Pain rehabilitation management system based on intelligent evaluation algorithm and full-process interaction

Abstract

The application relates to the technical field of action capture recognition, and provides a pain rehabilitation management system based on intelligent evaluation algorithms and full-process interaction. The application constructs a standardized information element library, uniformly codes and hierarchically labels field professional data, forms training samples including evaluation characteristic vectors, safety constraint vectors and target labels, and guarantees the professionalism and standardization of a rehabilitation plan. A rehabilitation evaluation plan generation deep learning model is constructed and trained to generate a personalized rehabilitation plan, so that excessive dependence on artificial experience is eliminated. A motion capture standardized comparison and scoring model is constructed, a standardized database is used to monitor a patient's rehabilitation training process in real time, the completion rate and operation accuracy of training actions can be accurately identified, and full-process interactive evaluation of doctors and patients is realized. Iterative feedback optimization is carried out by combining evaluation data and patient needs, a complete rehabilitation digital closed loop is formed, and the overall effect of pain rehabilitation is effectively improved.

Description

Technical Field

[0001] This application relates to the field of motion capture and recognition technology, and in particular to a pain rehabilitation management system based on intelligent assessment algorithms and full-process interaction. Background Technology

[0002] Pain rehabilitation aims to relieve pain, restore bodily functions, and improve patients' quality of life. It adheres to the principles of multidisciplinary, personalized, and non-pharmacological intervention, employing a model of precise etiological diagnosis, comprehensive intervention, priority functional restoration, dynamic program adjustment, and patient self-management. Currently, the core challenges and difficulties in implementing pain rehabilitation lie in standardized rehabilitation assessment, professional treatment, and comprehensive process management. Rehabilitation effectiveness highly depends on the therapist's professional competence and the patient's adherence to training.

[0003] With the digital development of the rehabilitation industry, various rehabilitation management systems have been gradually applied in clinical practice, realizing basic functions such as rehabilitation record management, process recording, data statistics, and doctor-patient collaboration, effectively standardizing rehabilitation processes and improving the efficiency of basic diagnosis and treatment. However, existing systems have significant technical shortcomings for pain-specific rehabilitation and have not yet formed a complete digital rehabilitation closed loop.

[0004] Existing technologies cannot intelligently generate standardized and professional rehabilitation plans based on pain classification assessment results, and the quality of the plans is easily affected by the professionalism of the personnel. At the same time, there is a lack of a two-way interactive evaluation mechanism between doctors and patients, making it impossible to accurately control the completion rate and operational accuracy of rehabilitation training. Furthermore, the efficacy assessment relies too much on patients' subjective scale scores and lacks objective quantitative evaluation basis for the recovery of bodily functions, resulting in inconsistent implementation effects of pain rehabilitation plans and insufficient accuracy in efficacy assessment, which seriously restricts the digital, standardized, and intelligent development of pain rehabilitation. Summary of the Invention

[0005] Therefore, it is necessary to provide a pain rehabilitation management system based on intelligent assessment algorithms and full-process interaction to address the aforementioned technical problems.

[0006] A pain rehabilitation management system based on intelligent assessment algorithms and full-process interaction, the system comprising: The training sample construction module is used to build a professional rehabilitation standardized information element database for pain diseases; it collects domain professional data from disease groups and health groups, and performs unified encoding, hierarchical labeling and safety constraint processing on the unstructured representations in the domain professional data to generate training samples, which include evaluation feature vectors, safety constraint vectors and target labels. The rehabilitation plan model building module is used to construct a deep learning model for generating rehabilitation assessment plans, which consists of a feature encoding network and a rule retrieval network, and to train the deep learning model for generating rehabilitation assessment plans using training samples. The motion capture comparison model building module is used to collect standard motion data samples, establish a standardized database of rehabilitation process operations and rehabilitation training movements related to pain diseases, and build and train a motion capture standardized comparison and scoring model composed of a motion capture evaluation model and a deep learning model. The execution module is used to collect patient information and rehabilitation-related information in real time, generate rehabilitation plans through a deep learning model of rehabilitation assessment plan generation, and dynamically monitor the execution of rehabilitation plans through motion capture standardization comparison and scoring models, generate evaluation data and output adjustment suggestions. The feedback adjustment module is used to adjust the rehabilitation plan for the next cycle based on evaluation data and patient needs.

[0007] In one embodiment, the professional rehabilitation standardized information element database for pain diseases includes a basic information sub-database, a pain characteristic sub-database, a functional assessment sub-database, a contraindication sub-database, a treatment goal sub-database, and a rehabilitation prescription sub-database.

