessment results.

[0007] Furthermore, the preprocessing of the collected motion data includes: A first noise threshold and a second noise threshold are preset, wherein the first noise threshold is greater than the second noise threshold; An adaptive filtering algorithm is used to denoise the collected multi-dimensional motion data. When the noise amplitude in the data is greater than the first noise threshold, a multi-level filtering strategy is used to perform time-domain filtering and frequency-domain filtering in sequence. When the noise amplitude is less than or equal to the first noise threshold and greater than the second noise threshold, only time-domain filtering is performed; When the noise amplitude is less than or equal to the second noise threshold, no additional noise reduction is required; When removing abnormal jumps, the rate of change of adjacent data points is calculated, and data segments with a rate of change greater than a preset jump threshold are identified as abnormal jumps. An interpolation algorithm is then used to complete the data in the abnormal jump segments. Numerical normalization uses a preset normalization interval to map the denoised and completed data to the preset normalization interval, thus obtaining standardized data. The upper and lower limits of the normalization interval are determined based on the historical valid range of the motion data.

[0008] Furthermore, when using interpolation algorithms to complete data for anomalous transition segments, the following steps are included: A first transition length threshold and a second transition length threshold are preset, wherein the first transition length threshold is greater than the second transition length threshold; Interpolation algorithms include two types: linear interpolation and cubic spline interpolation. When the number of data points in the abnormal jump segment exceeds the first jump length threshold, a cubic spline interpolation algorithm is used to complete the data points. During the completion process, the trend of a preset number of valid data points before and after the jump segment is referenced. When the number of data points in an abnormal jump segment is greater than the second jump length threshold but less than or equal to the first jump length threshold, a linear interpolation algorithm is used for completion. When the number of data points in the abnormal jump segment is less than or equal to the second jump length threshold, the abnormal data points are directly removed and replaced by the average of the adjacent valid data points. After completion, the data change rate needs to be checked again. If there are still cases where the change rate is greater than the preset threshold, the completion operation is repeated until the data is completed.

[0009] Furthermore, when calculating joint range of motion based on human upper limb kinematic modeling, the following is included: The degrees of freedom parameters of the shoulder, elbow, and wrist joints and the relationship of the length of the connecting rods between the joints are determined based on the aforementioned multi-joint kinematic model of the human upper limb. Using Euler angle data from standardized data as input, and combining it with forward kinematics algorithm, the flexion-extension angle, adduction-abduction angle and rotation angle of each joint are calculated sequentially; Set a reasonable range threshold for the range of motion angles of each joint, and determine the angle values ​​that exceed the reasonable range as invalid values. Use a smooth transition algorithm for the valid angles at adjacent time points to correct invalid values. The threshold for the reasonable range of joint motion angles is determined based on statistical data of upper limb joint mobility in healthy individuals, and is adjusted individually according to the patient's age and disease course.

[0010] Furthermore, when correcting invalid values ​​in the solution results, the following steps are included: A first correction weight and a second correction weight are preset, wherein the first correction weight is greater than the second correction weight; When the number of consecutive occurrences of invalid values ​​is less than the preset threshold, the weighted average of the valid angles at the previous moment and the valid angles at the next moment is used for correction. The weight of the angle at the previous moment is the first correction weight, and the weight of the angle at the next moment is the second correction weight. When the number of consecutive occurrences of invalid values ​​is greater than or equal to the preset threshold, the standard joint angle change curve corresponding to the action is invoked based on the type of rehabilitation action performed by the patient, and segmented correction is performed based on the trend of the curve. After correction, the smoothness of the joint angle change is calculated. If the smoothness is less than the preset smoothness threshold, the weight coefficient or standard curve matching degree is adjusted, and correction is performed again.

[0011] Furthermore, when dividing data segments according to preset window specifications and step sizes, the following is included: Pre-set the first window specification, the second window specification, the first step length, and the second step length. The window length corresponding to the first window specification is greater than the second window specification, and the first step length is greater than the second step length. The data segmentation method is determined based on the duration of the rehabilitation movements: When the duration of a rehabilitation movement exceeds a preset duration threshold, the first window size and the first step length are used for division. After division, each data segment contains a complete movement feature cycle. When the duration of the rehabilitation action is less than or equal to the preset duration threshold, the second window size and the second step size are used for division, and the window overlap rate is increased to the preset overlap ratio. If the length of the last data segment is less than half of the corresponding window size during the partitioning process, it will be merged with the previous data segment. If it is greater than or equal to half, it will be padded with zeros to complete the corresponding window size.

[0012] Furthermore, when adjusting window specifications and step sizes, the following should be included: A first feature complexity threshold and a second feature complexity threshold are preset, wherein the first feature complexity threshold is greater than the second feature complexity threshold; The feature complexity of the concatenated standardized data and joint angle data is calculated by combining the information entropy and variance of the data. When the feature complexity exceeds the first feature complexity threshold, the window size is adjusted to 1.2 times the corresponding size, while the step size remains unchanged. When the feature complexity is less than or equal to the first feature complexity threshold and greater than the second feature complexity threshold, the original window size and step size are maintained. When the feature complexity is less than or equal to the second feature complexity threshold, adjust the window size to 0.8 times the corresponding size and the step size to 1.2 times the corresponding step size.

[0013] Furthermore, when using deep learning networks with fused attention mechanisms for feature extraction and model training, the following steps are included: The deep learning network that integrates the attention mechanism is called the BiGRU-Attention network, which includes an input layer, a BiGRU layer, an attention layer, a fully connected layer, and an output layer. The BiGRU layer includes forward GRU units and backward GRU units, which are used to capture the temporal dependencies of data segments. The number of hidden layer nodes in the BiGRU layer is determined according to the dimension of the data segment. The attention layer assigns weights to the feature vectors output by the BiGRU layer and determines features with weight values ​​greater than a preset weight threshold as key features. The optimization algorithm selected is an intelligent optimization algorithm, which iteratively optimizes the weight parameters and bias parameters of the network by minimizing the loss function between the model's predicted value and the standard value of the rehabilitation assessment scale. During training, a cross-validation strategy is adopted, dividing the dataset into a training set, a validation set, and a test set. When the model accuracy on the validation set does not improve for a preset number of consecutive iterations, training is stopped, and the current best model is saved as the motor function level prediction model.

[0014] Furthermore, when iteratively optimizing model parameters using intelligent optimization algorithms, the following steps are included: A first iteration number threshold and a second iteration number threshold are preset, wherein the first iteration number threshold is greater than the second iteration number threshold; The optimization objectives of the intelligent optimization algorithm include two indicators: model accuracy and loss function value. Weight coefficients are set for the two indicators, and the comprehensive optimization objective value is calculated by weighting them. When the overall optimization target value is greater than the preset optimization target threshold, the iteration step size is adjusted to 0.8 times the initial step size, and the iteration continues until the first iteration number threshold is reached; When the comprehensive optimization target value is less than or equal to the preset optimization target threshold and greater than the preset minimum target threshold, maintain the initial iteration step size and iterate until the second iteration number threshold. When the overall optimization target value is less than or equal to the preset minimum target threshold, adjust the initial parameters of the optimization algorithm and restart the iterative optimization process; The weighting coefficients are determined based on the degree of influence of the two indicators on the model performance, and the sum of the weighting coefficients is 1.

