A stroke training method and training system based on deep learning

Abstract

The present application belongs to the technical field of medical rehabilitation information processing, and discloses a stroke training method and a training system based on deep learning, which comprises the following steps: acquiring multi-modal physiological and motion data in a training process, extracting motion coordination features representing multi-joint linkage relationship; calculating the coordination deviation degree between the motion coordination features and a target coordination reference model, and analyzing and acquiring deviation component information representing abnormal motion sources; inputting a multi-dimensional performance index sequence and a time period change sequence of the coordination deviation degree into a deep learning time sequence analysis model, outputting a dynamic performance mode and degree values of each influencing factor; and generating a multi-dimensional motion quality evaluation report according to the above results. The motion quality is evaluated from the multi-joint coordination mode level, the compensatory false compliance state can be identified and the abnormal joint source can be located, and the dynamic change of the performance in the training process can be subjected to time sequence mode recognition and multi-factor attribution analysis.

Description

Technical Field

[0001] This invention relates to the field of medical rehabilitation information processing technology, and more specifically, to a deep learning-based stroke training method and system. Background Technology

[0002] Stroke is a cerebrovascular disease with a high incidence and disability rate. Most survivors suffer from varying degrees of motor dysfunction, requiring long-term rehabilitation training. Accurate assessment of movement quality during rehabilitation training is crucial for developing and adjusting training programs. In recent years, deep learning technology has been introduced into the field of stroke rehabilitation training assessment. Its basic idea is to collect joint movement data during training using motion sensing devices, and then use a deep learning model to compare the time series of collected joint angles or distal positions with preset standard movement templates, using trajectory similarity as the scoring criterion for movement quality.

[0003] However, stroke patients often exhibit compensatory movements when performing training exercises due to decreased control of target joints caused by nerve damage. This means they unconsciously engage non-target joints or the trunk to complete the task. Compensatory movements can bring the distal end to the target position, making the trajectory resemble a standard movement and thus achieving a higher trajectory score. However, the actual source of the movement is abnormal. Existing trajectory-based assessment methods only extract motion data from the target joints, discarding motion information from non-target joints and the trunk. Therefore, they cannot perceive whether the coordinated movement patterns between multiple joints are normal, nor can they distinguish between genuine motor ability performance and compensatory false achievements, leading to inaccurate assessment results. Furthermore, during a complete rehabilitation training session, motor performance is not constant but dynamically changes due to various factors such as fatigue, changes in compensatory strategies, and neurological abnormalities. Current technologies typically perform simple mean aggregation or linear trend analysis on the scores of all movements during training, lacking the ability to identify and attribute temporal patterns of performance changes. For example, they cannot distinguish between performance decline caused by muscle fatigue and performance decline caused by the failure of compensatory strategies, as different causes lead to drastically different assessment conclusions.

[0004] In summary, how to assess motion quality from the perspective of multi-joint coordination patterns to identify compensatory abnormalities, and how to perform pattern recognition and attribution analysis on dynamic changes in performance during training, are problems that urgently need to be solved by those skilled in the art. Summary of the Invention

[0005] To overcome the aforementioned problems of the prior art, this invention proposes a deep learning-based stroke training method and system to solve the above problems.

[0006] This invention provides the following technical solution: A deep learning-based stroke training method includes: Acquire multimodal physiological and motion data during the training process, and extract motion coordination features representing multi-joint linkage relationships from the multimodal physiological and motion data; Calculate the coordination deviation between the motion coordination features and the preset target coordination reference model, and analyze the coordination deviation to obtain information on the deviation components that characterize the source of abnormal motion; The multidimensional performance index sequence of the training process in time series is obtained. The multidimensional performance index sequence and the time series change sequence of the co-deviation degree are input together into the pre-trained deep learning time series analysis model to output the dynamic performance pattern and the degree value of each influencing factor. Based on the aforementioned collaborative deviation degree, deviation component information, dynamic performance pattern, and the degree values ​​of each influencing factor, a multidimensional motion quality assessment report is generated and output.

[0007] Preferably, the multimodal physiological and motor data includes multi-joint motion angle data and multi-channel surface electromyography data; The extraction of motion coordination features characterizing multi-joint linkage relationships from the multimodal physiological and motion data includes: The multimodal physiological and motor data are preprocessed to obtain a multichannel motion-electromyography time series matrix; Within a preset sliding window, the covariance matrix is ​​calculated after removing the mean from the data segments in the multi-channel motion-electromyography time series matrix, and an inter-joint motion coordination matrix is ​​constructed as a motion coordination feature.

