Clinical and ultrasound-based methods and devices for the classification, identification, and auxiliary diagnosis of sarcopenia

Abstract

The method and device for grading and assisting diagnosis of sarcopenia based on clinical and ultrasound features can combine clinical and ultrasound features and be based on interpretable cascaded machine learning to achieve efficient diagnosis, grading and interpretable identification of sarcopenia and its early stages. The method includes: (1) standardizing the equipment, body position, positioning, scanning method, measurement definition and image acquisition of ultrasound examination; (2) constructing a dataset by combining multi-center clinical data and screening key features; (3) constructing a two-stage cascaded grading diagnosis model using the extreme gradient boosting algorithm XGBoost; and (4) using the SHAP algorithm for interpretability analysis.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method for the classification and auxiliary diagnosis of sarcopenia based on clinical and ultrasound methods, as well as a device for the classification and auxiliary diagnosis of sarcopenia based on clinical and ultrasound methods. Background Technology

[0002] Sarcopenia is a common syndrome affecting the health and quality of life of older adults. Current clinical assessments mainly rely on dual-energy X-ray absorptiometry (DXA), bioelectrical impedance analysis (BIA), and physical fitness tests (such as grip strength and walking speed). However, these methods are limited by the high cost of equipment, the complexity of operation, and the high environmental requirements, making them difficult to widely apply in primary care and bedside settings.

[0003] In contrast, ultrasound offers advantages such as being radiation-free, repeatable, portable, and low-cost. It has been recommended by numerous international guidelines and expert consensus for the quantitative assessment of muscle structure and echo intensity, providing a new technical approach for the non-invasive diagnosis of sarcopenia. In recent years, the development of artificial intelligence has driven the widespread application of machine learning in medical image analysis. By integrating multimodal features from clinical and ultrasound perspectives, it has significantly improved the accuracy and automation of disease identification. Simultaneously, interpretable model analysis methods (such as SHAP) are increasingly being applied in medical research, intuitively revealing the impact of various features on prediction results and enhancing model transparency and clinical readability.

[0004] However, existing studies are mostly limited to single-stage binary classification models, lacking the ability to classify and identify early stages of sarcopenia; the models lack interpretability and there is a lack of generalizable multicenter online diagnostic systems. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the technical problem to be solved by the present invention is to provide a method for the classification, identification and auxiliary diagnosis of sarcopenia based on clinical and ultrasound features, which can combine clinical and ultrasound features and be based on interpretable cascade machine learning to achieve efficient diagnosis, classification and interpretable identification of sarcopenia and its early stages.

[0006] The technical solution of this invention is: a method for the classification, identification, and auxiliary diagnosis of sarcopenia based on clinical and ultrasound methods, comprising the following steps: (1) Standardize the equipment, position, positioning, scanning method, measurement definition and image acquisition for ultrasound examination. The patient is in a supine position. The positioning point is the midpoint of the line connecting the right upper iliac spine and the upper edge of the patella. Obtain transverse and longitudinal images respectively. (2) A dataset was constructed by combining multi-center clinical data, including: gender, age, height, weight, body mass index, subcutaneous fat thickness, rectus femoris thickness, quadriceps femoris thickness, rectus femoris cross-sectional area, maximum, minimum, average, and range of echo intensity; key features were screened: first, minimum absolute contraction and selection operator regression were used to reduce some feature coefficients to zero through L1 regularization penalty, thereby achieving sparsity selection of variables and retaining features that significantly contribute to the predicted output; second, the Boruta feature screening algorithm was based on the random forest feature importance score, comparing the importance of the original features with the shadow features generated by random perturbation, filtering out unstable or near-random noise features, and retaining important features that are stable in multiple iterations; finally, recursive feature elimination (RFE) was based on a given learner, and iteratively deleted the features with the least contribution through training, sorting, and elimination, gradually converging to the optimal feature subset to improve the stability of the model under different data splits; (3) The extreme gradient boosting algorithm XGBoost is used to construct a two-stage cascaded grade diagnosis model: In the first stage, the model takes the key features selected in step (2) as input and establishes a binary classifier to distinguish between normal and abnormal muscle states. The output is the stage one discrimination label and the prediction probability. The discrimination label includes normal and abnormal. When it is judged to be normal, it directly outputs normal and terminates the process. When it is judged to be abnormal, the second stage model runs on the samples judged to be abnormal in the first stage. The stability of the model is evaluated by cross-validation during the model training stage, and the key hyperparameters are optimized by grid search. The optimal hyperparameter combination is selected on the validation set, and the model parameters of the first stage and the second stage are solidified. The two are encapsulated into a unified reasoning process according to the preset cascade logic. The cascade path of the first stage discrimination and the second stage grade is executed in sequence in the reasoning process, and the discrimination label and the conclusion of pre-sarcopenia and sarcopenia grade are output respectively, while providing the prediction probability of each stage. (4) Use the SHAP algorithm for interpretability analysis: At the overall level, a bee diagram is generated by the SHAP value of the sample to show the direction of influence, contribution distribution and importance ranking of each feature on the prediction results; at the individual level, the stage SHAP value is calculated for each subject and an individualized waterfall diagram is generated to show the positive or negative influence of each feature on the prediction probability or risk score according to the contribution.

