A stroke rehabilitation interpretable evaluation method based on a confidence rule base

Abstract

The application relates to a stroke rehabilitation interpretable evaluation method based on a confidence rule base, and belongs to the technical field of medical data processing and intelligent evaluation. The method acquires original motion data of stroke rehabilitation training, reference evaluation data and label information, carries out denoising filtering, time synchronization, action segment division and normalization processing, and constructs a multi-dimensional rehabilitation motion feature system; an initial soft score is generated by using a first evaluation model; a confidence rule base model is constructed based on the initial soft score, label information and prior rules of rehabilitation experts, confidence inference results and rule paths are output through rule matching, rule activation and confidence fusion; the appearance frequency and direction consistency of candidate feature relationships are statistically sampled by using self-service resampling, stable relationships are screened, and a motion mechanism diagram is constructed. The method outputs evaluation results, rule paths, mechanism profiles and rehabilitation training parameter adjustment prompt information, and is used for assisting in rehabilitation training data evaluation and explanation.

Description

Technical Field

[0001] This invention belongs to the field of medical and health data processing, intelligent assessment and knowledge reasoning technology, and specifically relates to an interpretable assessment method for stroke rehabilitation based on a confidence rule base. Background Technology

[0002] During stroke rehabilitation training, devices such as rehabilitation robots, wearable sensors, visual acquisition devices, force sensors, and encoders can collect multi-source motion data of the person being evaluated, including posture, position, angular velocity, acceleration, force, and trajectory. This data reflects the degree of completion of rehabilitation exercises, motion stability, trajectory control ability, limb coordination, and abnormal compensatory behaviors.

[0003] Current stroke rehabilitation assessment methods typically include manual scale assessments and automated assessments based on data-driven models. Manual scale assessments rely on the experience of rehabilitation physicians; while clinically understandable, they suffer from long evaluation cycles, significant subjective variability, and difficulty in fine-grained analysis of the training process. Data-driven assessment methods can utilize sensor data for automated scoring or grade prediction, but their outputs are usually presented as single classification results, scores, or probabilities, making it difficult to explain which specific motor characteristics, rule conditions, and training mechanisms led to the assessment result.

[0004] Furthermore, while some machine learning models can improve assessment accuracy, their internal decision-making processes are opaque, making it difficult to demonstrate to rehabilitation physicians the rule paths related to indicators such as movement completion rate, trajectory deviation, trajectory smoothness, and compensatory movement index, and also making it difficult to explain the stable correlations between different movement characteristics. Therefore, in rehabilitation training data analysis scenarios, simply relying on black-box models to output assessment levels cannot fully meet the needs for interpretable assessment, process tracking, and training parameter adjustment prompts.

[0005] Furthermore, existing methods typically use rehabilitation assessment results as the final output, lacking a structured expression of the reasons behind these results. For example, existing methods struggle to simultaneously output assessment results, rule pathways, mechanistic profiles, and rehabilitation training parameter adjustment prompts, which hinders data management during rehabilitation training, assistance in adjusting training programs, and tracking rehabilitation effects.

[0006] Therefore, there is an urgent need to provide an interpretable assessment method for stroke rehabilitation based on a confidence rule base, which can process stroke rehabilitation training data without directly outputting disease diagnosis conclusions, and form interpretable assessment results, rule paths, and motor mechanism diagrams, thereby improving the objectivity, traceability, and interpretability of rehabilitation training data assessment. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide an interpretable assessment method for stroke rehabilitation based on a confidence rule base, addressing the issues of insufficient interpretability of assessment results, unclear rule paths, lack of stability screening of motor feature relationships, and difficulty in generating training parameter adjustment prompts in existing stroke rehabilitation training data assessment methods.

[0008] The method described in this invention is used to process motion and assessment data generated during stroke rehabilitation training and generate interpretable assessment results. This method belongs to the category of data processing and auxiliary assessment methods; it does not directly output disease diagnostic conclusions, nor does it replace the diagnostic or treatment decisions of rehabilitation physicians.

[0009] To address the aforementioned technical problems, this invention employs the following technical solution: acquiring raw motion data, reference assessment data, and label information for stroke rehabilitation training to construct a training sample set; preprocessing the raw motion data to generate uniform-length posture sequences, velocity sequences, acceleration sequences, and trajectory sequences; constructing a multidimensional rehabilitation motion feature system based on the aforementioned sequences and reference assessment data; inputting the multidimensional rehabilitation motion feature system into a first assessment model to generate an initial soft score; constructing and training a confidence rule base model based on the initial soft score, label information, and prior rules from rehabilitation experts; generating confidence inference results, rule paths, decision logic, and threshold conditions through rule matching, rule activation, and confidence fusion; performing correlation mining on the multidimensional rehabilitation motion feature system, statistically analyzing the stability index of candidate feature relationships through self-sampling, and constructing a motion mechanism diagram by selecting candidate feature relationships that meet preset conditions; finally, processing the current rehabilitation training data of the subject to be assessed and outputting assessment results, rule paths, mechanism profiles, and rehabilitation training parameter adjustment prompts.

