Sinopenia assessment system and medical equipment

Abstract

The invention discloses a sarcopenia assessment system and medical equipment, and the system comprises a data processing module which is used for splitting and preprocessing clinical data; the feature screening module is used for determining an optimal feature subset; the model building and training module is used for building and optimizing a sarcopenia evaluation model; the model evaluation module is used for evaluating the generalization performance of the model; and the practical application module is used for evaluating a new patient. Based on a bioelectrical impedance analysis technology, four key characteristics of gender, body weight, extracellular fluid resistance and intracellular fluid resistance are automatically screened out, and a multi-model preferential strategy is adopted to construct a high-performance evaluation model. According to the method, the defects that a traditional method depends on expensive equipment, complicated processes, subjective indexes and the like are overcome, rapid, objective and accurate screening of sarcopenia is achieved, and an efficient and reliable auxiliary diagnosis tool is provided for clinic.

Description

Technical Field

[0001] This invention relates to the fields of smart healthcare and artificial intelligence-assisted diagnostic technology, and in particular to a sarcopenia assessment system and medical device. Background Technology

[0002] Sarcopenia is a syndrome characterized by progressive loss of skeletal muscle mass, decreased muscle strength, and impaired physical function. Patients undergoing maintenance hemodialysis (MHD) are at high risk for sarcopenia due to various pathophysiological mechanisms, including uremic toxin accumulation, chronic microinflammatory states, malnutrition, and metabolic acidosis, which lead to decreased protein synthesis and increased protein breakdown. Epidemiological surveys show that the prevalence of sarcopenia among MHD patients in my country remains high and is closely associated with adverse prognoses such as increased falls, hospitalization rates, and all-cause mortality. Therefore, early identification and diagnosis of sarcopenia in MHD patients is of significant clinical importance.

[0003] Currently, the gold standard for clinical diagnosis of sarcopenia typically relies on dual-energy X-ray absorptiometry (DXA) to accurately measure skeletal muscle mass. However, DXA equipment is expensive, inconvenient to operate, and poses radiation exposure risks, making it difficult to routinely implement and widely adopt in dialysis centers. Bioelectric Impedance Analysis (BIA) technology, due to its non-invasiveness, convenience, low cost, and good correlation with DXA measurements, has become a reliable alternative tool for assessing body composition and estimating skeletal muscle mass. The basic principle of BIA technology is to measure the body's impedance to a weak, high-frequency alternating current applied to the body. In human tissues, water- and electrolyte-containing tissues such as extracellular and intracellular fluids are good conductors of current, while tissues such as adipose tissue are non-conductors. Therefore, bioelectrical impedance values ​​can accurately reflect the body's water distribution. Specifically, extracellular fluid resistance (Re) and intracellular fluid resistance (Ri) are the most basic and fundamental electrophysiological parameters measured by BIA, directly corresponding to the extracellular and intracellular water content, respectively. Meanwhile, cell membranes exhibit capacitive properties electrically, and their integrity and health status affect the current conduction path and impedance value. Therefore, the values ​​of Re and Ri, and their interrelationship, are essentially a comprehensive reflection of the body's water distribution and the health status of cell membranes, which are closely related to the composition and physiological state of muscle tissue.

[0004] Although international guidelines have integrated indicators such as grip strength and physical fitness tests to improve diagnosis, the process remains relatively cumbersome, relying on multiple discrete thresholds and proving inefficient for large-scale clinical screening. In recent years, while some studies have attempted to construct simplified assessment models, such as using traditional statistical regression methods to screen for indicators like age, sex, weight, and grip strength, these still rely on grip strength testing—a subjective functional indicator requiring patient cooperation and limited by specialized equipment—failing to fully utilize the more original and objective electrophysiological parameters provided by BIA (bioassay induction) devices. Furthermore, traditional feature screening methods (such as univariate or multivariate regression) have limitations in handling multivariate and nonlinear relationships, and the types of models constructed are limited, with generalization ability and accuracy requiring further improvement.