[0007] This invention combines clinical and ultrasound features with interpretable cascaded machine learning to achieve efficient diagnosis, classification, and interpretable identification of sarcopenia and its early stages. The diagnosis is accurate, conforms to the concept of hierarchical medical treatment, supports online application in multi-level medical institutions, and has intelligent and clinical promotion value.

[0008] It also provides a clinical and ultrasound-based sarcopenia grading and auxiliary diagnostic device, which includes: The standardization module is configured to standardize the equipment, patient position, positioning, scanning method, measurement definition and image acquisition for ultrasound examination. The patient is in a supine position, and the positioning point is the midpoint of the line connecting the right upper iliac spine and the upper edge of the patella. Transverse and longitudinal images are acquired respectively. The feature selection module is configured to construct a dataset based on multi-center clinical data, including: gender, age, height, weight, body mass index, subcutaneous fat thickness, rectus femoris thickness, quadriceps femoris thickness, rectus femoris cross-sectional area, maximum, minimum, average, and range of echo intensity. Key features are selected through several steps: First, minimum absolute contraction and selection operator regression, through L1 regularization penalty, shrinks some feature coefficients to zero, thus achieving sparsity selection of variables and retaining features that significantly contribute to the predicted output. Second, the Boruta feature selection algorithm, based on random forest feature importance scoring, compares the importance of original features with shadow features generated through random perturbation, filtering out unstable or near-random noise features and retaining important features that remain stable over multiple iterations. Finally, recursive feature elimination (RFE), based on a given learner, iteratively removes the features with the least contribution through training, sorting, and elimination, gradually converging to the optimal feature subset to improve the model's stability under different data splits. The model module is configured to use the Extreme Gradient Boosting (XGBoost) algorithm to construct a two-stage cascaded graded diagnostic model: In the first stage, the model takes the selected key features as input and builds a binary classifier to distinguish between normal and abnormal muscle states. The output is a stage one discrimination label and a predicted probability. The discrimination label includes normal and abnormal. When the result is normal, the model outputs "normal" directly and terminates the process. When the result is abnormal, the second-stage model runs on the samples that were judged as abnormal in the first stage. During the model training stage, the stability is evaluated through cross-validation, and key hyperparameters are optimized using grid search. The optimal hyperparameter combination is selected on the validation set, and the model parameters for the first and second stages are solidified. The two are then encapsulated into a unified inference process according to the preset cascaded logic. In the inference process, the cascaded path of the first-stage discrimination and the second-stage grading is executed sequentially, outputting the discrimination label and the conclusions of pre-sarcopenia and sarcopenia grading, respectively, while providing the predicted probability for each stage. The analysis module is configured to use the SHAP algorithm for interpretability analysis: at the overall level, a bee diagram is generated based on the SHAP values ​​of the samples to show the direction of influence, contribution distribution, and importance ranking of each feature on the prediction results; at the individual level, a stage-specific SHAP value is calculated for each subject and an individualized waterfall chart is generated to show the positive or negative impact of each feature on the predicted probability or risk score based on its contribution. Attached Figure Description

[0009] Figure 1 is a feature selection diagram according to the present invention, used to intuitively present the selection logic and result basis from candidate features to final key features. (A) LASSO selection diagram, (B) Boruta selection diagram, (C) RFE selection diagram, (D) Feature overlap heatmap, (E) Venn diagram, (F) VIF analysis diagram.

[0010] Figure 2 This is a SHAP bee diagram according to the present invention. Wherein, (A) is a first-stage bee diagram, and (B) is a second-stage bee diagram.

[0011] Figure 3 This is a SHAP waterfall diagram according to the present invention. Among them, (A) analysis based on a two-step cascade model for normal population, (B) analysis based on a two-step cascade model for pre-sarcopenia, and (C) analysis based on a two-step cascade model for sarcopenia.

[0012] Figure 4 This is a schematic diagram of the input interface according to the present invention, showing the operation interface for a user to upload key clinical and ultrasound features on a website (http: / / sarcopeniapredict.icu). The system includes a data entry area and a submission module, supporting users to input clinical parameters and muscle ultrasound parameters and submit the data.

[0013] Figure 5 The cascaded two-stage prediction process and visualization results according to the present invention are illustrated. First, normal and abnormal individuals are screened; those with abnormalities automatically proceed to the second stage for sarcopenia diagnosis analysis; the analysis terminates for those with normal characteristics. The system outputs individualized prediction probabilities and a SHAP waterfall plot, demonstrating the contribution of each feature to the prediction results.

