A method and apparatus for extrapolating muscle function state based on visual pose estimation

By using a muscle function state extrapolation method based on visual posture estimation, the high cost and complex equipment issues of muscle function assessment in the elderly have been resolved, enabling its promotion in communities and accurate muscle function assessment, and providing a basis for individualized rehabilitation training.

CN123120215APending Publication Date: 2026-11-17PEKING UNION MEDICAL COLLEGE HOSPITAL
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611273445.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for assessing muscle function in the elderly are costly, require complex equipment, are difficult to promote in community or home settings, and lack quantitative analysis methods for the relationship between movement abnormalities and muscle function, thus failing to reflect the true state of movement.

Method used

A muscle function state inference method based on visual pose estimation is adopted. By acquiring continuous video data, abnormal gait is labeled and scored, muscle group mapping rules are established, and a trained muscle function scoring model is used to generate muscle function risk information.

Benefits of technology

It reduces assessment costs, is suitable for promotion in communities and primary healthcare institutions, improves the accuracy of abnormality identification and the interpretability of the assessment process, realizes the deduction from external motor performance to internal muscle function, and provides a basis for rehabilitation training and individualized intervention.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123120215A_ABST
    Figure CN123120215A_ABST
Patent Text Reader

Abstract

This application discloses a method and apparatus for inferring muscle function status based on visual pose estimation. The method involves: acquiring continuous video data to be analyzed; annotating and scoring abnormal gait in the video data to obtain abnormal gait data; acquiring preset muscle group mapping rules; generating abnormal gait variables based on the abnormal gait data and the preset muscle group mapping rules; acquiring a trained muscle function scoring model; and generating muscle function risk information for the subject based on the abnormal gait variables and the trained muscle function scoring model. This application uses ordinary video acquisition equipment and does not rely on expensive or invasive detection methods such as electromyography or imaging to assess the muscle function status of the elderly, significantly reducing application costs and implementation barriers, making it suitable for widespread use in community and primary healthcare institutions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of image processing technology, specifically to a method and apparatus for inferring muscle functional state based on visual pose estimation. Background Technology

[0002] Currently, muscle function assessment in the elderly mainly relies on electromyography (EMG), imaging examinations (MRI, CT), and clinical experience assessment. This approach has several drawbacks: high testing costs and complex equipment; difficulty in widespread adoption in communities or homes; primarily static testing that fails to reflect true motor function; and a lack of quantitative analysis methods for the relationship between motor abnormalities and muscle function. On the other hand, while computer vision-based gait analysis techniques exist, they primarily focus on kinematic parameter extraction and have not yet established a systematic model for extrapolating from gait abnormalities to specific muscle function states.

[0003] Therefore, it is necessary to propose a method for muscle function extrapolation based on real motion videos, combined with manual annotation and biomechanical rules. Summary of the Invention

[0004] The purpose of this invention is to provide a method for inferring muscle functional state based on visual pose estimation to at least solve one of the above-mentioned technical problems.

[0005] One aspect of the present invention provides a method for extrapolating muscle function state based on visual pose estimation, the method comprising: Step 1: Obtain the continuous video data to be analyzed; Step 2: Annotate and score the abnormal gait in the continuous video data to be analyzed, thereby obtaining abnormal gait data; Step 3: Obtain the preset muscle group mapping rules; Step 4: Generate abnormal gait variables based on the abnormal gait data and preset muscle group mapping rules; Step 5: Obtain the trained muscle function scoring model; Step 6: Generate the subject's muscle function risk information based on the abnormal gait variables and the trained muscle function scoring model.

[0006] Optionally, step 2 includes: Step 21: Observe the continuous video data to be analyzed frame by frame or segment by segment to identify abnormal gait; Step 22: Label the abnormal gait to obtain gait abnormality type data, which includes pelvic tilt or pelvic sinking, trunk compensatory swing, knee hyperextension or hypoflexion, foot drop or dragging, and insufficient gait propulsion. Step 23: Label the gait anomaly type data with occurrence features to obtain anomaly occurrence feature data, which includes the time point of anomaly occurrence, duration, and frequency of occurrence; Step 24: Classify the severity of the gait anomaly type data to obtain anomaly severity data. The anomaly severity is divided into three levels: mild, moderate, and severe. The abnormal gait data includes gait anomaly type data, anomaly occurrence characteristic data, and anomaly severity data.

[0007] Optionally, step 3 includes: Step 31: Obtain continuous video data of the training samples; Step 32: Perform pose estimation on the continuous video data of the training samples, extract the time series data of human key points, thereby constructing gait motion feature vectors and obtaining gait feature data of the training samples; Step 33: Perform abnormal gait annotation and scoring on the continuous video data of the training samples to obtain abnormal gait data of the training samples; Step 34: Based on the principles of biomechanics and muscle kinematics, perform correlation analysis on the gait feature data of the training samples and the abnormal gait data of the training samples to establish the correspondence between abnormal gait features and the functional state of the target muscle groups. Step 35: Based on the correspondence, organize and form a mapping rule from gait abnormality to muscle function, thereby obtaining a preset muscle group mapping rule; wherein, the target muscle group includes at least the gluteus medius, quadriceps femoris, triceps surae, and tibialis anterior.

[0008] Optionally, step 4 includes: Step 41: According to the preset muscle group mapping rules, map each abnormal gait in the abnormal gait data to the corresponding target muscle group, thereby obtaining the abnormal gait subset corresponding to each target muscle group. Step 42: Score and summarize the abnormal gait subsets corresponding to each target muscle group to obtain the abnormal gait score for each target muscle group; Step 43: Use the abnormal gait scores of each target muscle group as abnormal gait variables to generate abnormal gait variables; wherein, the abnormal gait scores of each target muscle group adopt binary classification or three-level scoring.

[0009] Optionally, step 5 includes: Step 51: Obtain the dataset of abnormal gait variables from the training samples; Step 52: Obtain the SPPB test data of the training samples, define SPPB ≤ 9 as impaired muscle function and SPPB > 9 as unimpaired, perform binary classification transformation to obtain the outcome variable dataset; Step 53: Using abnormal gait variables as independent variables and outcome variables as dependent variables, construct a multivariate logistic regression model and perform regression analysis to obtain the regression coefficients corresponding to each target muscle group; Step 54: Calculate the assigned weights based on the regression coefficients of each target muscle group to obtain the weight configuration for each muscle group; Step 55: Establish a formula for calculating the total muscle function risk score based on the weight configuration of each muscle group, thereby obtaining the scoring calculation rules; Step 56: Based on the total muscle function risk score and outcome variable of the training samples, determine the risk stratification threshold through the receiver operating characteristic curve and Youden index, thereby obtaining the risk stratification criteria; Step 57: The trained muscle function scoring model includes the weight configuration of each muscle group, scoring calculation rules, and risk stratification criteria.

[0010] Optionally, step 6 includes: Step 61: Calculate the total muscle function risk score based on the abnormal gait scores of each target muscle group; Step 62: Determine the functional status of each muscle group based on the abnormal gait score of each target muscle group. A score of 0 is considered normal, and a score greater than 0 is considered abnormal, thereby obtaining the functional status results of each muscle group. Step 63: Determine the subject's overall risk level based on the total muscle function risk score; wherein, the subject's muscle function risk information includes the functional status results of each muscle group, the abnormal gait score of each target muscle group, the total muscle function risk score, and the overall risk level.

[0011] Optionally, step 5 includes: Step 501: Construct a three-layer causal semantic ontology structure, which includes a gait abnormality layer, a muscle function layer, and a risk outcome layer. Each layer contains multiple semantic nodes, and there is a unidirectional causal dependency relationship between the layers. Step 502: Based on the three-layer causal semantic ontology structure, construct a weighted directed muscle-gait-risk heterogeneous graph. The node set of the weighted directed muscle-gait-risk heterogeneous graph includes all gait abnormality nodes, muscle function nodes, and risk outcome nodes, and the edge set includes inter-layer directed causal edges. Step 503: Execute the reverse causal decay propagation algorithm on the weighted directed muscle-gait-risk heterogeneous graph. Starting from the risk outcome node, propagate the causal influence backward along the causal edge. The final causal activation value of each node is obtained through iterative calculation. The iteration terminates when the maximum value of the change in activation value of all nodes is less than a preset threshold, thereby obtaining the set of node causal activation values. Step 504: Calculate the causal betweenness centrality of each muscle node based on the set of causal activation values ​​of the nodes, thereby obtaining the causal betweenness centrality of each muscle node; Step 505: Convert the causal betweenness centrality of each muscle node into assigned weights to obtain the weight configuration for each muscle group. Step 506: Establish a formula for calculating the total muscle function risk score based on the weight configuration of each muscle group, and determine the risk stratification threshold through the weighted Youden index, thereby obtaining the scoring calculation rules and risk stratification standards; Step 507: By combining the weight configuration of each muscle group, the scoring calculation rules, and the risk stratification criteria, a trained muscle function scoring model is obtained.

[0012] Optionally, the edge weight of the directed causal edge is obtained by the following formula: For the causal edge from the gait anomaly layer to the muscle function layer, the initial weights are: ; in, For the Sigmoid function, Gait abnormality With muscles Point-to-point information, This represents the biomechanical prior strength of the muscle. and The balance coefficients are satisfied. + =1; For the causal edge from the muscle function layer to the risk outcome layer, the initial weights are: ; in, For muscles The corresponding one-way ratio, This is the scaling adjustment parameter.

[0013] Optionally, the causal betweenness centrality is obtained by the following formula: ; in, To start from gait anomaly nodes To the risk outcome point The sum of the total weights of all causal paths, From arrive And passing through muscle nodes The sum of the total weights of all causal paths, and These are the final causal activation values ​​for gait nodes and risk nodes, respectively.

[0014] This application also provides a muscle function state inference device based on visual pose estimation, the muscle function state inference device based on visual pose estimation includes: A continuous video data acquisition module for analysis, wherein the continuous video data acquisition module is used to acquire continuous video data to be analyzed; An abnormal gait data acquisition module is used to annotate and score abnormal gait in the continuous video data to be analyzed, thereby acquiring abnormal gait data. A muscle group mapping rule acquisition module is used to acquire preset muscle group mapping rules. An abnormal gait variable acquisition module is used to generate abnormal gait variables based on the abnormal gait data and preset muscle group mapping rules. A muscle function scoring model acquisition module, which is used to acquire a trained muscle function scoring model; A muscle function risk generation module is used to generate muscle function risk information for the subject based on the abnormal gait variables and a trained muscle function scoring model.

[0015] This application utilizes ordinary video capture equipment, eliminating the need for expensive or invasive detection methods such as electromyography and imaging, to assess the muscle function status of the elderly. This significantly reduces application costs and implementation barriers, making it suitable for widespread use in community and primary healthcare institutions. In terms of method design, by introducing professional personnel to perform structured annotation of gait videos, clinical experience is combined with objective motion images. This not only improves the accuracy of anomaly identification but also enhances the interpretability of the assessment process, facilitating understanding and application by clinicians. Furthermore, this invention establishes a correspondence between gait characteristics and specific muscle function states, enabling the deduction from external motor performance to internal muscle function, thus providing a clear basis for rehabilitation training and individualized intervention. In addition, this method is designed for common low-intensity walking scenarios in the elderly, reflecting true motor ability in a natural state and avoiding biases caused by high-intensity testing. Moreover, the standardized annotation system and data structure can serve as a high-quality data source for subsequent artificial intelligence model training, providing a reliable foundation for automated assessment. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for extrapolating muscle function status based on visual pose estimation according to an embodiment of this application.

[0017] Figure 2 This is a schematic diagram of constructing gait motion feature vectors in one embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some, but not all, embodiments of this application. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0019] like Figure 1 The method for extrapolating muscle function status based on visual pose estimation, as shown, includes: Step 1: Obtain the continuous video data to be analyzed; Step 2: Perform abnormal gait annotation and scoring on the continuous video data to be analyzed, thereby obtaining abnormal gait data; specifically, perform pose estimation on the continuous video data to be analyzed, extract human key point time series data to construct gait motion feature vectors as annotation references; based on the gait motion feature vectors, perform abnormal gait annotation and scoring on the continuous video data to be analyzed, thereby obtaining abnormal gait data; Step 3: Obtain the preset muscle group mapping rules; Step 4: Generate abnormal gait variables based on the abnormal gait data and preset muscle group mapping rules; Step 5: Obtain the trained muscle function scoring model; Step 6: Generate the subject's muscle function risk information based on the abnormal gait variables and the trained muscle function scoring model.