[0005] Therefore, there is an urgent need in the field for an intelligent assessment solution that can deeply integrate clinical workflows, achieve an automated closed loop from data acquisition to diagnostic result output, and adopt optimized data processing strategies for MHD patients, in order to overcome the above-mentioned deficiencies of existing technologies. Summary of the Invention

[0006] Therefore, the technical problem to be solved by the present invention is to overcome the problems of sarcopenia assessment in the prior art, which relies on expensive equipment, has a complicated process, depends on subjective functional indicators, has limited feature screening methods, and has insufficient generalization ability due to a single model.

[0007] To address the aforementioned technical problems, this invention provides a sarcopenia assessment system, characterized by comprising the following modules: The data processing module is used to split the collected clinical data of multiple groups of maintenance hemodialysis patients into training set and test set, and to perform data preprocessing on the training set; The feature selection module is used to select features from the preprocessed training data and determine the optimal feature subset from multiple features. The model building and training module is used to select a model architecture from multiple machine learning algorithms based on the optimal feature subset to build a sarcopenia assessment model, and to train and optimize the sarcopenia assessment model to obtain an optimized sarcopenia assessment model. The model evaluation module is used to input the test set into the optimized sarcopenia assessment model and evaluate the generalization performance of the model. The practical application module is used to acquire feature data corresponding to the optimal feature subset of the patient to be evaluated during the application phase, and to obtain the evaluation results using the sarcopenia assessment model.

[0008] In one embodiment of the present invention, the model building and training module, when selecting a model architecture from multiple machine learning algorithms to build a sarcopenia assessment model, includes: Select and instantiate machine learning models from a pre-defined model library, which includes at least four classifiers: support vector machine, logistic regression, random forest, and extreme gradient boosting tree. Configure corresponding initialization parameters for different model types. For example, configure the kernel function type and penalty coefficient for support vector machine, configure the regularization strength and optimization algorithm for logistic regression, and configure the number and depth of trees for random forest and extreme gradient boosting tree.

[0009] In one embodiment of the present invention, the model building and training module, when training and hyperparameter optimizing the sarcopenia assessment model to obtain an optimized sarcopenia assessment model, includes: Extract feature columns corresponding to the optimal feature subset from the preprocessed training set to construct a simplified training set; The five-fold cross-validation method is adopted, and the simplified training set is randomly divided into five non-overlapping subsets. Four of these subsets are used as training data and the remaining subset is used as validation data for iterative training and validation. Based on the cross-validation process, different parameter combinations are systematically traversed in the preset hyperparameter space using grid search technology, and the hyperparameter configuration that optimizes model performance is found using the average validation set performance as an indicator. Using the optimal hyperparameter configuration, the model is retrained on the complete simplified training set to obtain the optimized sarcopenia assessment model.

[0010] In one embodiment of the present invention, when the data processing module splits the collected clinical data of multiple groups of maintenance hemodialysis patients into a training set and a test set, the following steps are included: dividing the clinical data into different levels, including at least two levels: sarcopenic patients and non-sarcopenic patients; within each level, using a random sampling method, sampling samples from each level according to a preset sampling ratio; merging the samples extracted from each level to form a training set, and merging the remaining samples from each level to form a test set.

[0011] In one embodiment of the present invention, the data processing module, when performing data preprocessing on the training set, includes: detecting missing feature values ​​in the training set, using Bayesian ridge regression as an estimator, performing iterative regression imputation on multiple variables through chain equations to fill in the missing feature values; and standardizing or normalizing the filled continuous numerical features.

[0012] In one embodiment of the present invention, when the feature selection module performs feature selection on the preprocessed training data and determines the optimal feature subset from multiple features, it includes: automatically finding the optimal regularization penalty coefficient through 5-fold cross-validation. ; and in the optimal regularization penalty coefficient Under the given conditions, the feature coefficients with low contribution are compressed to zero, and features with non-zero regression coefficients are automatically selected to form the optimal feature subset.