[0014] Figure 6 This is a flowchart of the method for grading, identification, and auxiliary diagnosis of sarcopenia based on clinical and ultrasound methods according to the present invention. Detailed Implementation

[0015] like Figure 6 As shown, this method for grading and assisting in the diagnosis of sarcopenia based on clinical and ultrasound methods includes the following steps: (1) Standardize the equipment, position, positioning, scanning method, measurement definition and image acquisition for ultrasound examination. The patient is in a supine position. The positioning point is the midpoint of the line connecting the right upper iliac spine and the upper edge of the patella. Obtain transverse and longitudinal images respectively. (2) A dataset was constructed by combining multi-center clinical data, including: gender, age, height, weight, body mass index, subcutaneous fat thickness, rectus femoris thickness, quadriceps femoris thickness, rectus femoris cross-sectional area, maximum, minimum, average, and range of echo intensity; key features were screened: first, minimum absolute contraction and selection operator regression were used to reduce some feature coefficients to zero through L1 regularization penalty, thereby achieving sparsity selection of variables and retaining features that significantly contribute to the predicted output; second, the Boruta feature screening algorithm was based on the random forest feature importance score, comparing the importance of the original features with the shadow features generated by random perturbation, filtering out unstable or near-random noise features, and retaining important features that are stable in multiple iterations; finally, recursive feature elimination (RFE) was based on a given learner, and iteratively deleted the features with the least contribution through training, sorting, and elimination, gradually converging to the optimal feature subset to improve the stability of the model under different data splits; (3) The extreme gradient boosting algorithm XGBoost is used to construct a two-stage cascaded grade diagnosis model: In the first stage, the model takes the key features selected in step (2) as input and establishes a binary classifier to distinguish between normal and abnormal muscle states. The output is the stage one discrimination label and the prediction probability. The discrimination label includes normal and abnormal. When it is judged to be normal, it directly outputs normal and terminates the process. When it is judged to be abnormal, the second stage model runs on the samples judged to be abnormal in the first stage. The stability of the model is evaluated by cross-validation during the model training stage, and the key hyperparameters are optimized by grid search. The optimal hyperparameter combination is selected on the validation set, and the model parameters of the first stage and the second stage are solidified. The two are encapsulated into a unified reasoning process according to the preset cascade logic. The cascade path of the first stage discrimination and the second stage grade is executed in sequence in the reasoning process, and the discrimination label and the conclusion of pre-sarcopenia and sarcopenia grade are output respectively, while providing the prediction probability of each stage. (4) Use the SHAP algorithm for interpretability analysis: At the overall level, a bee diagram is generated by the SHAP value of the sample to show the direction of influence, contribution distribution and importance ranking of each feature on the prediction results; at the individual level, the stage SHAP value is calculated for each subject and an individualized waterfall diagram is generated to show the positive or negative influence of each feature on the prediction probability or risk score according to the contribution.

[0016] This invention combines clinical and ultrasound features with interpretable cascaded machine learning to achieve efficient diagnosis, classification, and interpretable identification of sarcopenia and its early stages. The diagnosis is accurate, conforms to the concept of hierarchical medical treatment, supports online application in multi-level medical institutions, and has intelligent and clinical promotion value.

[0017] Preferably, the method further includes step (5), which integrates the trained two-stage cascaded XGBoost model and its parameters with the SHAP visualization module and deploys it to an online intelligent diagnostic platform. The platform includes a data input front-end, a model analysis back-end, and a result display interface. After the user uploads clinical features and ultrasound parameters, the first stage screening and the second stage grading are performed in sequence: if the first stage is judged to be normal, the normal conclusion is directly output and the process terminates; if it is judged to be abnormal, the second stage is entered to output the grading conclusion of pre-sarcopenia or sarcopenia; at the same time, the prediction probability of each stage is provided and an individualized SHAP waterfall chart is generated.

[0018] Preferably, in step (1), the probe applies minimum pressure perpendicular to the skin, and the cross-sectional image is used to measure key indicators such as subcutaneous fat thickness, rectus femoris thickness, quadriceps femoris thickness, rectus femoris cross-sectional area, and echo intensity. The subcutaneous fat thickness is the distance from the skin-fat interface to the fat-muscle boundary; the rectus femoris thickness is the distance from the fat-muscle interface to the boundary between the rectus femoris and vastus intermedius; the quadriceps femoris thickness is the distance from the subcutaneous fat layer to the anterior femoral cortex; the rectus femoris cross-sectional area is manually marked and calculated using electronic calipers; and the echo intensity is calculated using MATLAB image processing software to determine the maximum, minimum, and mean values ​​of the rectus femoris region to reflect the echo distribution and heterogeneity of the muscle tissue. At least three qualified images are collected from each subject, and the average value is used for analysis. All images are collected and measured by at least two physicians who have received uniform training and use a blinded grouping method, and the reliability of the measurement is verified by inter-observer consistency assessment.

[0019] Preferably, in step (2), the features selected by RFE include: age, weight, cross-sectional area of ​​rectus femoris, thickness of quadriceps femoris, average echo intensity, minimum echo intensity, thickness of rectus femoris, height, body mass index, gender, maximum echo intensity, echo intensity range, and subcutaneous fat thickness.