[0015] Furthermore, when generating upper limb motor function assessment results, the following are included: A first confidence threshold and a second confidence threshold are preset, wherein the first confidence threshold is greater than the second confidence threshold; at the same time, clinical information weight and action completion weight are set, wherein the sum of the clinical information weight and the action completion weight is 1; The evaluation results include confidence assessment, result correction, and verification. Obtain the preliminary assessment results and corresponding confidence values ​​output by the prediction model. When the confidence value is greater than or equal to the first confidence threshold, the preliminary assessment results are deemed reliable, and the process proceeds directly to the result verification stage. When the confidence value is greater than or equal to the second confidence threshold and less than the first confidence threshold, the patient's clinical information is retrieved, and the preliminary assessment result is weighted and corrected in combination with the weight of the clinical information. When the confidence value is less than the second confidence threshold, it is determined that the current data is insufficient to support a reliable assessment. A supplementary data collection prompt is output, requiring the patient to perform a preset number of rehabilitation actions to supplement the data. After repeating the data collection and data segmentation process, the data is input into the model again to generate preliminary assessment results. After the results are corrected, the action completion score is calculated. The action completion score is derived from the matching degree between the joint angle and the standard action angle. The corrected results are then further optimized by combining the action completion weight. The result verification process employs a dual verification mechanism: Compare the assessment results with the grade classification range of the rehabilitation assessment scale; The system retrieves historical assessment data of similar patients, calculates the similarity of the results, and confirms the validity of the assessment results when the similarity is greater than a preset similarity threshold. If the similarity is less than or equal to the preset similarity threshold, the weights of clinical information and action completion are readjusted, and the correction process is repeated until the result similarity meets the standard or the maximum number of corrections is reached. Among them, the types of rehabilitation movements collected in the supplementary collection are the same as those collected in the initial collection, and the number does not exceed 50% of the number of movements collected in the initial collection.

[0016] Compared with existing technologies, the advantages of this invention are as follows: It collects objective, multi-dimensional motion data through a wearable inertial sensing system, and accurately calculates joint angles by combining standardized preprocessing and kinematic modeling, replacing subjective manual observation and significantly improving the consistency and quantification of assessment results; it automatically completes feature extraction and grade prediction using a deep learning model that integrates attention mechanisms, significantly shortening assessment time and improving assessment efficiency; and it supports real-time data processing and result output throughout the entire process, dynamically capturing functional changes in the patient's rehabilitation process, providing a reliable basis for timely adjustments to the rehabilitation plan, and filling the gap in traditional assessments that lack real-time feedback. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of an intelligent assessment method for upper limb motor function based on kinematics and deep learning for stroke rehabilitation assessment provided in an embodiment of the present invention; Figure 2 A comparison chart showing the use of only raw IMU data for model training, provided for embodiments of the present invention; Figure 3 This is a schematic diagram illustrating the recognition rate of the model provided in this embodiment of the invention on the test set. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] Reference Figures 1-3As shown in some embodiments of this application, a method for intelligent assessment of upper limb motor function based on kinematics and deep learning for stroke rehabilitation includes: Step S100: Construct a wearable inertial sensing system to collect multi-dimensional motion data of the patient when performing rehabilitation movements; Step S200: Preprocess the collected motion data, including length normalization, removal of abnormal jumps and numerical normalization operations, to obtain standardized data; Step S300: Based on human upper limb kinematic modeling, calculate the joint movement angle from standardized data; Step S400: The multi-dimensional motion data and joint motion angle data are stitched together, and the data segments are divided according to the preset window size and step length; Step S500: Use a deep learning network with an attention mechanism to extract features from the data segments, and combine iterative optimization algorithms to optimize the model parameters to train a motor function level prediction model based on a rehabilitation assessment scale. Step S600: Perform preprocessing, joint angle calculation and data segmentation operations on the multidimensional motion data of the newly assessed patients, input the processed data into the trained prediction model, and generate upper limb motor function assessment results.

[0020] Specifically, step S100 uses seven inertial measurement units (IMUs) to build a wearable sensing system. A data acquisition scheme adapted for hospital, community, and home scenarios is designed for stroke patients. It supports flexible adjustment of the number of sensors (e.g., wearing only on the affected side or adding lower limb sensors to expand the assessment range) and optimization of the wearing position (selecting a position that is comfortable for the patient and minimizes muscle interference). It collects nine-axis motion data (three-axis acceleration, three-axis angular velocity, and three-axis Euler angles) when the patient performs six core rehabilitation movements, including pronation / supination, upper arm abduction, touching the ear, and touching the lumbar spine. Each movement corresponds to three functional levels on the Fugl-Meyer scale: 0, 1, and 2. By comprehensively capturing displacement, velocity, and posture changes in upper limb movements, it provides high-quality raw data for subsequent data preprocessing and joint angle calculation. Furthermore, through multi-dimensional data acquisition design, it reduces the dependence of traditional IMU raw data on the wearing position from the source, laying the foundation for improving the model's generalization ability and robustness.

[0021] The above embodiments collect objective, multi-dimensional motion data through wearable inertial sensing systems, and accurately calculate joint angles by combining standardized preprocessing and kinematic modeling, replacing manual subjective observation and significantly improving the consistency and quantification of assessment results. By using a deep learning model that integrates attention mechanisms to automatically complete feature extraction and level prediction, the assessment time is significantly shortened and the assessment efficiency is improved. The entire process supports real-time data processing and result output, which can dynamically capture functional changes in the patient's rehabilitation process, providing a reliable basis for timely adjustment of rehabilitation plans and filling the gap of traditional assessments lacking real-time feedback.

[0022] Specifically, the preprocessing of the collected motion data includes: A first noise threshold and a second noise threshold are preset, wherein the first noise threshold is greater than the second noise threshold; An adaptive filtering algorithm is used to denoise the collected multi-dimensional motion data. When the noise amplitude in the data is greater than the first noise threshold, a multi-level filtering strategy is used to perform time-domain filtering and frequency-domain filtering in sequence. When the noise amplitude is less than or equal to the first noise threshold and greater than the second noise threshold, only time-domain filtering is performed; When the noise amplitude is less than or equal to the second noise threshold, no additional noise reduction is required; When removing abnormal jumps, the rate of change of adjacent data points is calculated, and data segments with a rate of change greater than a preset jump threshold are identified as abnormal jumps. An interpolation algorithm is then used to complete the data in the abnormal jump segments. Numerical normalization uses a preset normalization interval to map the denoised and completed data to the preset normalization interval, thus obtaining standardized data. The upper and lower limits of the normalization interval are determined based on the historical valid range of the motion data.