[0020] In this embodiment, the continuous video data to be analyzed or the training video of this application can be obtained in the following ways: Semi-controlled test environment setup Within an indoor activity space longer than 40 m, a 30 m long straight marker is laid on the ground as a walking test path, clearly marking the start and end points. A monocular camera is positioned approximately 1 m directly in front of the starting point, connected to a computer for real-time video data acquisition and transmission. A physical shielding device is installed approximately 3 m perpendicularly from the marker line on one side of the test path to prevent unauthorized personnel from entering the filming area and interfering with the experiment.

[0021] In practice, the straight-line markers can be formed by pasting yellow anti-slip tape along the ground in a straight line, with a width of approximately 5 cm. It is essential to ensure the tape is flat and wrinkle-free to avoid affecting the subject's walking. The start and end points can be clearly marked with red square markers, approximately 20 cm × 20 cm in size, pasted at both ends of the straight-line markers. The monocular camera can be a high-definition network camera or a USB camera, mounted on an adjustable-height tripod. The camera height can be adjusted according to the subject's height, approximately level with the subject's hip joint, to obtain the optimal shooting angle for lower limb movements. The physical obstruction device can be a movable fabric screen or plastic barrier, placed along one side of the test path, covering the entire 30 m test area in length, with a height of no less than 1.5 m, to effectively block the view of unauthorized personnel and prevent them from entering the shooting area.

[0022] Before the trial began, participants were instructed to change into suitable walking shoes and undergo a basic vital sign assessment, including blood pressure and heart rate measurements. After a brief warm-up, participants stood in the starting position one by one and began the 6-minute walk test (6MWT) after the tester gave the command, while the entire process was continuously recorded on video.

[0023] 2. Video capture Continuous video data was collected from elderly subjects during the standard 6MWT procedure. The videos were acquired using a monocular camera and recorded using video capture software, in MP4 format. Specific acquisition parameters are as follows: Acquisition equipment: Monocular camera; Video resolution: no less than 1080P; Frame rate: No less than 25 frames per second; Shooting angle: Approximately 1 m directly in front of the starting point of the subject's walking direction; The video fully records the subject's continuous walking process in a natural state.

[0024] In practice, the video capture software can be OBS Studio, PotPlayer, or dedicated capture software for the camera. Before recording, the camera's white balance and focus need to be adjusted to ensure clear images and natural colors. During filming, the camera must be kept stationary to avoid shaking, and the tester must monitor the footage in real time via computer to ensure the subject remains within the effective frame throughout the walking process. If the subject moves out of the frame, the tester should promptly remind the subject to adjust their walking route.

[0025] The video fully records the subject's continuous walking process in a natural state, including starting from the starting point, walking back and forth along the straight line markers, and ending after 6 minutes. The video file can be named in the format of subject number + test date for easy subsequent management and retrieval.

[0026] In this embodiment, step 2 includes: Step 21: Observe the continuous video data to be analyzed frame by frame or segment by segment to identify abnormal gait. Specifically, two professionally trained assessors observe the video data frame by frame or segment by segment to identify abnormal gaits in the video.

[0027] After acquiring the continuous video data to be analyzed, the video is first preprocessed and gait feature extracted: image denoising, frame sequence segmentation, and human target detection are performed sequentially to accurately locate the subject; then, a pose estimation algorithm is used to extract key human point information frame by frame, including wrist, elbow, shoulder, hip, knee, and ankle joints; based on the time series data of the key points, a gait motion feature vector is constructed, covering the dynamic change features of each joint angle, the spatial motion trajectory of the key points, the gait cycle, and the time series features, providing an objective and quantitative visual reference for subsequent abnormal gait annotation, and realizing the combination of visual analysis and professional manual annotation.

[0028] In practice, the assessors must have a background in geriatrics, rehabilitation medicine, or sports biomechanics, and undergo standardized gait annotation training, mastering the definitions, clinical manifestations, and identification points of various abnormal gaits. The annotation training must include theoretical learning, sample annotation practice, and consistency assessments; only those who pass the assessments are eligible to participate in formal annotation work.

[0029] The annotation work can be carried out using professional video annotation software, such as ELAN, VGG Image Annotator, or self-developed annotation tools. The software should support functions such as frame-by-frame playback, slow motion, fast forward, precise time point positioning, region marking, and label addition to meet the needs of refined annotation.

[0030] The observation method can be flexibly selected according to the characteristics of the abnormal gait: for abnormal gaits that are long-lasting and obvious, segmented observation can be used to improve annotation efficiency; for abnormal gaits that are short-lasting and subtle, frame-by-frame observation is required to ensure the accuracy of identification. During the observation process, assessors need to focus on the movement status of the subject's pelvis, trunk, hip joints, knee joints, ankle joints, etc., and make a comprehensive judgment based on different phases of the gait cycle (stance phase, swing phase).

[0031] Two evaluators must independently complete the annotation work for the same video segment and must not discuss or refer to each other's annotation results to ensure the independence and objectivity of the annotation.

[0032] Step 22: Label the abnormal gait to obtain gait abnormality type data. The abnormal gait is labeled to obtain gait abnormality type data, which includes pelvic tilt or pelvic sinking, trunk compensatory swing, knee hyperextension or hypoflexion, foot drop or dragging, and insufficient gait propulsion.

[0033] In practice, the specific criteria for judging various abnormal gaits are as follows: (1) Pelvic tilt or pelvic sinking: refers to an abnormal tilt or sinking of the pelvis in the coronal plane during walking, manifested as the pelvis on one side being significantly lower than the other side during the stance phase, or the pelvis tilting to one side as a whole. It can be judged by observing the horizontality of the line connecting the two iliac crests. If the pelvic sinking exceeds the normal range at the end of the stance phase, it is judged as pelvic tilt or pelvic sinking.

[0034] (2) Compensatory trunk swaying: This refers to abnormal lateral swaying or back-and-forth swaying of the trunk during walking to compensate for insufficient lower limb function. It manifests as a significant tilting of the trunk on the side opposite to the support, or an increased amplitude of left-right swaying of the trunk throughout the walking process. It can be judged by observing the angle between the trunk's central axis and the vertical line to the ground.

[0035] (3) Knee hyperextension or insufficiency: Knee hyperextension refers to excessive extension of the knee joint at the end of the stance phase, exceeding the normal range of flexion and extension; insufficiency refers to insufficient knee flexion angle during the swing phase, leading to lower limb clearance impairment. This can be determined by observing the changes in knee joint angle at different stages of the gait cycle.

[0036] (4) Foot drop or dragging: Foot drop refers to the inability of the ankle joint to actively dorsiflex at the end of the swing phase, resulting in the toes drooping; dragging refers to the toes dragging forward during walking without being fully lifted. It is manifested by the sound of the toes scraping the ground, obvious wear on the front of the shoe sole, or insufficient toe height during the swing phase.

[0037] (5) Insufficient gait propulsion: This refers to insufficient force of the lower limbs pushing off the ground during walking, resulting in weakened forward propulsion. It manifests as a shorter stride length, slower walking speed, longer support phase time, and an overall weak walking posture.

[0038] During annotation, for each identified abnormal gait type, it is necessary to mark the corresponding time segment in the video and add an appropriate type label. If multiple abnormal gait types appear in the same video segment, they must be annotated separately.

[0039] Step 23: Perform occurrence feature annotation on the gait anomaly type data to obtain anomaly occurrence feature data. The gait anomaly type data is labeled with occurrence features to obtain anomaly occurrence feature data, which includes the time point of anomaly occurrence, duration, and frequency of occurrence.

[0040] In practice, the annotation methods for each occurrence feature are as follows: (1) Time of Abnormal Occurrence: This refers to the time point when the abnormal gait first appears, recorded based on the video timeline, accurate to 0.1 seconds. The time point corresponding to the start frame of the abnormal gait can be directly read through the timestamp function of the annotation software.

[0041] (2) Duration: refers to the length of time from the start to the end of a single abnormal gait, in seconds. For the same abnormal gait occurring multiple times, the duration of each occurrence must be recorded separately. The duration can be automatically calculated by marking the start and end frames of the abnormal gait using annotation software.

[0042] (3) Frequency of occurrence: refers to the number of times the abnormal gait occurs per unit of time, usually expressed as the number of occurrences per minute. It can be calculated by counting the total number of times the abnormal gait occurs in the entire 6-minute walking test video and dividing by the total duration (6 minutes). For non-continuous abnormal gait, the frequency index can better reflect its severity and impact on walking function.

[0043] After annotation, the occurrence characteristics of each abnormal gait type need to be summarized and organized to form structured feature data, which will provide a basis for subsequent analysis.

[0044] Step 24: Classify the severity of the gait anomaly type data to obtain anomaly severity data. The severity of the gait anomaly type data is classified into three levels: mild, moderate, and severe. The abnormal gait data includes gait anomaly type data, anomaly occurrence characteristic data, and anomaly severity data.

[0045] In practice, the severity of abnormalities should be graded by comprehensively considering factors such as the amplitude, duration, frequency, and impact on walking function of the abnormal gait. The reference standards for each level are as follows: (1) Mild: Abnormal gait occurs occasionally, with a low frequency (e.g., less than 3 times per minute), a short duration each time, and a small degree of abnormality. It has little impact on walking speed, stability, and other walking functions. The subjects are basically able to maintain a normal walking rhythm and posture.

[0046] (2) Moderate: Abnormal gait occurs frequently, with a moderate frequency (e.g., 3 to 10 times per minute), or a long duration of a single occurrence. The abnormal manifestations are more obvious and have a significant impact on walking function. The subject's walking speed decreases or stability decreases, and certain compensatory movements are needed to maintain walking.

[0047] (3) Severe: Abnormal gait persists or occurs frequently (e.g., more than 10 times per minute), with a large range of abnormal manifestations, seriously affecting walking function. Subjects have difficulty walking, their walking speed is significantly slowed down, their gait is unstable, and they may even need to use assistive devices to walk or be unable to complete the 6-minute walking test.

[0048] During the grading process, assessors need to make a comprehensive judgment based on the characteristics of the abnormal gait type. Different types of abnormal gait can have their grading criteria adjusted appropriately according to their clinical significance. For example, for the same frequency of occurrence, knee hyperextension may have a greater impact on function than mild pelvic tilt, and the grading can be raised by one level accordingly.

[0049] After two assessors complete their grading independently, a consistency comparison is required. If their grading results are consistent, they are taken as the final result. If their grading results differ by one level, the median value can be used or the two assessors can reach a consensus through discussion. If the difference is two levels or more, a third senior assessor must be introduced for arbitration, or a group discussion must be organized to determine the final grading.

[0050] Through the above steps of annotation and scoring, complete abnormal gait data is finally obtained, which includes gait abnormality type data, abnormality occurrence characteristic data, and abnormality severity data.

[0051] In this embodiment, step 3 establishes a correspondence between gait abnormalities and muscle function states through analysis of training sample data, forming a preset muscle group mapping rule to provide a mapping basis for the subsequent generation of abnormal gait variables. The preset muscle group mapping rule is pre-established and stored based on the training samples and can be directly called during actual evaluation without needing to be reconstructed each time.

[0052] In this embodiment, step 3 includes: Step 31: Obtain continuous video data for training samples: In practice, the selection of training samples must adhere to certain inclusion and exclusion criteria to ensure sample representativeness and data quality. Inclusion criteria may include: elderly individuals aged 60 years or older; ability to independently complete the 6-minute walk test; being conscious and able to cooperate with the test; and signing an informed consent form. Exclusion criteria may include: individuals with severe cardiopulmonary disease, musculoskeletal disease, or neurological disease affecting their walking; individuals with unhealed lower limb injuries or surgical history; individuals with severe cognitive impairment preventing cooperation; and individuals wearing prostheses or using walking aids.