[0010] The raw motion data includes at least one of position, attitude, angular velocity, acceleration, force, and trajectory data acquired by at least one of wearable sensors, force sensors, encoders, vision acquisition devices, and rehabilitation robots.

[0011] The reference evaluation data includes at least one of standard action templates, reference trajectories, and historical evaluation records.

[0012] The labeling information includes at least one of the following: clinical scale grades, staged functional scores, and prior rule information annotated by rehabilitation physicians. The clinical scale grades or staged functional scores can be generated by at least one of the following: the Fugl-Meyer Motor Function Assessment Scale, the Brunnstrom Recovery Stage Assessment, the Modified Ashworth Scale, and manual scoring by rehabilitation physicians.

[0013] The preprocessing includes denoising filtering, time synchronization, outlier removal, motion segment segmentation, time normalization, and amplitude normalization. The denoising filtering includes at least one of low-pass filtering, moving average filtering, wavelet denoising, and Kalman filtering. The motion segment segmentation is performed based on at least one of velocity thresholds, trajectory start and end points, task trigger markers, and time windows.

[0014] The multidimensional rehabilitation exercise feature system includes at least time-domain features, frequency-domain features, kinematic features, dynamic features, and exercise quality features. Specifically, the time-domain features include average velocity, peak velocity, exercise duration, and pause time; the frequency-domain features include dominant frequency, frequency energy, and energy concentration; the kinematic features include joint angles, angular velocity, trajectory length, and trajectory deviation; the dynamic features include acceleration, impact, inertial force, and rate of change of force; and the exercise quality features include trajectory smoothness, coordination index, movement completion rate, path efficiency, and compensation movement index.

[0015] The first evaluation model is any one of the following: a one-dimensional convolutional neural network, a recurrent neural network, a long short-term memory network, a gated recurrent unit network, and a convolutional recurrent combination network. The first evaluation model has a length of... And the feature dimension is The time series feature matrix is ​​taken as input, and the output is... The probability distribution or continuous assessment score corresponding to each assessment level.

[0016] Each rule in the confidence rule base model includes a prerequisite attribute, a reference level for the prerequisite attribute, a rule weight, an attribute weight, and a posterior confidence level. The prerequisite attributes include at least two of the following: average speed, trajectory smoothness, coordination index, action completion rate, and compensating motion index. Rule activation is obtained by calculating the matching degree of the sample to be evaluated relative to each prerequisite attribute reference level, and the activation weights of each rule are normalized. Confidence fusion is performed by synthesizing the normalized rule activation weights with the posterior confidence levels of each rule to obtain the final confidence distribution.

[0017] The motion mechanism graph is constructed from stable candidate feature relationships. These stable candidate feature relationships are obtained through self-service resampling statistics, and the stability indicators include the frequency of occurrence and directional consistency of the candidate feature relationships. Candidate feature relationships that meet preset conditions are retained as edges in the motion mechanism graph, and candidate features serve as nodes in the motion mechanism graph.

[0018] The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the stroke rehabilitation interpretable assessment method based on confidence rule base described in the present invention.

[0019] The present invention also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the stroke rehabilitation interpretable assessment method based on a confidence rule base as described in this invention.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] First, by preprocessing and constructing a unified length sequence of multi-source motion data during stroke rehabilitation training, this invention can improve the comparability of data from different acquisition devices, different training movements, and different training durations.

[0022] Second, this invention constructs a multidimensional rehabilitation exercise feature system that includes time-domain features, frequency-domain features, kinematic features, dynamic features, and exercise quality features. This system can characterize the rehabilitation training process from aspects such as movement speed, trajectory deviation, movement smoothness, movement completion rate, and compensatory movements.

[0023] Third, this invention uses the first assessment model to generate an initial soft score, and then constructs a confidence rule base model based on the initial soft score, label information and prior rules of rehabilitation experts, which can take into account both the fitting ability of the data-driven model and the interpretability of the rule model.

[0024] Fourth, this invention outputs confidence inference results and rule paths through rule matching, rule activation, and confidence fusion, which can show users the key feature conditions, rule paths, and confidence distributions corresponding to the evaluation results, thereby improving the traceability of the evaluation process.

[0025] Fifth, this invention uses self-service resampling to statistically analyze the frequency and direction consistency of candidate feature relationships and constructs a motion mechanism graph, which can screen stable motion feature relationships and reduce the impact of accidental correlations on the interpretation results.

[0026] Sixth, this invention outputs assessment results, rule paths, mechanism profiles, and rehabilitation training parameter adjustment prompts, which can be used for rehabilitation training data-assisted assessment, training process recording, and training parameter adjustment reference, but does not directly output disease diagnosis conclusions. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall architecture of the stroke rehabilitation interpretable assessment system based on a confidence rule base, as presented in this invention.

[0028] Figure 2 This is a schematic diagram of the interpretable assessment method for stroke rehabilitation based on a confidence rule base, as described in this invention.

[0029] Figure 3 This is a schematic diagram of the motion mechanism diagram construction process of the present invention.