[0008] Preferably, calculating the coordination deviation between the motion coordination features and the preset target coordination reference model includes: The joint motion coordination matrices corresponding to each of the multiple sliding windows are expanded and stacked in time order, and non-negative matrix decomposition is performed to extract the coordination basis vectors that represent the current joint coordination motion mode. Based on the current rehabilitation stage, select the target collaboration template corresponding to the stage from the pre-built multi-stage target collaboration template library as the target collaboration reference model; Singular value decomposition is performed on the cooperative basis vectors and the target cooperative template to obtain orthogonal bases, and the mean value of the main characters between the two sets of orthogonal bases is calculated as the cooperative deviation.

[0009] Preferably, the step of analyzing the coordinated deviation to obtain deviation component information characterizing the source of abnormal motion includes: The cooperative deviation is spatially decomposed to obtain the deviation component vector; Extract the joint coupling pair with the largest deviation value from the deviation component vector, and use the joint coupling pair as the deviation component information.

[0010] Preferably, the method for constructing the multi-stage target collaborative template library includes: Multi-joint motion and physiological data were collected from healthy subjects and subjects with high recovery levels at various rehabilitation stages during training tasks. The collected data are used to construct a coordination matrix in the same way as the inter-joint motion coordination matrix and perform matrix decomposition. The coordination basis vectors corresponding to each rehabilitation stage are extracted as the target coordination templates for that stage. The goal-coordination templates for each rehabilitation stage are compiled into a multi-stage goal-coordination template library.

[0011] Preferably, the method for obtaining the multidimensional performance index sequence includes: The current training process is divided into multiple time periods according to preset time intervals; For each time period, the task completion rate, the motion smoothness index derived from kinematic data, the mean of the coordination deviation, and the relative rate of change of physiological signal amplitude were calculated. After standardizing each indicator, they are arranged in chronological order to construct a multidimensional performance indicator sequence.

[0012] Preferably, the training method for the deep learning time series analysis model includes: Construct a training dataset, in which the input of each sample is a sequence of multidimensional performance indicators for a training process and a sequence of changes in co-variance for the corresponding time period. The annotation includes the type annotation of the dynamic performance mode and the degree annotation of each influencing factor. The influencing factors include fatigue factors, compensatory change factors, and neurological abnormality factors; A weighted combination of pattern classification loss and degree regression loss is used as the joint loss function for training.

[0013] Preferably, the method for determining the type labeling of the dynamic performance pattern of each sample in the training dataset includes: The linear regression slope along the time period is calculated for each dimension of the multidimensional performance index sequence in the sample, and the overall performance change rate is determined based on the slope of each dimension. The slope of the linear regression is calculated for the sequence of changes in the degree of coordination deviation in the sample, and is used as the slope of the change in deviation. Based on the combined relationship between the rate of change of the overall performance and the slope of the deviation, the type label of the dynamic performance mode of the sample is determined.

[0014] Preferably, the generation of the multidimensional motion quality assessment report includes: The collaboration deviation is mapped to a preset multi-level rating interval to generate an overall collaboration quality rating. Based on the degree of deviation of each joint coupling pair in the deviation component information, sort them in descending order, extract the top-ranked preset number of joint coupling pairs and their corresponding degree of deviation values, and generate an abnormal joint positioning list. Based on the dynamic performance pattern, determine the state evolution type identifier of the current training process, and associate it with the degree value proportion distribution of each influencing factor corresponding to the state evolution type to generate the training process state evaluation result.

[0015] This invention also provides a deep learning-based stroke training method, which is used to implement a deep learning-based stroke training method, comprising: The collaborative feature extraction module is used to acquire multimodal physiological and motion data during the training process, and extract motion collaborative features representing multi-joint linkage relationships from the multimodal physiological and motion data; The deviation analysis module is used to calculate the degree of coordination between the motion coordination features and the preset target coordination reference model, and to analyze the degree of coordination deviation to obtain information on the deviation components that characterize the source of abnormal motion. The temporal attribution module is used to obtain a multi-dimensional performance index sequence of the training process over time. The multi-dimensional performance index sequence and the time-period change sequence of the co-deviation degree are input together into a pre-trained deep learning temporal analysis model, and the dynamic performance pattern and the degree value of each influencing factor are output. The assessment report generation module is used to generate and output a multidimensional motion quality assessment report based on the coordination deviation, deviation component information, dynamic performance pattern, and the degree values ​​of each influencing factor.