[0053] The number of training samples should meet the requirements of statistical analysis, and it is generally recommended to have no less than 200 cases. The samples should also cover subjects of different genders, ages and muscle function levels to ensure the universality of the mapping rule.

[0054] The method for collecting continuous video data of the training samples is the same as described in steps 1 and 2, that is, in the same semi-controlled experimental environment, using the same acquisition equipment and parameters, continuous video data of the subjects during the standard 6MWT process is collected to ensure the consistency between the training sample data and the actual evaluation data.

[0055] Step 32: Perform pose estimation on the continuous video data of the training samples, extract the time series data of human key points, thereby constructing gait motion feature vectors and obtaining gait feature data of the training samples; The acquired video data undergoes preprocessing, including image denoising, frame sequence segmentation, and human target detection, to achieve accurate localization of the subject. Based on this, a pose estimation algorithm is used to extract key human point information frame by frame from the video. These key points include the wrist, elbow, shoulder, hip, knee, and ankle joints. A gait feature vector is constructed based on the time-series data of these key points. This feature vector includes dynamic changes in the angles of each joint, the spatial trajectory of the key points, and the gait cycle and temporal features. This feature vector dynamically represents the subject's gait, focusing on the movement states of the hip, knee, ankle, and pelvic regions.

[0056] In practical implementation, image denoising can employ methods such as Gaussian filtering, median filtering, or bilateral filtering to remove noise interference from the video and improve image quality. Frame sequence segmentation can be appropriately set according to the video frame rate and walking speed, for example, extracting one frame every 1 to 3 frames for analysis, thus improving processing efficiency while maintaining temporal resolution. Human target detection can utilize target detection algorithms such as YOLO and Faster R-CNN to automatically identify and locate the subject region in the video.

[0057] Pose estimation algorithms can employ mature human pose estimation models such as OpenPose, MediaPipe Pose, and AlphaPose. These algorithms can automatically detect and output the pixel coordinates of key points on the human body from monocular video. For monocular video, pixel coordinates can also be converted into actual three-dimensional spatial coordinates or two-dimensional world coordinates through camera calibration and geometric transformation for subsequent kinematic analysis.

[0058] The method for constructing gait motion feature vectors is as follows: (1) Dynamic change characteristics of joint angles: Based on the positional relationship of adjacent key points, calculate the angles of each joint in each frame, including shoulder joint angle, elbow joint angle, hip joint angle, knee joint angle, ankle joint angle, etc., to form a time series of joint angles, and extract their maximum value, minimum value, average value, standard deviation, rate of change and other statistical characteristics.

[0059] (2) Spatial motion trajectory of key points: Extract the spatial coordinate changes of each key point during the entire walking process to form a motion trajectory, and calculate the length, amplitude, smoothness, symmetry and other features of the trajectory.

[0060] (3) Gait cycle and timing characteristics: By detecting gait events such as heel touch or toe lift, the gait cycle is divided, and gait parameters such as step frequency, step length, step speed, support phase time, swing phase time, and left and right symmetry indicators are calculated.

[0061] By combining the above features, a gait motion feature vector with uniform dimensions is formed, which is used to characterize the subject's gait kinematics. The composition structure and construction logic of the above gait motion feature vector are as follows: Figure 2 As shown.

[0062] Step 33: Perform abnormal gait annotation and scoring on the continuous video data of the training samples to obtain abnormal gait data of the training samples; Abnormal gait is labeled and scored on the continuous video data of the training samples to obtain abnormal gait data of the training samples. The labeling and scoring method is the same as described in step 2, that is, two professionally trained evaluators observe the video data frame by frame or segment by segment, and label and score the abnormal gait. The labeling content includes the abnormality type, abnormality occurrence characteristics, and abnormality severity level.

[0063] In practice, the annotation of training samples requires more stringent quality control standards to ensure the reliability of the training data. The following quality control measures can be taken: (1) Training of annotation personnel: Systematic training will be provided to the evaluation personnel involved in annotation, including theoretical knowledge learning, interpretation of annotation specifications, analysis of typical cases, and practical exercises, to ensure that all annotation personnel have a consistent understanding of the definition and judgment criteria of abnormal gait.

[0064] (2) Consistency check: Before formal annotation, a certain number of samples are selected for pre-annotation, and the consistency coefficient (such as Kappa value) among the annotators is calculated. Only when the consistency reaches the preset requirements (such as Kappa≥0.8) can the formal annotation stage be entered.

[0065] (3) Sampling review: During the labeling process, a certain percentage of the labeled samples are periodically sampled for review. Senior experts conduct a second review. If labeling errors or inconsistencies are found, they are corrected in a timely manner and the relevant labeling personnel are retrained.

[0066] (4) Discussion of difficult cases: For difficult cases with atypical presentation and difficult judgment, multiple assessors are organized to conduct a collective discussion to jointly determine the final labeling result.

[0067] Through rigorous quality control, the accuracy and consistency of abnormal gait data in the training samples are ensured, providing a reliable data foundation for subsequent correlation analysis and mapping rule establishment.

[0068] Step 34: Based on the principles of biomechanics and muscle kinematics, perform correlation analysis on the gait feature data of the training samples and the abnormal gait data of the training samples to establish the correspondence between abnormal gait features and the functional state of the target muscle groups. Based on the principles of biomechanics and muscle kinesiology, a correspondence between abnormal gait characteristics and muscle functional states is established. The muscle groups include at least the gluteus medius, quadriceps femoris, triceps surae, and tibialis anterior.

[0069] In practice, correlation analysis needs to combine relevant theories from biomechanics, kinesiology, and clinical rehabilitation medicine to analyze the formation mechanism of each abnormal gait and determine its corresponding responsible muscle group. The specific correspondences are as follows: (1) Quadriceps-related gait abnormalities: The quadriceps is the main muscle for knee extension and plays a role in stabilizing the knee joint during the stance phase. When the quadriceps muscle strength is weakened, it can lead to knee hyperextension (maintaining stability through the knee locking mechanism) or knee insufficiency (insufficient knee flexion during the swing phase). Therefore, knee hyperextension and knee insufficiency correspond to quadriceps dysfunction.

[0070] (2) Gluteus medius-related gait abnormalities: The gluteus medius is the main muscle for lateral pelvic stability and plays a role in maintaining pelvic level during the single-leg stance phase. When the gluteus medius muscle strength is weakened, it can lead to pelvic tilt or pelvic sinking on one side during the stance phase, which can result in a waddling gait in severe cases. Therefore, pelvic tilt or pelvic sinking corresponds to gluteus medius dysfunction.

[0071] (3) Abnormal gait related to the tibialis anterior muscle: The tibialis anterior muscle is the main muscle for dorsiflexion of the ankle joint and plays a role in lifting the toes during the swing phase. When the strength of the tibialis anterior muscle is weakened, it will lead to foot drop, that is, the toes cannot be lifted during the swing phase, resulting in dragging. Therefore, foot drop and dragging correspond to abnormal tibialis anterior muscle function.

[0072] (4) Gastrocnemius-related gait abnormalities: The gastrocnemius muscle is the main muscle for plantar flexion of the ankle joint, providing the force to push off the ground at the end of the stance phase, and is an important source of propulsion for walking. When the strength of the gastrocnemius muscle is weakened, it will lead to insufficient gait propulsion, shortened stride length, and slowed walking speed. Therefore, insufficient gait propulsion corresponds to abnormal gastrocnemius muscle function.

[0073] (5) Compensatory trunk swing: Compensatory trunk swing is usually a comprehensive manifestation of multiple lower limb muscle dysfunctions. It is a compensatory movement that occurs in order to maintain balance and forward movement. The main responsible muscle group can be judged based on the direction and characteristics of the swing, combined with other abnormal gait manifestations.

[0074] In the process of association analysis, quantitative analysis can also be performed by combining gait feature data of training samples. For example, the correlation between the muscle strength of a certain muscle group and the corresponding gait feature parameters can be calculated to verify the rationality of the correspondence. Statistical methods such as Pearson correlation analysis and Spearman rank correlation analysis can be used to calculate the correlation coefficient and test the significance of the association.

[0075] Step 35: Based on the correspondence, organize and form a mapping rule from gait abnormality to muscle function, thereby obtaining the preset muscle group mapping rule; Based on the correspondence, a mapping rule from gait abnormality to muscle function is formed, thereby obtaining a preset muscle group mapping rule; wherein, the target muscle group includes at least the gluteus medius, quadriceps femoris, triceps surae, and tibialis anterior.

[0076] In practice, mapping rules can be stored in the form of a mapping table. Rows in the table correspond to different abnormal gait types, columns correspond to different target muscle groups, and the values ​​in the table represent the strength of the association between the abnormal gait and the corresponding muscle group. The association strength can be divided into three levels: strong association, moderate association, and weak association, each corresponding to a different weighting coefficient.

[0077] The specific form of the mapping rule is shown in the following example: Knee hyperextension → quadriceps (strong correlation); Insufficient knee flexion → quadriceps muscle (strong correlation); Pelvic tilt or pelvic depression → gluteus medius (strong correlation); Foot drop → Tibialis anterior muscle (strong correlation). Drag step → Tibialis anterior muscle (strong correlation); Insufficient gait propulsion → gastrocnemius muscle (strong correlation); Compensatory trunk swaying → Multi-muscle group syndrome (judged based on specific manifestations); After the mapping rules are established, they need to be validated and optimized. Independent test samples can be selected, and the muscle function assessment results obtained based on the mapping rules can be compared with the clinical gold standard (such as muscle strength test results). Indicators such as accuracy, sensitivity, and specificity can be calculated to evaluate the effectiveness of the mapping rules. If the results are not ideal, the mapping rules can be adjusted and optimized until they meet the requirements for clinical application.

[0078] Once verified, the muscle group mapping rules can be stored in a database or configuration file and can be directly read and called as preset mapping rules when actually performing muscle function state simulation.

[0079] In this embodiment, step 4 includes: Step 41: According to the preset muscle group mapping rules, map each abnormal gait in the abnormal gait data to the corresponding target muscle group, thereby obtaining the abnormal gait subset corresponding to each target muscle group. According to the preset muscle group mapping rules, each abnormal gait in the abnormal gait data is mapped to the corresponding target muscle group, thereby obtaining the abnormal gait subset corresponding to each target muscle group.

[0080] In practice, the mapping process is carried out item by item according to the preset muscle group mapping rules. For each abnormal gait type in the abnormal gait data, the target muscle group corresponding to it in the mapping rules is found, and all relevant information of the abnormal gait (including occurrence characteristics, severity, etc.) is assigned to the subset of the corresponding muscle group.

[0081] The mapping process must follow these principles: (1) One-to-one mapping: For abnormal gait that clearly corresponds to a single muscle group, it is directly mapped to that muscle group. For example, knee hyperextension is directly mapped to the quadriceps femoris, and pelvic tilt is directly mapped to the gluteus medius.

[0082] (2) Many-to-one mapping: When multiple abnormal gaits correspond to the same muscle group, all of them are classified into a subset of that muscle group. For example, foot drop and dragging both correspond to the tibialis anterior muscle, so both are classified into the abnormal gait subset of the tibialis anterior muscle.

[0083] (3) One-to-many mapping: For abnormal gait that may involve multiple muscle groups (such as compensatory trunk swing), it is necessary to combine the specific manifestations and other abnormal gait to make a comprehensive judgment on the main responsible muscle group and classify it into the corresponding subset; if it is difficult to determine the main responsible muscle group, it can be assigned to multiple related muscle groups based on clinical experience, or treated as a separate category.

[0084] (4) No corresponding mapping: For rare abnormal gait not covered by the mapping rules, the assessor can make a manual judgment based on professional knowledge and classify it into the most relevant muscle group, or mark it as an unknown type and handle it separately.

[0085] After mapping, each target muscle group corresponds to a subset of abnormal gait patterns. This subset contains detailed information on all abnormal gait patterns related to that muscle group, including the type of abnormality, the time of occurrence, duration, frequency, and severity. Target muscle groups include at least the gluteus medius, quadriceps femoris, triceps surae, and tibialis anterior.

