A rehabilitation training data explainable evaluation method based on a double-path causality framework
By constructing a rehabilitation training data evaluation method based on a dual-pathway causal framework, the problems of structured processing of rehabilitation training data and interpretability of evaluation results are solved, enabling multidimensional evaluation and refined management of training data, and providing interpretable rule paths and training suggestions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHANGCHUN UNIV OF TECH
- Filing Date
- 2026-04-27
- Publication Date
- 2026-05-29
AI Technical Summary
Existing methods for processing rehabilitation training data suffer from insufficient structuring, weak interpretability of assessment results, and limited output dimensions, making it difficult to achieve continuous and refined training management.
We construct an interpretable assessment method for rehabilitation training data based on a dual-pathway causal framework. By acquiring training data and historical assessment labels, we construct a three-domain, nine-item structured feature, generate soft labels using a teacher model, extract rule paths using a rule tree model, and construct a robust association graph using causal relationship mining and Bootstrap resampling. The method outputs assessment results, rule paths, and training suggestions for the training phase.
It enables structured processing of rehabilitation training data, improves the interpretability of assessment results and the refinement of training management, and provides identifiable rule paths and training suggestions.
Smart Images
Figure CN122114205A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of healthcare informatics, rehabilitation training data processing, and medical artificial intelligence, and specifically relates to an interpretable assessment method for rehabilitation training data based on a dual-pathway causal framework. Background Technology
[0002] In the context of information technology-based rehabilitation training, patients continuously generate multidimensional training data related to the training process when performing training tasks on training equipment, interactive terminals, or digital training platforms. This multidimensional training data typically includes trajectory data, time-series data, task feedback data, and historical evaluation data.
[0003] In existing technologies, the processing of patient-related rehabilitation training data mainly relies on manual observation, single-indicator statistics, or black-box model prediction, which has problems such as insufficient structuring, weak interpretability of training stage assessment results, and single output dimension.
[0004] While some automatic evaluation methods can achieve high prediction accuracy based on training data, they often fail to provide clear rule paths and cannot clearly indicate the influence of different features on the evaluation results during the training phase.
[0005] In addition, existing methods typically only output a single assessment label, making it difficult to further develop a mechanism profile for training management and output training recommendations, which is not conducive to continuous and refined management of patients' specific training states.
[0006] Therefore, there is an urgent need to provide an interpretable assessment method for patient-related rehabilitation training data to achieve structured processing of training data, assessment of training phases, rule generation, and output of training recommendations. Summary of the Invention
[0007] To address the shortcomings of existing technologies, such as insufficient structured processing of patient-related rehabilitation training data, weak interpretability of training phase assessment results, and insufficient ability to output training suggestions, this invention provides an interpretable assessment method for rehabilitation training data based on a dual-pathway causal framework.
[0008] The technical solution adopted by this invention to solve the above-mentioned technical problems is as follows: by acquiring the patient's training data, reference task data and historical assessment labels during the rehabilitation training process, a three-domain nine-item structured feature is constructed; on this basis, a teacher model is used to generate soft label probabilities, a rule tree model is used to extract rule paths, and a robust association graph is constructed by combining causal relationship mining and Bootstrap resampling, thereby outputting the training stage assessment results, rule paths, dual-pathway mechanism profiles and training suggestion information.
[0009] This invention presents an interpretable assessment method for rehabilitation training data based on a dual-pathway causal framework. The method includes training data acquisition, data preprocessing, feature construction, teacher model training, rule distillation, causal relationship mining, robust association graph construction, and result output. First, relevant patient training data and historical assessment labels are acquired. Second, the training data is denoised, smoothed, and discretely differentiated to obtain velocity and acceleration sequences. Third, nine structured features are constructed, including kinematic stability features, trajectory accuracy features, and cognitive-motor response features. Then, the teacher model is used to perform training-stage predictions and generate soft labels. Rule trees are used to distill and train the soft labels to extract rule paths. Finally, causal relationship mining and Bootstrap resampling aggregation are used to construct a robust association graph, forming a dual-pathway mechanism explanation. The method outputs training-stage assessment results, rule paths, dual-pathway mechanism profiles, and training suggestions.