[0008] In one embodiment, domain-specific data from disease and health groups are collected, and unstructured representations in the domain-specific data are uniformly encoded, hierarchically labeled, and subjected to security constraints to generate training samples, including: Collect domain-specific data from disease and health groups, and construct a pain rehabilitation terminology mapping table to map unstructured expressions in the domain-specific data into unified field codes; The data indicators in the domain-specific data are uniformly converted according to preset units and value ranges, and outlier checks are performed. For missing data in the domain-specific data, the missing type is identified. For single-point missing data in short-term continuous follow-up, interpolation between adjacent time points of the same patient is used for completion. For non-continuous missing data across patients, similarity-weighted completion is used for completion, and the missing marker field is retained. A standard timeline is established based on the initial assessment time. The enrollment assessment, treatment records, motion capture results, and reassessment data in the domain professional data are aligned according to a preset period window to form a time series sample that can be used for longitudinal analysis. Based on the preset labeling specifications, the disease group data are stratified and labeled according to pain type, disease stage, pain severity, degree of functional limitation and risk level, and an assessment feature vector is generated. The standardized rehabilitation plan list confirmed by experts is used as the target label to obtain the training sample.

[0009] In one embodiment, a deep learning model for generating rehabilitation assessment plans is constructed, consisting of a feature encoding network and a rule retrieval network. The deep learning model is then trained using training samples, including: A deep learning model for generating rehabilitation assessment plans, consisting of a feature encoding network and a rule retrieval network, is constructed. The evaluation feature vector is input into the feature encoding network to generate patient representation vectors. The rule retrieval network is based on patient representation vectors The top K rule entries and case plans with the highest similarity were retrieved from the normative knowledge base and historical case database, and the rule representation vector was obtained by similarity-weighted aggregation. ; Patient representation vector , rule representation vector and interactive features Perform splicing and fully connected transformations to obtain the fused representation. ; Based on fusion representation It generates a rehabilitation plan and calls a security constraint mask during the generation process to block items that conflict with the contraindications sub-library.

[0010] In one embodiment, the loss function used when training the deep learning model to generate the rehabilitation assessment plan is:

[0011] in, Generate a loss function for a deep learning model for the rehabilitation assessment plan; , , , For non-negative weights, satisfying ; Classify losses for planned projects; For the regression loss of treatment parameters; To constrain and punish losses for safety reasons; This results in a loss of consistency with the guidelines.

[0012] In one embodiment, standard movement data samples are collected to establish a standardized database of rehabilitation procedures and rehabilitation training movements associated with pain disorders, including: Standard motion data samples were collected, and human key points were extracted, motion start and end points were segmented, body coordinate system was normalized, left and right side mirroring was unified, noise points were removed, and abnormal frames were corrected to obtain a standard skeleton sequence. The dynamic time warping method is used to align the time sequence of multiple standard samples of the same action, extract template features, form standard action templates and allowable deviation ranges, and construct a standardized database of rehabilitation process operations and rehabilitation training actions related to pain diseases.

[0013] In one embodiment, the motion capture evaluation model is constructed according to the following steps: Acquire the video to be evaluated and perform motion recognition and motion phase segmentation; The key point sequence in the video to be evaluated is aligned frame by frame with the standard motion template. During the motion phase k, the angle deviations of the key joints are calculated respectively. Key trajectory spatial deviation Movement rhythm deviation and compensatory action deviation The values ​​are then mapped to the [0,1] interval using a standardized function, and the stage score is calculated.

[0014]

[0015] in, For stage scoring; For standardized functions; , , , For non-negative weights, satisfying ; The total score is calculated; Assign importance weights to each action phase.

[0016] In one embodiment, the deep learning model is constructed according to the following steps: Using domain-specific data, standard action data samples, and expert rating data from disease and health groups as training samples, a spatiotemporal feature extraction network consisting of skeleton sequence branches, video frame branches, cross-modal fusion modules, and multi-task output heads is constructed to obtain a deep learning model.

[0017] In one embodiment, the loss function used for training the motion capture standardization comparison and scoring model is:

[0018] in, The loss function for the motion capture standardized comparison and scoring model; , , , For non-negative weights, satisfying ; Loss is identified for action categories; Locate the losses during the critical phase; The regression loss for motion quality scoring; The loss is due to the constraint of consistency in expert scoring.

[0019] In one embodiment, patient information and rehabilitation-related information are collected in real time; a rehabilitation plan is generated through a deep learning model that generates a rehabilitation assessment plan; the implementation of the rehabilitation plan is dynamically monitored through motion capture standardization comparison and scoring models, evaluation data is generated, and adjustment suggestions are output; the method also includes: The rehabilitation plan is broken down into therapist execution forms and patient self-check-in forms.