[0030] Figure 4 This is a schematic diagram of the rule reasoning process of the confidence rule base of this invention. Detailed Implementation

[0031] The present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the following specific embodiments are used to illustrate the technical solution of the present invention and are not intended to limit the scope of protection of the present invention. Without departing from the principles of the present invention, those skilled in the art can make adaptive adjustments to the type of data acquisition equipment, preprocessing method, evaluation model type, confidence rule base parameter training method, motion mechanism graph construction method, and output format.

[0032] The method described in this invention is executed by an electronic device to process patient motion data and assessment data generated during stroke rehabilitation training and generate interpretable assessment results. The assessment results, rule paths, mechanism profiles, and rehabilitation training parameter adjustment prompts output by the method are used for auxiliary assessment of rehabilitation training data and management of the training process, and do not directly output disease diagnosis conclusions.

[0033] like Figure 1 As shown, in one optional embodiment, the functional units executed by the electronic device of the present invention include a data acquisition unit 101, a preprocessing unit 102, a feature construction unit 103, a model processing unit 104, a rule reasoning unit 105, a mechanism graph construction unit 106, and a result output unit 107.

[0034] The data acquisition unit 101 is used to collect raw motion data, reference assessment data, and label information for stroke rehabilitation training. The preprocessing unit 102 is used to perform noise reduction filtering, time synchronization, outlier removal, motion segmentation, time normalization, and amplitude normalization on the raw motion data. The feature construction unit 103 is used to construct a multidimensional rehabilitation motion feature system. The model processing unit 104 is used to input the multidimensional rehabilitation motion feature system into the first assessment model and output an initial soft score. The rule reasoning unit 105 is used to construct and train a confidence rule base model and output confidence reasoning results, rule paths, decision logic, and threshold conditions. The mechanism graph construction unit 106 is used to statistically analyze the stability index of candidate feature relationships through self-sampling and construct a motion mechanism graph. The result output unit 107 is used to output assessment results, rule paths, mechanism profiles, and rehabilitation training parameter adjustment prompts.

[0035] like Figure 2 As shown, the method flow of this invention sequentially includes stroke rehabilitation training data acquisition, raw motion data preprocessing, construction of a multidimensional rehabilitation motion feature system, initial soft score generation, confidence rule base model construction and inference, motion mechanism diagram construction, and assessment output of the object to be assessed. These processes together form a complete data processing chain from rehabilitation training data collection to interpretable assessment result output.

[0036] In one optional implementation, raw motion data, reference assessment data, and labeling information for stroke rehabilitation training are first acquired to construct a training sample set. The raw motion data includes at least one of position, posture, angular velocity, acceleration, force, and trajectory data acquired by at least one of wearable sensors, force sensors, encoders, vision acquisition devices, and rehabilitation robots.

[0037] The reference assessment data includes at least one of standard movement templates, reference trajectories, and historical assessment records. The labeling information includes at least one of clinical scale grades, staged functional scores, and prior rule information annotated by the rehabilitation physician. The clinical scale grades or staged functional scores can be generated from at least one of the Fugl-Meyer Motor Function Assessment Scale, the Brunnstrom Recovery Stage Assessment, the Modified Ashworth Scale, and manual scoring by the rehabilitation physician.

[0038] In one alternative implementation, the first The original motion data sequence of each training sample is represented as:

[0039] (1)

[0040] in, Indicates the first The original motion data sequence of each training sample; Indicates the first The training sample at the th ... Motion data vectors at each sampling time; Indicates the training sample number; Indicates the sampling time sequence number; Indicates the first The total number of original sampling points for each training sample.

[0041] In one alternative implementation, the first The training sample at the th ... The motion data vector at each sampling time is represented as:

[0042] (2)

[0043] in, Indicates the first The training sample at the th ... Motion data vectors at each sampling time; Represents location data; Represents attitude data; Represents angular velocity data; Represents acceleration data; Represents force data; Represents trajectory data; Indicates the training sample number; Indicates the sampling time sequence number.

[0044] In one alternative implementation, the training sample set is represented as:

[0045] (3)

[0046] in, Represents the training sample set; Indicates the first The original motion data sequence of each training sample; Indicates the first Reference evaluation data corresponding to each training sample; Indicates the first Label information corresponding to each training sample; Indicates the first Prior rule information of rehabilitation experts corresponding to each training sample; Indicates the training sample number; This represents the total number of training samples.

[0047] After obtaining the training sample set, the original motion data is preprocessed to generate a uniform-length attitude sequence, velocity sequence, acceleration sequence, and trajectory sequence. The preprocessing includes noise reduction filtering, time synchronization, outlier removal, motion segment segmentation, time normalization, and amplitude normalization.

[0048] The denoising filtering includes at least one of low-pass filtering, moving average filtering, wavelet denoising, and Kalman filtering. The action segmentation is performed based on at least one of velocity threshold, trajectory start and end points, task trigger markers, and time windows. For rehabilitation training data with task trigger markers, the task trigger markers are used as the start or end point of the action; for data without task trigger markers, action segments are determined jointly based on velocity changes, trajectory changes, and duration windows.