[0016] This invention provides a deep learning-based stroke training method and system, which has the following beneficial effects: By constructing a covariance matrix from multi-joint motion angle data and electromyography data, and extracting collaborative basis vectors through non-negative matrix decomposition, and then performing subspace protagonist measurement with the target collaborative template constructed according to rehabilitation stages, information on collaborative deviation and deviation components is obtained, enabling the assessment of movement quality at the level of multi-joint collaborative patterns. Compared with existing assessment methods that only compare target joint trajectories with standard movement templates, this scheme expands the assessment dimension from single-joint trajectory similarity to the structural differences in multi-joint collaborative movement patterns. It can identify compensatory false attainment states where the movement trajectory seems to meet the standard but the multi-joint collaborative pattern is abnormal. By locating the specific joint coupling source causing the collaborative abnormality through deviation component information, the assessment results not only reflect whether the movement is completed, but also whether the movement source of the movement is normal.

[0017] By inputting a multidimensional performance index sequence and a time-series variation sequence of co-deviation into a deep learning time-series analysis model, the model outputs dynamic performance patterns and the degree values ​​of each influencing factor. This enables time-series pattern recognition and multi-factor attribution analysis of dynamic performance changes during training. Compared with existing methods that aggregate the mean or perform linear trend analysis on all action scores during training, this approach can distinguish performance change patterns of different time series and quantify the contribution of fatigue factors, compensatory change factors, and neural abnormality factors to performance changes. This allows the evaluation results to reflect not only whether performance has declined, but also why performance has declined. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a deep learning-based stroke training method according to the present invention. Figure 2 This is a schematic diagram of a deep learning-based stroke training method according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0020] Please see Figure 1 In this embodiment, a deep learning-based stroke training method includes: S1. Acquire multimodal physiological and motion data during the training process, and extract motion coordination features representing multi-joint linkage relationships from the multimodal physiological and motion data; The multimodal physiological and motor data includes multi-joint motion angle data and multi-channel surface electromyography data; The extraction of motion coordination features characterizing multi-joint linkage relationships from the multimodal physiological and motion data includes: The multimodal physiological and motor data are preprocessed to obtain a multichannel motion-electromyography time series matrix; Within a preset sliding window, the covariance matrix is ​​calculated after removing the mean from the data segments in the multi-channel motion-electromyography time series matrix, and an inter-joint motion coordination matrix is ​​constructed as a motion coordination feature.

[0021] In this embodiment, it should be noted that one of the core characteristics of post-stroke motor dysfunction is the abnormality of coordinated movement patterns between multiple joints, rather than simply limited range of motion of a single joint. To assess movement quality at the level of multi-joint linkage, this approach collects multi-joint motion angle data and multi-channel surface electromyography (EMG) data as multimodal physiological and motor data. Multi-joint motion angle data can be acquired using depth cameras (such as Azure Kinect), inertial measurement units, or optical motion capture systems. Multi-channel EMG data can be collected using EMG sensors attached to relevant muscle groups, typically selecting the major muscle groups related to the training movement; for example, during upper limb training, channels for the deltoid, biceps, triceps, and trapezius muscles can be selected.

[0022] It should be noted that the raw data collected needs to be preprocessed to construct a multi-channel motion-electromyography time series matrix. Preprocessing usually includes conventional operations in the field, such as applying low-pass filtering to motion data to eliminate sensor noise, band-pass filtering and envelope extraction to electromyography data. The preprocessed data of each channel is organized into a matrix form according to time alignment.

[0023] It should be noted that, in order to capture the aforementioned multi-dimensional linkage relationships within a local time frame, this scheme calculates the covariance matrix after removing the mean from all channels in the multi-channel motion-electromyography time series matrix within a preset sliding window, constructing an inter-joint motion coordination matrix as a coordination representation. The sliding window length can be 0.5 to 2 seconds, and the sliding step size is usually 30% to 50% of the window length. If the number of joint angle channels involved in the analysis is M and the number of electromyography channels is P, then the resulting coordination matrix is ​​a symmetric matrix of (M+P)×(M+P), where the M×M sub-block reflects the motion coupling strength between joints, the P×P sub-block reflects the coordination activation relationship between muscles, and the M×P sub-block reflects the coupling relationship between joint motion and muscle activation, thereby encoding the linkage relationship of multiple joints and multiple muscle groups into matrix-based motion coordination features.