[0086] To ensure the accuracy of the mapping, a verification process can be implemented, whereby professionals conduct random checks on the automatically mapped results. If errors are found, manual corrections are made promptly. For cases where the mapping results are questionable, an expert panel is organized to determine the final mapping outcome.

[0087] Step 42: Score and summarize the abnormal gait subsets corresponding to each target muscle group to obtain the abnormal gait score for each target muscle group; The abnormal gait subsets corresponding to each target muscle group are scored and summarized to obtain the abnormal gait score for each target muscle group. The abnormal gait score for each muscle group can be 0 or 1; 0 indicates that no abnormal gait pattern related to that muscle group was observed, and 1 indicates that an abnormal gait pattern related to that muscle group was observed. If a graded scoring method is used, the abnormal gait score for each muscle group can also be set to level 0, 1, and 2, representing no abnormality, mild abnormality, and obvious abnormality, respectively.

[0088] In practice, the scoring summary can be done using either a two-category scoring method or a three-level scoring method, depending on the actual needs.

[0089] The summary method for binary classification scoring is as follows: For each target muscle group, check if its corresponding abnormal gait subset is empty. If the subset does not contain any abnormal gait, the abnormal gait score for that muscle group is 0, indicating that no abnormal gait pattern related to that muscle group has been observed; if the subset contains at least one abnormal gait, the abnormal gait score for that muscle group is 1, indicating that an abnormal gait pattern related to that muscle group has been observed. The binary classification method is simple and intuitive, and suitable for rapid screening scenarios.

[0090] The summary method of the three-level scoring method is as follows: For each target muscle group, a comprehensive assessment is made based on the severity of abnormalities within its subset of abnormal gait patterns. If the subset contains no abnormal gait patterns, the score is 0, indicating no abnormality; if the abnormalities in the subset are all mild, or only occasionally moderate abnormalities occur, the score is 1, indicating mild abnormality; if the subset contains moderate or higher abnormalities with a high frequency, or contains severe abnormalities, the score is 2, indicating significant abnormality.

[0091] The summary of the three-level scores needs to comprehensively consider factors such as the severity, frequency, and duration of the abnormality. A weighted scoring method can be used, for example: assigning different base scores according to the severity of the abnormality (mild 1 point, moderate 2 points, severe 3 points), then multiplying by a frequency coefficient (occasional occurrence multiplied by 0.5, frequent occurrence multiplied by 1, persistent occurrence multiplied by 1.5), and finally determining the final 0, 1, or 2 level score based on the total score range. The specific scoring criteria can be adjusted and calibrated according to clinical needs.

[0092] For example, the specific calculation formula is as follows: First, for the i-th abnormal gait in the abnormal gait subset of a certain muscle group, calculate its individual weighted score: ; in, To score the severity, mild abnormality is scored as 1, moderate abnormality as 2, and severe abnormality as 3. For frequency scores, 0.5 is given for occasional occurrences, 1 for frequent occurrences, and 1.5 for continuous occurrences. For duration, a score of 0.8 is given for short duration, 1 for moderate duration, and 1.2 for longer duration. , , The weighting coefficients for severity, frequency, and duration are respectively, satisfying... + + = 1, which can be adjusted according to clinical needs. The default values ​​are 0.5, 0.3, and 0.2 respectively.

[0093] Then, calculate the overall score for this muscle group: ; in, Let be the overall score of the j-th muscle group, and n be the number of abnormal gait types corresponding to that muscle group. The maximum value among all individual scores. This is the average of all individual scores. The formula takes into account both the impact of the most severe anomalies and the overall level of anomalies, highlighting the main problems while also considering the overall situation.

[0094] Finally, based on the overall score Determine the final three-level rating. : ; in, The final abnormal gait score for the j-th muscle group is given. This is the cutoff threshold between mild and significant abnormalities, which can be calibrated based on clinical sample data; the default value is 1.5. A score of 0 indicates no abnormality, a score of 1 indicates a mild abnormality, and a score of 2 indicates a significant abnormality.

[0095] The specific method for threshold calibration is as follows: using the results of manual muscle testing (MMT) as the gold standard, the comprehensive score corresponding to the maximum Youden index is selected as T1; the calibration sample should include subjects of different ages, genders, and muscle strength levels, with a sample size of no less than 100 cases to ensure the universality of the threshold. After the scores are summarized, each target muscle group corresponds to a final abnormal gait score, which can quantitatively reflect the degree of functional abnormality of the muscle group.

[0096] Step 43: Use the abnormal gait scores of each target muscle group as abnormal gait variables to generate abnormal gait variables. Abnormal gait scores for each target muscle group are used as abnormal gait variables to generate abnormal gait variables; among them, the abnormal gait scores for each target muscle group are obtained by binary classification or tertiary scoring.

[0097] In practice, abnormal gait variables are organized as vectors or arrays, with each element corresponding to an abnormal gait score for a target muscle group. The order of the variables corresponds one-to-one with the order of the target muscle groups, for example, arranged in the order of quadriceps femoris, gluteus medius, tibialis anterior, and triceps surae, forming a four-dimensional abnormal gait variable vector.

[0098] An example of the representation of abnormal gait variables is as follows: Binary classification values: X=[x1,x2,x3,x4], where x1, x2, x3, and x4 correspond to the abnormal gait scores of the quadriceps femoris, gluteus medius, tibialis anterior, and triceps surae muscles, respectively, and are either 0 or 1.

[0099] The three-level score is calculated as follows: X = [x1, x2, x3, x4], where x1, x2, x3, and x4 correspond to the abnormal gait scores of the four muscle groups, and the values ​​are 0, 1, or 2.

[0100] The vector form of abnormal gait variables can be represented as: ; in, (Right now (Corresponds to the quadriceps femoris muscle) (Right now (Corresponds to the gluteus medius) (Right now (Corresponds to the tibialis anterior muscle) (Right now This corresponds to the triceps surae muscle. If a binary classification method is used, then... If a three-level rating system is used, then , = 1, 2, 3, 4.

[0101] The generated abnormal gait variables can be stored in structured data formats such as JSON, XML, or CSV for easy model retrieval and calculation. In addition to the scores of each muscle group, the variables can also include metadata such as basic subject information, test time, and data source for easy data management and traceability.

[0102] Once abnormal gait variables are generated, they can be used as input to the muscle function scoring model for subsequent muscle function risk assessment.

[0103] Step 5 uses the abnormal gait variable data from the training samples and the SPPB gold standard data to determine the correlation strength between abnormal gait and muscle function impairment for each target muscle group through multivariate logistic regression analysis. Then, it calculates the scoring weights for each muscle group, establishes scoring formulas and risk stratification criteria, and finally forms a trained muscle function scoring model. This model is a pre-defined model and can be directly used for subsequent muscle function risk assessment of subjects after its establishment.

[0104] In this embodiment, step 5 includes: Step 51: Obtain the abnormal gait variable dataset of the training samples. In specific implementation, the abnormal gait variable dataset of the training samples is obtained through the methods described in steps 3 and 4, that is, by performing pose estimation, abnormal gait annotation, and muscle group mapping on the continuous video data of the training samples, and finally obtaining the abnormal gait variables corresponding to each training sample.

[0105] Strict data quality control is required during the dataset construction process: (1) Completeness check: Check whether the abnormal gait variables of each sample are complete, whether the scores of the four target muscle groups are all available, and samples with missing data should be removed or supplemented.

[0106] (2) Consistency check: The consistency of the labeling results is checked to ensure that the consistency of the labeling results of different evaluators meets the preset requirements. Samples that do not meet the consistency requirements need to be relabeled or removed.

[0107] (3) Outlier detection: Statistical methods are used to detect outliers in the data. For data points that deviate significantly from the normal range, the original data must be verified and appropriate processing is required.

[0108] (4) Data Standardization: Standardize data from different sources and batches to ensure uniformity in data format and scoring criteria. The number of training samples should meet the requirements of statistical modeling, generally recommended to be 10 to 20 times the number of independent variables to ensure the stability and reliability of the model. For a model with four independent variables, it is recommended that the number of training samples be no less than 200. The dataset can be divided into training and validation sets according to a certain ratio, for example, 70% as the training set for model building and 30% as the validation set for model validation to evaluate the model's generalization ability.

[0109] Step 52: Obtain SPPB test data for the training samples. Define SPPB ≤ 9 points as impaired muscle function and SPPB > 9 points as no impairment. Perform binary classification transformation to obtain the outcome variable dataset. Use SPPB impairment as the model outcome variable, preferably defining SPPB ≤ 9 points as impaired muscle function and SPPB > 9 points as no impairment. Obtain SPPB test data for the training samples and perform binary classification transformation to obtain the outcome variable dataset. In specific implementation, SPPB, or Short Physical Performance Battery, is a commonly used tool for assessing lower limb function in older adults. It includes three parts: balance test, walking speed test, and seated stand test. The total score ranges from 0 to 12 points, with lower scores indicating poorer physical fitness. SPPB testing must be conducted by trained professionals according to standard operating procedures to ensure the accuracy and reliability of the test results. The test content includes: (1) Balance test: The subjects’ static balance ability was assessed by testing their maintenance time in lateral parallel stance, semi-tandem stance and full tandem stance.

[0110] (2) Walking speed test: The time it takes for the subject to walk over a distance of 4 meters is measured, the walking speed is calculated, and the walking ability is assessed.

[0111] (3) Seat-to-Stand Test: The time required for the subject to stand up from the chair 5 times consecutively was measured to assess lower limb muscle strength and explosive power. The scores for the three parts ranged from 0 to 4 points, and the total score was the sum of the scores for the three parts, ranging from 0 to 12 points. After obtaining the total SPPB score for each training sample, a binary classification transformation was performed according to a preset threshold: SPPB ≤ 9 points was defined as impaired muscle function and assigned a value of 1; SPPB > 9 points was defined as unimpaired and assigned a value of 0. After the transformation, an outcome variable dataset was formed, with each sample corresponding to a binary outcome label. During the data transformation process, the SPPB data needed to be quality checked to ensure the integrity and accuracy of the test data. Data with missing information or non-standard testing procedures needed to be removed to ensure the reliability of the outcome variables.

[0112] Step 53: Construct a multivariate logistic regression model with abnormal gait variables as independent variables and outcome variables as dependent variables, and perform regression analysis to obtain the regression coefficients corresponding to each target muscle group. Using the abnormal gait variables corresponding to each target muscle group as independent variables, construct a multivariate logistic regression model to estimate the association strength between abnormal gait patterns related to each target muscle group and SPPB damage. The model form is as follows: ; Where X1 represents the quadriceps femoris-related abnormal gait pattern, X2 represents the gluteus medius-related abnormal gait pattern, X3 represents the tibialis anterior-related abnormal gait pattern, and X4 represents the triceps surae-related abnormal gait pattern; β1, β2, β3, and β4 represent the regression coefficients corresponding to each abnormal gait pattern.

[0113] Multivariate logistic regression models are a common statistical modeling method in this field. This application does not take the logistic regression model itself as its inventive point, but rather as a means of determining the weights of abnormal gait patterns. In practice, regression analysis can be performed using statistical software such as SPSS, R, and Python. The modeling process includes the following steps: (1) Univariate analysis: First, perform univariate logistic regression analysis on each independent variable to preliminarily screen variables that are statistically associated with the outcome variable. Generally, variables with P values ​​less than 0.1 or 0.2 are included in multivariate analysis.

[0114] (2) Multivariate analysis: The selected independent variables were simultaneously included in the model for multivariate logistic regression analysis. Stepwise regression or forced entry method was used to select variables, and finally the regression coefficients, standard errors, Wald values, p-values, OR values ​​(odds ratios), and 95% confidence intervals of each variable were obtained. This application classifies abnormal gait patterns obtained from video posture estimation according to the functions of the main muscle groups of the lower limbs, and constructs an EGPS scoring system for identifying muscle function risks in the elderly based on the strength of its association with SPPB damage.