[0049] In one alternative implementation, the original motion data sequence is time-normalized to a uniform length. Motion data sequence:

[0050] (4)

[0051] in, Indicates the first A time-normalized sequence of motion data from each training sample; Indicates the first The training sample at the th ... Motion data vectors at each normalized sampling time; Indicates the training sample number; Indicates the normalized sampling time sequence number; This represents the normalized sequence length.

[0052] In one alternative implementation, the location sequence, after time normalization, is represented as:

[0053] (5)

[0054] in, Indicates the first The time-normalized position sequence of each training sample; Indicates the first The training sample at the th ... The position vector at each normalized sampling time; Indicates the training sample number; Indicates the normalized sampling time sequence number; This represents the normalized sequence length.

[0055] In one alternative implementation, the velocity sequence is obtained by differencing the position sequence:

[0056] (6)

[0057] in, Indicates the first The training sample at the th ... The velocity vector corresponding to each normalized sampling time; Indicates the first The training sample at the th ... The position vector at each normalized sampling time; Indicates the first The training sample at the th ... The position vector at each normalized sampling time; This represents the time interval between adjacent normalized sampling times; Indicates the training sample number; This represents the normalized sampling time sequence number.

[0058] In one alternative implementation, the acceleration sequence is obtained by differencing the velocity sequence:

[0059] (7)

[0060] in, Indicates the first The training sample at the th ... The acceleration vector corresponding to each normalized sampling time; Indicates the first The training sample at the th ... The velocity vector corresponding to each normalized sampling time; Indicates the first The training sample at the th ... The velocity vector corresponding to each normalized sampling time; This represents the time interval between adjacent normalized sampling times; Indicates the training sample number; This represents the normalized sampling time sequence number.

[0061] In one optional implementation, outliers are removed or corrected. When the position, attitude, velocity, acceleration, or force data of a sampling point exceeds the outlier threshold of the corresponding channel, processing is performed using methods such as interpolation of nearby valid sampling points, sliding window value replacement, or sample fragment removal. After preprocessing, attitude sequences, velocity sequences, acceleration sequences, and trajectory sequences of uniform length are obtained.

[0062] After preprocessing, a multidimensional rehabilitation exercise feature system is constructed based on the posture sequence, velocity sequence, acceleration sequence, trajectory sequence, and reference evaluation data. This multidimensional rehabilitation exercise feature system includes at least time-domain features, frequency-domain features, kinematic features, dynamic features, and exercise quality features.

[0063] The time-domain features include average velocity, peak velocity, motion duration, and pause time. The frequency-domain features include dominant frequency, frequency-domain energy, and energy concentration. The kinematic features include joint angles, angular velocity, trajectory length, and trajectory deviation. The dynamic features include acceleration, impact force, inertial force, and rate of change of force. The motion quality features include trajectory smoothness, coordination index, motion completion rate, path efficiency, and compensation motion index.

[0064] In one alternative implementation, the average speed is expressed as:

[0065] (8)

[0066] in, Indicates the first The average speed of the training samples; This represents the normalized sequence length; Indicates the normalized sampling time sequence number; Indicates the first The training sample at the th ... The velocity vector corresponding to each normalized sampling time; Represents the velocity vector The modulus length; This indicates the training sample number.

[0067] In one alternative implementation, the trajectory deviation is expressed as:

[0068] (9)

[0069] in, Indicates the first Trajectory deviation of each training sample; This represents the normalized sequence length; Indicates the normalized sampling time sequence number; Indicates the first The training sample at the th ... The position vector at each normalized sampling time; Indicates the first The training sample at the th ... The reference trajectory position vector corresponding to each normalized sampling time; This indicates the distance between the actual position and the reference trajectory position; This indicates the training sample number.

[0070] In one alternative implementation, path efficiency is expressed as:

[0071] (10)

[0072] in, Indicates the first Path efficiency of each training sample; Indicates the first The training sample at the th ... The position vector at each normalized sampling time; Indicates the first The training sample at the th ... The position vector at each normalized sampling time; Indicates the first The training sample at the th ... The position vector at each normalized sampling time; Indicates the first The training sample at the th ... The position vector at each normalized sampling time; This represents the normalized sequence length; Indicates the normalized sampling time sequence number; This indicates a positive number used to avoid a denominator of zero; This indicates the training sample number.

[0073] In one alternative implementation, the trajectory smoothness is characterized by the discrete second-order difference of the velocity sequence:

[0074] (11)

[0075] in, Indicates the first The trajectory smoothness index of each training sample; This represents the normalized sequence length; Indicates the normalized sampling time sequence number; Indicates the first The training sample at the th ... The velocity vector corresponding to each normalized sampling time; Indicates the first The training sample at the th ... The velocity vector corresponding to each normalized sampling time; Indicates the first The training sample at the th ... The velocity vector corresponding to each normalized sampling time; The magnitude of the discrete second-order difference of the velocity sequence; This indicates the training sample number.