[0013] In one embodiment of the present invention, the optimal feature subset includes sex, weight, extracellular fluid resistance, and intracellular fluid resistance.

[0014] In one embodiment of the present invention, the model evaluation module, when evaluating the generalization performance of the model, includes: The test set is input into the optimized sarcopenia assessment model to obtain the model's prediction results for each test sample; The prediction results are compared with the corresponding real labels in the test set, and multiple performance indicators, including accuracy, sensitivity, specificity, area under the receiver operating characteristic curve, and F1 score, are calculated. Multiple sample subsets are generated by random sampling with replacement from the test set, and the performance metrics are recalculated on each sample subset. Based on the performance index distribution obtained from multiple resampling, the confidence interval of each performance index is calculated to quantify the statistical reliability of the evaluation results.

[0015] In one embodiment of the present invention, the sarcopenia assessment system further includes a model storage module for serializing the optimized sarcopenia assessment model and its parameters into a persistently storable model file; and storing the model file in a non-volatile storage medium of a server or terminal device.

[0016] The present invention also provides a sarcopenia assessment medical device, the sarcopenia assessment medical device including the sarcopenia assessment system described above.

[0017] The technical solution of the present invention has the following advantages compared with the prior art: This invention provides a highly efficient and objective sarcopenia assessment system and medical device. The system processes clinical data through an automated workflow, automatically screening key objective indicators such as extracellular fluid resistance and intracellular fluid resistance, without relying on traditional subjective functional tests. Employing a multi-model selection and intelligent optimization strategy, it ensures that the final assessment model possesses both high accuracy and strong generalization ability. Its complete closed-loop design allows the system to be directly integrated into medical devices, enabling plug-and-play rapid screening and providing clinicians with a low-cost, efficient, and reliable sarcopenia assessment solution. Attached Figure Description

[0018] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0019] Figure 1 This is a schematic diagram of the sarcopenia assessment system in one embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the process of constructing and training a sarcopenia assessment model in one embodiment of the present invention; Figure 3 This is a schematic diagram of the data annotation process in one embodiment of the present invention; Figure 4 This is a schematic diagram of the process of obtaining the optimal feature subset through feature filtering in one embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the path of how the feature coefficients change with the regularization penalty coefficient during Lasso regression in one embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the principle of selecting a support vector machine as a machine learning classifier in one embodiment of the present invention; Figure 7 This is a schematic diagram illustrating the principle of selecting a logistic regression model as a machine learning classifier in one embodiment of the present invention; Figure 8 These are the ROC curves of different model test sets in Experiment 1 of this invention.

[0020] Explanation of the reference numerals in the instruction manual: 100, Sarcopenia Assessment System; 110, Data Processing Module; 120, Feature Selection Module; 130, Model Building and Training Module; 140, Model Evaluation Module; 150, Practical Application Module. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0022] Example 1: like Figure 1 As shown, the present invention provides a sarcopenia assessment system 100, comprising the following modules: The data processing module 110 is used to split the collected clinical data of multiple groups of maintenance hemodialysis patients into training set and test set, and to perform data preprocessing on the training set; The feature selection module 120 is used to perform feature selection on the preprocessed training data and determine the optimal feature subset from multiple features. The model building and training module 130 is used to select a model architecture from multiple machine learning algorithms based on the optimal feature subset to build a sarcopenia assessment model, and to train and optimize the sarcopenia assessment model to obtain an optimized sarcopenia assessment model. The model evaluation module 140 is used to input the test set into the optimized sarcopenia assessment model and evaluate the generalization performance of the model. The practical application module 150 is used to acquire feature data corresponding to the optimal feature subset of the patient to be evaluated during the application phase, and to obtain the evaluation result using the sarcopenia assessment model.

[0023] This invention provides a sarcopenia assessment system 100. The system first performs standardized cleaning and segmentation of clinical data of maintenance hemodialysis patients through a data processing module 110. Then, it uses a feature screening module 120 to automatically identify the most discriminative feature combination. Subsequently, it constructs an assessment model from a variety of machine learning algorithms through a model building and training module 130. After rigorous performance verification, the generalization performance of the model is evaluated in a model evaluation module 140. Finally, it realizes rapid assessment of new patients in a practical application module 150.