[0010] The method includes the following steps:
[0011] Step 1: Obtain the training sample set and historical evaluation labels:
[0012] Acquire training data generated by patients when performing preset interactive training tasks on rehabilitation training equipment, and establish the correlation between the training data, reference task data, and historical assessment labels to form a training sample set for subsequent assessment.
[0013] Step 2, Training Data Preprocessing:
[0014] The training data obtained in step 1 is denoised and smoothed to reduce the impact of sampling noise and high-frequency disturbances on feature calculation. After smoothing, the training data is discretely differentiated to generate velocity and acceleration sequences, which serve as the basis for subsequent feature construction.
[0015] Step 3: Construct three-domain, nine-item structured features:
[0016] Based on the training data, velocity sequences, and acceleration sequences obtained in step 2, a three-domain, nine-item structured feature set is constructed, comprising kinematic stability features, trajectory accuracy features, and cognitive-motor response features. The kinematic stability features include velocity coefficient of variation, number of sub-motions, and pause ratio; the trajectory accuracy features include normalized root mean square error, path ratio, closure error, and gain error; and the cognitive-motor response features include time lag and hit rate. These three-domain, nine-item structured features are used to form a multidimensional feature vector of patient-related rehabilitation training data.
[0017] Step 4: Train the teacher model and generate soft tags:
[0018] The three-domain, nine-item structured features obtained in step 3 are input into the teacher model to perform predictions during the training phase, and the soft label probability corresponding to each sample is output. The teacher model is used to learn the nonlinear mapping relationship between the structured features and the training phase.
[0019] Step 5: Distill the rule tree and extract the rule paths:
[0020] Using the soft labels generated by the teacher model in step 4, the rule tree model is distilled and trained to extract rule paths. These rule paths are used to form interpretable decision logic, enabling the evaluation results during the training phase to correspond to identifiable feature thresholds and branch conditions.
[0021] Step 6: Causal Relationship Mining and Bootstrap Resampling Aggregation:
[0022] Based on the three-domain, nine-item structured features in step 3, causal relationship mining is performed. At the same time, Bootstrap is used to resample the samples multiple times, and the frequency of candidate association edges in multiple results is counted to perform resampling aggregation and stability screening of the association structure.
[0023] Step 7: Construct a robust relationship graph:
[0024] Based on the frequency of candidate association edges in step 6, candidate association edges that meet the pre-set confidence conditions are retained to construct a robust association graph. This robust association graph forms a kinematic stability pathway and a cognitive accuracy pathway at the mechanism explanation level. The kinematic stability pathway uses the velocity variation coefficient to indicate the number of sub-motions and the pause ratio, while the cognitive accuracy pathway uses the normalized root mean square error to indicate the hit rate, and the hit rate to indicate the time delay.
[0025] Step 8: Obtain training data for the object to be evaluated and generate evaluation features:
[0026] Obtain the training data and corresponding reference task data of the object to be evaluated during the current training process, and generate the three-domain nine-item structured features of the object to be evaluated according to steps 2 and 3.
[0027] Step 9: Output the evaluation results:
[0028] Based on the training phase prediction results in step 4, the rule paths in step 5, and the robust association graph in step 7, the training phase evaluation results, rule paths, dual-pathway mechanism profiles, and training suggestions are output. The training suggestions include at least one of the following: training parameter adjustment suggestions, training task difficulty adjustment suggestions, or training scheme recommendations.
[0029] Compared with existing technologies, the present invention has the following beneficial effects: by constructing a three-domain, nine-item structured feature, it can realize the structured processing of patient-related rehabilitation training data; by combining teacher model with rule tree distillation training, it can obtain interpretable rule paths while maintaining high predictive performance; by using causal relationship mining and Bootstrap resampling aggregation to construct a robust association graph, it can improve the mechanism interpretation ability of the assessment results; by outputting training phase assessment results, rule paths, dual-pathway mechanism profiles, and training suggestion information, it can improve the refinement of patient-specific training management. Attached Figure Description
[0030] Figure 1 This is a schematic diagram of the overall architecture of the rehabilitation training data interpretable assessment system of the present invention.