[0076] In one alternative implementation, the synergy index is characterized by the correlation coefficient between the upper limb joint angle sequence and the trunk angle sequence:

[0077] (12)

[0078] in, Indicates the first The synergistic index of each training sample; This represents the function for calculating the correlation coefficient. Indicates the first Upper limb joint angle sequences of training samples; Indicates the first A sequence of torso angles from a training sample; This indicates the training sample number.

[0079] In one alternative implementation, the action completion rate is expressed as:

[0080] (13)

[0081] in, Indicates the first Action completion rate of each training sample; Indicates the first The actual completed trajectory length of each training sample; Indicates the first The length of the reference trajectory for each training sample; Indicates the first The target arrival state variable for each training sample; Indicates the weight of the trajectory length term; Indicates the weight of the target arrival status item; This indicates a positive number used to avoid a denominator of zero; This indicates the training sample number.

[0082] In one alternative implementation, the compensating motion index is expressed as:

[0083] (14)

[0084] in, Indicates the first Compensation motion index for each training sample; Indicates the first The amplitude of trunk displacement in each training sample; Indicates the first Shoulder abduction angle of each training sample; Indicates the first Duration of abnormal coordinated actions in each training sample; The weights corresponding to the amplitude of trunk displacement; This indicates the weight corresponding to the shoulder joint abduction angle; The weights corresponding to the duration of abnormal coordinated actions; This indicates the training sample number.

[0085] In one alternative implementation, the above features are combined to form the first Multidimensional rehabilitation exercise feature vectors of training samples:

[0086] (15)

[0087] in, Indicates the first The multidimensional rehabilitation exercise feature vector corresponding to each training sample; Indicates average speed; Indicates trajectory deviation; Indicates path efficiency; Indicators representing trajectory smoothness; Indicates the synergy index; Indicates the completion rate of the action; The ellipsis indicates the compensating motion index; the ellipsis indicates other time-domain features, frequency-domain features, kinematic features, dynamic features, or motion quality features. This indicates the training sample number.

[0088] In one alternative implementation, the feature vectors of multiple training samples are combined to form a feature matrix:

[0089] (16)

[0090] in, This represents the feature matrix corresponding to the training samples; Indicates the first Multidimensional rehabilitation exercise feature vectors of training samples; Indicates the first Multidimensional rehabilitation exercise feature vectors of training samples; Indicates the first Multidimensional rehabilitation exercise feature vectors of training samples; Indicates the total number of training samples; superscript This indicates transpose.

[0091] After obtaining the multidimensional rehabilitation exercise feature system, the system is input into the first assessment model to output an initial soft score. The initial soft score includes the probability distribution of different assessment levels or continuous assessment scores.

[0092] The first evaluation model is any one of the following: a one-dimensional convolutional neural network, a recurrent neural network, a long short-term memory network, a gated recurrent unit network, and a convolutional recurrent combination network. The first evaluation model has a length of... And the feature dimension is The time series feature matrix is ​​taken as input, and the output is... The probability distribution corresponding to each assessment level.

[0093] In one alternative implementation, the time series feature matrix input to the first evaluation model is represented as:

[0094] (17)

[0095] in, Indicates the first Each training sample is input into the time series feature matrix of the first evaluation model; Represents the real number field; This represents the normalized sequence length; This represents the feature dimension corresponding to each normalized sampling time. This indicates the training sample number.

[0096] In one alternative implementation, the first evaluation model outputs an initial soft score:

[0097] (18)

[0098] in, Indicates the first Initial soft scores for each training sample; The parameter is The first evaluation model; This represents the trainable parameters of the first evaluation model; Indicates the first Each training sample is input into the time series feature matrix of the first evaluation model; This indicates the training sample number.

[0099] In one alternative implementation, when the initial soft score is a probability distribution, the initial soft score is expressed as:

[0100] (19)

[0101] in, Indicates the first Initial soft scores for each training sample; Indicates the first The training sample belongs to the first... The probability of each assessment level; Indicates the assessment level number; Indicates the total number of assessment levels; This indicates the training sample number.

[0102] In one optional implementation, the first evaluation model is trained based on the label information of the training samples, such that the difference between the initial soft score output by the first evaluation model and the label information meets preset training conditions. The initial soft score is used to provide a data-driven evaluation reference for the confidence rule base model.

[0103] After obtaining the initial soft score, a confidence rule base model is constructed and trained based on the initial soft score, label information, and prior rules of rehabilitation experts to generate confidence inference results, rule paths, decision logic, and threshold conditions.

[0104] Each rule in the confidence rule base model includes a prerequisite attribute, a reference level for the prerequisite attribute, a rule weight, an attribute weight, and a posterior confidence level. The prerequisite attribute includes at least two of the following: average speed, trajectory smoothness, coordination index, action completion rate, and compensating motion index.

[0105] In one alternative implementation, the first A confidence rule is expressed as:

[0106] (20)

[0107] in, Indicates the first Confidence rules; Indicates the rule number; Indicates the first The first rule Each prerequisite attribute references the level condition; Indicates the prerequisite attribute number; Indicates the first The number of prerequisite attributes in a rule; This indicates a logical AND relationship; This represents the rule mapping relationship from the reference level conditions of the premise attributes to the posterior confidence distribution; Indicates the first One assessment level; Indicates the first Rule number 1 Posterior confidence level of each assessment level; Indicates the assessment level number; This indicates the total number of assessment levels.