[0023] Specifically, when using interpolation algorithms to complete data for anomalous transition segments, the following steps are included: A first transition length threshold and a second transition length threshold are preset, wherein the first transition length threshold is greater than the second transition length threshold; Interpolation algorithms include two types: linear interpolation and cubic spline interpolation. When the number of data points in the abnormal jump segment exceeds the first jump length threshold, a cubic spline interpolation algorithm is used to complete the data points. During the completion process, the trend of a preset number of valid data points before and after the jump segment is referenced. When the number of data points in an abnormal jump segment is greater than the second jump length threshold but less than or equal to the first jump length threshold, a linear interpolation algorithm is used for completion. When the number of data points in the abnormal jump segment is less than or equal to the second jump length threshold, the abnormal data points are directly removed and replaced by the average of the adjacent valid data points. After completion, the data change rate needs to be checked again. If there are still cases where the change rate is greater than the preset threshold, the completion operation is repeated until the data is completed.

[0024] Specifically, when preprocessing the collected motion data, a first noise threshold and a second noise threshold are preset. The first noise threshold for acceleration is set to 0.8g and the second noise threshold is set to 0.3g. The first noise threshold for angular velocity is set to 20° / s and the second noise threshold is set to 8° / s. The first noise threshold for Euler angle is set to 15° and the second noise threshold is set to 5°. Preset jump thresholds are set according to the data type: the acceleration rate of change threshold is 5g / s, the angular velocity rate of change threshold is 30° / s², and the Euler angle rate of change threshold is 25° / s². An adaptive filtering algorithm is used to denoise the collected nine-axis multi-dimensional motion data (three-axis acceleration, three-axis angular velocity, and three-axis Euler angles). When the noise amplitude in the data is greater than the first noise threshold of the corresponding dimension, a multi-level filtering strategy is used to sequentially perform time-domain low-pass filtering (cutoff frequency set to 10Hz) and frequency-domain wavelet filtering (using the db4 wavelet basis). When the noise amplitude is less than or equal to the first noise threshold but greater than the second noise threshold, only time-domain low-pass filtering (cutoff frequency 10Hz) is performed. When the noise amplitude is less than or equal to the second noise threshold, no additional denoising is required. When removing abnormal jumps, the instantaneous rate of change of adjacent data points is calculated. Data segments with a rate of change greater than the preset jump threshold of the corresponding dimension are identified as abnormal jumps, and an interpolation algorithm is used to complete the data for the abnormal jump segments. Numerical normalization uses a preset normalization interval [0,1]. The upper and lower limits of this interval are determined based on the historical effective range of motion data. The historical effective range is based on the statistical analysis of nine-axis motion data collected from four stroke patients performing six rehabilitation movements. The effective range of acceleration is -3g to 3g, the effective range of angular velocity is -180° / s to 180°, and the effective range of Euler angles is -180° to 180°. Based on this, the denoised and completed data is mapped to the interval [0,1] to obtain standardized data.

[0025] Specifically, when using interpolation algorithms to complete data in anomalous transition segments, a first transition length threshold of 15 data points and a second transition length threshold of 5 data points are pre-set. The interpolation algorithms include linear interpolation and cubic spline interpolation. When the number of data points in an anomalous transition segment is greater than 15, cubic spline interpolation is used for completion, referencing the changing trends of the 10 valid data points before and after the transition segment to ensure the continuity of the completed data with the preceding and following data. When the number of data points in an anomalous transition segment is greater than 5 but less than or equal to 15, linear interpolation is used, constructing a linear function using adjacent valid data points before and after the transition segment as endpoints to fill the missing data. When the number of data points in an anomalous transition segment is less than or equal to 5, the anomalous data points in that segment are directly removed, and the average of the 3 valid data points before and after the transition segment is used as a replacement. After completion, the rate of change of adjacent data points needs to be recalculated for verification. If there are still cases where the rate of change is greater than the preset threshold for the corresponding dimension, the completion operation is repeated, with a maximum of 3 repetitions, until the data meets the rate of change requirement.

[0026] The above embodiments, by setting graded noise thresholds and employing adaptive filtering strategies, achieve precise graded processing of noise of different intensities. This efficiently filters out interference signals in motion data while preserving effective motion features to the greatest extent possible, avoiding feature distortion or incomplete noise removal caused by a single filtering method. For abnormal jumps, linear interpolation, cubic spline interpolation, or mean substitution schemes are selected based on length, combined with a post-completion verification mechanism to ensure the accuracy and continuity of abnormal data repair. This addresses the shortcomings of traditional anomaly processing methods, such as their crudeness and the disconnect between completed data and the original trend. A normalization interval is set based on the historical effective range of motion data, making the standardized data more closely match the actual distribution of upper limb movement characteristics, improving data consistency and comparability, providing high-quality data support for subsequent joint angle calculations, and reducing the negative impact of poor data quality on deep learning model training. This indirectly improves the sufficiency of model feature extraction and the accuracy of evaluation results. Furthermore, the entire process is automated without manual intervention, further improving the overall efficiency of rehabilitation assessment.

[0027] Specifically, when calculating joint range of motion based on human upper limb kinematics modeling, the following are included: The degrees of freedom parameters of the shoulder, elbow, and wrist joints and the relationship of the length of the connecting rods between the joints were determined based on the kinematic model of the human upper limb multi-joint joint. Using Euler angle data from standardized data as input, and combining it with forward kinematics algorithm, the flexion-extension angle, adduction-abduction angle and rotation angle of each joint are calculated sequentially; Set a reasonable range threshold for the range of motion angles of each joint, and determine the angle values ​​that exceed the reasonable range as invalid values. Use a smooth transition algorithm for the valid angles at adjacent time points to correct invalid values. The threshold for the reasonable range of joint motion angles is determined based on statistical data of upper limb joint mobility in healthy individuals, and is adjusted individually according to the patient's age and disease course.

[0028] Specifically, when correcting invalid values ​​in the solution results, the following is included: A first correction weight and a second correction weight are preset, with the first correction weight being greater than the second correction weight; When the number of consecutive occurrences of invalid values ​​is less than the preset threshold, the weighted average of the valid angles at the previous moment and the valid angles at the next moment is used for correction. The weight of the angle at the previous moment is the first correction weight, and the weight of the angle at the next moment is the second correction weight. When the number of consecutive occurrences of invalid values ​​is greater than or equal to the preset threshold, the standard joint angle change curve corresponding to the action is invoked based on the type of rehabilitation action performed by the patient, and segmented correction is performed based on the trend of the curve. After correction, the smoothness of the joint angle change is calculated. If the smoothness is less than the preset smoothness threshold, the weight coefficient or standard curve matching degree is adjusted, and correction is performed again.