[0115] The core of the muscle function scoring model in this application is not simply the application of a Logistic regression model, but rather a continuous technical process that uses video abnormal gait recognition, lower limb target muscle group mapping, SPPB damage association modeling, integer weighting, and EGPS total score output to achieve an interpretable quantitative assessment of the risk of muscle function impairment in the elderly.

[0116] (3) Model testing: Perform goodness-of-fit tests on the constructed model, such as the Hosmer-Lemeshow test, to evaluate the model's fit; at the same time, calculate the model's C statistic (AUC value) to evaluate the model's discriminative ability.

[0117] (4) Multicollinearity diagnosis: Check whether there is multicollinearity among the independent variables. This can be determined by calculating the variance inflation factor (VIF). If the VIF is greater than 5 or 10, it indicates severe multicollinearity, and the variables need to be treated accordingly. After the regression analysis is completed, the regression coefficient β value and related statistics corresponding to each target muscle group are obtained, which provide a basis for the subsequent calculation of scoring weights.

[0118] Step 54: Calculate the assigned weights based on the regression coefficients of each target muscle group to obtain the weight allocation for each muscle group. After obtaining the regression coefficients of each variable, select the regression coefficient with the smallest absolute value and positive risk significance in the model as the benchmark coefficient, and set its corresponding weight to 1. The weights of other variables are converted according to the ratio of their regression coefficients to the benchmark coefficient, and the nearest integer is taken to obtain the assigned weights for each target muscle group. The calculation formula is as follows: ; in, This represents the scoring weight corresponding to the i-th target muscle group. This represents the regression coefficient of abnormal gait variables related to the target muscle group. This represents the smallest positive regression coefficient among the variables included in the model, and "round" indicates rounding to the nearest integer.

[0119] If the regression coefficient of a variable is inconsistent with the risk of muscle function impairment, or if its statistical stability does not meet the preset requirements, it may be excluded from the final scoring model. For example, in one embodiment, multivariate logistic regression analysis showed that the regression coefficient of the quadriceps-related abnormal gait pattern was the highest, the regression coefficients of the gluteus medius and tibialis anterior-related abnormal gait patterns were in the middle, and the regression coefficient of the triceps surae-related abnormal gait pattern was the lowest. Using the regression coefficient of the triceps surae-related abnormal gait pattern as the baseline coefficient and assigning it a value of 1, after integer conversion, the scoring weights for the quadriceps, gluteus medius, tibialis anterior, and triceps surae were determined to be 5, 2, 2, and 1, respectively.

[0120] See Table 1 below: Table 1 shows the univariate and multivariate logistic regression analyses and the weighted scores for each muscle group.

[0121] In practical implementation, the following points should be noted in the weight calculation process: (1) Positive risk screening: Only variables with positive regression coefficients are included in the weight calculation. Negative coefficients indicate that the abnormal gait is associated with a reduced risk of muscle function impairment, which is not in line with expectations and is generally not included in the scoring model.

[0122] (2) Statistical significance requirement: Variables included in the model should meet certain statistical significance requirements, such as a p-value less than 0.05, to ensure the reliability of the weights. For variables that are close to the significance level, it can be decided whether to include them based on professional judgment.

[0123] (3) Integerization: After rounding the weights, the relative size relationship between each variable should be retained to ensure that the weights can reflect the degree of contribution of different muscle group abnormalities to the risk of muscle function impairment.

[0124] (4) Weight range control: The weight range should not be too large, generally controlled between 1 and 10, to facilitate clinical use and understanding. Through the above calculations, the final scoring weight configuration for each target muscle group is obtained, which is used to establish the subsequent scoring formula.

[0125] Step 55: Establish a formula for calculating the total muscle function risk score based on the weight configuration of each muscle group, thereby obtaining the scoring calculation rules. The formula for calculating the total muscle function risk score is thus obtained: The total muscle function risk score is calculated as follows: 5 × quadriceps-related abnormal gait score + 2 × gluteus medius-related abnormal gait score + 2 × tibialis anterior-related abnormal gait score + 1 × gastrocnemius-related abnormal gait score. The abnormal gait score for each muscle group can be either 0 or 1; 0 indicates no abnormal gait pattern was observed for that muscle group, and 1 indicates an abnormal gait pattern was observed for that muscle group.

[0126] If a rating scale is used, the abnormal gait scores related to each muscle group can be set to levels 0, 1, and 2, representing no abnormality, mild abnormality, and significant abnormality, respectively. The total EGPS score for each subject is obtained by multiplying the abnormal gait scores related to each muscle group by their respective weights and then summing them up. The higher the score, the higher the risk of muscle function impairment.

[0127] In practical implementation, the scoring calculation rules must clarify the following: (1) Scoring formula: Clarify the calculation method of the total score, that is, multiply the abnormal gait scores of each muscle group by their corresponding weights and then sum them.

[0128] (2) Variable values: Clarify the value method of abnormal gait scores for each muscle group. Depending on the actual needs, binary classification (0 / 1) or three-level scoring (0 / 1 / 2) can be selected.

[0129] (3) Weight correspondence: Clearly define the muscle group corresponding to each weight to ensure that the correspondence is correct during calculation.

[0130] (4) Score range: Clearly define the possible range of the total score to facilitate understanding and interpretation. For example, when using a binary scoring system, the total score range is 0 to 10 points; when using a three-level scoring system, the total score range is 0 to 20 points.

[0131] (5) Scoring Interpretation: Clarify the relationship between high and low scores and the risk of muscle function impairment, i.e., the higher the score, the higher the risk of impairment. After the scoring calculation rules are established, automated calculation can be achieved through programming to improve assessment efficiency. For example, a simple calculation tool can be developed to automatically calculate the total score and risk level after inputting the abnormal gait scores of each muscle group. Step 56: Based on the total muscle function risk score and outcome variable of the training sample, determine the risk stratification threshold through the subject working characteristic curve and Youden index to obtain the risk stratification standard. Further, the subjects can be risk-stratified based on the EGPS total score. For example, those with a total score below the first preset threshold are judged as low risk, those with a total score between the first and second preset thresholds are judged as medium risk, and those with a total score above the second preset threshold are judged as high risk. The threshold can be determined based on the SPPB impairment status in the subject sample through a combination of subject working characteristic curve, Youden index, sensitivity, and specificity.

[0132] In practical implementation, the process for determining the risk stratification threshold is as follows: (1) Calculate the total score of the training samples: Calculate the total muscle function risk score for each training sample according to the established scoring formula.

[0133] (2) Plotting ROC curves: Using SPPB damage status as the gold standard and total muscle function risk score as the test variable, plot the receiver operating characteristic (ROC) curve, calculate the area under the curve (AUC), and evaluate the discrimination ability of the scoring model.

[0134] (3) Calculate the Youden Index: Youden Index = Sensitivity + Specificity - 1, which reflects the effectiveness of the screening method. The total score corresponding to the maximum Youden Index is the optimal cutoff point.

[0135] (4) Determine the stratification thresholds: Based on clinical needs and ROC curve analysis results, determine two risk stratification thresholds. The first threshold can be the score corresponding to the maximum Youden index, used to distinguish between low risk and medium-to-high risk; the second threshold can be determined according to specificity requirements, for example, the score corresponding to a specificity of 85% or 90%, used to distinguish between medium risk and high risk.

[0136] (5) Validation threshold: The determined threshold is validated on the validation set, and the sensitivity, specificity, positive predictive value, negative predictive value and other indicators corresponding to each risk stratification are calculated to evaluate the stratification effect.

[0137] Once the risk stratification criteria are established, the clinical significance and treatment recommendations for each stratum need to be clarified, for example: Low risk: Muscle function is basically normal. It is recommended to maintain a healthy lifestyle and have regular physical examinations.

[0138] Medium risk: There is a certain risk of muscle function decline. Targeted exercise intervention and regular follow-up examinations are recommended.

[0139] High risk: There is a high risk of muscle function impairment. It is recommended to seek medical attention promptly for further professional evaluation and rehabilitation treatment.

[0140] Step 57: The trained muscle function scoring model includes the weight configuration of each muscle group, scoring calculation rules, and risk stratification criteria.

[0141] In practice, the muscle function scoring model is stored and managed in a structured format, mainly comprising the following components: (1) Basic information of the model: including metadata such as model name, version number, establishment time, training sample source, and sample size.

[0142] (2) Weighting of each muscle group: The weighting of the four target muscle groups (quadriceps femoris, gluteus medius, tibialis anterior, and triceps surae).

[0143] (3) Scoring calculation rules: the calculation formula for the total muscle function risk score, the variable value selection method, the score range, etc.

[0144] (4) Risk stratification criteria: the total score threshold range corresponding to the three levels of low risk, medium risk and high risk, and the clinical interpretation of each level.

[0145] (5) Model performance metrics: AUC, sensitivity, specificity, Youden index, and other performance evaluation metrics of the model on the training and validation sets. The model can be stored in a database, configuration file, or dedicated model file, and can be directly loaded and called when performing muscle function assessment. After the model is built, it should be updated and optimized regularly. With the accumulation of new data, the model can be retrained or calibrated to maintain the accuracy and applicability of the model.

[0146] In another embodiment, step 5 of this application uses a self-developed scoring model, the specific acquisition method of which is as follows: Step 501: Construct a three-layer causal semantic ontology structure, which includes a gait abnormality layer, a muscle function layer, and a risk outcome layer. Each layer contains multiple semantic nodes, and the layers are connected by a unidirectional causal dependency. A three-layer causal semantic ontology structure is constructed, which includes a gait abnormality layer, a muscle function layer, and a risk outcome layer. Each layer contains multiple semantic nodes, and there is a unidirectional causal dependency relationship between the layers.

[0147] In practice, the three-layer causal semantic ontology is constructed according to the causal logic of cause → intermediate mechanism → result, with each layer corresponding to a link in the causal chain.

[0148] First layer (cause layer): Gait abnormality layer G The gait anomaly layer contains all observable anomalous gait pattern nodes, denoted as G={g1,g2,...,g...}. m}, where m is the total number of abnormal gait types. In this embodiment, gait abnormality nodes include: pelvic tilt, pelvic depression, compensatory trunk swing, knee hyperextension, knee hypoflexion, foot drop, dragging gait, and insufficient gait propulsion. Each node represents a specific gait abnormality and is the starting cause of the causal chain.

[0149] Second layer (middle layer): Muscle function layer M The muscle function layer contains the functional state nodes of each target muscle group, denoted as M={m1,m2,...,m n} where n is the number of target muscle groups. In this embodiment, muscle function nodes include quadriceps femoris dysfunction, gluteus medius dysfunction, tibialis anterior dysfunction, and triceps surae dysfunction. Each node represents the functional state of a muscle or muscle group and serves as an intermediate link between gait abnormalities and risk outcomes.

[0150] Third layer (outcome layer): Risk outcome layer R The risk outcome layer contains the outcome node of impaired muscle function, denoted as R={r0,r1}. Here, r0 represents no muscle function impairment, and r1 represents impaired muscle function. This layer is the final result of the causal chain and is the core outcome of clinical concern.

[0151] Interlayer causal relationships are defined as unidirectional causal dependencies: nodes in the gait abnormality layer are the causes, influencing nodes in the muscle function layer and ultimately leading to the outcome in the risk result layer. That is, gait abnormality → muscle function abnormality → muscle function impairment, forming a complete causal transmission pathway. Only top-down causal edges exist between layers; there are no intra-layer or reverse edges.

[0152] The construction of the ontology requires the integration of knowledge from biomechanics, kinematics, and clinical medicine, and must be completed collaboratively by domain experts to ensure the rationality and accuracy of causal relationships. The completed causal semantic ontology provides a semantic framework and structural foundation for the subsequent construction of heterogeneous graphs.

[0153] Step 502: Based on the three-layer causal semantic ontology structure, construct a weighted directed muscle-gait-risk heterogeneous graph. The node set of the weighted directed muscle-gait-risk heterogeneous graph includes all gait abnormality nodes, muscle function nodes, and risk outcome nodes, and the edge set includes inter-layer directed causal edges. Based on the aforementioned three-layer causal semantic ontology structure, a weighted directed muscle-gait-risk heterogeneous graph (MGR-HG) is constructed. The node set of the weighted directed muscle-gait-risk heterogeneous graph includes all gait abnormality nodes, muscle function nodes, and risk outcome nodes, and the edge set includes inter-layer directed causal edges.