[0108] In one alternative implementation, the sample to be evaluated is relative to the first... The first prerequisite attribute The matching degree of each reference level is expressed as:

[0109] (twenty one)

[0110] in, Indicates the sample to be evaluated relative to the first The first prerequisite attribute The degree of matching for each reference level; Indicates the sample to be evaluated is in the 1st month. Feature values ​​on each prerequisite attribute; Indicates the first The first prerequisite attribute One reference level center value; Indicates the first The first prerequisite attribute One reference level width parameter; This indicates a positive number used to avoid a denominator of zero; Indicates the prerequisite attribute number; Indicates the reference level number.

[0111] In one alternative implementation, the first The unnormalized activation weights of the rule are represented as follows:

[0112] (twenty two)

[0113] in, Indicates the first The unnormalized activation weights of the rule; Indicates the first The rule weight of each rule; Indicates the rule number; Indicates the prerequisite attribute number; Indicates the first The number of prerequisite attributes in a rule; Indicates the sample to be evaluated is in the 1st month. Rule number 1 The degree of matching on the prerequisite attributes; Indicates the first The attribute weight of each prerequisite attribute; This indicates a series of multiplication operations.

[0114] In one alternative implementation, the rule activation weights are normalized:

[0115] (twenty three)

[0116] in, Indicates the first Normalized activation weights for each rule; Indicates the first The unnormalized activation weights of the rule; Indicates the first The unnormalized activation weights of the rule; Indicates the rule number; This represents the total number of rules in the confidence rule base; This indicates a positive number used to avoid a denominator of zero.

[0117] In one alternative implementation, the confidence fusion result is expressed as:

[0118] (twenty four)

[0119] in, Indicates the first confidence fusion. The confidence level corresponding to each assessment level; Indicates the rule number; This represents the total number of rules in the confidence rule base; Indicates the first Normalized activation weights for each rule; Indicates the first Rule number 1 Posterior confidence level of each assessment level; This indicates the evaluation level number.

[0120] In one alternative implementation, the final evaluation grade is expressed as:

[0121] (25)

[0122] in, Indicates the final assessment level; This operation represents the selection of the assessment rank number that maximizes the corresponding confidence level. Indicates the assessment level number; Indicates the first confidence fusion. The confidence level corresponding to each assessment level.

[0123] In one alternative implementation, the training objective of the confidence rule base model includes making the confidence fusion result approximate the label information and ensuring that the confidence fusion result is consistent with the initial soft score output by the first evaluation model:

[0124] (26)

[0125] in, This represents the total loss function of the confidence rule base model; Indicates the label-supervised loss weights; This represents the weight of the soft scoring consistency loss; This represents the loss weights constrained by the rule parameters; Indicates the confidence fusion result With tag information The losses between; Indicates the confidence fusion result Compared with the initial soft score The losses between; This represents the loss due to rule parameter constraints. This represents the set of parameters for the confidence rule base model.

[0126] In one optional implementation, the rule path includes the prerequisite attributes of the activated rule, the reference level of the prerequisite attributes, the rule weight, the attribute weight, the activation weight, and the posterior confidence level. The rule path is used to interpret the process of forming the evaluation results. For example, when the action completion rate is at a low reference level, the trajectory deviation is at a high reference level, and the compensation movement index is at a high reference level, the corresponding rule path points to a low rehabilitation movement performance level; when the trajectory smoothness is at a high reference level, the path efficiency is at a high reference level, and the compensation movement index is at a low reference level, the corresponding rule path points to a high rehabilitation movement performance level.

[0127] like Figure 4 As shown, the confidence rule base rule reasoning process includes current evaluation feature input, premise attribute reference level matching, rule base call, rule activation, activation weight normalization, confidence fusion, confidence reasoning result output, and rule path output.

[0128] While constructing the confidence rule base model, the association mining is performed on the multidimensional rehabilitation exercise feature system. The stability index of candidate feature relationships is statistically analyzed through self-service resampling, and candidate feature relationships that meet preset conditions are selected to construct the exercise mechanism diagram.

[0129] In one alternative implementation, the training sample set is executed. Each autopilot resampling step yields a subset of samples, and candidate feature relationships are computed on each subset. These candidate feature relationships can be obtained based on correlation analysis, conditional correlation analysis, structure learning, or directed relation screening methods.

[0130] In one alternative implementation, the frequency of occurrence of candidate feature relationships is expressed as:

[0131] (27)

[0132] in, Representing candidate feature relationships The frequency of occurrence; Indicates the starting feature node of the candidate relationship; Represents the target feature node of the candidate relationship; Indicates the number of self-service resampling attempts; Indicates the first Second self-service resampling; Indicates an indicator function; Represents feature nodes With feature nodes Candidate relationships between them; Indicates the first The candidate relation set obtained by the second self-sampling.