[0024] S2. Calculate the coordination deviation between the motion coordination feature and the preset target coordination reference model, and analyze the coordination deviation to obtain information on the deviation components that characterize the source of abnormal motion. The calculation of the coordination deviation between the motion coordination features and the preset target coordination reference model includes: The joint motion coordination matrices corresponding to each of the multiple sliding windows are expanded and stacked in time order, and non-negative matrix decomposition is performed to extract the coordination basis vectors that represent the current joint coordination motion mode. Based on the current rehabilitation stage, select the target collaboration template corresponding to the stage from the pre-built multi-stage target collaboration template library as the target collaboration reference model; Singular value decomposition is performed on the cooperative basis vectors and the target cooperative template to obtain orthogonal bases, and the mean value of the main characters between the two sets of orthogonal bases is calculated as the cooperative deviation.

[0025] The step of analyzing the coordinated deviation to obtain deviation component information characterizing the source of abnormal motion includes: The cooperative deviation is spatially decomposed to obtain the deviation component vector; Extract the joint coupling pair with the largest deviation value from the deviation component vector, and use the joint coupling pair as the deviation component information.

[0026] The method for constructing the multi-stage target collaborative template library includes: Multi-joint motion and physiological data were collected from healthy subjects and subjects with high recovery levels at various rehabilitation stages during training tasks. The collected data are used to construct a coordination matrix in the same way as the inter-joint motion coordination matrix and perform matrix decomposition. The coordination basis vectors corresponding to each rehabilitation stage are extracted as the target coordination templates for that stage. The goal-coordination templates for each rehabilitation stage are compiled into a multi-stage goal-coordination template library.

[0027] In this embodiment, it should be noted that each sliding window in step S1 generates a coordination matrix. A single matrix only reflects the instantaneous coordination relationship within that time window and is insufficient to stably represent the overall coordination motion pattern of the current training action. To extract stable coordination features, this scheme expands the coordination matrices corresponding to multiple sliding windows into row vectors and stacks them in temporal order to form a matrix, and then performs nonnegative matrix factorization (NMF). Each extracted coordination basis vector represents a basic inter-joint coordination motion pattern. The number of basis vectors is typically 3 to 5, which can be determined according to the complexity of the training action.

[0028] It should be noted that different stages of stroke rehabilitation have different normal synergistic movement patterns. Taking Brunnstrom as an example, in Brunnstrom stage 3 (spastic stage), flexor or extensor synergy is dominant, with a high degree of inter-joint coupling; in stage 5 (dissociative movement stage), joints can achieve more independent control, and the degree of coupling is reduced. The same degree of joint coupling may be normal in stage 3, but indicates an abnormal synergistic pattern in stage 5. It is unreasonable to use a uniform standard to measure the synergistic deviation of all stages. Therefore, this scheme pre-constructs a multi-stage target synergistic template library to provide a comparative benchmark for each stage. The construction method is as follows: collect multi-joint movement data and physiological data of healthy subjects and subjects with high recovery levels in each rehabilitation stage when performing training tasks, construct the synergistic matrix in the same way as in step S1 and perform NMF decomposition, extract the synergistic basis vectors corresponding to each stage as the target synergistic templates for that stage, and collect them to form a multi-stage target synergistic template library. The amount of data required to construct the template library is usually at least 20 to 30 subjects' movement data for each stage. When using it, select the target collaboration template corresponding to the current rehabilitation stage.

[0029] It should be noted that the collaborative basis vectors of the current training movement and the target collaborative template each constitute a set of vectors. The distance between the subspaces spanned by these two vectors reflects the structural differences in collaborative patterns more stably than vector-wise matching. This scheme performs Singular Value Decomposition (SVD) on both sets of vectors to obtain orthogonal bases for their respective column spaces. Then, it calculates the principal actors between the two sets of orthogonal bases and takes the mean of each principal actor as the collaborative deviation. Both SVD and principal actor calculation are existing linear algebra methods in this field. The larger the collaborative deviation, the further the current inter-joint collaborative movement pattern deviates from the target pattern. For example, if the collaborative deviation of the upper limb extension movement is high, it usually means that non-target joints such as the trunk and shoulder are excessively involved in the movement, i.e., there is a compensatory phenomenon.

[0030] It should be noted that the coordination deviation, as a scalar, can only reflect the overall degree of deviation and cannot pinpoint which specific joint coordination relationships are abnormal. To accurately locate the source of the anomaly, this scheme further decomposes the coordination deviation into spatial components. Specifically, both the coordination basis vector and the target coordination template originate from the decomposition of the expanded coordination matrix. Each component corresponds to the coupling relationship of each channel pair in the original coordination matrix. Therefore, the differences between the coordination basis vector and the target coordination template can be analyzed at the component level, decomposing the overall deviation into the coupling dimensions of each channel to obtain the deviation component vector. Each component represents the degree of deviation of the corresponding joint coupling pair. Extracting the joint coupling pair with the largest deviation value from the deviation component vector can determine the main source of the coordination anomaly. For example, if the "shoulder joint-trunk" coupling pair has the largest deviation component in the deviation component vector, it indicates that trunk compensation is the main source of the current coordination anomaly.