[0024] This system design specifically addresses several prominent issues in the background technology. Through automated feature screening and model optimization processes, it overcomes the limitations of traditional statistical methods in handling multivariate nonlinear relationships. By replacing subjective functional indicators that require active patient cooperation with objective physiological parameters, it significantly simplifies the assessment process. The multi-model selection mechanism ensures that the final sarcopenia assessment model has both high accuracy and strong generalization ability. It provides clinicians with a sarcopenia screening solution that is not dependent on expensive equipment, is easy to operate, and has reliable results, and has significant clinical application value.

[0025] like Figure 2 As shown, the system described in this embodiment includes a complete technical process from acquiring raw clinical data to generating a final deployable evaluation model file.

[0026] Specifically, complete clinical data of 185 MHD patients admitted to the Department of Nephrology of Wuxi People's Hospital were collected through the Hospital Information System (HIS) and the electronic medical record system of the dialysis center to construct the original dataset for model training. This original dataset comprehensively covers the following four categories of indicators: 1. Basic demographic and routine measurement information, including gender, age, height, and weight; 2. BIA parameters, specifically primitive electrophysiological indicators such as extracellular fluid resistance (Re), intracellular fluid resistance (Ri), and cell membrane capacitance (Cm); 3. Laboratory test indicators, including albumin (Alb), hemoglobin (Hb), serum calcium (Ca), serum phosphorus (P), and parathyroid hormone (PTH); 4. Core diagnostic indicators of sarcopenia, including skeletal muscle mass (ASM), skeletal muscle mass index (SMI), and grip strength.

[0027] like Figure 3As shown, the data annotation strictly followed the 2024 edition of the "Chinese Guidelines for the Diagnosis and Treatment of Sarcopenia": First, the patient's grip strength was measured; men <28kg or women <18kg were defined as having low muscle strength; when SMI was measured by DXA method, men <7.0kg / m 2 Or female <5.4kg / m 2 Determined to have low muscle mass, when SMI is measured by the BIA method, males <7.0 kg / m² 2 Or female <5.7kg / m 2 Patients who meet both the criteria for low muscle strength and low muscle mass are diagnosed with sarcopenia and assigned a binary classification label of "1". Non-sarcopenic patients are labeled with "0", thus providing a gold standard label for supervised learning.

[0028] Furthermore, when the data processing module 110 splits the collected clinical data of multiple groups of maintenance hemodialysis patients into training and test sets, it includes: dividing the clinical data into different levels, including at least two levels: sarcopenic patients and non-sarcopenic patients; within each level, using a random sampling method, sampling samples from each level according to a preset sampling ratio; merging the samples extracted from each level to form a training set, and merging the remaining samples from each level to form a test set.

[0029] Specifically, the data processing module 110 performs data splitting operations on the collected clinical data. To avoid model training bias caused by the imbalance between sarcopenic and non-sarcopenic patients in the original data, this embodiment adopts a stratified random sampling strategy. The stratified random sampling strategy adopted in this embodiment is a specific optimization scheme proposed by this invention for the sarcopenic assessment scenario of MHD patients. Its necessity lies in the fact that the ratio of sarcopenic to non-sarcopenic samples in the MHD patient population may be unbalanced, and the physiological data distribution of the two types of patients differs significantly. Using simple random sampling is very likely to lead to model training bias.

[0030] To address this, the strategy first divides all clinical data into two independent tiers based on sarcopenia diagnosis results: a sarcopenia patient tier and a non-sarcopenia patient tier. This stratification based on sarcopenia diagnostic labels is a key technical prerequisite for building an unbiased and highly generalizable diagnostic model. Its purpose is to ensure that the disease spectrum distribution in the training and test sets is highly consistent with the distribution of real-world clinical data, thereby effectively avoiding the technical problem of overfitting to the majority class due to sample imbalance and significantly improving the model's reliability in real-world clinical scenarios.