[0154] In practical implementation, a weighted directed heterogeneous graph is denoted as H=(V,E,W), where: The node set V: V=G∪M∪R contains all gait abnormality nodes, muscle function nodes, and risk outcome nodes, which are three different types of nodes, hence it is called a heterogeneous graph.

[0155] Edge set E: E= ∪ It contains only directed causal edges between layers, and no intra-layer edges. It is the set of causal edges pointing from the gait abnormality layer to the muscle function layer, indicating that gait abnormality is the cause of muscle function abnormality; This is the set of causal edges pointing from the muscle function layer to the risk outcome layer, indicating that abnormal muscle function is a cause of the risk outcome. Each edge e( )∈ Indicates gait abnormality May lead to muscle Functional malfunction; each edge e( , )∈ Indicates muscles Functional abnormalities may lead to risky outcomes .

[0156] Weight function W: Assigns a weight value within the interval [0,1] to each directed causal edge, representing the strength of the causal relationship. The edge weights are initialized using the following formula: For the causal edge from the gait anomaly layer to the muscle function layer, the initial weights are:

[0157] in, This is the Sigmoid function, used to normalize the calculation result to the interval [0, 1]. Its form is: (x) = 1 / (1+e) -x ); PMI ( ) indicates gait abnormality With muscles The point mutual information is obtained from the statistics of the training samples, and the calculation formula is: ; in, Gait abnormality With muscles The joint probability of simultaneous functional abnormalities. Gait abnormality The marginal probability of occurrence For muscles Marginal probability of functional abnormality; For muscles The biomechanical prior strength is assigned by domain experts based on biomechanical knowledge, with a value range of [0, 1], reflecting the importance of the muscle in walking; and The balancing coefficient is used to adjust the weights of data-driven statistical associations and prior knowledge, satisfying the following conditions: + = 1, in this embodiment Take 0.6, Take 0.4.

[0158] For the causal edge from the muscle function layer to the risk outcome layer, the initial weights are: ; in, For muscles The corresponding univariate odds ratio, obtained from the univariate logistic regression analysis of the training samples, reflects the strength of the association between the abnormal muscle function and the muscle function impairment outcome. This is a scaling parameter used to control the decay rate of the exponential function. In this embodiment... Take 0.5; this formula converts the ratio into a weight value in the interval [0, 1]. The larger the ratio, the higher the corresponding edge weight, indicating a stronger causal relationship.

[0159] in The value range is 0.3 to 0.8, with a default value of 0.5. This parameter maps the numerical range of the OR value to the weighted range of [0,1). It is adjusted downwards when the overall single-factor OR value is too high. When the overall level is low, adjust upwards. This ensures the distinguishability of edge weights.

[0160] For undamaged ending nodes The edge weight can be set as the complement of the corresponding damaged edge weight, or calculated separately according to the actual situation.

[0161] Using the above method, an initialized weighted directed muscle-gait-risk heterogeneous graph is constructed, providing a graph structure foundation for subsequent causal propagation and centrality calculation.

[0162] Step 503: Execute the reverse causal decay propagation algorithm on the weighted directed muscle-gait-risk heterogeneous graph. Starting from the risk outcome node, propagate the causal influence backward along the causal edges. Iterate to obtain the final causal activation value of each node. The iteration terminates when the maximum value of the change in activation value of all nodes is less than a preset threshold, thereby obtaining the set of node causal activation values. The weighted directed muscle-gait-risk heterogeneous graph is subjected to the Reverse Causal Decay Propagation (RCDP) algorithm. Starting from the risk outcome node, the causal influence is propagated backward along the causal edges. During the propagation process, a path decay coefficient is used to attenuate the causal influence of long paths. The path decay coefficient ranges from 0.7 to 0.95. The final causal activation value of each node is obtained through iterative calculation. The iteration terminates when the maximum value of the change in activation value of all nodes is less than a preset threshold, thereby obtaining the set of node causal activation values.

[0163] In practical implementation, the basic idea of ​​the reverse causal decay propagation algorithm is: starting from the risk outcome node, trace back layer by layer along the opposite direction of the causal edge, propagate the causal influence in reverse to the upstream muscle node and gait node, and finally obtain the activation degree of each node in the whole causal network.

[0164] initialization: Define the causal activation value A(v) for each node v, representing the activity level of that node in the causal pathway. Initially: The activation value of the risk-damaged node is set to 1, i.e., A( )=1 indicates that the risk outcome is the source of causal propagation; The activation value of nodes unaffected by risk is set to 0, i.e., A( )=0; The initial activation value of all gait abnormality nodes and muscle function nodes is set to 0, i.e., A( )=0, A( )=0.

[0165] Iterative propagation process: The algorithm iterates in rounds, propagating the causal influence backward from the result layer to the cause layer once in each round. In the t-th iteration: Step 1: Update the activation values ​​of the muscle function layer nodes For each muscle node Its new activation value equals the original activation value plus the influence propagated back from the risk layer node. The calculation formula is: ; in, Risk node in round t-1 Activation value; For muscle nodes To the risk point Causal edge weights; To start from the risk node To muscle nodes The shortest path length is given here. Since the paths are directly connected, the path length is always 1. Therefore, the 1 / d term equals 1 and is mainly used for attenuation of longer paths.

[0166] Step 2: Update the activation values ​​of the gait anomaly layer nodes. For each gait anomaly node The new activation value is equal to the original activation value plus the influence propagated back from the muscle layer nodes. The calculation formula is: ; in, For the muscle nodes at round t-1 Activation value; Gait nodes To muscle nodes Causal edge weights; The path attenuation coefficient has a value range of (0, 1). In this embodiment, it is set to 0.85 to simulate the characteristic that the causal influence decreases as the path length increases. To from muscle nodes To gait node The shortest path length is always 1, since the paths are directly connected. This value is determined by a grid search on the training set, with a search interval of [0.7, 0.95] and a step size of 0.05. The parameter value corresponding to the highest AUC value of the model is selected. When the causal link hierarchy expands, the value can be appropriately reduced. To avoid excessive attenuation of long-path influence.

[0167] Iteration termination condition: After each iteration, calculate the change in the activation values ​​of all nodes, i.e., | - The iteration stops when the maximum change in activation value of all nodes is less than a preset threshold ε. In this embodiment, ε is set to 10. -6 To ensure convergence accuracy.

[0168] The verification rule for iterative convergence is: convergence is determined when the maximum change in activation value of all nodes in three consecutive iterations is less than ε; the selectable range of ε is... The higher the precision requirement, the smaller the value should be, and the corresponding number of iterations should be increased accordingly.

[0169] After iterative convergence, the final causal activation value A*(v) of each node is obtained, forming a set of node causal activation values. The higher the final activation value, the more important the node is in the causal network and the greater its contribution to the risk outcome.

[0170] Step 504: Based on the set of causal activation values ​​of the nodes, calculate the causal betweenness centrality of each muscle node, thereby obtaining the causal betweenness centrality of each muscle node. Based on the set of causal activation values ​​of the nodes, the causal betweenness centrality (CBC) of each muscle node is calculated, thereby obtaining the causal betweenness centrality of each muscle node.

[0171] In practical implementation, causal betweenness centrality is an indicator of a node's pivotal role in a causal pathway. It comprehensively considers the frequency of a node's occurrence across all causal paths and the activation strength of nodes at both ends of the path. For a muscle node m... j The formula for calculating its causal betweenness centrality is: ; in, To start from gait anomaly nodes To the risk outcome point The sum of the total weights of all causal paths; To start from gait anomaly nodes To the risk outcome point And passing through muscle nodes The sum of the total weights of all causal paths; Nodes with gait abnormalities The final causal activation value; Risk outcome node The final causal activation value.

[0172] Methods for calculating path weights: For a causal path P = from a gait node to a risk node → → Its path weight is defined as the product of the weights of all edges on the path multiplied by the path decay factor, that is: ; in, λ is the path length (number of edges), which is 2 here; λ is the path decay coefficient, consistent with the propagation algorithm.

[0173] For cases with multiple paths (such as one gait node corresponding to multiple muscle nodes, or one muscle node corresponding to multiple gait nodes), calculate the weight of each path separately and then sum them.

[0174] Calculation process: First, enumerate all causal paths from gait anomaly nodes to risk outcome nodes, calculate the weight of each path, and obtain... ; Then for each muscle node Statistics of all processes The sum of the path weights is obtained. ; Calculate the ratio / This ratio represents the ratio from arrive Of all the causal influences, those passing through muscle nodes The proportion; Multiply this ratio by the activation values ​​of the two endpoints. and This yields the path to the muscle nodes. The contribution of causal betweenness centrality; Summing all combinations of gait nodes and risk nodes yields the final muscle nodes. causal betweenness centrality .

[0175] The higher the causal betweenness centrality, the more important the muscle node is in the causal transmission pathway from gait abnormality to risk outcome, the stronger its pivotal role, and the greater its contribution to muscle function impairment.

[0176] Step 505: Convert the causal betweenness centrality of each muscle node into assigned weights to obtain the weight configuration for each muscle group. The causal betweenness centrality of each muscle node is transformed into a weight assignment, thereby obtaining the weight configuration of each muscle group.

[0177] In practice, the transformation process consists of three steps: normalization, nonlinear mapping, and integerization. Step 1: Normalization First, the causal betweenness centrality values ​​of all muscle nodes are normalized and mapped to the interval [0,1]. The calculation formula is as follows: ; in, This represents the maximum causal betweenness centrality among all muscle nodes. After normalization, the most important muscle nodes have a normalization value of 1, while other nodes are scaled down proportionally.

[0178] Step 2: Nonlinear Mapping A nonlinear mapping is applied to the normalized centrality values ​​to amplify the difference between high-centrality and low-centrality muscles, making the weight distribution more consistent with clinical reality. The calculation formula is as follows: ; in, The non-linear adjustment index is set to 0.7 in this embodiment. When the index is less than 1, it will amplify the relative difference between the high and low values. The maximum weight is set to 5 in this embodiment to ensure that the highest weight does not exceed 5 points, which is convenient for clinical use and understanding. The value range is 0.5~1.0, with a default value of 0.7. This parameter is used to adjust the dispersion of the weight distribution. A smaller value is used when it is necessary to enhance the differentiation of the core muscle groups, and a larger value is used when it is necessary to smooth the weight differences. The value is determined jointly through clinical expert surveys and model performance verification.

[0179] Step 3: Integerization The weights after nonlinear mapping are rounded to the nearest integer to obtain the final integer weights. The calculation formula is as follows: ; The round (•) function is used for rounding to the nearest integer.

[0180] Through the above transformation, the integer weights of each target muscle group are finally obtained. For example, in one embodiment, the quadriceps femoris has the highest causal betweenness centrality, corresponding to a weight of 5 points; the gluteus medius and tibialis anterior have moderate centrality, corresponding to a weight of 2 points; and the triceps surae has the lowest centrality, corresponding to a weight of 1 point.

[0181] Step 506: Establish a formula for calculating the total muscle function risk score based on the weight configuration of each muscle group, and determine the risk stratification threshold through the weighted Youden index, thereby obtaining the scoring calculation rules and risk stratification standards. A formula for calculating the total risk score of muscle function is established based on the weight configuration of each muscle group, and the risk stratification threshold is determined by the weighted Youden index, thereby obtaining the scoring calculation rules and risk stratification standards.

[0182] In practice, the process of establishing scoring calculation rules and risk stratification standards is as follows: 1. Establish a formula for calculating the total risk score of muscle function. The total muscle function risk score is equal to the sum of the abnormal gait scores of each target muscle group multiplied by their corresponding weights. The calculation formula is: ; in, Assign a weight to the j-th muscle group. The abnormal gait score is for the j-th muscle group.

[0183] Abnormal gait scores for each muscle group can be obtained using a binary or ternary scoring system: in binary scoring, 0 indicates no abnormality and 1 indicates abnormality; in ternary scoring, 0 indicates no abnormality, 1 indicates mild abnormality, and 2 indicates significant abnormality.