[0031] It should be noted that, through the above process, this step quantifies the coordination anomaly from two dimensions: overall severity and specific location. The coordination deviation reflects the overall severity of compensatory movement, while the deviation component information locates the specific joint from which the compensatory movement originates. The combination of these two factors enables the assessment to identify compensatory false compliance states where the movement trajectory appears to meet the standard but the movement source is abnormal, and to pinpoint the specific joint source of the compensation.

[0032] S3. Obtain the multidimensional performance index sequence of the training process in time series, and input the multidimensional performance index sequence and the time series change sequence of the co-deviation degree into the pre-trained deep learning time series analysis model to output the dynamic performance pattern and the degree value of each influencing factor. The method for obtaining the multidimensional performance index sequence includes: The current training process is divided into multiple time periods according to preset time intervals; For each time period, the task completion rate, the motion smoothness index derived from kinematic data, the mean of the coordination deviation, and the relative rate of change of physiological signal amplitude were calculated. After standardizing each indicator, they are arranged in chronological order to construct a multidimensional performance indicator sequence.

[0033] The training method for the deep learning time series analysis model includes: Construct a training dataset, in which the input of each sample is a sequence of multidimensional performance indicators for a training process and a sequence of changes in co-variance for the corresponding time period. The annotation includes the type annotation of the dynamic performance mode and the degree annotation of each influencing factor. The influencing factors include fatigue factors, compensatory change factors, and neurological abnormality factors; A weighted combination of pattern classification loss and degree regression loss is used as the joint loss function for training.

[0034] The method for determining the type labeling of the dynamic performance pattern of each sample in the training dataset includes: The linear regression slope along the time period is calculated for each dimension of the multidimensional performance index sequence in the sample, and the overall performance change rate is determined based on the slope of each dimension. The slope of the linear regression is calculated for the sequence of changes in the degree of coordination deviation in the sample, and is used as the slope of the change in deviation. Based on the combined relationship between the rate of change of the overall performance and the slope of the deviation, the type label of the dynamic performance mode of the sample is determined.

[0035] In this embodiment, it should be noted that during a complete rehabilitation training session, motor performance is dynamically affected by various factors such as fatigue, changes in compensatory strategies, and neurological abnormalities. Simple mean aggregation or linear trend analysis is insufficient to fully reveal the temporal patterns of performance changes and their underlying causes. To capture this dynamic change, this scheme divides the training process into multiple time periods according to preset time intervals, typically ranging from 2 to 5 minutes.

[0036] It should be noted that multidimensional performance indicators are calculated separately for each time period. Each indicator reflects training performance from different dimensions: the task completion rate is the ratio of the number of times the target position is reached to the number of times required by the task, reflecting task execution ability; the motion smoothness indicator can be calculated using common methods in this field, such as normalized jerk, reflecting the precision of motor control; the mean of coordination deviation is calculated from the data in step S2 for that time period, reflecting the deviation of the coordination pattern in that time period; the relative change rate of physiological signal amplitude, for example, can be the ratio of the change in the root mean square amplitude of electromyography signals relative to the previous time period, reflecting the change in physiological load. After standardization (such as z-score standardization) to eliminate dimensional differences, all indicators are arranged in time period order to form a multidimensional performance indicator sequence.

[0037] It should be noted that the temporal patterns of performance changes during training may exhibit various forms, and the underlying causes of different forms are different. For example, a sustained, slow decline in performance may be primarily due to fatigue; a decline in performance accompanied by a rapid increase in coordinating deviation suggests that compensatory strategies are intensifying or failing; and a sudden, significant drop in performance may indicate neurological abnormalities such as spasms. Such complex temporal pattern recognition and multi-factor attribution are difficult to achieve through simple statistical rules. Therefore, this approach uses a multi-dimensional performance index sequence and a time-varying sequence of coordinating deviation as multi-channel temporal inputs, which are then fed into a pre-trained deep learning temporal analysis model for processing. The model can employ temporal feature extraction structures such as Temporal Convolutional Networks (TCNs) or Long Short-Term Memory Networks (LSTMs), with classification and regression branches at the output to output the type of dynamic performance pattern and the degree values ​​of each influencing factor, respectively.