[0031] Optionally, within each level, the same preset sampling ratio (preferably 70%) is used for independent sampling using computer-generated random numbers to ensure the randomness of the sampling process. Subsequently, the samples drawn from the two levels are merged to form the training set, while the remaining samples from each level are merged to form an independent test set. This method ensures that the distribution ratio of sarcopenic patients in the training and test sets is basically consistent with the original dataset, thereby effectively preventing prediction bias caused by class imbalance and providing representative data partitioning for subsequent model training.

[0032] Furthermore, after data splitting, the training set undergoes refined data preprocessing. First, missing value handling is performed: an automatic detection program is initiated to identify missing feature values ​​in the training set. For any missing data, a chain equation multiple imputation method based on a Bayesian Ridge regression estimator is used for imputation. This method treats the dataset containing missing values ​​as a system composed of multiple variables. Through iterative loops, one feature variable is used as the output each time, and the other feature variables are used as inputs to construct a Bayesian Ridge regression model to predict and impute missing values. This process is repeated until all missing values ​​are imputed and the results converge. Compared to traditional mean imputation or deletion methods, this iterative regression imputation method can more effectively preserve the complex correlations and uncertainties between variables, thus providing a more accurate and reliable data foundation for subsequent feature selection.

[0033] After missing value imputation, all continuous numerical features are standardized or normalized. By scaling continuous features with different dimensions and numerical ranges, they are transformed into a unified numerical interval, effectively eliminating the negative impact of differences in dimensions between features on model training, accelerating model convergence, and improving the performance and stability of subsequent machine learning algorithms.

[0034] Through a systematic and meticulous data preprocessing process, not only was the quality and consistency of the training data significantly improved, but more importantly, a solid and reliable data foundation was laid for subsequent feature selection and model building.

[0035] Furthermore, when the feature selection module 120 selects features from multiple features in the preprocessed training data, it includes: automatically finding the optimal regularization penalty coefficient through 5-fold cross-validation. ; and in the optimal regularization penalty coefficient Under the given conditions, the feature coefficients with low contribution are compressed to zero, and features with non-zero regression coefficients are automatically selected to form the optimal feature subset.

[0036] Specifically, during feature selection, after receiving the preprocessed training data, the core issue faced by the feature selection module 120 is how to identify the feature combination with the highest discriminative power and lowest redundancy from multiple features, including demographic information, BIA parameters, and laboratory indicators.

[0037] Traditional feature selection methods have limitations in handling multivariate collinearity and complex nonlinear relationships. Furthermore, excessively high dimensionality of the original feature set not only increases computational complexity but also makes the model more prone to overfitting, severely impacting its generalization ability in clinical practice.

[0038] In this embodiment, the feature selection process is based on an automated feature selection method based on Lasso regression (Least Absolute Shrinkage and Selection Operator with Cross-Validation, LassoCV), which effectively solves the technical challenge of selecting key indicators from high-dimensional clinical features.

[0039] Lasso regression is an algorithm that adds an L1 regularization penalty term to the standard linear regression loss function. It can precisely compress the coefficients of features with low contribution to zero, thereby achieving automatic feature selection.

[0040] like Figure 4 As shown, Lasso regression is used as the core algorithm for feature selection. The objective function of this algorithm consists of the loss function of standard linear regression and the L1 regularization penalty term, and its objective function is: , Where m is the number of samples and p is the number of features. It is the true label of the i-th sample. It is the j-th feature value of the i-th sample. It is the regression coefficient of the j-th feature. This is the regularization penalty coefficient, which controls the severity of the penalty on model complexity. L1 penalty term. The coefficients of features with low contribution can be used to... It is compressed directly to 0, thereby achieving automatic feature selection.