[0031] Figure 2 This is a flowchart of the rehabilitation training data interpretability assessment method of the present invention.
[0032] Figure 3 This is a schematic diagram of the correlation results of the dual-pathway mechanism of the present invention.
[0033] Figure 4 This is a schematic diagram of the distillation determination path of the present invention. Detailed Implementation
[0034] The present invention will now be described in detail with reference to the accompanying drawings.
[0035] like Figure 1 As shown, the rehabilitation training data interpretable assessment system based on a dual-pathway causal framework in this embodiment includes a training data acquisition unit 101, a preprocessing unit 102, a feature construction unit 103, a teacher model unit 104, a distillation rule unit 105, an association mining unit 106, and a result output unit 107. The training data acquisition unit 101 acquires patient-related training data and reference task data, and constructs historical assessment labels; the preprocessing unit 102 performs denoising, smoothing, and discrete differentiation on the training data; the feature construction unit 103 constructs three-domain, nine-item structured features; the teacher model unit 104 performs training phase prediction and generates soft labels; the distillation rule unit 105 constructs rule paths based on classification and regression trees and extracts rule thresholds; the association mining unit 106 performs causal relationship mining, Bootstrap resampling aggregation, and stable association screening; and the result output unit 107 outputs training phase assessment results, rule paths, dual-pathway mechanism profiles, and training suggestion information.
[0036] In one embodiment, the feature construction unit 103 further includes a kinematic stability feature subunit 1031, a trajectory accuracy feature subunit 1032, and a cognitive motion response feature subunit 1033. The kinematic stability feature subunit 1031 is used to construct the velocity variation coefficient, the number of sub-motions, and the pause ratio; the trajectory accuracy feature subunit 1032 is used to construct the normalized root mean square error, path ratio, closure error, and gain error; and the cognitive motion response feature subunit 1033 is used to construct the time delay and hit rate.
[0037] like Figure 2 As shown, the method flow of the present invention includes: acquiring a training sample set and historical evaluation labels, training data preprocessing, constructing three-domain nine-item structured features, training a teacher model and generating soft labels, distilling rule trees and extracting rule paths, causal relationship mining and Bootstrap resampling aggregation, constructing a robust association graph, acquiring training data of the object to be evaluated and generating evaluation features, and outputting the evaluation results, rule paths, dual-pathway mechanism profiles and training suggestion information of the training stage.
[0038] In one embodiment, in step S1, the patient-related training data may originate from rehabilitation training equipment, interactive training terminals, or digital training platforms. In one specific embodiment, the rehabilitation training equipment may be a five-bar parallel rehabilitation robot, but the present invention is not limited to this specific device form.
[0039] In one embodiment, after time synchronization or equal-length resampling, the actual trajectory corresponding to the training data can be represented as:
[0040] (1)
[0041] in, Represents the actual set of trajectory points; Indicates the first One actual trajectory point; and They represent the first The x and y coordinates of each actual trajectory point; This indicates the total number of sampling points.
[0042] In one embodiment, the reference trajectory in the reference task data corresponding to the training data can be represented as:
[0043] (2)
[0044] in, For reference trajectory point set; For the first One reference trajectory point; and The first The x and y coordinates of each reference trajectory point; This represents the total number of sampling points after alignment with the actual trajectory.
[0045] In one embodiment, the velocity sequence in step S2 can be obtained by discretizing the derivatives of adjacent sampling points, as follows:
[0046] (3)
[0047] in, Indicates the first The velocity amplitude corresponding to each sampling moment; This indicates the sampling time interval between two adjacent sampling points; This represents the displacement difference between adjacent sampling points in the horizontal direction; This represents the displacement difference between adjacent sampling points along the vertical axis.