[0038] It should be noted that training a deep learning time series analysis model requires constructing a training dataset. The input for each sample is a sequence of multidimensional performance indicators for a training process and a sequence of co-variance changes for the corresponding time period. Labels include the type of dynamic performance pattern and the degree of each influencing factor, including fatigue factors, compensatory change factors, and neural abnormality factors. During training, a weighted combination of pattern classification loss (e.g., cross-entropy loss) and degree regression loss (e.g., mean squared error loss) is used as the joint loss function. The weights of these two losses can be adjusted based on performance on the validation set.

[0039] It should be noted that the dynamic performance pattern type labeling of each sample in the training dataset is determined through a dual-slope combination analysis: The linear regression slope along the time period is calculated for each dimension of the multi-dimensional performance index sequence in the samples, and the overall performance change rate is determined based on the slopes of each dimension; for example, a weighted average of the slopes of each dimension can be used. Simultaneously, the linear regression slope of the coordination deviation change sequence in the samples is calculated as the deviation change slope. The overall performance change rate reflects the direction and magnitude of the overall performance change, while the deviation change slope reflects the trend of the coordination pattern change. Different combinations of these two factors can distinguish performance change patterns caused by different reasons.

[0040] For example, a negative threshold for the rate of change in overall performance and a positive threshold for the slope of deviation can be set. Based on the relationship between these two thresholds and their respective thresholds, dynamic performance patterns can be classified into the following types: When the rate of change in overall performance is lower than the negative threshold and the slope of deviation does not exceed the positive threshold, it indicates that overall performance is declining but the synergistic pattern has not deteriorated significantly, with fatigue as the dominant factor, and can be labeled as a gradual decline type; When the rate of change in overall performance is lower than the negative threshold and the slope of deviation exceeds the positive threshold, it indicates that performance is declining while the synergistic pattern is rapidly deteriorating, suggesting that compensatory strategies are intensifying or failing, and can be labeled as a compensatory deterioration type; When the absolute value of the rate of change in overall performance does not exceed the absolute value of the negative threshold but the slope of deviation exceeds the positive threshold, it indicates that overall performance remains relatively stable but the synergistic pattern continues to deviate, suggesting that the surface performance is being maintained through increasingly severe compensation, and can be labeled as a stable deviation type; When the change amplitude of any dimension in the multidimensional performance index sequence exceeds the preset mutation amplitude threshold between adjacent time periods, it indicates that a sudden performance collapse has occurred, which may be related to neurological abnormal events such as spasms, and can be labeled as a mutation type. The degree of each influencing factor can be labeled by rehabilitation professionals based on training records and clinical observations.

[0041] It should be noted that, through the above process, pattern recognition and multi-factor attribution of dynamic changes in the training process are achieved in the time dimension, so that the evaluation results can distinguish the performance changes caused by different reasons, such as the performance decline caused by muscle fatigue and the performance decline caused by the failure of compensatory strategies.

[0042] S4. Based on the aforementioned coordination deviation, deviation component information, dynamic performance pattern, and degree values ​​of each influencing factor, generate and output a multidimensional motion quality assessment report.

[0043] The generated multidimensional motion quality assessment report includes: The collaboration deviation is mapped to a preset multi-level rating interval to generate an overall collaboration quality rating. Based on the degree of deviation of each joint coupling pair in the deviation component information, sort them in descending order, extract the top-ranked preset number of joint coupling pairs and their corresponding degree of deviation values, and generate an abnormal joint positioning list. Based on the dynamic performance pattern, determine the state evolution type identifier of the current training process, and associate it with the degree value proportion distribution of each influencing factor corresponding to the state evolution type to generate the training process state evaluation result.

[0044] In this embodiment, it should be noted that the aforementioned steps have obtained the degree of collaborative deviation, information on deviation components, dynamic performance patterns, and degree values ​​of each influencing factor. This step integrates the above multidimensional analysis results into a structured multidimensional motion quality assessment report.

[0045] It should be noted that the coordination deviation is mapped to a preset multi-level rating interval to generate an overall coordination quality rating. For example, five levels can be set to intuitively reflect the overall degree of deviation between the current movement mode and the target coordination mode. The boundary values ​​of the rating interval can be determined based on the statistical distribution of historical data, such as using the percentile of the coordination deviation samples of each period in the target coordination template library as the dividing basis.