[0133] In one alternative implementation, the directional consistency of candidate feature relationships is represented as:

[0134] (28)

[0135] in, Representing candidate feature relationships Consistency in direction; The direction is indicated by the feature nodes. Pointing to feature nodes The number of times it appears; The direction is indicated by the feature nodes. Pointing to feature nodes The number of times it appears; This represents the function that takes the maximum value. This indicates a positive number used to avoid a denominator of zero.

[0136] In one optional implementation, when the frequency of occurrence of a candidate feature relationship is not less than a first preset threshold and the directional consistency is not less than a second preset threshold, the candidate feature relationship is retained as a stable relationship in the motion mechanism diagram.

[0137] (29)

[0138] in, Represents feature nodes With feature nodes Candidate relationships between them; This represents the set of stable relations obtained through filtering. Representing candidate feature relationships The frequency of occurrence; Representing candidate feature relationships Consistency in direction; This indicates the first preset threshold. This indicates the second preset threshold. Indicates the starting feature node of the candidate relationship; This represents the target feature node of the candidate relationship.

[0139] In one alternative implementation, the motion mechanism diagram is represented as follows:

[0140] (30)

[0141] in, Diagram showing the motion mechanism; This represents the set of feature nodes in the motion mechanism graph; This represents the set of stable relationships in the motion mechanism diagram.

[0142] The feature nodes in the motion mechanism graph include at least two of the following: average velocity, peak velocity, motion duration, pause time, trajectory deviation, trajectory smoothness, path efficiency, coordination index, action completion rate, and compensating motion index. Stable relationship edges are used to represent stable candidate feature relationships obtained through self-service resampling.

[0143] like Figure 3 As shown, the process of constructing the motion mechanism diagram includes input of a multidimensional rehabilitation motion feature system, self-service resampling, candidate relationship mining, stability statistics, preset condition judgment, stable relationship screening, and output of the motion mechanism diagram.

[0144] After completing the construction of the first assessment model, the confidence rule base model, and the motion mechanism diagram, the motion data of the subject under assessment during the current rehabilitation training process is obtained, and the current assessment features are generated according to the same preprocessing method as the training stage. The current assessment features are input into the first assessment model and the confidence rule base model, and the assessment results, rule paths, mechanism profiles, and rehabilitation training parameter adjustment prompts are output in combination with the motion mechanism diagram.

[0145] In one alternative implementation, the current evaluation feature is represented as:

[0146] (31)

[0147] in, This indicates the current assessment characteristics corresponding to the current rehabilitation training process of the subject to be assessed; Indicates the current average speed; Indicates the current trajectory deviation; Indicates the efficiency of the current path; This indicates the smoothness index of the current trajectory; Indicates the current synergy index; Indicates the current action completion rate; The ellipsis indicates the current compensating motion index; the ellipsis indicates other current time-domain features, frequency-domain features, kinematic features, dynamic features, or motion quality features.

[0148] In one alternative implementation, the result output is represented as follows:

[0149] (32)

[0150] in, Indicates the set of results to be output; Indicates the evaluation results; Indicate the confidence distribution; Indicates the rule path; Representation of the mechanism profile; This indicates a prompt message for adjusting rehabilitation training parameters.

[0151] The evaluation results include evaluation levels, continuous evaluation scores, or confidence distributions for each evaluation level. The rule path includes the main activation rule, corresponding prerequisite attributes, reference level, rule weight, attribute weight, activation weight, and posterior confidence. The mechanism profile includes the feature states corresponding to the nodes in the motion mechanism graph in the current sample, and the association explanations corresponding to the stable relation edges.

[0152] The rehabilitation training parameter adjustment prompts include adjustments for training speed, target trajectory range, number of repetitions, resistance level, or assistive force level. These prompts are for auxiliary evaluation of rehabilitation training data and management of the training process; they are not intended as disease diagnoses, nor do they replace the professional judgment of rehabilitation physicians.

[0153] In one alternative implementation, when the action completion rate is lower than the corresponding reference level, the trajectory deviation is higher than the corresponding reference level, and the compensation motion index is higher than the corresponding reference level, the result output unit 107 generates prompt information to reduce the difficulty of the target trajectory, reduce the training speed, or increase the level of assistance.

[0154] In one alternative implementation, when the trajectory smoothness is lower than the corresponding reference level, the pause time is higher than the corresponding reference level, and the speed change is unstable, the result output unit 107 outputs a mechanism profile related to motion stability and generates prompt information on training parameters that focus on motion continuity.

[0155] In one alternative implementation, when the path efficiency is higher than the corresponding reference level, the trajectory deviation is lower than the corresponding reference level, and the compensation exercise index is lower than the corresponding reference level, the result output unit 107 outputs the evaluation result corresponding to the higher rehabilitation exercise performance level and generates a prompt message to maintain the current training parameters or appropriately increase the training difficulty.

[0156] This invention can be applied to data processing of different rehabilitation training movements, including at least one of upper limb extension, elbow flexion and extension, shoulder abduction, wrist dorsiflexion and palmar flexion, forearm pronation and supination, and target reaching tasks.