[0048] In one embodiment, the acceleration sequence in step S2 can be obtained by discretizing the velocity sequence, as follows:
[0049] (4)
[0050] in, Indicates the first The acceleration amplitude corresponding to each sampling time; This represents the difference in velocity amplitude between adjacent sampling times; Indicates taking the absolute value; Indicates the sampling time interval.
[0051] In one embodiment, step S2 involves low-pass filtering and smoothing the training data to reduce the impact of measurement noise and high-frequency disturbances; subsequently, discrete differentiation is performed on the training data to obtain velocity and acceleration sequences. This preprocessing step provides basic data for subsequent features such as velocity variation coefficient, number of sub-motions, pause ratio, and time delay.
[0052] In one embodiment, the three-domain, nine-item structured features constructed in step S3 include: a first category, kinematic stability features, which include velocity variation coefficient, number of sub-motions, and pause ratio, used to characterize velocity regulation capability, motion fragmentation degree, and near-zero velocity state proportion; a second category, trajectory accuracy features, which include normalized root mean square error, path ratio, closure error, and gain error, used to characterize trajectory spatial deviation, path efficiency, start-end point drift, and amplitude modulation capability; and a third category, cognitive motion response features, which include time delay and hit rate, used to characterize reaction delay and task hit rate.
[0053] In one embodiment, the velocity variation coefficient can be expressed as:
[0054] (5)
[0055] in, Indicates the velocity variation coefficient; The standard deviation of the velocity sequence; This represents the average value of the velocity sequence.
[0056] In one embodiment, the mean and standard deviation of the velocity sequence can be expressed as follows:
[0057] (6)
[0058] (7)
[0059] in, This represents the summation operation; Indicates the first The velocity amplitude corresponding to each sampling moment; Indicates the total number of sampling points; The standard deviation of the velocity sequence; This represents the average value of the velocity sequence.
[0060] In one embodiment, the number of sub-motions can be expressed as:
[0061] (8)
[0062] in, Indicates the number of sub-movements; This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. This indicates a logical AND operation. Indicates the peak threshold; , and These represent the velocity amplitude at adjacent sampling times.
[0063] In one embodiment, the pause ratio can be expressed as:
[0064] (9)
[0065] in, Indicates the pause ratio; Indicates the pause speed threshold; Indicates an indicator function; Indicates the first The velocity amplitude at each sampling time.
[0066] In one embodiment, the actual training path length and the reference path length can be expressed as follows:
[0067] (10)
[0068] (11)
[0069] in, Indicates the actual training path length; Indicates the length of the reference path; , They represent the first The x and y coordinates of each actual trajectory point; , They represent the first The x and y coordinates of each reference trajectory point; This indicates the total number of sampling points.
[0070] In one embodiment, the normalized root mean square error can be expressed as:
[0071] (12)
[0072] in, This represents the normalized root mean square error. Indicates the length of the reference path; Indicates the first The deviation of the sampling point in the horizontal axis direction; Indicates the first The deviation of the sampling point in the vertical axis direction; This indicates the total number of sampling points.
[0073] In one embodiment, the path ratio can be expressed as:
[0074] (13)
[0075] in, Indicates the path ratio; Indicates the length of the reference path; This indicates the actual training path length.
[0076] In one embodiment, the closure error can be expressed as:
[0077] (14)
[0078] in, Indicates closure error; Indicates the coordinates of the starting point of the actual trajectory; Indicates the coordinates of the actual endpoint of the trajectory; Indicates the length of the reference path.
[0079] In one embodiment, the gain error can be expressed as:
[0080] (15)
[0081] in, Indicates gain error; Indicates the actual amplitude of motion; Indicates the reference motion amplitude; This indicates taking the absolute value.
[0082] In one embodiment, the time delay can be expressed as:
[0083] (16)
[0084] in, Indicates the time delay; Indicates the time when the prompt signal was issued; Indicates the moment when the motion begins.