[0029] Specifically, when calculating joint range of motion based on human upper limb kinematic modeling, a multi-joint kinematic model of the human upper limb is first constructed to clarify the degrees of freedom parameters of the shoulder, elbow, and wrist joints and the relationship of the lengths of the inter-joint links: the shoulder joint is set to 3 degrees of freedom (flexion-extension, adduction-abduction, rotation), the elbow joint is set to 1 degree of freedom (flexion-extension), and the wrist joint is set to 3 degrees of freedom (flexion-extension, adduction-abduction, rotation); the lengths of the inter-joint links are determined based on human anatomical statistics. The upper arm link length is taken as the average upper arm length of healthy individuals, 40±5cm; the forearm link length is taken as 30±4cm; and the wrist link length is taken as 15±3cm. The link lengths can be adjusted proportionally according to the patient's actual height. Using the triaxial Euler angle data from the standardized data as input, the forward kinematic model is constructed using the DH parameter method to calculate the key angles of each joint in sequence: shoulder joint flexion-extension angle, adduction-abduction angle, and rotation angle; elbow joint flexion-extension angle; and wrist joint flexion-extension angle, adduction-abduction angle, and rotation angle. Reasonable range thresholds for the range of motion of each joint were set. Specifically, the reasonable range for shoulder joint flexion-extension angle was -10° to 180°, adduction-abduction angle was -45° to 135°, and rotation angle was -90° to 90°; the reasonable range for elbow joint flexion-extension angle was 0° to 150°; and the reasonable range for wrist joint flexion-extension angle was -60° to 80°, adduction-abduction angle was -30° to 30°, and rotation angle was -45° to 45°. These reasonable range thresholds were determined based on statistical data of upper limb joint range of motion from 1000 healthy individuals aged 18-60 years. For patients over 60 years of age, the reasonable range thresholds for each joint were reduced by 10% to 15%; for stroke patients with a disease duration of more than one year, the reasonable range thresholds for each joint were further adjusted by 5% to 10% based on their previous rehabilitation assessment records to achieve personalized adaptation. Angle values ​​exceeding the above reasonable ranges in the calculation results were judged as invalid values ​​and corrected using a smooth transition algorithm of effective angles at adjacent time points.

[0030] Specifically, when correcting invalid values ​​in the solution results, a first correction weight of 0.7 and a second correction weight of 0.3 are preset, with a preset threshold of 3 occurrences. When the number of consecutive invalid values ​​is less than 3, a weighted average of the effective angles at the previous and subsequent moments is used for correction, calculated as: Corrected angle = Effective angle at the previous moment × 0.7 + Effective angle at the subsequent moment × 0.3. When the number of consecutive invalid values ​​is greater than or equal to 3, the standard joint angle change curve for the corresponding movement is retrieved based on the type of rehabilitation movement performed by the patient (e.g., pronation / supination, upper arm abduction, touching the ear, touching the lumbar spine, etc.). The standard curve is fitted based on joint angle data collected from 100 healthy individuals performing the same movement. Segmented correction is performed according to the movement execution stage (starting stage, execution stage, completion stage), with each stage matching the trend characteristics of the corresponding interval of the standard curve. After correction, the smoothness of the joint angle change is calculated. The smoothness is represented by the standard deviation of the angle change rate between two adjacent time points. The preset smoothness threshold is 5° / frame. If the calculated smoothness is less than 5° / frame, it indicates that there is a sudden change in the angle. At this time, the first correction weight is adjusted to 0.6, the second correction weight is adjusted to 0.4, or the standard curve matching threshold is increased to 85%, and the correction operation is repeated until the smoothness meets the preset requirements.

[0031] The above embodiments clarify the relationship between the degrees of freedom parameters of the shoulder, elbow, and wrist joints and the length of the inter-joint links by constructing a multi-joint kinematic model of the human upper limb. Combined with forward kinematics algorithms, the flexion-extension, adduction-abduction, and rotation angles of each joint are accurately calculated. Based on statistical data of upper limb joint range of motion in healthy individuals, the threshold values ​​of reasonable joint range of motion are adjusted according to the patient's age and disease duration, effectively reducing the probability of invalid values. Simultaneously, the number of consecutive occurrences of invalid values ​​is graded using a weighted average of the effective angles at the previous and subsequent moments (with the first correction weight greater than the second correction weight) or the corresponding standard joint for rehabilitation movements. The strategy of segmented correction of angle change curves, combined with the correction smoothness verification and weight coefficient and standard curve matching adjustment mechanism, not only solves the problems of IMU raw data being easily affected by wearing position and insufficient validity and continuity of joint angle data in traditional rehabilitation assessment, but also avoids the error accumulation caused by a single correction method. It provides high-quality and reliable feature data for subsequent deep learning model training, helps to improve the accuracy of motor function assessment and model generalization ability, and supports quantitative and automated assessment processes, reduces reliance on professional physician experience, and further ensures the objectivity and consistency of assessment results.

[0032] Specifically, when dividing data segments according to preset window sizes and step sizes, the following is included: Pre-set the first window specification, the second window specification, the first step length, and the second step length. The window length corresponding to the first window specification is greater than the second window specification, and the first step length is greater than the second step length. The data segmentation method is determined based on the duration of the rehabilitation movements: When the duration of a rehabilitation movement exceeds a preset duration threshold, the first window size and the first step length are used for division. After division, each data segment contains a complete movement feature cycle. When the duration of the rehabilitation action is less than or equal to the preset duration threshold, the second window size and the second step size are used for division, and the window overlap rate is increased to the preset overlap ratio. If the length of the last data segment is less than half of the corresponding window size during the partitioning process, it will be merged with the previous data segment. If it is greater than or equal to half, it will be padded with zeros to complete the corresponding window size.

[0033] Specifically, adjusting window size and step size includes: A first feature complexity threshold and a second feature complexity threshold are preset, wherein the first feature complexity threshold is greater than the second feature complexity threshold; The feature complexity of the concatenated standardized data and joint angle data is calculated by combining the information entropy and variance of the data. When the feature complexity exceeds the first feature complexity threshold, the window size is adjusted to 1.2 times the corresponding size, while the step size remains unchanged. When the feature complexity is less than or equal to the first feature complexity threshold and greater than the second feature complexity threshold, the original window size and step size are maintained. When the feature complexity is less than or equal to the second feature complexity threshold, adjust the window size to 0.8 times the corresponding size and the step size to 1.2 times the corresponding step size.

[0034] Specifically, when dividing data segments according to preset window specifications and step lengths, the first window specification is preset to 50 data points, the second window specification to 30 data points, the first step length to 25 data points, and the second step length to 15 data points. The window length of the first window specification and the first step length are greater than the second window specification and the second step length, respectively. The preset duration threshold is 8 seconds (based on a sampling frequency of 10Hz, corresponding to 80 data points, matching the temporal characteristics of rehabilitation action acquisition), and the preset overlap ratio is 80%. The data segmentation method is strictly determined based on the duration of the rehabilitation movements: when the duration of rehabilitation movements such as pronation, supination, and upper arm abduction is greater than 8 seconds, the first window size (50 data points) and the first step size (25 data points) are used for segmentation to ensure that each data segment contains the complete movement initiation-execution-completion characteristic cycle; when the duration of rehabilitation movements is less than or equal to 8 seconds, the second window size (30 data points) and the second step size (15 data points) are used for segmentation, and the window overlap rate is increased to a preset 80% to fully preserve the detailed features of short-duration movements. During the partitioning process, the length of the last data segment needs to be verified: If the first window specification is used for partitioning, and the length of the last data segment is less than 25 data points (half of the corresponding window specification), then it is merged with the previous data segment into a complete segment; if the length is greater than or equal to 25 data points, then it is padded with zeros to reach the first window specification of 50 data points; if the second window specification is used for partitioning, if the length of the last data segment is less than 15 data points (half of the corresponding window specification), then it is merged with the previous segment; if it is greater than or equal to 15 data points, then it is padded with zeros to reach the second window specification of 30 data points.