[0020] Preferably, in step (2), the variance inflation factor (VIF) is calculated to diagnose collinearity of key features, and variables with VIF exceeding a preset threshold are removed to finally obtain a set of key features for modeling, including: gender, age, weight, quadriceps area, quadriceps thickness, minimum echo intensity, and average echo intensity.

[0021] Preferably, in step (3), the key hyperparameters include: learning rate, tree depth, number of iterations, subsampling ratio, column sampling ratio, and minimum leaf sample weight.

[0022] Preferably, in step (4), at the overall level, in the first stage of normal and abnormal discrimination, body weight, quadriceps thickness and echo intensity contribute significantly to abnormal identification; in the second stage of sarcopenia pre- and sarcopenia grading discrimination, echo intensity, age and quadriceps thickness contribute significantly; at the individual level, the individualized waterfall plot shows the positive or negative impact of each feature on the predicted probability and risk score according to the contribution degree.

[0023] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium. When executed, the program includes the steps of the methods of the above embodiments. The storage medium can be ROM / RAM, magnetic disk, optical disk, memory card, etc. Therefore, corresponding to the method of the present invention, the present invention also includes a clinical and ultrasound-based sarcopenia grading identification and auxiliary diagnostic device. This device is typically represented in the form of functional modules corresponding to the steps of the method. The device includes: The standardization module is configured to standardize the equipment, patient position, positioning, scanning method, measurement definition and image acquisition for ultrasound examination. The patient is in a supine position, and the positioning point is the midpoint of the line connecting the right upper iliac spine and the upper edge of the patella. Transverse and longitudinal images are acquired respectively. The feature selection module is configured to construct a dataset based on multi-center clinical data, including: gender, age, height, weight, body mass index, subcutaneous fat thickness, rectus femoris thickness, quadriceps femoris thickness, rectus femoris cross-sectional area, maximum, minimum, average, and range of echo intensity. Key features are selected through several steps: First, minimum absolute contraction and selection operator regression, through L1 regularization penalty, shrinks some feature coefficients to zero, thus achieving sparsity selection of variables and retaining features that significantly contribute to the predicted output. Second, the Boruta feature selection algorithm, based on random forest feature importance scoring, compares the importance of original features with shadow features generated through random perturbation, filtering out unstable or near-random noise features and retaining important features that remain stable over multiple iterations. Finally, recursive feature elimination (RFE), based on a given learner, iteratively removes the features with the least contribution through training, sorting, and elimination, gradually converging to the optimal feature subset to improve the model's stability under different data splits. The model module is configured to use the Extreme Gradient Boosting (XGBoost) algorithm to construct a two-stage cascaded graded diagnostic model: In the first stage, the model takes the selected key features as input and builds a binary classifier to distinguish between normal and abnormal muscle states. The output is a stage one discrimination label and a predicted probability. The discrimination label includes normal and abnormal. When the result is normal, the model outputs "normal" directly and terminates the process. When the result is abnormal, the second-stage model runs on the samples that were judged as abnormal in the first stage. During the model training stage, the stability is evaluated through cross-validation, and key hyperparameters are optimized using grid search. The optimal hyperparameter combination is selected on the validation set, and the model parameters for the first and second stages are solidified. The two are then encapsulated into a unified inference process according to the preset cascaded logic. In the inference process, the cascaded path of the first-stage discrimination and the second-stage grading is executed sequentially, outputting the discrimination label and the conclusions of pre-sarcopenia and sarcopenia grading, respectively, while providing the predicted probability for each stage. The analysis module is configured to use the SHAP algorithm for interpretability analysis: at the overall level, a bee diagram is generated based on the SHAP values ​​of the samples to show the direction of influence, contribution distribution, and importance ranking of each feature on the prediction results; at the individual level, a stage-specific SHAP value is calculated for each subject and an individualized waterfall chart is generated to show the positive or negative impact of each feature on the predicted probability or risk score based on its contribution.

[0024] Preferably, the device further includes an integration module configured to integrate the trained two-stage cascaded XGBoost model and its parameters with the SHAP visualization module and deploy it to an online intelligent diagnostic platform. This platform includes a data input front-end, a model analysis back-end, and a result display interface. After the user uploads clinical features and ultrasound parameters, the system sequentially performs a first-stage screening and a second-stage grading: if the first stage is determined to be normal, a normal conclusion is directly output and the process terminates; if it is determined to be abnormal, the system proceeds to the second stage and outputs a grading conclusion for pre-sarcopenia or sarcopenia; simultaneously, it provides the predicted probability for each stage and generates an individualized SHAP waterfall chart.