[0184] A higher total score indicates a higher risk of impaired muscle function in the subject. For example, when using a binary scoring system, the total score ranges from 0 to 10; when using a three-level scoring system, the total score ranges from 0 to 20.

[0185] Risk stratification thresholds were determined using the weighted Youden index. The traditional Youden index places equal emphasis on sensitivity and specificity. However, in clinical applications, the emphasis on sensitivity and specificity may differ depending on the specific scenario. Therefore, this embodiment uses a weighted Youden index to determine the risk stratification threshold, allowing for flexible adjustment of the weights of sensitivity and specificity according to clinical needs.

[0186] The formula for calculating the weighted Youden index is: ; in, The sensitivity corresponding to the threshold t is the proportion of actual damage that is correctly identified as damage. is the specificity corresponding to the threshold t, that is, the proportion of those that are actually undamaged and are correctly identified as undamaged; Weights for sensitivity; The weights are for specificity; generally, they satisfy... + = 1, and the ratio of the two can be adjusted according to clinical needs.

[0187] In this embodiment, considering the screening attributes of muscle function assessment, the weight of sensitivity can be appropriately increased, for example, by taking... = 0.6, = 0.4, in order to minimize missed diagnoses.

[0188] Determine the threshold for three-level risk stratification First, the total muscle function risk score for each subject in the training sample was calculated, and then the ROC curve was plotted using the SPPB gold standard as a reference.

[0189] The total score that maximizes the weighted Youden index WYI(t) is selected as the first threshold T1 to distinguish between low-risk and medium-to-high-risk individuals. The comprehensive screening efficiency is optimal at this threshold.

[0190] Then, the lowest total score that achieves the preset specificity requirement (e.g., specificity ≥ 0.85) is selected as the second threshold T2, used to distinguish between medium-risk and high-risk individuals. This threshold has a high positive predictive value and is suitable for diagnosing high-risk individuals.

[0191] The final three-tier risk stratification standard was formed: Total score < T1 indicates low risk; T1 ≤ Total Score ≤ T2 indicates medium risk; Total score > T2 indicates high risk.

[0192] Step 507: Combining the weight configuration of each muscle group, the scoring calculation rules, and the risk stratification criteria, a trained muscle function scoring model is obtained. By combining the weight configuration of each muscle group, the scoring calculation rules, and the risk stratification criteria, a trained muscle function scoring model is obtained.

[0193] In practice, the trained muscle function scoring model is a complete and independently operating scoring system, mainly consisting of the following components: (1) Basic information of the model: including metadata such as model name, version number, establishment time, source of training samples, sample size, and algorithm type, which are used for model identification and management.

[0194] (2) Weighting of each muscle group: The assigned weights of the four target muscle groups (quadriceps femoris, gluteus medius, tibialis anterior, and triceps surae) are the core parameters of the model.

[0195] (3) Scoring calculation rules: including the calculation formula of the total muscle function risk score, the scoring method of abnormal gait of each muscle group (two-category or three-level scoring), the range of total score and the meaning of the score.

[0196] (4) Risk stratification criteria: the total score threshold range corresponding to the three levels of low risk, medium risk and high risk, as well as the clinical interpretation and recommended treatment measures for each risk level.

[0197] (5) Model performance metrics: various performance evaluation metrics of the model on the training set and validation set, such as AUC value, sensitivity, specificity, weighted Youden index, positive predictive value, negative predictive value, etc., are used to evaluate the effectiveness and reliability of the model.

[0198] The model can be stored as structured data in a database, configuration file, or dedicated model file, and can be directly loaded and invoked during muscle function assessment. After the model is established, it should be maintained and updated regularly. As new clinical data accumulates, the model can be retrained or its parameters calibrated to maintain its accuracy and applicability.

[0199] The model designed using this application has the following advantages: Existing scoring models based on logistic regression are flattened linear statistical models that can only establish a direct statistical association between "muscle group abnormality variables and risk outcomes". They cannot reveal the hierarchical causal transmission process of external gait abnormalities, internal muscle function abnormalities, and overall functional impairment, nor can they quantify the contribution paths of different gait abnormalities to outcomes through different muscle pathways. The interpretability of the model is limited to the level of variable association strength and lacks causal logic support.

[0200] This approach constructs a three-layer causal semantic ontology structure—gait abnormality layer, muscle function layer, and risk outcome layer—to structurally express the causal logic of clinical biomechanics. Based on this ontology, a weighted directed muscle-gait-risk heterogeneous graph is built, treating gait patterns, muscle function, and risk outcomes as different types of nodes, connected by unidirectional causal edges between layers. This fully reconstructs the causal transmission path from observed phenomena to underlying mechanisms and then to clinical outcomes. This approach overcomes the limitations of traditional regression models, which are flat and lack causal structure, making the scoring model interpretable and enabling a clear tracing of the upstream transmission pathways of risk outcomes.

[0201] The weighting of existing logistic regression relies entirely on the regression coefficients of the sample data, which is a purely data-driven statistical method. When the sample size is limited or the sample distribution is biased, the weights are prone to statistical fluctuations. Furthermore, it cannot incorporate prior knowledge from biomechanics and rehabilitation medicine to constrain and correct the results, resulting in insufficient model stability in small community sample scenarios.

[0202] This scheme achieves a fusion of statistical correlation and biomechanical prior in the edge weight initialization stage of heterogeneous graphs: For causal edges from the gait abnormality layer to the muscle function layer, it simultaneously introduces point mutual information (PMI) based on sample co-occurrence statistics and biomechanical prior strength (BioPri) based on domain expert knowledge. By adjusting the weights of the two through a balancing coefficient, it utilizes the statistical regularity of real data and constrains the rationality of the weights through professional priors. Even in small sample scenarios, it can obtain weight results that conform to biomechanical logic, significantly improving the model's small sample robustness and clinical adaptability. For causal edges from the muscle function layer to the risk outcome layer, it achieves weight normalization through exponential mapping of single-factor odds ratios, preserving the strength of statistical correlation while adapting to the propagation calculation logic of graph networks.

[0203] The regression coefficients of existing logistic regressions only reflect the marginal association strength between a single variable and the outcome. They cannot reflect the mediating role of muscle nodes in the entire causal transmission network, nor can they distinguish the structural importance of different muscles in the causal pathway. The weighting results are only statistically significant and lack support at the biomechanical mechanism level.

[0204] This scheme employs a reverse causal decay propagation algorithm, starting from the risk outcome node and tracing the causal influence backward along the causal edges. Iterative calculation yields the final causal activation value of all nodes, quantifying the activity and contribution of each node within the entire causal network. Building upon this, a causal betweenness centrality index is introduced, comprehensively considering the proportion of muscle nodes traversing all gait-risk causal paths, as well as the activation intensity of nodes at both ends of the path. This quantifies the pivotal core value of muscle nodes in causal transmission, and then transforms this into weighted scores. This weighting method goes beyond univariate statistical associations, assessing the core contribution of muscles from the structural perspective of the entire causal network. The physical meaning of the weights aligns more closely with the clinical mechanism by which muscle dysfunction is the core mediator between gait abnormalities and functional impairment.

[0205] The existing scheme uses the traditional Youden index to determine the risk stratification threshold, giving equal weight to sensitivity and specificity. It cannot adjust the assessment focus according to the needs of different clinical scenarios and is difficult to adapt to different application scenarios such as large-scale community screening and precise clinical diagnosis.

[0206] This scheme uses the weighted Youden index to calculate the risk stratification threshold, and can be flexibly adapted to different application scenarios by adjusting the weight coefficients of sensitivity and specificity: in large-scale screening of elderly people in the community, the sensitivity weight can be increased to reduce the missed diagnosis rate of muscle function impairment; in clinical rehabilitation assessment scenarios, the specificity weight can be increased to reduce misdiagnosis and over-intervention. At the same time, combined with the dual-threshold three-level stratification mechanism, it takes into account both the efficacy of initial screening and diagnostic efficacy, and greatly improves the model's adaptability to clinical scenarios.

[0207] Existing Logistic regression models require aggregating multiple gait abnormalities into a single muscle group variable before modeling, which cannot handle the complex many-to-many mapping relationship where one gait abnormality is associated with multiple muscle groups or one muscle group corresponds to multiple gait abnormalities. When adding new gait abnormality types, expanding target muscle groups, or changing outcome indicators, it is necessary to rebuild the variable system and completely retrain the model, resulting in poor scalability.

[0208] This approach is based on a modular structure of heterogeneous graphs, which allows for the flexible addition of gait abnormality nodes, muscle function nodes, and corresponding causal edges without reconstructing the overall model framework. At the same time, the computational framework for causal propagation and causal betweenness centrality has universal applicability and can be directly extended to more muscle groups such as the upper limbs and trunk, or more clinical outcome indicators such as fall risk and decline in daily activity ability. The scalability and generalization ability of the model are far superior to traditional statistical regression models.

[0209] In this embodiment, step 6 takes the abnormal gait variables generated in step 4 as input and loads the muscle function scoring model constructed in step 5 (containing three core parameters: weight configuration for each muscle group, scoring calculation rules, and risk stratification criteria). It then sequentially performs total score calculation, single muscle group status assessment, comprehensive risk stratification, and structured output of results, ultimately generating muscle function risk information that can be directly used for clinical assessment and intervention guidance. The calculation logic of this step is compatible with both multivariate logistic regression-based scoring models and causal heterogeneity graph-based scoring models; only the weights and threshold parameters of the corresponding models need to be loaded to complete the assessment.

[0210] Step 61 takes as input the abnormal gait scores of each target muscle group output in Step 4, and the pre-stored scoring weights of each muscle group in the scoring model. According to the model's preset scoring calculation rules, the abnormal gait score of each target muscle group is multiplied by its corresponding scoring weight and then summed to obtain the subject's total muscle function risk score (EGPS total score).

[0211] The calculation formula is: ; in, , , , The assigned weights for the quadriceps femoris, gluteus medius, tibialis anterior, and triceps surae muscles are determined by the model training in step 5. , , , These are the abnormal gait scores for the corresponding muscle groups.

[0212] The scoring supports two modes: one is a binary classification mode, with a score of 0 or 1, where 0 indicates that no abnormal gait pattern related to the muscle group was observed, and 1 indicates that an abnormal gait pattern related to the muscle group was observed, which is suitable for large-scale rapid screening in the community; the other is a three-level scoring mode, with scores of 0, 1, and 2, corresponding to no abnormality, mild abnormality, and obvious abnormality, respectively, which is suitable for refined clinical assessment scenarios.

[0213] In the implementation of the scoring model based on logistic regression, the weights assigned to the quadriceps femoris, gluteus medius, tibialis anterior, and triceps surae are 5, 2, 2, and 1, respectively. The total score ranges from 0 to 10 in the binary classification mode and from 0 to 20 in the tertiary scoring mode. A higher score indicates a higher risk of muscle function impairment.

[0214] Step 62: Determine the functional status of each target muscle group Based on the abnormal gait scores of each target muscle group, the functional status of each muscle group is assessed: When the abnormal gait score of a muscle group is 0, the functional status of that muscle group is considered normal, that is, no abnormal gait manifestations related to the function of that muscle group are detected. When a binary classification score is used and the score is 1, the functional status of the muscle group is determined to be abnormal. When using a three-level scoring system, a score of 1 indicates mild functional impairment, and a score of 2 indicates significant functional impairment. The degree of impairment is positively correlated with the frequency, duration, and magnitude of gait abnormalities.

[0215] The system outputs assessment results for each muscle group, forming a set of functional status data for the lower limb core muscle groups, including the quadriceps femoris, gluteus medius, tibialis anterior, and triceps surae. This data visually presents the functional impairment of individual muscle groups, providing precise targets for subsequent targeted interventions.