[0041] Therefore, the penalty coefficient The selection of the optimal feature selection method is crucial to the effectiveness of feature selection. This invention employs 5-fold cross-validation to automatically find the optimal method. The specific implementation process includes: randomly and uniformly dividing the preprocessed training dataset into 5 disjoint subsets; initiating an iterative optimization process, traversing the pre-defined candidate subsets. Value sequence; in each Under the given conditions, perform 5 rounds of training and validation. In each round, use 4 subsets as the training set to build a Lasso regression model, and validate the model performance on the remaining subset, recording evaluation metrics (such as mean squared error). After completing all iterations, calculate each... The value corresponds to the average performance of the 5 rounds of validation results; select the value that optimizes the average performance. The value serves as the optimal regularization penalty coefficient. ).

[0042] like Figure 5 As shown, using the optimal The Lasso model is retrained on the complete training set, and features with non-zero regression coefficients are automatically selected as the optimal feature subset. In a preferred embodiment of the present invention, the optimal feature subset successfully selected through this technical approach includes four key indicators: sex, weight, extracellular fluid resistance (Re), and intracellular fluid resistance (Ri).

[0043] Those skilled in the art will understand that the optimal feature subset {gender, weight, extracellular fluid resistance (Re), intracellular fluid resistance (Ri)} selected by this invention constitutes a preferred embodiment of the invention, but should not be considered as a limitation of the invention. The core of this invention lies in identifying a simplified combination of objective indicators representing the patient's basic physical parameters and objective indicators representing the core bioelectrical impedance state through a systematic data processing and screening process.

[0044] Therefore, any equivalent substitution or derivative transformation of the above features should fall within the protection scope of this invention. For example, features representing basic physical parameters, besides weight, can also be derived features such as the Body Mass Index (BMI) calculated from the combination of height and weight; features representing core bioelectrical impedance parameters, besides directly using extracellular fluid resistance (Re) and intracellular fluid resistance (Ri), can also be their derived indices, such as their ratio, original multi-frequency bioelectrical impedance measurements, extracellular water (ECW) and intracellular water (ICW) calculated based on them. These derived features can also reflect the essential information of the body's water distribution and cell membrane health status.

[0045] Furthermore, the model building and training module 130, when selecting a model architecture from multiple machine learning algorithms to build a sarcopenia assessment model, includes: Select and instantiate machine learning models from a pre-defined model library, which includes at least four classifiers: support vector machine, logistic regression, random forest, and extreme gradient boosting tree. Configure corresponding initialization parameters for different model types. For example, configure the kernel function type and penalty coefficient for support vector machine, configure the regularization strength and optimization algorithm for logistic regression, and configure the number and depth of trees for random forest and extreme gradient boosting tree.

[0046] Specifically, the model building and training process in this embodiment embodies systematic and diversified technical characteristics. After obtaining the optimal feature subset determined by the feature selection module 120, the model building and training module 130 initiates a multi-model collaborative optimization building strategy. This module pre-defines a model library containing various machine learning algorithms, including at least four classifiers with different mathematical foundations and performance characteristics: Support Vector Machine (SVM), Logistic Regression, Random Forest, and Extreme Gradient Boosting Tree (EXTP-BR). The system configures corresponding initialization parameters for different algorithm types: for Support Vector Machine (SVM), the focus is on configuring the kernel function type (e.g., linear kernel, polynomial kernel, or radial basis kernel) and the penalty coefficient C; for Logistic Regression, the focus is on configuring the regularization strength (e.g., L1 or L2 regularization) and the optimization algorithm (e.g., quasi-Newton method or stochastic gradient descent); for the two ensemble learning algorithms, Random Forest and Extreme Gradient Boosting Tree, the focus is on configuring key parameters such as the number of trees, maximum depth, and splitting criteria.

[0047] Optionally, when SVM is chosen as the machine learning classifier, SVM is a machine learning classification algorithm based on statistical learning theory. The core idea of ​​SVM is to find an optimal hyperplane in a high-dimensional feature space that can separate sample points of different classes most effectively. This optimal hyperplane can maximize the margin between the nearest sample points (i.e., support vectors) of the two classes, thus achieving excellent generalization ability, especially suitable for classification problems with small samples, nonlinearity, and high dimensionality.