[0020] The aforementioned pain rehabilitation management system, based on intelligent assessment algorithms and full-process interaction, builds a standardized information element database for professional rehabilitation of pain diseases through a training sample construction module. It uniformly encodes and hierarchically labels domain-specific data, forming training samples including assessment feature vectors, safety constraint vectors, and target labels. This solves the problem of inconsistent standards and varying levels of professionalism in traditional rehabilitation programs due to differences in personnel expertise, ensuring the professionalism and standardization of rehabilitation plans. By constructing and training a deep learning model for rehabilitation assessment plan generation that integrates feature encoding networks and rule retrieval networks, it intelligently generates personalized and standardized rehabilitation plans based on patient pain assessment characteristics, thereby reducing over-reliance on human experience. By constructing a standardized motion capture comparison and scoring model, it monitors the patient's rehabilitation training process in real time based on a standardized database. This accurately identifies the completion rate and operational accuracy of training movements, enabling full-process interactive evaluation between doctors and patients. This addresses the pain points of traditional rehabilitation execution processes lacking quantitative supervision and uncontrollable training effects, significantly improving patient compliance and the standardization of rehabilitation operations. By combining evaluation data with patient needs for iterative feedback optimization, a complete digital rehabilitation closed loop is formed, effectively improving the overall effectiveness of pain rehabilitation. Attached Figure Description

[0021] Figure 1 This is a structural block diagram of a pain rehabilitation management system based on intelligent assessment algorithms and full-process interaction in one embodiment; Figure 2 This is a framework diagram of a pain rehabilitation management system based on intelligent assessment algorithms and full-process interaction in one embodiment; Figure 3 This is a schematic diagram illustrating the role of a deep learning model in generating a rehabilitation assessment plan in one embodiment. Figure 4 This is a schematic diagram illustrating the role of the motion capture standardization comparison and scoring model in one embodiment. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0023] In one embodiment, such as Figure 1, Figure 2 As shown, a pain rehabilitation management system based on intelligent assessment algorithms and full-process interaction is provided. The system includes: The training sample construction module 901 is used to construct a professional rehabilitation standardized information element database for pain diseases; it collects domain professional data from disease groups and health groups, and performs unified encoding, hierarchical labeling and security constraint processing on unstructured representations in the domain professional data to generate training samples, which include evaluation feature vectors, security constraint vectors and target labels.

[0024] The rehabilitation plan model construction module 902 is used to construct a deep learning model for generating rehabilitation assessment plans, which consists of a feature encoding network and a rule retrieval network, and to train the deep learning model for generating rehabilitation assessment plans using training samples.

[0025] The motion capture comparison model construction module 903 is used to collect standard motion data samples, establish a standardized database of rehabilitation process operations and rehabilitation training movements associated with pain diseases, and construct and train a motion capture standardized comparison and scoring model composed of a motion capture evaluation model and a deep learning model.

[0026] The execution module 904 is used to collect patient information and rehabilitation-related information in real time, generate a rehabilitation plan through a deep learning model generated by the rehabilitation assessment plan, and dynamically monitor the execution of the rehabilitation plan through motion capture standardization comparison and scoring model, generate evaluation data and output adjustment suggestions.

[0027] The feedback adjustment module 905 is used to adjust the rehabilitation plan for the next cycle based on evaluation data and patient needs.

[0028] The aforementioned pain rehabilitation management system, based on intelligent assessment algorithms and full-process interaction, establishes a standardized information element database for professional rehabilitation of pain diseases through a training sample construction module. This database unifies and hierarchically labels domain-specific data, forming training samples that include assessment feature vectors, safety constraint vectors, and target labels. This addresses the issues of inconsistent standards and varying levels of professionalism in traditional rehabilitation programs due to differences in personnel expertise, ensuring the professionalism and standardization of rehabilitation plans. By constructing and training a deep learning model that integrates feature encoding and rule retrieval networks, a personalized and standardized rehabilitation plan is intelligently generated based on the patient's pain assessment characteristics, thus reducing over-reliance on human experience. Furthermore, by constructing a standardized motion capture comparison and scoring model, the system monitors the patient's rehabilitation training process in real time based on a standardized database. This accurately identifies the completion rate and operational accuracy of training movements, enabling full-process interactive evaluation between doctors and patients. This addresses the pain points of traditional rehabilitation execution processes lacking quantitative supervision and uncontrollable training outcomes, significantly improving patient compliance and the standardization of rehabilitation operations. Through iterative feedback optimization combining evaluation data and patient needs, a complete digital rehabilitation closed loop is formed, effectively improving the overall effectiveness of pain rehabilitation.

[0029] In one embodiment, the professional rehabilitation standardized information element database for pain diseases includes a basic information sub-database, a pain characteristic sub-database, a functional assessment sub-database, a contraindication sub-database, a treatment goal sub-database, and a rehabilitation prescription sub-database.

[0030] Specifically, the basic information sub-database includes information such as age, gender, disease course, and complications; the pain characteristics sub-database includes information on pain location, pain nature, triggering factors, relieving factors, and pain level; the functional assessment sub-database includes information on joint range of motion, muscle strength, gait / posture, daily activity ability, and scale scores; the contraindications sub-database includes information on acute injury, postoperative restrictions, and risks of comorbidities; the treatment goals sub-database includes short-term, medium-term, and long-term goals; and the rehabilitation prescription sub-database includes treatment items, frequency, intensity, duration, and precautions.