[0046] It should be noted that, based on the descending order of the deviation degree of each joint coupling pair in the deviation component information, a preset number of joint coupling pairs and their corresponding deviation degree values ​​are extracted to generate an abnormal joint location list. Usually, the first 2 to 3 are extracted to clearly indicate the main source of the coordination anomaly and the degree of their respective anomalies.

[0047] It should be noted that the current training process state evolution type is identified based on the dynamic performance pattern, and the training process state assessment result is generated by associating the degree value distribution of each influencing factor corresponding to this state evolution type. The degree value distribution can be calculated by dividing the degree value of each factor by the total degree value, reflecting the contribution of fatigue, compensatory changes, and neural abnormalities to the current training state change. The state evolution type identifier reflects the temporal morphological characteristics of performance changes. For example, a gradual decay type indicates a continuous downward trend in performance, a stable shift type indicates that the overall performance remains stable but the synergistic pattern shifts, and a sudden change type indicates a sudden performance change. The degree value distribution of influencing factors reflects the causal composition of this temporal morphology. For example, even within the gradual decay type, the causal composition may be mainly fatigue factors, or it may be driven by both fatigue factors and compensatory change factors. The combination of these two aspects allows the training process state assessment result to not only present how the performance changes, but also reveal why it changes in this way.

[0048] It should be noted that the final multidimensional motion quality assessment report is presented at two levels: the action level and the process level. The action level includes an overall coordination quality rating and a list of abnormal joint locations, reflecting the overall quality of the motion coordination pattern and the specific sources of abnormalities, respectively, enabling the identification of compensatory abnormalities at the multi-joint coordination pattern level. The process level presents the training process state assessment results, reflecting the evolution trend of the state throughout the training process and the contribution degree of each influencing factor, thus addressing the problem of pattern recognition and attribution analysis of dynamic changes in performance during training. Example

[0049] Please see Figure 2 This invention provides a deep learning-based stroke training method, which includes: The collaborative feature extraction module is used to acquire multimodal physiological and motion data during the training process, and extract motion collaborative features representing multi-joint linkage relationships from the multimodal physiological and motion data; The deviation analysis module is used to calculate the degree of coordination between the motion coordination features and the preset target coordination reference model, and to analyze the degree of coordination deviation to obtain information on the deviation components that characterize the source of abnormal motion. The temporal attribution module is used to obtain a multi-dimensional performance index sequence of the training process over time. The multi-dimensional performance index sequence and the time-period change sequence of the co-deviation degree are input together into a pre-trained deep learning temporal analysis model, and the dynamic performance pattern and the degree value of each influencing factor are output. The assessment report generation module is used to generate and output a multidimensional motion quality assessment report based on the coordination deviation, deviation component information, dynamic performance pattern, and the degree values ​​of each influencing factor.

[0050] In the several embodiments provided by this invention, 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 units is only one method, and other division methods may exist in actual implementation. 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 mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or units may be electrical, mechanical, or other forms.

[0051] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

[0052] Finally, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based stroke training method, characterized in that, include: Acquire multimodal physiological and motion data during the training process, and extract motion coordination features representing multi-joint linkage relationships from the multimodal physiological and motion data; Calculate the coordination deviation between the motion coordination features and the preset target coordination reference model, and analyze the coordination deviation to obtain information on the deviation components that characterize the source of abnormal motion; The multidimensional performance index sequence of the training process in time series is obtained. The multidimensional performance index sequence and the time series change sequence of the co-deviation degree are input together into the pre-trained deep learning time series analysis model to output the dynamic performance pattern and the degree value of each influencing factor. Based on the aforementioned collaborative deviation degree, deviation component information, dynamic performance pattern, and the degree values ​​of each influencing factor, a multidimensional motion quality assessment report is generated and output. 2.The stroke training method based on deep learning according to claim 1, wherein, The multimodal physiological and motor data includes multi-joint motion angle data and multi-channel surface electromyography data; The extraction of motion coordination features characterizing multi-joint linkage relationships from the multimodal physiological and motion data includes: The multimodal physiological and motor data are preprocessed to obtain a multichannel motion-electromyography time series matrix; Within a preset sliding window, the covariance matrix is ​​calculated after removing the mean from the data segments in the multi-channel motion-electromyography time series matrix, and an inter-joint motion coordination matrix is ​​constructed as a motion coordination feature.