[0085] In one embodiment, the start time of the motion can be expressed as:
[0086] (17)
[0087] in, This indicates taking the minimum value; Indicates the first Each sampling time; Indicates the threshold for initiating movement; Indicates the first The velocity amplitude corresponding to each sampling moment.
[0088] In one embodiment, the hit rate can be expressed as:
[0089] (18)
[0090] in, Indicates hit rate; Indicates the number of times the target was hit; Indicates the total number of times the target is reached.
[0091] In one embodiment, step S4 employs an extreme gradient boosting model as the teacher model. This teacher model learns the mapping relationship between the three domains and nine structured features and the training stages, and outputs the soft label probability for each candidate training stage.
[0092] In one embodiment, step S5 uses a classification and regression tree as the distillation rule model. By inputting the soft labels generated by the teacher model into the classification and regression tree, rule paths can be obtained. In these rule paths, a gain error threshold of 0.55 is used to distinguish between candidates in the initial training stage and candidates in the advanced training stage, and a path ratio threshold of 0.85 is used to further distinguish different training paths. In subsequent subdivisions, hit rate, normalized root mean square error, time delay, number of sub-motions, and closure error can be used as key decision metrics.
[0093] like Figure 4 As shown, the rule distillation decision path in this embodiment includes: root node 401, with the judgment condition being whether the gain error is less than or equal to 0.55; if the root node condition is met, then proceed to node 402 to form a candidate for the primary training stage, and continue to perform subdivision based on the hit rate and normalized root mean square error; if the root node condition is not met, then proceed to node 403 to form a candidate for the advanced training stage; in the advanced training stage candidate, node 404, with the judgment condition being whether the path ratio is less than or equal to 0.85; if the condition of node 404 is met, then continue to perform subdivision based on the time delay, number of sub-motions, and closure error; if the condition of node 404 is not met, then proceed to the inefficient training path. Thus, the rule tree can convert the main decision logic of the teacher model into an interpretable hierarchical rule path.
[0094] In one embodiment, causal relationship mining is performed in step S6. First, candidate association structures are constructed based on conditional independence judgment. Then, the samples are resampled multiple times using Bootstrap, and the frequency of occurrence of candidate association edges in multiple results is counted to evaluate the stability of the association edges.
[0095] In one embodiment, the frequency of occurrence of candidate associated edges can be expressed as:
[0096] (19)
[0097] in, Indicates the frequency of occurrence of candidate associated edges; Indicates a candidate associated edge in The number of times it appears in the Bootstrap resampling results; This indicates the total number of Bootstrap resampling attempts.
[0098] In one embodiment, the directional consistency of candidate associated edges can be represented as:
[0099] (20)
[0100] in, This indicates the consistency of the direction of candidate associated edges; Indicates direction as The number of times the edge appears in the multiple resampling results; Indicates direction as The number of times the edge appears in the multiple resampling results; This indicates taking the maximum value.
[0101] In one embodiment, a candidate association edge that satisfies a preset trust condition can be represented as:
[0102] (twenty one)
[0103] in, This indicates a preset threshold value; This represents a preset direction threshold; when a candidate associated edge simultaneously meets both the occurrence frequency threshold and the direction consistency threshold, it is retained for constructing a robust association graph.
[0104] In one embodiment, step S7 constructs a robust correlation graph based on the resampling results of step S6 and interprets it using a dual-pathway mechanism. For example... Figure 3 As shown, the dual-pathway mechanism association results of the robust association graph include the kinematic stability pathway 301, the cognitive accuracy pathway 302, and the joint output result 303.
[0105] In the kinematic stability pathway 301, the velocity variation coefficient points to the number of sub-movements and the pause ratio, respectively, and is used to explain the changes in motion fragmentation and pause behavior caused by abnormal velocity regulation, thus forming an explanation of the kinematic stability mechanism.
[0106] In the cognitive accuracy pathway 302, the normalized root mean square error points to the hit rate, which in turn points to the time lag, and is used to explain the cognitive accuracy mechanism composed of trajectory spatial error, hit level, and reaction delay.