[0025] Preferably, in the standardized module, the probe applies minimal pressure perpendicular to the skin. The cross-sectional image is used to measure key indicators such as subcutaneous fat thickness, rectus femoris thickness, quadriceps femoris thickness, rectus femoris cross-sectional area, and echo intensity. Subcutaneous fat thickness is the distance from the skin-fat interface to the fat-muscle boundary; rectus femoris thickness is the distance from the fat-muscle interface to the boundary between the rectus femoris and vastus intermedius; quadriceps femoris thickness is the distance from the subcutaneous fat layer to the anterior femoral cortex; the rectus femoris cross-sectional area is manually marked and calculated using electronic calipers; and the echo intensity is calculated using MATLAB image processing software to determine the maximum, minimum, and mean values ​​of the rectus femoris region to reflect the echo distribution and heterogeneity of the muscle tissue. At least three qualified images are collected from each subject, and the average value is analyzed. All images are collected and measured by at least two physicians who have received standardized training and use a blinded grouping method. The reliability of the measurements is verified through inter-observer consistency assessment.

[0026] The method of the present invention will be described in more detail below.

[0027] (1) Standardization of ultrasound assessment and measurement procedures

[0028] To ensure the comparability and repeatability of ultrasound features obtained from different centers and devices, this invention standardizes the equipment, patient position, localization, scanning method, measurement definition, and image acquisition for ultrasound examinations. This study used an ultrasound system equipped with a linear array probe. The patient was placed in a supine position with both lower limbs relaxed and toes pointing upwards. The localization point was the midpoint of the line connecting the anterior superior iliac spine of the right thigh and the upper edge of the patella. Transverse and longitudinal images were acquired. Minimum pressure was applied perpendicular to the skin using the probe to avoid tissue compression that could cause thickness and echo deviations. Cross-sectional images serve as the primary basis for quantitative measurements, used to measure key indicators such as subcutaneous fat thickness, rectus femoris thickness, quadriceps femoris thickness, rectus femoris cross-sectional area, and echo intensity. Subcutaneous fat thickness refers to the distance from the skin-fat interface to the fat-muscle boundary; rectus femoris thickness is the distance from the fat-muscle interface to the boundary between the rectus femoris and vastus intermedius; quadriceps femoris thickness is the distance from the subcutaneous fat layer to the anterior femoral cortex. The rectus femoris cross-sectional area is manually marked using electronic calipers and automatically calculated by the device's built-in software. Echo intensity is calculated using MATLAB image processing software to determine the maximum, minimum, and mean values ​​of the rectus femoris region, reflecting the echo distribution and heterogeneity of the muscle tissue. To ensure data accuracy, at least three qualified images were collected from each subject, and the average value was used for analysis. All images were collected and measured by at least two uniformly trained physicians using a blinded grouping method, and the reliability of the measurements was verified through inter-observer consistency assessment, thereby ensuring the quality and stability of the input data.

[0029] (2) Multicenter data and key feature screening

[0030] Based on the standardized ultrasound quantitative indicators acquired in step (1), this invention constructs a dataset using multi-center clinical data, including gender, age, height, weight, body mass index, subcutaneous fat thickness, rectus femoris thickness, quadriceps femoris thickness, rectus femoris cross-sectional area, maximum, minimum, and average echo intensity, and echo intensity range (maximum-minimum). To determine key predictive indicators, this invention employs multiple feature selection algorithms: First, Least Absolute Shrinkage and Selection Operator (LASSO) uses L1 regularization penalty to shrink some feature coefficients to zero, thereby achieving sparsity selection of variables and retaining features that significantly contribute to the predicted output. Second, the Boruta feature selection algorithm, based on random forest feature importance scoring, compares the importance of original features with "shadow features" generated through random perturbation, filtering out unstable or near-random noise features and retaining important features that remain stable over multiple iterations. Finally, Recursive Feature Elimination (RFE), based on a given learner, iteratively removes the least contributing features through a "training-ranking-elimination" process, gradually converging to the optimal feature subset to improve the model's stability under different data splits. Features prioritized by RFE include age, weight, rectus femoris cross-sectional area, quadriceps thickness, average echo intensity, minimum echo intensity, rectus femoris thickness, height, body mass index, gender, maximum echo intensity, echo intensity range, and subcutaneous fat thickness. These three methods complement each other from the perspectives of "sparse selection—stability testing—iterative elimination," ultimately selecting features commonly retained by all three as key prediction indicators, aiming to reduce the bias of a single method and enhance reproducibility. To avoid parameter instability and interpretative bias caused by multicollinearity, this invention further diagnoses multicollinearity of key features by calculating the Variance Inflation Factor (VIF), eliminating variables with VIF exceeding a preset threshold to ensure controllable correlation within the feature set entering the model. Through this series of screening and diagnostic processes, a set of key features for modeling is finally obtained, including gender, age, weight, quadriceps area, quadriceps thickness, minimum echo intensity, and average echo intensity. Figure 1 illustrates the feature screening process.

[0031] (3) A two-stage cascaded model was constructed using the XGBoost algorithm to achieve hierarchical diagnosis.

[0032] To achieve a clinical pathway-based discrimination that first identifies abnormalities and then performs graded diagnosis, this invention employs the Extreme Gradient Boosting (XGBoost) algorithm to construct a two-stage cascaded graded diagnosis model. XGBoost is an ensemble learning algorithm based on gradient boosting trees. By constructing multiple decision trees and weighting their outputs, it progressively fits the error and corrects the classification results, thereby effectively modeling complex nonlinear relationships and feature interactions.