[0035] Specifically, when adjusting the window size and step size, the first feature complexity threshold is preset to 0.8 and the second feature complexity threshold is preset to 0.4, with the first feature complexity threshold being greater than the second feature complexity threshold. First, the feature complexity after concatenating the standardized data and joint angle data is calculated. The feature complexity is obtained by comprehensively calculating the information entropy and variance of the data: the information entropy is calculated using the Shannon entropy formula, with a value range of 0-3 (based on the dimensional characteristics of the nine-axis sensor data and joint angle data); the variance is calculated by normalizing the overall variance of the concatenated data matrix to the 0-1 range. Finally, the comprehensive feature complexity = 0.5 × information entropy + 0.5 × normalized variance (both indicators are weighted at 0.5 to balance the uncertainty and dispersion of the data). Adjust the window size and step size based on the calculation results of the comprehensive feature complexity: When the feature complexity is greater than 0.8, adjust the currently selected window size to 1.2 times the corresponding size (if the first window size was used, adjust it to 60 data points; if the second window size was used, adjust it to 36 data points), and keep the step size unchanged (still 25 or 15 data points); when the feature complexity is less than or equal to 0.8 and greater than 0.4, maintain the original window size and step size; when the feature complexity is less than or equal to 0.4, adjust the window size to 0.8 times the corresponding size (the first window size is adjusted to 40 data points, and the second window size is adjusted to 24 data points), and at the same time adjust the step size to 1.2 times the corresponding step size (the first step size is adjusted to 30 data points, and the second step size is adjusted to 18 data points), to ensure that the data segmentation is adapted to the feature complexity.

[0036] The above embodiments, by pre-setting hierarchical window specifications (first window specification, second window specification) and step lengths (first step length, second step length), flexibly select the division method according to the duration of rehabilitation actions. For long-duration actions, the complete feature cycle is retained, and for short-duration actions, the window overlap rate is increased to ensure that key features of actions of different durations are not lost. At the same time, based on the feature complexity calculated by combining data information entropy and variance, the window specifications and step length are dynamically adjusted to adapt to the data requirements of different feature richness, avoiding feature redundancy or insufficient extraction caused by fixed parameters. Furthermore, the data integrity and consistency are ensured by merging or zero-padding mechanisms of the last data segment, providing high-quality and highly adaptable input data for subsequent deep learning network feature extraction. This effectively solves the defects of traditional data division lacking specificity and poor adaptability. It not only improves the model's adaptability to different action scenarios and data with different feature complexities, but also lays a solid foundation for the generalization ability and assessment accuracy of the motor function level prediction model. At the same time, the automated division and adjustment process reduces manual intervention and further improves the overall efficiency of rehabilitation assessment.

[0037] Specifically, when using deep learning networks with fused attention mechanisms for feature extraction and model training, the following are included: A deep learning network that incorporates an attention mechanism is called a BiGRU-Attention network. The BiGRU-Attention network includes an input layer, a BiGRU layer, an attention layer, a fully connected layer, and an output layer. The BiGRU layer includes forward GRU units and backward GRU units, which are used to capture the temporal dependencies of data segments. The number of hidden layer nodes in the BiGRU layer is determined according to the dimension of the data segment. The attention layer assigns weights to the feature vectors output by the BiGRU layer and identifies features with weight values ​​greater than a preset weight threshold as key features. The optimization algorithm selected is an intelligent optimization algorithm, which iteratively optimizes the weight parameters and bias parameters of the network by minimizing the loss function between the model's predicted value and the standard value of the rehabilitation assessment scale. During training, a cross-validation strategy is adopted, dividing the dataset into a training set, a validation set, and a test set. When the model accuracy on the validation set does not improve for a preset number of consecutive iterations, training is stopped, and the current best model is saved as the motor function level prediction model.

[0038] Specifically, when using intelligent optimization algorithms to iteratively optimize model parameters, the following are included: A first iteration threshold and a second iteration threshold are preset, wherein the first iteration threshold is greater than the second iteration threshold; The optimization objectives of the intelligent optimization algorithm include two indicators: model accuracy and loss function value. Weight coefficients are set for the two indicators, and the comprehensive optimization objective value is calculated by weighting them. When the overall optimization target value is greater than the preset optimization target threshold, the iteration step size is adjusted to 0.8 times the initial step size, and the iteration continues until the first iteration number threshold is reached; When the comprehensive optimization target value is less than or equal to the preset optimization target threshold and greater than the preset minimum target threshold, maintain the initial iteration step size and iterate until the second iteration number threshold. When the overall optimization target value is less than or equal to the preset minimum target threshold, adjust the initial parameters of the optimization algorithm and restart the iterative optimization process; The weighting coefficients are determined based on the degree of influence of the two indicators on the model performance, and the sum of the weighting coefficients is 1.

[0039] Specifically, when using a deep learning network with an attention fusion mechanism for feature extraction and model training, the deep learning network with the attention fusion mechanism is clearly defined as a BiGRU-Attention network. This network consists of an input layer, a BiGRU layer, an attention layer, a fully connected layer, and an output layer connected in series. The input layer dimension is determined based on the features of the concatenated data. Since multi-dimensional motion data includes three-axis acceleration, three-axis angular velocity, and three-axis Euler angles (a total of 9 dimensions), and joint motion angles include flexion-extension, adduction-abduction, and rotation angles of the shoulder, elbow, and wrist joints (a total of 9 dimensions), the concatenated single data point has an 18-dimensional dimension. Combined with a preset window size of 50 data points, the input layer ultimately receives a 50×18-dimensional matrix of data. The BiGRU layer comprises one set of forward GRU units and one set of backward GRU units. Each set has 128 hidden layer nodes, determined based on the dimensionality and feature complexity of the data segment. The forward GRU unit captures temporal dependencies from the beginning of the data sequence backward, while the backward GRU unit captures them from the end of the data sequence forward. The output feature vectors of both are concatenated to form a 256-dimensional fused feature vector. The attention layer employs an additive attention mechanism to assign weights to the 256-dimensional feature vector output by the BiGRU layer. A preset weight threshold of 0.6 is used; features with weights greater than 0.6 are considered key features, and the corresponding vector dimensions are retained for subsequent calculations. The optimization algorithm uses the Frostbite algorithm, employing the cross-entropy loss function. By minimizing the cross-entropy loss between the model's predicted motor function level and the Fugl-Meyer scale standard values ​​(0, 1, and 2 points), the network's weight and bias parameters are iteratively optimized. During training, a 5-fold cross-validation strategy is adopted, dividing the dataset into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively. The preset threshold for the number of consecutive iterations is 10. When the model accuracy on the validation set does not improve for 10 consecutive iterations, training is stopped, and the network parameters of the current iteration are saved as the motion function level prediction model. The output layer uses the Softmax activation function to output three probability values ​​corresponding to three function levels.