[0107] In the joint output 303, the kinematic stability pathway 301 and the cognitive accuracy pathway 302 work together to evaluate the training phase, the velocity-driven motion instability profile, the cognitive-accuracy trade-off control deficit profile, and training suggestion information, thereby expanding the single training phase label into a multi-dimensional output for training management.
[0108] In one embodiment, the result output unit 107 simultaneously outputs the training phase evaluation results, rule paths, dual-pathway mechanism profiles, and training suggestion information. Thus, the system can not only provide training phase judgments but also identify the key characteristic paths leading to those judgments, the mechanism explanations, and corresponding training management suggestions.
[0109] In one specific embodiment, Figure 1 The system overall architecture shown Figure 2 The method flow shown Figure 3 The correlation results of the dual-pathway mechanism shown and Figure 4 The distillation determination paths shown together constitute a complete embodiment of the present invention, wherein, Figure 1 Reflecting the division of system units, Figure 2 Reflects the execution order of the method. Figure 3 The mechanism explanation and joint output results reflecting the robustness of the correlation graph are presented. Figure 4 The key decision paths of the rule tree are reflected. The four diagrams collectively correspond to the same technical solution, rather than being independent implementation methods.
[0110] This invention is not limited to the embodiments described above. Various modifications and improvements can be made by those skilled in the art without departing from the principles of this invention, and these modifications and improvements should also be considered within the scope of protection of this invention.
Claims
1. A method for interpretable assessment of rehabilitation training data based on a dual-pathway causal framework, characterized in that, Performed by electronic devices, this process involves processing patient-related rehabilitation training data and generating training phase assessment results, including the following steps: S1. Acquire training data, reference task data and historical evaluation labels, and construct a training sample set. Each sample in the training sample set includes: training data generated by the patient when performing a preset interactive training task on the rehabilitation training device, reference task data corresponding to the preset interactive training task, and historical evaluation labels corresponding to the training data. S2. Preprocess the training data in the training sample set, including filtering and smoothing, and obtain the velocity sequence and acceleration sequence according to the sampling time interval. S3. Based on the training data, velocity sequence, acceleration sequence and reference task data, construct a three-domain nine-item structured feature, wherein the three-domain nine-item structured feature includes at least kinematic stability feature, trajectory accuracy feature and cognitive motion response feature; S4. Input the three-domain nine-item structured features into the teacher model, perform training phase prediction, and generate soft label probabilities corresponding to each candidate training phase. S5. Using the soft label probability as the distillation target, train the rule tree model and extract the rule paths used to characterize the decision logic in the training phase. S6. Perform causal relationship mining on the three domains and nine structured features in the training sample set, and statistically analyze the frequency of candidate association edges in multiple resampling results based on Bootstrap resampling. S7. Filter candidate association edges that meet the preset confidence conditions according to the occurrence frequency, construct a robust association graph, and determine the kinematic stability pathway and cognitive accuracy pathway according to the robust association graph; S8. Obtain training data of the object to be evaluated and generate evaluation features, specifically including: obtaining training data and corresponding reference task data of the object to be evaluated in the current training process, and generating three-domain nine-item structured features of the object to be evaluated in accordance with steps S2 and S3. S9. Input the three-domain, nine-item structured features of the object to be evaluated into the teacher model and the rule tree model, and combine them with the robust association graph to output the training phase evaluation results, rule paths, dual-pathway mechanism profiles, and training suggestions.
2. The rehabilitation training data interpretability assessment method according to claim 1, characterized in that, In step S1, the historical assessment label is generated from existing rehabilitation assessment records, manually labeled records, or historical training phase records.
3. The rehabilitation training data interpretability assessment method according to claim 1, characterized in that, In step S2, the filtering and smoothing process includes at least one of low-pass filtering, moving average smoothing, and Savitzky-Golay smoothing, and the velocity sequence and acceleration sequence are obtained by discretely differentiating the smoothed training data.