[0031] In one embodiment, domain-specific data from disease and health groups are collected, and unstructured representations in the domain-specific data are uniformly encoded, hierarchically labeled, and subjected to security constraints to generate training samples, including: Collect domain-specific data from disease and health groups, and construct a pain rehabilitation terminology mapping table to map unstructured expressions in the domain-specific data into unified field codes; The data indicators in the domain-specific data are uniformly converted according to preset units and value ranges, and outlier checks are performed. For missing data in the domain-specific data, the missing type is identified. For single-point missing data in short-term continuous follow-up, interpolation between adjacent time points of the same patient is used for completion. For non-continuous missing data across patients, similarity-weighted completion is used for completion, and the missing marker field is retained. A standard timeline is established based on the initial assessment time. The enrollment assessment, treatment records, motion capture results, and reassessment data in the domain professional data are aligned according to a preset period window to form a time series sample that can be used for longitudinal analysis. Based on the preset labeling specifications, the disease group data are stratified and labeled according to pain type, disease stage, pain severity, degree of functional limitation and risk level, and an assessment feature vector is generated. The standardized rehabilitation plan list confirmed by experts is used as the target label to obtain the training sample.

[0032] It should be noted that unstructured expressions include synonyms and abbreviations in electronic medical record texts; data indicators include scale scores, joint range of motion, muscle strength grades, and examination indicators. Weighted completion was performed using similarity-based methods for cases with similar pain types, disease stages, age ranges, and levels of functional limitation.

[0033] In this embodiment, by standardizing, cleaning, completing, aligning, and hierarchically labeling the domain-specific data sequentially, high-quality and high-precision training samples are provided for model training, significantly improving the model's accuracy and stability. By constructing a pain rehabilitation terminology mapping table, unified field encoding of unstructured professional data is achieved, and data indicators are uniformly converted and outlier checked, effectively eliminating invalid and erroneous data and ensuring the accuracy and standardization of basic data. Differentiated interpolation and similarity-weighted completion schemes are used for missing data in different scenarios, while retaining missing marker fields. This maximizes the restoration of the temporal characteristics of real rehabilitation data while preserving the original data features, avoiding data distortion caused by a single completion method. By constructing a standard timeline to align the temporal sequence of multiple types of rehabilitation data, a longitudinal temporal sample is formed, which can accurately support the dynamic analysis of the entire patient rehabilitation process, solving the shortcomings of traditional data's chaotic time dimension and inability to achieve dynamic tracking of the disease progression. Simultaneously, a standardized rehabilitation plan list confirmed by experts is used as the target label, thereby achieving precise quantification of the disease characteristics of pain patients.

[0034] In one embodiment, reference Figure 3 A deep learning model for generating rehabilitation assessment plans, consisting of a feature encoding network and a rule retrieval network, is constructed. This model is then trained using training samples, including: A deep learning model for generating rehabilitation assessment plans, consisting of a feature encoding network and a rule retrieval network, is constructed. The evaluation feature vector is input into the feature encoding network to generate patient representation vectors. The rule retrieval network is based on patient representation vectors The top K rule entries and case plans with the highest similarity were retrieved from the normative knowledge base and historical case database, and the rule representation vector was obtained by similarity-weighted aggregation. ; Patient representation vector , rule representation vector and interactive features Perform splicing and fully connected transformations to obtain the fused representation. ; Based on fusion representation It generates a rehabilitation plan and calls a security constraint mask during the generation process to block items that conflict with the contraindications sub-library.

[0035] Specifically, the rule retrieval network includes inputs of treatment guidelines, expert consensus, and historical high-quality case plans. The feature encoding network converts pain location, pain intensity, functional indicators, disease stage, and risk level into patient representation vectors; the plan item classification head outputs a set of treatment item candidates, and the parametric regression head outputs frequency, intensity, duration, and reassessment cycle.

[0036] In this embodiment, multi-dimensional representation fusion and weighted calculation are used to fully integrate individual patient characteristics and standardized rehabilitation rules, significantly improving the personalization and adaptability of the rehabilitation plan. Simultaneously, a safety constraint masking mechanism is introduced to automatically block contraindication and conflicting items, thereby mitigating rehabilitation risks and balancing the accuracy, standardization, and safety of the rehabilitation plan, achieving intelligent and safe generation of the rehabilitation plan.

[0037] In one embodiment, the loss function used when training the deep learning model to generate the rehabilitation assessment plan is:

[0038] in, Generate a loss function for a deep learning model for the rehabilitation assessment plan; , , , For non-negative weights, satisfying ; Classify losses for planned projects; For the regression loss of treatment parameters; To constrain and punish losses for safety reasons; This results in a loss of consistency with the guidelines.

[0039] Specifically Used to ensure the accuracy of treatment program selection; Used to constrain output errors in frequency, intensity, and duration; Used to penalize program items that conflict with contraindications, risk levels, or safety boundaries; It is used to penalize deviations from pre-set treatment guidelines, expert consensus, and high-quality case plans.