[0033] In the first stage, the model uses the key features selected in step (2) as input to build a binary classifier to distinguish between "normal muscle state" and "abnormal muscle state". Its output is the stage one discrimination label (normal / abnormal) and the predicted probability; when it is judged to be normal, the system directly outputs the normal conclusion and terminates the process. The second stage model only runs on the samples judged to be abnormal in the first stage. The goal is to further distinguish between "pre-sarcopenia" and "sarcopenia" in the abnormal population, to provide more detailed stratification results for clinical practice, and to support early intervention and personalized treatment decisions.

[0034] During the model training phase, stability is evaluated through cross-validation, and key hyperparameters, including learning rate, tree depth, number of iterations, subsampling ratio, column sampling ratio, and minimum leaf sample weight, are optimized using grid search. The optimal combination of hyperparameters is selected on the validation set, and the model parameters for the first and second phases are solidified. The two are then encapsulated into a unified inference process according to a pre-defined cascade logic.

[0035] During the inference phase, the system sequentially executes the cascaded path of "first-stage discrimination—second-stage grading," outputting normal / abnormal conclusions and pre-sarcopenia / sarcopenia grading conclusions, respectively, while providing the predicted probability for each stage for clinical reference. Compared to a one-time three-class classification model, the two-stage cascaded structure of this invention can effectively reduce the risk of misjudgment caused by overlapping category boundaries. Especially in the scenario of "early identification within abnormal populations," it is more in line with the actual screening—triage—further evaluation process, improving the detection capability of pre-sarcopenia and strengthening early management support.

[0036] (4) Use the SHAP algorithm for interpretability analysis

[0037] To address the interpretability issues of traditional machine learning models, this invention introduces the Shapley Additive Interpretation (SHAP) algorithm on top of the cascaded model, providing a quantified explanation of feature contributions. Based on Shapley values ​​in game theory, SHAP decomposes a single prediction result into the marginal contribution value of each input feature to the output, clearly demonstrating how each feature drives the model to the current conclusion. At the overall level, the system generates a bee diagram using the SHAP values ​​of samples, showing the direction of influence, contribution distribution, and importance ranking of each feature on the prediction results, helping doctors understand which features play a major role and their trends at different stages. In the first stage of normal / abnormal discrimination, features such as weight, quadriceps thickness, and echogenicity contribute significantly to abnormal identification; while in the second stage of pre-sarcopenia / sarcopenia grading, features such as echogenicity, age, and quadriceps thickness contribute significantly to sarcopenia identification. At the individual level, the system calculates the stage-specific SHAP value for each subject and generates a personalized waterfall chart, showing the positive or negative impact of each feature on the predicted probability or risk score based on its contribution. Figure 2 and Figure 3 The images show the SHAP bee diagram and three typical examples of SHAP waterfall diagrams.

[0038] (5) The system is deployed on a network platform

[0039] To facilitate clinical application, this invention integrates a trained two-stage cascaded XGBoost model and its parameters with a SHAP visualization module and deploys it on an online intelligent diagnostic platform. This platform includes a data input front-end, a model analysis back-end, and a results display interface. After the user uploads clinical features and ultrasound parameters, the system automatically performs a first-stage screening and a second-stage grading process sequentially: if the first stage is determined to be normal, the system directly outputs a normal conclusion and terminates; if it is determined to be abnormal, it proceeds to the second stage and outputs a grading conclusion of "pre-sarcopenia" or "sarcopenia." The platform also provides the predicted probability for each stage and automatically generates an individualized SHAP waterfall chart.

[0040] The specific arrangements are as follows: (1) Community health service centers or tertiary medical institutions collect information on the muscle ultrasound characteristics and related clinical characteristics of the examinees; (2) The automatic identification system of this invention deployed on the website automatically analyzes and predicts the uploaded data and generates visual prediction results; (3) Different outputs of the cascade model are applied by medical institutions at all levels according to their needs to achieve the goals of early screening at the community level and precise diagnosis and treatment at the tertiary hospital level.