4. The rehabilitation training data interpretability assessment method according to claim 1, characterized in that, In step S3, the three-domain, nine-item structured features are used to form a multidimensional feature vector of patient-related rehabilitation training data. The three-domain, nine-item structured features include: The velocity variation coefficient, number of sub-motions, and pause ratio are kinematic stability characteristics. The characteristics of trajectory accuracy include normalized root mean square error, path ratio, closure error, and gain error; The time delay and hit rate in cognitive motor response characteristics.
5. The rehabilitation training data interpretability assessment method according to claim 4, characterized in that: The velocity variation coefficient is characterized by the ratio of the standard deviation of the velocity sequence to the mean of the velocity sequence; The number of sub-motions is characterized by the number of local peaks in the velocity sequence that exceed the peak threshold; The pause ratio is characterized by the proportion of the number of sampling points whose speed is lower than the pause speed threshold to the total number of sampling points. The normalized root mean square error is obtained by normalizing the deviation between the training data and the reference task data using the reference path length. The path ratio is characterized by the ratio of the reference path length to the training path length; The closure error is obtained by normalizing the distance between the start and end points of the training trajectory using the reference path length. The gain error is characterized by the relative deviation between the actual motion amplitude and the reference motion amplitude; The time delay is characterized by the time difference between the moment the prompt signal is issued and the moment the movement begins. The hit rate is characterized by the ratio of the number of times the target is hit to the total number of times the target is hit.
6. The rehabilitation training data interpretability assessment method according to claim 1, characterized in that, In step S4, the teacher model is an extreme gradient boosting model; in step S5, the rule tree model is a classification and regression tree model.
7. The rehabilitation training data interpretability assessment method according to claim 1, characterized in that, In step S5, the rule path includes: using gain error as the first-level judgment feature and path ratio as the second-level judgment feature; when the gain error is less than or equal to a first preset threshold, the sample is judged as a candidate for the primary training stage, and further subdivided according to the hit rate and normalized root mean square error; when the gain error is greater than the first preset threshold, the sample is judged as a candidate for the advanced training stage, and further judged according to the path ratio; when the path ratio is less than or equal to a second preset threshold, further subdivided according to the time delay, number of sub-motions, and closure error; when the path ratio is greater than the second preset threshold, it is judged as an inefficient training path.
8. The rehabilitation training data interpretability assessment method according to claim 1, characterized in that, In step S7, the preset confidence condition is: the frequency of occurrence of candidate associated edges is not lower than the preset confidence threshold, and the directional consistency is not lower than the preset directional threshold.
9. The rehabilitation training data interpretability assessment method according to claim 1, characterized in that, In step S7, the kinematic stability pathway includes a directed relationship between the velocity variation coefficient and the number of sub-motions and the pause ratio, respectively. The cognitive accuracy pathway includes a directed relationship between the normalized root mean square error and the hit rate, and between the hit rate and the time delay.
10. A rehabilitation training data interpretability assessment system based on a dual-pathway causal framework, characterized in that, include: The training data acquisition unit is used to acquire patient-related training data and reference task data and to construct historical assessment labels. It is also used to acquire training data and corresponding reference task data of the subject to be assessed during the current training process. The preprocessing unit is used to denoise, smooth, and discretely differentiate the training data to generate velocity and acceleration sequences. Feature building unit, used to construct three-domain, nine-item structured features; The teacher model unit is used to perform predictions and generate soft labels during the training phase. Distillation rule unit, used to construct rule paths and extract rule thresholds based on rule tree model; The association mining unit is used to perform causal relationship mining, Bootstrap resampling aggregation, and stable association filtering to build a robust association graph; The results output unit is used to output the evaluation results of the training phase, rule paths, dual-pathway mechanism profiles, and training suggestions.
11. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, which, when executed by the processor, implements the rehabilitation training data interpretable assessment method according to any one of claims 1 to 9.
12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the rehabilitation training data interpretable assessment method according to any one of claims 1 to 9.