[0040] Regarding the training process, it should be noted that the recall, parameter error, and security conflict rate on the validation set are adjusted within a preset range. Once adjusted, these values ​​are used for model training and deployment. Model training is considered complete when the planned recall, parameter error, and security conflict rate on the validation set reach preset thresholds.

[0041] The rehabilitation plan checklist should include at least the order of treatment items, treatment frequency, duration of each training session, training difficulty level, follow-up assessment time, and precautions. To ensure feasibility, the system simultaneously outputs both therapist execution forms and patient home training forms, and retains key characteristics and constraints that trigger the plan for subsequent tracking and manual verification.

[0042] In one embodiment, standard movement data samples are collected to establish a standardized database of rehabilitation procedures and rehabilitation training movements associated with pain disorders, including: Standard motion data samples were collected, and human key points were extracted, motion start and end points were segmented, body coordinate system was normalized, left and right side mirroring was unified, noise points were removed, and abnormal frames were corrected to obtain a standard skeleton sequence. The dynamic time warping method is used to align the time sequence of multiple standard samples of the same action, extract template features, form standard action templates and allowable deviation ranges, and construct a standardized database of rehabilitation process operations and rehabilitation training actions related to pain diseases.

[0043] It should be noted that the template features include joint angles, angular velocity, trajectory smoothness, symmetry, compensation amplitude, stability, and completion time.

[0044] In this embodiment, by forming standard action templates and allowable deviation ranges, a standardized database of rehabilitation process operations and rehabilitation training actions related to pain diseases is constructed, thereby providing a basis for subsequent standard comparison.

[0045] In one embodiment, the motion capture evaluation model is constructed according to the following steps: Acquire the video to be evaluated and perform motion recognition and motion phase segmentation; The key point sequence in the video to be evaluated is aligned frame by frame with the standard motion template. During the motion phase k, the angle deviations of the key joints are calculated respectively. Key trajectory spatial deviation Movement rhythm deviation and compensatory action deviation The values ​​are then mapped to the [0,1] interval using a standardized function, and the stage score is calculated.

[0046]

[0047] in, For stage scoring; For standardized functions; , , , The weights are non-negative and satisfy the following conditions: ; The total score is calculated; Assign importance weights to each action phase, satisfying... .

[0048] Specifically, the critical joint angle deviation can be expressed as the average difference between the target joint angle and the standard template angle within this stage; the critical trajectory spatial deviation can be expressed as the average Euclidean distance between the normalized critical point spatial trajectory and the standard trajectory; the motion rhythm deviation can be expressed as the relative deviation between the actual stage duration and the standard duration; and the compensatory motion deviation can be expressed as the normalized deviation of abnormal displacement of non-target joints, trunk tilt, or abnormal changes in support points.

[0049] It should be noted that for the stage score, when the calculated result is less than 0, it is counted as 0, and when it is greater than 100, it is counted as 100.

[0050] The motion capture evaluation model constructed in this embodiment quantifies the deviations of joint angles, key trajectories, movement rhythms, and compensatory movements, and maps them uniformly to a standard range, which can comprehensively cover the core evaluation dimensions of pain rehabilitation movements. By combining an adaptive weight allocation mechanism, it distinguishes different types of deviations and the importance of each movement stage, and generates refined and quantifiable stage scores and overall scores. This effectively solves the problems of strong subjectivity, single evaluation dimensions, and inability to quantify the degree of deviation in traditional manual evaluation, and achieves objective, accurate, standardized, and quantitative assessment of the quality of rehabilitation movements.

[0051] The construction of the motion capture evaluation model also includes: Each deviation is compared with its corresponding threshold.

[0052] Specifically, when or When the threshold is exceeded, it is marked as an incomplete action; when the sequence of action stages is inconsistent with the standard procedure, it is marked as an incorrect action sequence; when When the speed exceeds the threshold, it is marked as abnormal; when the support point stability index is below the threshold, it is marked as unstable support; when If the threshold is exceeded, it is marked as excessive compensation.

[0053] In this embodiment, by comparing each deviation with the corresponding threshold, error type labels are formed, thereby accurately identifying various non-standard problems in rehabilitation training, such as incomplete movements, incorrect sequence, abnormal speed, unstable support, and excessive compensation, so as to achieve refined, standardized, and digital judgment of rehabilitation movement deviations.

[0054] In one embodiment, the deep learning model is constructed according to the following steps: Using domain-specific data, standard action data samples, and expert rating data from disease and health groups as training samples, a spatiotemporal feature extraction network consisting of skeleton sequence branches, video frame branches, cross-modal fusion modules, and multi-task output heads is constructed to obtain a deep learning model.