[0216] Step 63: Classify the overall risk level of the subjects Load the pre-stored risk stratification thresholds in the scoring model, compare the calculated total muscle function risk score with the thresholds, and classify the subject's overall muscle function impairment risk level: When the total score is lower than the first preset threshold, it is judged as low risk, indicating that the overall muscle function is intact, there is no obvious functional impairment, and the walking ability is basically normal. When the total score is between the first preset threshold and the second preset threshold, it is judged as medium risk, indicating that there is a certain degree of decline in muscle function, compensatory gait changes have occurred, and walking function has been slightly affected. When the total score is higher than the second preset threshold, it is judged as high risk, indicating a high risk of impaired muscle function, significant gait abnormalities, and a marked impact on walking function, with an increased risk of falls.

[0217] The first and second preset thresholds are determined by step 5 based on training sample data, through a combination of receiver operating characteristic curves, Youden index (the weighted Youden index is used in the causal heterogeneity plot scheme), sensitivity and specificity, to ensure that the stratification results are consistent with the SPPB gold standard.

[0218] Step 64: Generate structured muscle function risk information Integrating the aforementioned calculations and judgments, a complete risk information profile of the subject's muscle function is formed, specifically including four core categories: Results of single muscle group function: Functional status of quadriceps femoris, gluteus medius, tibialis anterior, and triceps surae (normal / mildly abnormal / significantly abnormal). Abnormal gait scores for each muscle group: Quantitative scores for the four target muscle groups can be used to trace the severity of the corresponding gait abnormalities; Muscle function risk score: The subject's comprehensive EGPS quantitative score, which can be used for cross-group comparison and longitudinal follow-up monitoring; Overall risk level: The stratification of low risk / medium risk / high risk directly reflects the overall risk level of muscle function impairment.

[0219] All information is output in a structured data format, which can be directly stored in the assessment database or imported into the subsequent rehabilitation intervention decision system for further analysis. It also supports the generation of standardized assessment reports for clinical staff to review.

[0220] Step 65: Output clinical intervention-oriented assessment description By combining comprehensive risk levels with information on affected muscle groups, corresponding clinical guidelines are generated to guide the development of individualized exercise interventions or rehabilitation training programs. Low-risk subjects: It is recommended to maintain existing exercise habits, conduct routine muscle function monitoring regularly, and no targeted rehabilitation intervention is required; For medium-risk subjects: corresponding muscle strengthening and functional training programs will be developed for the target muscle groups that are identified as abnormal. For example, for those with quadriceps abnormalities, the focus will be on knee extension muscle strength training, and for those with gluteus medius abnormalities, the focus will be on pelvic stability training. Gait and muscle function will be reviewed regularly. High-risk subjects: It is recommended to go to a medical institution in a timely manner for professional muscle strength testing and rehabilitation assessment, so that clinical staff can develop a standardized rehabilitation treatment plan. At the same time, fall prevention should be strengthened to reduce the risk of functional decline and accidental injury.

[0221] The EGPS scoring system developed in this scheme enables interpretable quantitative assessment of the risk of muscle function impairment in older adults, providing a standardized tool for muscle function screening in grassroots settings. This application also provides a muscle function state inference device based on visual pose estimation, the muscle function state inference device based on visual pose estimation includes: A continuous video data acquisition module for analysis, wherein the continuous video data acquisition module is used to acquire continuous video data to be analyzed; An abnormal gait data acquisition module is used to annotate and score abnormal gait in the continuous video data to be analyzed, thereby acquiring abnormal gait data. A muscle group mapping rule acquisition module is used to acquire preset muscle group mapping rules. An abnormal gait variable acquisition module is used to generate abnormal gait variables based on the abnormal gait data and preset muscle group mapping rules. A muscle function scoring model acquisition module, which is used to acquire a trained muscle function scoring model; A muscle function risk generation module is used to generate muscle function risk information for the subject based on the abnormal gait variables and a trained muscle function scoring model.

[0222] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for extrapolating muscle functional state based on visual pose estimation, characterized in that, The method for extrapolating muscle function status based on visual pose estimation includes: Step 1: Obtain the continuous video data to be analyzed; Step 2: Annotate and score the abnormal gait in the continuous video data to be analyzed, thereby obtaining abnormal gait data; Step 3: Obtain the preset muscle group mapping rules; Step 4: Generate abnormal gait variables based on the abnormal gait data and preset muscle group mapping rules; Step 5: Obtain the trained muscle function scoring model; Step 6: Generate the subject's muscle function risk information based on the abnormal gait variables and the trained muscle function scoring model.

2. The method for extrapolating muscle function state based on visual pose estimation as described in claim 1, characterized in that, Step 2 includes: Step 21: Observe the continuous video data to be analyzed frame by frame or segment by segment to identify abnormal gait; Step 22: Label the abnormal gait to obtain gait abnormality type data, which includes pelvic tilt or pelvic sinking, trunk compensatory swing, knee hyperextension or hypoflexion, foot drop or dragging, and insufficient gait propulsion. Step 23: Label the gait anomaly type data with occurrence features to obtain anomaly occurrence feature data, which includes the time point of anomaly occurrence, duration, and frequency of occurrence; Step 24: Classify the severity of the gait anomaly type data to obtain anomaly severity data. The anomaly severity is divided into three levels: mild, moderate, and severe. The abnormal gait data includes gait anomaly type data, anomaly occurrence characteristic data, and anomaly severity data.

3. The muscle function state deduction based on visual pose estimation as described in claim 2, characterized in that, Step 3 includes: Step 31: Obtain continuous video data of the training samples; Step 32: Perform pose estimation on the continuous video data of the training samples, extract the time series data of human key points, thereby constructing gait motion feature vectors and obtaining gait feature data of the training samples; Step 33: Perform abnormal gait annotation and scoring on the continuous video data of the training samples to obtain abnormal gait data of the training samples; Step 34: Based on the principles of biomechanics and muscle kinematics, perform correlation analysis on the gait feature data of the training samples and the abnormal gait data of the training samples to establish the correspondence between abnormal gait features and the functional state of the target muscle groups. Step 35: Based on the correspondence, organize and form a mapping rule from gait abnormality to muscle function, thereby obtaining a preset muscle group mapping rule; wherein, the target muscle group includes at least the gluteus medius, quadriceps femoris, triceps surae, and tibialis anterior.

4. The method for inferring muscle function state based on visual pose estimation as described in claim 3, characterized in that, Step 4 includes: Step 41: According to the preset muscle group mapping rules, map each abnormal gait in the abnormal gait data to the corresponding target muscle group, thereby obtaining the abnormal gait subset corresponding to each target muscle group. Step 42: Score and summarize the abnormal gait subsets corresponding to each target muscle group to obtain the abnormal gait score for each target muscle group; Step 43: Use the abnormal gait scores of each target muscle group as abnormal gait variables to generate abnormal gait variables; wherein, the abnormal gait scores of each target muscle group adopt binary classification or three-level scoring.

5. The method for extrapolating muscle function state based on visual pose estimation as described in claim 4, characterized in that, Step 5 includes: Step 51: Obtain the dataset of abnormal gait variables from the training samples; Step 52: Obtain the SPPB test data of the training samples, define SPPB ≤ 9 as impaired muscle function and SPPB > 9 as unimpaired, perform binary classification transformation to obtain the outcome variable dataset; Step 53: Using abnormal gait variables as independent variables and outcome variables as dependent variables, construct a multivariate logistic regression model and perform regression analysis to obtain the regression coefficients corresponding to each target muscle group; Step 54: Calculate the assigned weights based on the regression coefficients of each target muscle group to obtain the weight configuration for each muscle group; Step 55: Establish a formula for calculating the total muscle function risk score based on the weight configuration of each muscle group, thereby obtaining the scoring calculation rules; Step 56: Based on the total muscle function risk score and outcome variable of the training samples, determine the risk stratification threshold through the receiver operating characteristic curve and Youden index, thereby obtaining the risk stratification criteria; Step 57: The trained muscle function scoring model includes the weight configuration of each muscle group, scoring calculation rules, and risk stratification criteria.

6. The method for extrapolating muscle function state based on visual pose estimation as described in claim 5, characterized in that, Step 6 includes: Step 61: Calculate the total muscle function risk score based on the abnormal gait scores of each target muscle group; Step 62: Determine the functional status of each muscle group based on the abnormal gait score of each target muscle group. A score of 0 is considered normal, and a score greater than 0 is considered abnormal, thereby obtaining the functional status results of each muscle group. Step 63: Determine the subject's overall risk level based on the total muscle function risk score; wherein, the subject's muscle function risk information includes the functional status results of each muscle group, the abnormal gait score of each target muscle group, the total muscle function risk score, and the overall risk level.

7. The method for extrapolating muscle function state based on visual pose estimation as described in claim 4, characterized in that, Step 5 includes: Step 501: Construct a three-layer causal semantic ontology structure, which includes a gait abnormality layer, a muscle function layer, and a risk outcome layer. Each layer contains multiple semantic nodes, and there is a unidirectional causal dependency relationship between the layers. Step 502: Based on the three-layer causal semantic ontology structure, construct a weighted directed muscle-gait-risk heterogeneous graph. The node set of the weighted directed muscle-gait-risk heterogeneous graph includes all gait abnormality nodes, muscle function nodes, and risk outcome nodes, and the edge set includes inter-layer directed causal edges. Step 503: Execute the reverse causal decay propagation algorithm on the weighted directed muscle-gait-risk heterogeneous graph. Starting from the risk outcome node, propagate the causal influence backward along the causal edge. The final causal activation value of each node is obtained through iterative calculation. The iteration terminates when the maximum value of the change in activation value of all nodes is less than a preset threshold, thereby obtaining the set of node causal activation values. Step 504: Calculate the causal betweenness centrality of each muscle node based on the set of causal activation values ​​of the nodes, thereby obtaining the causal betweenness centrality of each muscle node; Step 505: Convert the causal betweenness centrality of each muscle node into assigned weights to obtain the weight configuration for each muscle group. Step 506: Establish a formula for calculating the total muscle function risk score based on the weight configuration of each muscle group, and determine the risk stratification threshold through the weighted Youden index, thereby obtaining the scoring calculation rules and risk stratification standards; Step 507: By combining the weight configuration of each muscle group, the scoring calculation rules, and the risk stratification criteria, a trained muscle function scoring model is obtained.

8. The method for extrapolating muscle function state based on visual pose estimation as described in claim 7, characterized in that, The edge weight of the directed causal edge is obtained by the following formula: For the causal edge from the gait anomaly layer to the muscle function layer, the initial weights are: ; in, For the Sigmoid function, Gait abnormality With muscles Point-to-point information, For muscles Biomechanical prior strength and The balance coefficients are satisfied. + =1; in The value range is 0.5 to 0.

7. The corresponding value range is 0.3 to 0.5; this range was determined through 5-fold cross-validation, and can be increased when the sample size is greater than 300 cases. The proportion is used to strengthen data-driven weighting; it can be increased when the sample size is less than 100 cases. The proportion is used to strengthen the constraint of prior knowledge. For the causal edge from the muscle function layer to the risk outcome layer, the initial weights are: ; in, For muscles The corresponding one-way ratio, This is the scaling adjustment parameter.

9. The method for extrapolating muscle function state based on visual pose estimation as described in claim 8, characterized in that, The causal betweenness centrality is obtained by the following formula: ; in, To start from gait anomaly nodes To the risk outcome point The sum of the total weights of all causal paths, From arrive And passing through muscle nodes The sum of the total weights of all causal paths, and These are the final causal activation values ​​for gait nodes and risk nodes, respectively.

10. A device for inferring muscle function state based on visual pose estimation, characterized in that, The muscle function state inference device based on visual pose estimation includes: A continuous video data acquisition module for analysis, wherein the continuous video data acquisition module is used to acquire continuous video data to be analyzed; An abnormal gait data acquisition module is used to annotate and score abnormal gait in the continuous video data to be analyzed, thereby acquiring abnormal gait data. A muscle group mapping rule acquisition module is used to acquire preset muscle group mapping rules. An abnormal gait variable acquisition module is used to generate abnormal gait variables based on the abnormal gait data and preset muscle group mapping rules. A muscle function scoring model acquisition module, which is used to acquire a trained muscle function scoring model; A muscle function risk generation module is used to generate muscle function risk information for the subject based on the abnormal gait variables and a trained muscle function scoring model.