[0041] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for grading and assisting in the diagnosis of sarcopenia based on clinical and ultrasound methods, characterized by: It includes the following steps: (1) Standardize the equipment, position, positioning, scanning method, measurement definition and image acquisition for ultrasound examination. The patient is in a supine position. The positioning point is the midpoint of the line connecting the right upper iliac spine and the upper edge of the patella. Obtain transverse and longitudinal images respectively. (2) A dataset was constructed by combining multi-center clinical data, including: gender, age, height, weight, body mass index, subcutaneous fat thickness, rectus femoris thickness, quadriceps femoris thickness, rectus femoris cross-sectional area, maximum, minimum, average, and range of echo intensity; key features were screened: first, minimum absolute contraction and selection operator regression were used to reduce some feature coefficients to zero through L1 regularization penalty, thereby achieving sparsity selection of variables and retaining features that significantly contribute to the predicted output; second, the Boruta feature screening algorithm was based on the random forest feature importance score, comparing the importance of the original features with the shadow features generated by random perturbation, filtering out unstable or near-random noise features, and retaining important features that are stable in multiple iterations; finally, recursive feature elimination (RFE) was based on a given learner, and iteratively deleted the features with the least contribution through training, sorting, and elimination, gradually converging to the optimal feature subset to improve the stability of the model under different data splits; (3) The extreme gradient boosting algorithm XGBoost is used to construct a two-stage cascaded grade diagnosis model: In the first stage, the model takes the key features selected in step (2) as input and establishes a binary classifier to distinguish between normal and abnormal muscle states. The output is the stage one discrimination label and the prediction probability. The discrimination label includes normal and abnormal. When it is judged to be normal, it directly outputs normal and terminates the process. When it is judged to be abnormal, the second stage model runs on the samples judged to be abnormal in the first stage. The stability of the model is evaluated by cross-validation during the model training stage, and the key hyperparameters are optimized by grid search. The optimal hyperparameter combination is selected on the validation set, and the model parameters of the first stage and the second stage are solidified. The two are encapsulated into a unified reasoning process according to the preset cascade logic. The cascade path of the first stage discrimination and the second stage grade is executed in sequence in the reasoning process, and the discrimination label and the conclusion of pre-sarcopenia and sarcopenia grade are output respectively, while providing the prediction probability of each stage. (4) Use the SHAP algorithm for interpretability analysis: At the overall level, a bee diagram is generated by the SHAP value of the sample to show the direction of influence, contribution distribution and importance ranking of each feature on the prediction results; at the individual level, the stage SHAP value is calculated for each subject and an individualized waterfall diagram is generated to show the positive or negative influence of each feature on the prediction probability or risk score according to the contribution.

2. The method for grading and assisting diagnosis of sarcopenia based on clinical and ultrasound methods according to claim 1, characterized in that: The method also includes step (5), which integrates the trained two-stage cascaded XGBoost model and its parameters with the SHAP visualization module and deploys it to an online intelligent diagnostic platform. The platform includes a data input front-end, a model analysis back-end, and a result display interface. After the user uploads clinical features and ultrasound parameters, the first stage screening and the second stage grading are performed in sequence: if the first stage is judged to be normal, the normal conclusion is directly output and the process terminates; if it is judged to be abnormal, the second stage is entered to output the grading conclusion of pre-sarcopenia or sarcopenia; at the same time, the predicted probability of each stage is provided and an individualized SHAP waterfall chart is generated.

3. The method for grading and assisting diagnosis of sarcopenia based on clinical and ultrasound methods according to claim 2, characterized in that: In step (1), the probe applies minimum pressure perpendicular to the skin. The cross-sectional image is used to measure key indicators such as subcutaneous fat thickness, rectus femoris thickness, quadriceps femoris thickness, rectus femoris cross-sectional area, and echo intensity. Subcutaneous fat thickness is the distance from the skin-fat interface to the fat-muscle boundary. Rectus femoris thickness is the distance from the fat-muscle interface to the boundary between the rectus femoris and vastus intermedius. Quadriceps femoris thickness is the distance from the subcutaneous fat layer to the anterior femoral cortex. The rectus femoris cross-sectional area is manually marked and calculated using electronic calipers. The echo intensity is calculated using MATLAB image processing software to determine the maximum, minimum, and mean values ​​of the rectus femoris region to reflect the echo distribution and heterogeneity of the muscle tissue. At least three qualified images are collected from each subject, and the average value is analyzed. All images are collected and measured by at least two physicians who have received uniform training and use a blinded grouping method. The reliability of the measurements is verified by inter-observer consistency assessment.

4. The method for grading and assisting diagnosis of sarcopenia based on clinical and ultrasound methods according to claim 3, characterized in that: In step (2), the features selected by RFE include: age, weight, cross-sectional area of ​​rectus femoris, thickness of quadriceps femoris, average echo intensity, minimum echo intensity, thickness of rectus femoris, height, body mass index, gender, maximum echo intensity, echo intensity range, and subcutaneous fat thickness.

5. The method for grading and assisting diagnosis of sarcopenia based on clinical and ultrasound methods according to claim 4, characterized in that: In step (2), the variance inflation factor (VIF) is calculated to diagnose collinearity of key features, and variables with VIF exceeding a preset threshold are removed. Finally, a set of key features for modeling is obtained, including: gender, age, weight, quadriceps area, quadriceps thickness, minimum echo intensity, and average echo intensity.

6. The method for grading and assisting diagnosis of sarcopenia based on clinical and ultrasound methods according to claim 5, characterized in that: In step (3), the key hyperparameters include: learning rate, tree depth, number of iterations, subsampling ratio, column sampling ratio, and minimum leaf sample weight.