[0055] Specifically, the skeleton sequence branch takes human keypoint coordinates, joint connections, and keypoint confidence as input, and employs at least three layers of spatiotemporal graph convolutional modules for encoding. These spatiotemporal graph convolutional modules are used to extract joint coordination variations, motion amplitude, and continuous motion evolution features. Each spatiotemporal graph convolutional module includes a spatial graph convolutional unit based on the topological relationship of the human skeleton, a temporal convolutional unit along the time dimension, a residual connection unit, and a normalization unit. The video frame branch takes the sequence of motion video frames as input, employs a 3D convolutional encoding module to extract local spatiotemporal texture features, and then uses a temporal attention module to extract motion phase transitions and keyframe features. The cross-modal fusion module first converts the skeleton feature vector... With video feature vectors Projecting onto the same dimension, and then generating fused features through stitching, gating weight allocation, and cross-attention operations. Among them, the gating weights are used to automatically adjust the contributions of two types of features based on the confidence level of key point recognition and video sharpness. The multi-task output head is based on... The system outputs the probability of the action category, the location result of the action stage, the action quality score, and the risk warning result, thereby realizing the integrated processing of action recognition, execution quality evaluation, and risk warning.

[0056] In one embodiment, the loss function used for training the motion capture standardization comparison and scoring model is:

[0057] in, The loss function for the motion capture standardized comparison and scoring model; , , , For non-negative weights, satisfying ; Loss is identified for action categories; Locate the losses during the critical phase; The regression loss for motion quality scoring; Constraint loss for expert rating consistency.

[0058] The specific process of motion capture standardization comparison and scoring model is as follows: Figure 4 As shown.

[0059] Specifically Cross-entropy loss is used to calculate the difference between the predicted action category and the actual action category; Used to constrain the prediction error of the start point, end point and critical stage boundary of the action; The difference between the predicted score and the standard score is calculated using the mean squared error loss model. It can be expressed as the average of the squared differences between the model's predicted score and the independent scores from multiple experts, and is used to reduce the impact of subjective bias from a single rater.

[0060] For training the motion capture standardized comparison and scoring model, it should be noted that, based on the action recognition accuracy, stage localization error, scoring mean square error, and expert consistency index on the validation set, the weights are adjusted within a preset range and then fixed for final training. Training is considered complete when the model's action recognition accuracy, scoring mean square error, and consistency index with expert scoring on the validation set reach preset thresholds.

[0061] In the execution module, the implementation of the rehabilitation plan is dynamically monitored through motion capture standardization comparison and scoring models. Specifically, the professional rehabilitation standardization information element database for pain diseases is linked with the standardized database of rehabilitation process operations and rehabilitation training movements associated with pain diseases. The actual operation of therapists, the training movements of patients, and the phased functional indicators are compared to generate therapist execution completion rate, patient movement completion rate, risk movement warning results, and phased efficacy evaluation results. The therapist execution completion rate can be expressed as:

[0062] in, Execution completion rate for therapists; The number of therapist actions completed within the specified time and achieving a process score threshold; The number of therapist interventions required by the plan; The patient's completion rate of the action can be expressed as:

[0063] in, For patients' completion rate of actions; The number of training sessions completed within a specified period and in which the movement quality score reaches a threshold. To ensure the completion of training sessions, the phased efficacy evaluation results were calculated using a multimodal comprehensive scoring method. The system first standardized changes in pain scale, functional recovery, movement quality, and compliance to [specific parameters]. , , , Four categories of scores, and calculation of risk penalty items. ; This can be obtained based on the decrease in pain scale scores before and after the reassessment. Results can be obtained based on improvements in joint range of motion, muscle strength, gait, or daily activity ability. It can be obtained based on the total score of the movement, the stage score, or the decline in compensatory movements. According to , And the rate of timely attendance. It can be obtained based on high-risk compensatory movements, the number of contraindicated triggers, and the number of abnormal pain responses. The comprehensive score for each stage of efficacy can be expressed as:

[0064] in, The overall score for the stage of efficacy; , , , , To preset weights, , Used for security deduction. When When the safety threshold is exceeded, the system prioritizes outputting risk action warnings and reassessment prompts, rather than directly providing intensity upgrade suggestions.

[0065] Understandably, the results of the phased efficacy evaluation are not based solely on the score of a single action, but are comprehensively calculated by combining changes in the scale, changes in the quality of the action, the trend of functional recovery, adherence to the procedure, and the risk status.

[0066] In one embodiment, patient information and rehabilitation-related information are collected in real time; a rehabilitation plan is generated through a deep learning model that generates a rehabilitation assessment plan; the implementation of the rehabilitation plan is dynamically monitored through motion capture standardization comparison and scoring models, evaluation data is generated, and adjustment suggestions are output; the method also includes: The rehabilitation plan is broken down into therapist execution forms and patient self-check-in forms.

[0067] It should be noted that in the pain rehabilitation management system based on intelligent assessment algorithms and full-process interaction, the deep learning model for generating rehabilitation assessment plans and the standardized comparison and scoring model for motion capture are deployed on a unified system platform, realizing full-process interactive management from enrollment assessment, plan generation, execution reminders, motion evaluation to efficacy review and plan regeneration.