[0040] Specifically, when using the intelligent optimization algorithm (Frost Optimization Algorithm) to iteratively optimize model parameters, the threshold for the first iteration is pre-set to 200 times and the threshold for the second iteration is set to 100 times, with the first iteration threshold being greater than the second iteration threshold. The initial step size of the intelligent optimization algorithm is set to 0.01. The optimization objectives include two indicators: model accuracy and loss function value. The weight coefficient for model accuracy is set to 0.6, and the weight coefficient for loss function value is set to 0.4. The sum of the two weight coefficients is 1, and the weight is determined based on the degree of influence of the two indicators on model performance (accuracy directly reflects the evaluation effect and has a higher weight). The comprehensive optimization objective value is obtained through weighted calculation, and the specific calculation formula is: Comprehensive optimization objective value = 0.6 × model accuracy + 0.4 × (1 - normalized loss function value), where the normalized loss function value is obtained by mapping the cross-entropy loss function result to the [0,1] interval to ensure that the two indicators participate in the calculation at the same order of magnitude. When the overall optimization target value is greater than the preset optimization target threshold of 0.95, the iteration step size is adjusted to 0.8 times the initial step size (i.e., 0.008), and iteration continues until the first iteration threshold of 200 times is reached; when the overall optimization target value is less than or equal to 0.95 and greater than the preset minimum target threshold of 0.85, the initial iteration step size is maintained at 0.01, and iteration continues until the second iteration threshold of 100 times is reached; when the overall optimization target value is less than or equal to 0.85, the initial parameters of the frost optimization algorithm are adjusted, specifically including adjusting the population size from 30 to 40, expanding the search range by 20%, and restarting the iterative optimization process to ensure that the algorithm can fully search the optimal parameter space.

[0041] The above embodiments accurately capture the temporal dependencies of motion data through the forward and backward GRU units of the BiGRU-Attention network. Combined with the attention layer's weight allocation of feature vectors and key feature selection (weight values ​​greater than a preset threshold), it completely eliminates the reliance on manual feature extraction in traditional machine learning, solving the defects of insufficient feature extraction and ignoring temporal differences in data. The Frostbite optimization algorithm is selected with model accuracy and loss function value as dual optimization objectives (the sum of weight coefficients is 1). By dynamically adjusting the iteration step size, number of iterations, and initial algorithm parameters, it enhances global search capabilities, accelerates convergence speed, and ensures targeted parameter optimization. To ensure both accuracy and effectiveness, and to avoid performance imbalances caused by a single optimization objective, the model employs a cross-validation strategy and an early stopping mechanism based on validation set accuracy. This effectively prevents overfitting and improves the model's generalization ability, ultimately achieving a test set recognition rate of 99.23%. This represents a significant improvement over models trained solely using raw IMU data. It addresses the limitations in assessment accuracy and poor generalization ability in traditional rehabilitation assessments, while also enabling efficient training and optimal parameter configuration for motor function level prediction models. This provides core technical support for the objective, accurate, and automated assessment of upper limb motor function in stroke patients, further reducing reliance on the experience of professional physicians and improving assessment efficiency and consistency.

[0042] Specifically, when generating upper limb motor function assessment results, the following are included: A first confidence threshold and a second confidence threshold are preset, with the first confidence threshold being greater than the second confidence threshold; at the same time, clinical information weight and action completion weight are set, with the sum of the clinical information weight and the action completion weight being 1; The evaluation results include confidence assessment, result correction, and verification. Obtain the preliminary assessment results and corresponding confidence values ​​output by the prediction model. When the confidence value is greater than or equal to the first confidence threshold, the preliminary assessment results are deemed reliable, and the process proceeds directly to the result verification stage. When the confidence value is greater than or equal to the second confidence threshold and less than the first confidence threshold, the patient's clinical information is retrieved, and the preliminary assessment result is weighted and corrected in combination with the weight of the clinical information. When the confidence value is less than the second confidence threshold, it is determined that the current data is insufficient to support a reliable assessment. A supplementary data collection prompt is output, requiring the patient to perform a preset number of rehabilitation actions to supplement the data. After repeating the data collection and data segmentation process, the data is input into the model again to generate preliminary assessment results. After the results are corrected, the action completion score is calculated. The action completion score is derived from the matching degree between the joint angle and the standard action angle. The corrected results are then further optimized by combining the action completion weight. The result verification process employs a dual verification mechanism: Compare the assessment results with the grade classification range of the rehabilitation assessment scale; The system retrieves historical assessment data of similar patients, calculates the similarity of the results, and confirms the validity of the assessment results when the similarity is greater than a preset similarity threshold. If the similarity is less than or equal to the preset similarity threshold, the weights of clinical information and action completion are readjusted, and the correction process is repeated until the result similarity meets the standard or the maximum number of corrections is reached. Among them, the types of rehabilitation movements collected in the supplementary collection are the same as those collected in the initial collection, and the number does not exceed 50% of the number of movements collected in the initial collection.

[0043] Specifically, when generating upper limb motor function assessment results, a first confidence threshold of 0.9 and a second confidence threshold of 0.7 are pre-set, with the first confidence threshold being greater than the second. Simultaneously, a clinical information weight of 0.3 and a movement completion weight of 0.7 are set, with the sum of these two weights being 1. The clinical information includes key information such as the patient's disease course, age, previous rehabilitation assessment records, and whether other motor function disorders are present. The assessment result generation process sequentially performs confidence assessment, result correction, and verification: First, the preliminary assessment results output by the BiGRU-Attention prediction model (corresponding to three functional levels of the Fugl-Meyer scale: 0, 1, and 2 points) and their corresponding confidence values ​​(ranging from 0 to 1) are obtained. When the confidence value is greater than or equal to 0.9, the preliminary assessment result is considered reliable and proceeds directly to the result verification stage. When the confidence value is greater than or equal to 0.7 but less than 0.9, the aforementioned clinical information of the patient is retrieved, and the preliminary assessment result is weighted and corrected according to a clinical information weight of 0.3. The clinical information matching score is converted to 0-2 points based on the degree of fit between the patient information and the typical characteristics of the corresponding functional level. The correction formula is: Corrected result = Preliminary assessment result × 0.7 + Clinical information matching score × 0.3. When the confidence value is less than 0.7, it is determined that the current data is insufficient to support a reliable assessment. A supplementary data collection prompt is output, requiring the patient to perform 3 rehabilitation actions again (consistent with the type of the initial 6 collected actions, and the number does not exceed 50% of the initial number) to supplement the data. After repeating the data collection and data division process, the data is input into the model again to generate the preliminary assessment result. After the results are corrected, the action completion score is calculated. This score is calculated by the cosine similarity between the patient's actual joint angle and the standard action angle. A similarity of ≥0.95 corresponds to 100 points, 0.8-0.95 corresponds to 80 points, 0.6-0.8 corresponds to 60 points, and <0.6 corresponds to 40 points. Then, the corrected results are further optimized by combining the action completion weight of 0.7. The optimization formula is: Secondary optimization result = Corrected result × 0.3 + (Action completion score / 100 × 2) × 0.7. The results verification process employs a dual verification mechanism: first, it checks whether the assessment results fall within the 0-2 level range of the Fugl-Meyer scale; then, it retrieves historical assessment data (sample size no less than 50 cases) of patients with similar age ± 5 years, disease duration ± 3 months, and consistent action completion type, and calculates the similarity of the results using the Pearson correlation coefficient. The preset similarity threshold is 0.85. When the similarity is greater than 0.85, the assessment results are confirmed as valid; if the similarity is less than or equal to 0.85, the weights of clinical information (adjustment range 0.2-0.4) and action completion (corresponding adjustment to 0.8-0.6) are readjusted, and the correction process is repeated until the result similarity reaches the standard or the maximum number of corrections (3 times) is reached.