7. The method for grading and assisting diagnosis of sarcopenia based on clinical and ultrasound methods according to claim 6, characterized in that: In step (4), at the overall level, in the first stage of normal and abnormal discrimination, body weight, quadriceps thickness and echo intensity contribute significantly to abnormal identification. In the second stage of sarcopenia pre- and sarcopenia grading discrimination, echo intensity, age and quadriceps thickness contribute significantly. At the individual level, the individualized waterfall plot shows the positive or negative impact of each feature on the predicted probability and risk score based on the degree of contribution.

8. A clinical and ultrasound-based device for the classification, identification, and auxiliary diagnosis of sarcopenia, characterized in that: It includes: The standardization module is configured to standardize the equipment, patient position, positioning, scanning method, measurement definition and image acquisition for ultrasound examination. The patient is in a supine position, and the positioning point is the midpoint of the line connecting the right upper iliac spine and the upper edge of the patella. Transverse and longitudinal images are acquired respectively. The feature selection module is configured to construct a dataset based on multi-center clinical data, including: gender, age, height, weight, body mass index, subcutaneous fat thickness, rectus femoris thickness, quadriceps femoris thickness, rectus femoris cross-sectional area, maximum, minimum, average, and range of echo intensity. Key features are selected through several steps: First, minimum absolute contraction and selection operator regression, through L1 regularization penalty, shrinks some feature coefficients to zero, thus achieving sparsity selection of variables and retaining features that significantly contribute to the predicted output. Second, the Boruta feature selection algorithm, based on random forest feature importance scoring, compares the importance of original features with shadow features generated through random perturbation, filtering out unstable or near-random noise features and retaining important features that remain stable over multiple iterations. Finally, recursive feature elimination (RFE), based on a given learner, iteratively removes the features with the least contribution through training, sorting, and elimination, gradually converging to the optimal feature subset to improve the model's stability under different data splits. The model module is configured to use the Extreme Gradient Boosting (XGBoost) algorithm to construct a two-stage cascaded graded diagnostic model: In the first stage, the model takes the selected key features as input and builds a binary classifier to distinguish between normal and abnormal muscle states. The output is a stage one discrimination label and a predicted probability. The discrimination label includes normal and abnormal. When the result is normal, the model outputs "normal" directly and terminates the process. When the result is abnormal, the second-stage model runs on the samples that were judged as abnormal in the first stage. During the model training stage, the stability is evaluated through cross-validation, and key hyperparameters are optimized using grid search. The optimal hyperparameter combination is selected on the validation set, and the model parameters for the first and second stages are solidified. The two are then encapsulated into a unified inference process according to the preset cascaded logic. In the inference process, the cascaded path of the first-stage discrimination and the second-stage grading is executed sequentially, outputting the discrimination label and the conclusions of pre-sarcopenia and sarcopenia grading, respectively, while providing the predicted probability for each stage. The analysis module is configured to use the SHAP algorithm for interpretability analysis: at the overall level, a bee diagram is generated based on the SHAP values ​​of the samples to show the direction of influence, contribution distribution, and importance ranking of each feature on the prediction results; at the individual level, a stage-specific SHAP value is calculated for each subject and an individualized waterfall chart is generated to show the positive or negative impact of each feature on the predicted probability or risk score based on its contribution.

9. The sarcopenia grading and auxiliary diagnostic device based on clinical and ultrasound methods according to claim 8, characterized in that: It also includes an integration module configured to integrate the trained two-stage cascaded XGBoost model and its parameters with the SHAP visualization module and deploy it to an online intelligent diagnostic platform. This platform includes a data input front-end, a model analysis back-end, and a results display interface. After the user uploads clinical features and ultrasound parameters, the system sequentially performs the first stage screening and the second stage grading: if the first stage is judged to be normal, it directly outputs a normal conclusion and terminates; if it is judged to be abnormal, it enters the second stage and outputs a grading conclusion of pre-sarcopenia or sarcopenia; at the same time, it provides the prediction probability for each stage and generates an individualized SHAP waterfall chart.

10. The sarcopenia grading identification and auxiliary diagnostic device based on clinical and ultrasound methods according to claim 9, characterized in that: In the standardized module, the probe applies minimal pressure perpendicular to the skin. Cross-sectional images are used to measure key indicators such as subcutaneous fat thickness, rectus femoris thickness, quadriceps femoris thickness, rectus femoris cross-sectional area, and echo intensity. Subcutaneous fat thickness is the distance from the skin-fat interface to the fat-muscle boundary; rectus femoris thickness is the distance from the fat-muscle interface to the boundary between the rectus femoris and vastus intermedius; quadriceps femoris thickness is the distance from the subcutaneous fat layer to the anterior femoral cortex; the rectus femoris cross-sectional area is manually marked and calculated using electronic calipers; and the echo intensity is calculated using MATLAB image processing software to determine the maximum, minimum, and mean values ​​of the rectus femoris region to reflect the echo distribution and heterogeneity of the muscle tissue. At least three qualified images are collected from each subject, and the average value is analyzed. All images are collected and measured by at least two uniformly trained physicians using a blinded grouping method, and the reliability of the measurements is verified through inter-observer consistency assessment.

Images