[0068] Patient information and rehabilitation-related information: Specifically, patient information includes basic patient information, medical history, pain characteristics, imaging examinations, scale results, and functional examination results; rehabilitation-related information includes the diagnosis type, risk level, and rehabilitation goals jointly confirmed by doctors, therapists, and nursing staff.

[0069] A deep learning model is used to generate rehabilitation plans based on the rehabilitation assessment plan. These plans include standardized short-term, long-term, and rehabilitation management plans corresponding to the type, stage, and risk level of the pain disorder. The rehabilitation plan is broken down into therapist execution forms and patient self-check-in forms. These forms are detailed down to the name of the movement, the number of movements, the rhythm of the movement, precautions, contraindications, and review points. Based on the therapist execution forms and patient self-check-in forms, the system dynamically monitors and generates interactive execution reminders and evaluations for both therapists and patients. Specifically, for cases of late execution, insufficient execution frequency, decreased movement quality, or the occurrence of high-risk compensatory movements, the system automatically sends reminders and records the improvement in execution after the reminder, thus forming compliance evaluation data.

[0070] The implementation of rehabilitation plans is dynamically monitored through standardized motion capture comparison and scoring models. Evaluation data is generated, and adjustment suggestions are output. Specifically, the standardized motion capture comparison and scoring model analyzes and evaluates the standardization of rehabilitation therapists' procedures and the execution of patients' rehabilitation training movements, and outputs targeted adjustment suggestions. These suggestions include reducing or increasing training intensity, replacing high-risk movements, increasing assistive support, extending warm-up time, and conducting earlier re-evaluations. These adjustment suggestions enable intervention during the implementation process.

[0071] Based on the output of the motion capture evaluation model, standardized comparison scoring results, and data from the scale assessment system platform, a multimodal evaluation of pain rehabilitation efficacy is conducted. This multimodal evaluation includes dimensions of pain improvement, functional recovery, movement quality, and compliance. Pain improvement is represented by the rate of change in pain scale data; functional recovery is represented by the rate of change in joint range of motion, muscle strength, gait, or daily activity ability; movement quality is represented by the standard movement matching score and the decline in compensatory movements; and compliance is represented by the completion rate and on-time attendance rate. The system weights and integrates these dimensions and sets safety thresholds. When a dimension improves but the safety threshold is triggered, a risk warning is prioritized over an intensity escalation suggestion, thereby improving the objectivity and safety of the efficacy evaluation.

[0072] The feedback adjustment module adjusts the rehabilitation plan for the next cycle based on evaluation data and patient needs. Specifically, it attributees the causes based on multimodal evaluation results, distinguishing whether the inadequate efficacy is due to insufficient compliance, incorrect movements, mismatched intensity settings, or contraindication risks. Then, it adjusts the training frequency, movement difficulty, treatment program combination, and re-evaluation cycle respectively, and retains the previous and previous versions of the plan and the corresponding efficacy results to form a traceable closed-loop optimization record.

[0073] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0074] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A pain rehabilitation management system based on intelligent assessment algorithms and full-process interaction, characterized in that, The system includes: The training sample construction module is used to build a professional rehabilitation standardized information element database for pain diseases; it collects domain professional data from disease groups and health groups, and performs unified encoding, hierarchical labeling and safety constraint processing on the unstructured representations in the domain professional data to generate training samples, which include evaluation feature vectors, safety constraint vectors and target labels. The rehabilitation plan model building module is used to construct a deep learning model for generating rehabilitation assessment plans, which consists of a feature encoding network and a rule retrieval network, and to train the deep learning model for generating rehabilitation assessment plans using training samples. The motion capture comparison model building module is used to collect standard motion data samples, establish a standardized database of rehabilitation process operations and rehabilitation training movements related to pain diseases, and build and train a motion capture standardized comparison and scoring model composed of a motion capture evaluation model and a deep learning model. The execution module is used to collect patient information and rehabilitation-related information in real time, generate rehabilitation plans through a deep learning model of rehabilitation assessment plan generation, and dynamically monitor the execution of rehabilitation plans through motion capture standardization comparison and scoring models, generate evaluation data and output adjustment suggestions. The feedback adjustment module is used to adjust the rehabilitation plan for the next cycle based on evaluation data and patient needs.

2. The pain rehabilitation management system based on intelligent assessment algorithm and full-process interaction as described in claim 1, characterized in that, The professional rehabilitation standardization information element database for pain diseases includes a basic information sub-database, a pain characteristic sub-database, a functional assessment sub-database, a contraindication sub-database, a treatment goal sub-database, and a rehabilitation prescription sub-database.