[0044] The above embodiments, by pre-setting graded confidence thresholds (the first confidence threshold is greater than the second confidence threshold), employ differentiated processing for preliminary assessment results of different confidence levels, including direct verification, clinical information weighted correction (the sum of clinical information weight and movement completion weight is 1), or supplementary data collection (the type of rehabilitation movement is consistent with the initial one, and the number does not exceed 50% of the initial one). This not only solves the problem of unreliable results due to insufficient data support in traditional assessments, but also achieves secondary optimization through movement completion scoring (based on the matching degree between joint angles and standard movement angles), improving the accuracy of correction. Combined with a dual verification mechanism of "comparison between assessment results and rehabilitation scale level range + similarity verification of historical similar patient data," along with dynamic weight adjustment and maximum correction limit control, the subjectivity and individual differences of traditional assessments are effectively avoided, ensuring the consistency and effectiveness of results. At the same time, the automated grading, correction, and verification process reduces the intervention costs of professional physicians, further enhancing the objectivity, accuracy, and reliability of upper limb motor function assessment in stroke patients, and providing a more credible decision-making basis for rehabilitation diagnosis and treatment.

[0045] In other embodiments of this application, the number of sensors can be adjusted according to actual assessment needs. This can involve adding lower limb sensors to expand the assessment range, retaining only the sensor on the affected side for unilateral assessment, changing the sensor wearing position to select a location that is more comfortable for the patient and has less interference from muscle movement during exercise; different deep learning network structures (such as LSTM, Transformer) and optimization algorithms (such as genetic algorithms, gray wolf optimization algorithms) can be replaced to adapt to the model training needs in different scenarios; the size and step size of the sliding window can be appropriately adjusted to select more suitable parameter configurations, and different normalization methods such as Z-Score normalization can be selected to optimize the data preprocessing effect; the assessment range can also be expanded to simultaneous rehabilitation assessment of the upper and lower limbs or separate lower limb rehabilitation assessment, while also adding motor function assessment of other joints such as elbow and hand joints.

[0046] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0047] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0048] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0049] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0050] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for intelligent assessment of upper limb motor function based on kinematics and deep learning for stroke rehabilitation evaluation, characterized in that, include: Construct a wearable inertial sensing system to collect multi-dimensional motion data of patients when performing rehabilitation movements; The collected motion data is preprocessed, including length normalization, removal of anomalous jumps, and numerical normalization, to obtain standardized data. Based on human upper limb kinematic modeling, the joint movement angles are calculated from standardized data; Multidimensional motion data and joint range of motion data are stitched together, and the data segments are divided according to preset window specifications and stride length; A deep learning network with an attention fusion mechanism is used to extract features from data segments. The model parameters are then iteratively optimized using an optimization algorithm to train a motor function level prediction model based on a rehabilitation assessment scale. Preprocessing, joint angle calculation, and data segmentation are performed on the multidimensional motion data of newly assessed patients. The processed data is then input into the trained prediction model to generate upper limb motor function assessment results.

2. The intelligent assessment method for upper limb motor function based on kinematics and deep learning for stroke rehabilitation assessment according to claim 1, characterized in that, Preprocessing the collected motion data includes: A first noise threshold and a second noise threshold are preset, wherein the first noise threshold is greater than the second noise threshold; An adaptive filtering algorithm is used to denoise the collected multi-dimensional motion data. When the noise amplitude in the data is greater than the first noise threshold, a multi-level filtering strategy is used to perform time-domain filtering and frequency-domain filtering in sequence. When the noise amplitude is less than or equal to the first noise threshold and greater than the second noise threshold, only time-domain filtering is performed; When the noise amplitude is less than or equal to the second noise threshold, no additional noise reduction is required; When removing abnormal jumps, the rate of change of adjacent data points is calculated, and data segments with a rate of change greater than a preset jump threshold are identified as abnormal jumps. An interpolation algorithm is then used to complete the data in the abnormal jump segments. Numerical normalization uses a preset normalization interval to map the denoised and completed data to the preset normalization interval, thus obtaining standardized data. The upper and lower limits of the normalization interval are determined based on the historical valid range of the motion data.

3. The intelligent assessment method for upper limb motor function based on kinematics and deep learning for stroke rehabilitation assessment as described in claim 2, characterized in that, When using interpolation algorithms to complete data for anomalous transition segments, the following steps are included: A first transition length threshold and a second transition length threshold are preset, wherein the first transition length threshold is greater than the second transition length threshold; Interpolation algorithms include two types: linear interpolation and cubic spline interpolation. When the number of data points in the abnormal jump segment exceeds the first jump length threshold, a cubic spline interpolation algorithm is used to complete the data points. During the completion process, the trend of a preset number of valid data points before and after the jump segment is referenced. When the number of data points in an abnormal jump segment is greater than the second jump length threshold but less than or equal to the first jump length threshold, a linear interpolation algorithm is used for completion. When the number of data points in the abnormal jump segment is less than or equal to the second jump length threshold, the abnormal data points are directly removed and replaced by the average of the adjacent valid data points. After completion, the data change rate needs to be checked again. If there are still cases where the change rate is greater than the preset threshold, the completion operation is repeated until the data is completed.

4. The intelligent assessment method for upper limb motor function based on kinematics and deep learning for stroke rehabilitation assessment as described in claim 1, characterized in that, When calculating joint range of motion based on human upper limb kinematics modeling, the following are included: The degrees of freedom parameters of the shoulder, elbow, and wrist joints and the relationship of the length of the connecting rods between the joints are determined based on the aforementioned multi-joint kinematic model of the human upper limb. Using Euler angle data from standardized data as input, and combining it with forward kinematics algorithm, the flexion-extension angle, adduction-abduction angle and rotation angle of each joint are calculated sequentially; Set a reasonable range threshold for the range of motion angles of each joint, and determine the angle values ​​that exceed the reasonable range as invalid values. Use a smooth transition algorithm for the valid angles at adjacent time points to correct invalid values. The threshold for the reasonable range of joint motion angles is determined based on statistical data of upper limb joint mobility in healthy individuals, and is adjusted individually according to the patient's age and disease course.