[0157] The present invention also provides an electronic device, which includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the above-described interpretable assessment method for stroke rehabilitation based on a confidence rule base.

[0158] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described interpretable assessment method for stroke rehabilitation based on a confidence rule base.

[0159] The above specific embodiments are only used to illustrate the technical solution of the present invention and are not intended to limit the scope of protection of the present invention. Any equivalent substitutions, modifications, or improvements made to the data acquisition method, preprocessing method, feature construction method, first evaluation model type, confidence rule base model parameters, motion mechanism graph construction method, and result output method within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for interpretable assessment of stroke rehabilitation based on a confidence rule base, characterized in that, Performed by electronic devices, this process handles stroke rehabilitation training data and generates interpretable assessment results without directly outputting a disease diagnosis. It includes the following steps: S1. Obtain raw motion data, reference evaluation data, and label information to construct a training sample set; S2. Preprocess the raw motion data to generate a uniform-length attitude sequence, velocity sequence, acceleration sequence, and trajectory sequence; S3. Construct a multidimensional rehabilitation exercise feature system based on the sequence and reference evaluation data; S4. Input the multidimensional rehabilitation exercise feature system into the first assessment model and output the initial soft score; S5. Based on the initial soft score, label information and prior rules of rehabilitation experts, construct and train a confidence rule base model to generate confidence inference results and rule paths; S6. Perform correlation mining on the multidimensional rehabilitation exercise feature system, screen stable candidate feature relationships through self-service resampling and construct an exercise mechanism graph; S7. Obtain the current motion data of the object to be evaluated, generate the current evaluation features, input the current evaluation features into the first evaluation model and the confidence rule base model, and output the evaluation results, rule paths, mechanism profiles and rehabilitation training parameter adjustment prompts in combination with the motion mechanism diagram.

2. The method according to claim 1, characterized in that, In step S1, the raw motion data includes at least one of position, posture, angular velocity, acceleration, force, and trajectory data acquired by at least one of wearable sensors, force sensors, encoders, vision acquisition devices, and rehabilitation robots; the reference assessment data includes at least one of standard movement templates, reference trajectories, and historical assessment records; the labeling information includes at least one of clinical scale grades, staged functional scores, and prior rule information annotated by rehabilitation physicians, wherein the clinical scale grades or staged functional scores are generated by at least one of the Fugl-Meyer Motor Function Rating Scale, the Brunnstrom Recovery Stage Assessment, the Modified Ashworth Scale, and manual scoring by rehabilitation physicians.

3. The method according to claim 1, characterized in that, In step S2, the preprocessing includes denoising filtering, time synchronization, outlier removal, action segment segmentation, time normalization, and amplitude normalization. The denoising filtering includes at least one of low-pass filtering, moving average filtering, wavelet denoising, and Kalman filtering. The action segment segmentation is performed based on at least one of velocity threshold, trajectory start and end points, task trigger markers, and time windows.

4. The method according to claim 1, characterized in that, In step S3, the multidimensional rehabilitation exercise feature system includes at least time-domain features, frequency-domain features, kinematic features, dynamic features, and exercise quality features; wherein, the time-domain features include average velocity, peak velocity, exercise duration, and pause time; the frequency-domain features include dominant frequency, frequency-domain energy, and energy concentration; the kinematic features include joint angle, angular velocity, trajectory length, and trajectory deviation; the dynamic features include acceleration, impact, inertial force, and rate of change of force; and the exercise quality features include trajectory smoothness, synergy index, movement completion rate, path efficiency, and compensatory movement index.

5. The method according to claim 4, characterized in that, The trajectory smoothness is characterized by the discrete second-order difference of the velocity sequence or the integral value of the jerk; the coordination index is characterized by the correlation coefficient between the upper limb joint angle sequence and the trunk angle sequence; the action completion rate is determined by the actual completed trajectory length, the reference trajectory length, and the target arrival state; the compensation motion index is characterized by at least one of the trunk displacement amplitude, shoulder joint abduction angle, and abnormal coordination action duration.

6. The method according to claim 1, characterized in that, In step S4, the first evaluation model is any one of a one-dimensional convolutional neural network, a recurrent neural network, a long short-term memory network, a gated recurrent unit network, and a convolutional recurrent combination network; the first evaluation model takes a time series feature matrix of length T and feature dimension d as input and outputs the probability distributions corresponding to K evaluation levels.

7. The method according to claim 1, characterized in that, In step S5, each rule in the confidence rule base model includes a premise attribute, a premise attribute reference level, a rule weight, an attribute weight, and a posterior confidence level. The premise attribute includes at least two of the following: average speed, trajectory smoothness, coordination index, action completion rate, and compensating motion index. Rule activation is obtained by calculating the matching degree of the sample to be evaluated relative to the reference level of each premise attribute, and the activation weight of each rule is normalized. Confidence fusion synthesizes evidence by combining the normalized rule activation weights with the posterior confidence of each rule to obtain the final confidence distribution.

8. 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 method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 7.

Images