3. The pain rehabilitation management system based on intelligent assessment algorithm and full-process interaction as described in claim 1, characterized in that, Domain-specific data from disease and health groups were collected, and unstructured representations in the domain-specific data were uniformly encoded, hierarchically labeled, and subjected to security constraints to generate training samples, including: Collect domain-specific data from disease and health groups, and construct a pain rehabilitation terminology mapping table to map unstructured expressions in the domain-specific data into unified field codes; The data indicators in the domain-specific data are uniformly converted according to preset units and value ranges, and outlier checks are performed. For missing data in the domain-specific data, the missing type is identified. For single-point missing data in short-term continuous follow-up, interpolation between adjacent time points of the same patient is used for completion. For non-continuous missing data across patients, similarity-weighted completion is used for completion, and the missing marker field is retained. A standard timeline is established based on the initial assessment time. The enrollment assessment, treatment records, motion capture results, and reassessment data in the domain professional data are aligned according to a preset period window to form a time series sample that can be used for longitudinal analysis. Based on the preset labeling specifications, the disease group data are stratified and labeled according to pain type, disease stage, pain severity, degree of functional limitation and risk level, and an assessment feature vector is generated. The standardized rehabilitation plan list confirmed by experts is used as the target label to obtain the training sample.

4. The pain rehabilitation management system based on intelligent assessment algorithm and full-process interaction according to claim 3, characterized in that, A deep learning model for generating rehabilitation assessment plans, consisting of a feature encoding network and a rule retrieval network, is constructed. This model is then trained using training samples, including: A deep learning model for generating rehabilitation assessment plans, consisting of a feature encoding network and a rule retrieval network, is constructed. The evaluation feature vector is input into the feature encoding network to generate patient representation vectors. The rule retrieval network is based on patient representation vectors The top K rule entries and case plans with the highest similarity were retrieved from the normative knowledge base and historical case database, and the rule representation vector was obtained by similarity-weighted aggregation. ; Patient representation vector , rule representation vector and interactive features Perform splicing and fully connected transformations to obtain the fused representation. ; Based on fusion representation It generates a rehabilitation plan and calls a security constraint mask during the generation process to block items that conflict with the contraindications sub-library.

5. The pain rehabilitation management system based on intelligent assessment algorithm and full-process interaction according to claim 4, characterized in that, The loss function used when training the deep learning model generated by the rehabilitation assessment plan is: in, Generate a loss function for a deep learning model for the rehabilitation assessment plan; , , , The weights are non-negative and satisfy the following conditions: ; Classify losses for planned projects; For the regression loss of treatment parameters; To constrain and punish losses for safety reasons; This results in a loss of consistency with the guidelines.

6. The pain rehabilitation management system based on intelligent assessment algorithm and full-process interaction according to claim 1, characterized in that, Collect standard movement data samples to establish a standardized database of rehabilitation procedures and training movements related to pain disorders, including: Standard motion data samples were collected, and human key points were extracted, motion start and end points were segmented, body coordinate system was normalized, left and right side mirroring was unified, noise points were removed, and abnormal frames were corrected to obtain a standard skeleton sequence. The dynamic time warping method is used to align the time sequence of multiple standard samples of the same action, extract template features, form standard action templates and allowable deviation ranges, and construct a standardized database of rehabilitation process operations and rehabilitation training actions related to pain diseases.

7. The pain rehabilitation management system based on intelligent assessment algorithm and full-process interaction according to claim 6, characterized in that, The motion capture evaluation model is constructed according to the following steps: Acquire the video to be evaluated and perform motion recognition and motion phase segmentation; The key point sequence in the video to be evaluated is aligned frame by frame with the standard motion template. During the motion phase k, the angle deviations of the key joints are calculated respectively. Key trajectory spatial deviation Movement rhythm deviation and compensatory action deviation The values ​​are then mapped to the [0,1] interval using a standardized function, and the stage score is calculated. in, For stage scoring; For standardized functions; , , , The weights are non-negative and satisfy the following conditions: ; The total score is calculated; Assign importance weights to each action phase.

8. The pain rehabilitation management system based on intelligent assessment algorithm and full-process interaction according to claim 7, characterized in that, The deep learning model is constructed according to the following steps: Using domain-specific data, standard action data samples, and expert rating data from disease and health groups as training samples, a spatiotemporal feature extraction network consisting of skeleton sequence branches, video frame branches, cross-modal fusion modules, and multi-task output heads is constructed to obtain a deep learning model.

9. The pain rehabilitation management system based on intelligent assessment algorithm and full-process interaction according to claim 8, characterized in that, The loss function used for training the motion capture standardization comparison and scoring model is: in, The loss function for the motion capture standardized comparison and scoring model; , , , The weights are non-negative and satisfy the following conditions: ; Loss is identified for action categories; Locate the losses during the critical phase; The regression loss for motion quality scoring; The loss is due to the constraint of consistency in expert scoring.

10. The pain rehabilitation management system based on intelligent assessment algorithm and full-process interaction according to claim 1, characterized in that, The process involves collecting patient and rehabilitation-related information, and generating a rehabilitation plan using a deep learning model generated from the rehabilitation assessment plan. The rehabilitation plan implementation is dynamically monitored through standardized motion capture comparison and scoring models, generating evaluation data and outputting adjustment suggestions. It also includes: The rehabilitation plan is broken down into therapist execution forms and patient self-check-in forms.

Images