5. The intelligent assessment method for upper limb motor function based on kinematics and deep learning for stroke rehabilitation assessment according to claim 4, characterized in that, When correcting invalid values ​​in the solution results, the following is included: A first correction weight and a second correction weight are preset, wherein the first correction weight is greater than the second correction weight; When the number of consecutive occurrences of invalid values ​​is less than the preset threshold, the weighted average of the valid angles at the previous moment and the valid angles at the next moment is used for correction. The weight of the angle at the previous moment is the first correction weight, and the weight of the angle at the next moment is the second correction weight. When the number of consecutive occurrences of invalid values ​​is greater than or equal to the preset threshold, the standard joint angle change curve corresponding to the action is invoked based on the type of rehabilitation action performed by the patient, and segmented correction is performed based on the trend of the curve. After correction, the smoothness of the joint angle change is calculated. If the smoothness is less than the preset smoothness threshold, the weight coefficient or standard curve matching degree is adjusted, and correction is performed again.

6. The intelligent assessment method for upper limb motor function based on kinematics and deep learning for stroke rehabilitation assessment according to claim 1, characterized in that, When dividing data segments according to preset window specifications and step sizes, the following is included: Pre-set the first window specification, the second window specification, the first step length, and the second step length. The window length corresponding to the first window specification is greater than the second window specification, and the first step length is greater than the second step length. The data segmentation method is determined based on the duration of the rehabilitation movements: When the duration of a rehabilitation movement exceeds a preset duration threshold, the first window size and the first step length are used for division. After division, each data segment contains a complete movement feature cycle. When the duration of the rehabilitation action is less than or equal to the preset duration threshold, the second window size and the second step size are used for division, and the window overlap rate is increased to the preset overlap ratio. If the length of the last data segment is less than half of the corresponding window size during the partitioning process, it will be merged with the previous data segment. If it is greater than or equal to half, it will be padded with zeros to complete the corresponding window size.

7. The intelligent assessment method for upper limb motor function based on kinematics and deep learning for stroke rehabilitation assessment according to claim 6, characterized in that, When adjusting window size and step size, the following applies: A first feature complexity threshold and a second feature complexity threshold are preset, wherein the first feature complexity threshold is greater than the second feature complexity threshold; The feature complexity of the concatenated standardized data and joint angle data is calculated by combining the information entropy and variance of the data. When the feature complexity exceeds the first feature complexity threshold, the window size is adjusted to 1.2 times the corresponding size, while the step size remains unchanged. When the feature complexity is less than or equal to the first feature complexity threshold and greater than the second feature complexity threshold, the original window size and step size are maintained. When the feature complexity is less than or equal to the second feature complexity threshold, adjust the window size to 0.8 times the corresponding size and the step size to 1.2 times the corresponding step size.

8. The intelligent assessment method for upper limb motor function based on kinematics and deep learning for stroke rehabilitation assessment according to claim 1, characterized in that, When using deep learning networks with fused attention mechanisms for feature extraction and model training, the following are included: The deep learning network that integrates the attention mechanism is called the BiGRU-Attention network, which includes an input layer, a BiGRU layer, an attention layer, a fully connected layer, and an output layer. The BiGRU layer includes forward GRU units and backward GRU units, which are used to capture the temporal dependencies of data segments. The number of hidden layer nodes in the BiGRU layer is determined according to the dimension of the data segment. The attention layer assigns weights to the feature vectors output by the BiGRU layer and determines features with weight values ​​greater than a preset weight threshold as key features. The optimization algorithm selected is an intelligent optimization algorithm, which iteratively optimizes the weight parameters and bias parameters of the network by minimizing the loss function between the model's predicted value and the standard value of the rehabilitation assessment scale. During training, a cross-validation strategy is adopted, dividing the dataset into a training set, a validation set, and a test set. When the model accuracy on the validation set does not improve for a preset number of consecutive iterations, training is stopped, and the current best model is saved as the motor function level prediction model.

9. The intelligent assessment method for upper limb motor function based on kinematics and deep learning for stroke rehabilitation assessment according to claim 8, characterized in that, When iteratively optimizing model parameters using intelligent optimization algorithms, the following are included: A first iteration number threshold and a second iteration number threshold are preset, wherein the first iteration number threshold is greater than the second iteration number threshold; The optimization objectives of the intelligent optimization algorithm include two indicators: model accuracy and loss function value. Weight coefficients are set for the two indicators, and the comprehensive optimization objective value is calculated by weighting them. When the overall optimization target value is greater than the preset optimization target threshold, the iteration step size is adjusted to 0.8 times the initial step size, and the iteration continues until the first iteration number threshold is reached; When the comprehensive optimization target value is less than or equal to the preset optimization target threshold and greater than the preset minimum target threshold, maintain the initial iteration step size and iterate until the second iteration number threshold. When the overall optimization target value is less than or equal to the preset minimum target threshold, adjust the initial parameters of the optimization algorithm and restart the iterative optimization process; The weighting coefficients are determined based on the degree of influence of the two indicators on the model performance, and the sum of the weighting coefficients is 1.

10. The intelligent assessment method for upper limb motor function based on kinematics and deep learning for stroke rehabilitation assessment according to claim 1, characterized in that, When generating upper limb motor function assessment results, the following should be included: A first confidence threshold and a second confidence threshold are preset, wherein the first confidence threshold is greater than the second confidence threshold; at the same time, clinical information weight and action completion weight are set, wherein the sum of the clinical information weight and the action completion weight is 1; The evaluation results include confidence assessment, result correction, and verification. Obtain the preliminary assessment results and corresponding confidence values ​​output by the prediction model. When the confidence value is greater than or equal to the first confidence threshold, the preliminary assessment results are deemed reliable, and the process proceeds directly to the result verification stage. When the confidence value is greater than or equal to the second confidence threshold and less than the first confidence threshold, the patient's clinical information is retrieved, and the preliminary assessment result is weighted and corrected in combination with the weight of the clinical information. When the confidence value is less than the second confidence threshold, it is determined that the current data is insufficient to support a reliable assessment. A supplementary data collection prompt is output, requiring the patient to perform a preset number of rehabilitation actions to supplement the data. After repeating the data collection and data segmentation process, the data is input into the model again to generate preliminary assessment results. After the results are corrected, the action completion score is calculated. The action completion score is derived from the matching degree between the joint angle and the standard action angle. The corrected results are then further optimized by combining the action completion weight. The result verification process employs a dual verification mechanism: Compare the assessment results with the grade classification range of the rehabilitation assessment scale; The system retrieves historical assessment data of similar patients, calculates the similarity of the results, and confirms the validity of the assessment results when the similarity is greater than a preset similarity threshold. If the similarity is less than or equal to the preset similarity threshold, the weights of clinical information and action completion are readjusted, and the correction process is repeated until the result similarity meets the standard or the maximum number of corrections is reached. Among them, the types of rehabilitation movements collected in the supplementary collection are the same as those collected in the initial collection, and the number does not exceed 50% of the number of movements collected in the initial collection.

Images