3. The stroke training method based on deep learning according to claim 2, characterized in that, The calculation of the coordination deviation between the motion coordination features and the preset target coordination reference model includes: The joint motion coordination matrices corresponding to each of the multiple sliding windows are expanded and stacked in time order, and non-negative matrix decomposition is performed to extract the coordination basis vectors that represent the current joint coordination motion mode. Based on the current rehabilitation stage, select the target collaboration template corresponding to the stage from the pre-built multi-stage target collaboration template library as the target collaboration reference model; Singular value decomposition is performed on the cooperative basis vectors and the target cooperative template to obtain orthogonal bases, and the mean value of the main characters between the two sets of orthogonal bases is calculated as the cooperative deviation. 4.The stroke training method based on deep learning according to claim 3, characterized in that, The step of analyzing the coordinated deviation to obtain deviation component information characterizing the source of abnormal motion includes: The cooperative deviation is spatially decomposed to obtain the deviation component vector; Extract the joint coupling pair with the largest deviation value from the deviation component vector, and use the joint coupling pair as the deviation component information. 5.The stroke training method based on deep learning according to claim 4, characterized in that, The method for constructing the multi-stage target collaborative template library includes: Multi-joint motion and physiological data were collected from healthy subjects and subjects with high recovery levels at various rehabilitation stages during training tasks. The collected data are used to construct a coordination matrix in the same way as the inter-joint motion coordination matrix and perform matrix decomposition. The coordination basis vectors corresponding to each rehabilitation stage are extracted as the target coordination templates for that stage. The goal-coordination templates for each rehabilitation stage are compiled into a multi-stage goal-coordination template library. 6.The stroke training method based on deep learning according to claim 1, wherein, The method for obtaining the multidimensional performance index sequence includes: The current training process is divided into multiple time periods according to preset time intervals; For each time period, the task completion rate, the motion smoothness index derived from kinematic data, the mean of the coordination deviation, and the relative rate of change of physiological signal amplitude were calculated. After standardizing each indicator, they are arranged in chronological order to construct a multidimensional performance indicator sequence. 7.The stroke training method based on deep learning according to claim 6, characterized in that, The training method for the deep learning time series analysis model includes: Construct a training dataset, in which the input of each sample is a sequence of multidimensional performance indicators for a training process and a sequence of changes in co-variance for the corresponding time period. The annotation includes the type annotation of the dynamic performance mode and the degree annotation of each influencing factor. The influencing factors include fatigue factors, compensatory change factors, and neurological abnormality factors; A weighted combination of pattern classification loss and degree regression loss is used as the joint loss function for training. 8.The stroke training method based on deep learning according to claim 7, characterized in that, The method for determining the type labeling of the dynamic performance pattern of each sample in the training dataset includes: The linear regression slope along the time period is calculated for each dimension of the multidimensional performance index sequence in the sample, and the overall performance change rate is determined based on the slope of each dimension. The slope of the linear regression is calculated for the sequence of changes in the degree of coordination deviation in the sample, and is used as the slope of the change in deviation. Based on the combined relationship between the rate of change of the overall performance and the slope of the deviation, the type label of the dynamic performance mode of the sample is determined. 9.The stroke training method based on deep learning of claim 4, wherein, The generated multidimensional motion quality assessment report includes: The collaboration deviation is mapped to a preset multi-level rating interval to generate an overall collaboration quality rating. Based on the degree of deviation of each joint coupling pair in the deviation component information, sort them in descending order, extract the top-ranked preset number of joint coupling pairs and their corresponding degree of deviation values, and generate an abnormal joint positioning list. Based on the dynamic performance pattern, determine the state evolution type identifier of the current training process, and associate it with the degree value proportion distribution of each influencing factor corresponding to the state evolution type to generate the training process state evaluation result.

10. A deep learning-based stroke training method, used to implement the deep learning-based stroke training method as described in any one of claims 1-9, characterized in that, include: The collaborative feature extraction module is used to acquire multimodal physiological and motion data during the training process, and extract motion collaborative features representing multi-joint linkage relationships from the multimodal physiological and motion data; The deviation analysis module is used to calculate the degree of coordination between the motion coordination features and the preset target coordination reference model, and to analyze the degree of coordination deviation to obtain information on the deviation components that characterize the source of abnormal motion. The temporal attribution module is used to obtain a multi-dimensional performance index sequence of the training process over time. The multi-dimensional performance index sequence and the time-period change sequence of the co-deviation degree are input together into a pre-trained deep learning temporal analysis model, and the dynamic performance pattern and the degree value of each influencing factor are output. The assessment report generation module is used to generate and output a multidimensional motion quality assessment report based on the coordination deviation, deviation component information, dynamic performance pattern, and the degree values ​​of each influencing factor.

Images