[0048] Given a training sample set D with class labels: , Where m is the total number of patients in the training sample set. This is the n-dimensional feature vector of the i-th patient (26 terms in a preferred embodiment of the invention). The corresponding category label is set, with +1 indicating that the patient was diagnosed with sarcopenia and -1 indicating that the patient was not diagnosed with sarcopenia.

[0049] In the feature space, the partitioning hyperplane can be described by the following linear equation: , in, is the normal vector, which determines the direction of the hyperplane; b is the displacement term, which determines the distance between the hyperplane and the origin.

[0050] The basic form of SVM is solving a convex quadratic programming problem, that is, minimizing the objective function under constraints: , Its constraints are .

[0051] To solve the nonlinear separable problem, SVM introduces a kernel function. This implicitly maps the sample points in the original feature space to a higher-dimensional feature space, making the originally linearly inseparable problem linearly separable in this higher-dimensional space. Simultaneously, to increase the model's tolerance to noise, slack variables are introduced. ≥ 0. The optimization objective function is transformed into: , in The penalty parameter is used to balance model complexity and classification error rate. Through the Lagrange multiplier method and dual transformation, the final classification decision function is: , in They are Lagrange multipliers, and only support vectors correspond to them. It won't be zero. It is a symbolic function.

[0052] Preferably, such as Figure 6 As shown, the input feature vector in this embodiment It is a four-dimensional vector, that is .

[0053] Kernel function The preferred approach is a radial basis function, whose expression is: , in It is a free parameter, generally defined as a new parameter. The kernel function parameters are used to control the range of influence of the samples. The above formula can be rewritten as: .

[0054] This preferred example of the SVM diagnostic model is... , The optimal performance is achieved through the following calculation process: Input the four optimal features that have been filtered into the input layer. , representing gender, weight, Re, and Ri, respectively; The input vector is processed at the kernel layer. Compared with all support vectors in the training set Perform kernel function operations In this example, a Gaussian radial basis kernel is selected; In the weighted summation layer, the output of the kernel function layer is multiplied by its corresponding weight. Then accumulate them; In the output layer, the accumulated result is added with a bias term b, and then passed through a sign function. The final output is a classification result of 1 (diagnosed as sarcopenia) or 0 (diagnosed as non-sarcopenia).

[0055] Optionally, the logistic regression model is a classic linear model widely used for binary classification problems. When the logistic regression model is chosen as the machine learning classifier, the model maps the output value of the linear regression to the (0, 1) interval through the Sigmoid function, thereby obtaining the probability that the sample belongs to a certain category.

[0056] like Figure 7 As shown, the logistic regression model first performs a linear combination of the input features: , Then, the result of the linear combination Input into the Sigmoid function: , in, This is the model's predicted output. For the Sigmoid function, The output value is the probability that the predicted sample class is 1 (e.g., sarcopenia). The goal of model training is to find the optimal parameters. The values ​​of b and b minimize the loss function (usually log-likelihood loss or cross-entropy loss).

[0057] The calculation process of the logistic regression diagnostic model is as follows: The same four optimal features as those in the SVM model are input into the input layer. ; Input features are incorporated into the linear combination layer. Weights corresponding to the model Multiply, then sum, and add the bias term b to obtain a linear output. ; In the Sigmoid activation layer The value is non-linearly transformed by the Sigmoid function, mapping it to the interval (0, 1); The value output by the output layer represents the probability that the patient being evaluated has sarcopenia. In clinical applications, a threshold (such as 0.5) can be set. When the probability is greater than the threshold, sarcopenia is diagnosed.

[0058] Furthermore, when the model building and training module 130 trains and optimizes the hyperparameters of the sarcopenia assessment model to obtain the optimized sarcopenia assessment model, it includes: Extract feature columns corresponding to the optimal feature subset from the preprocessed training set to construct a simplified training set; The five-fold cross-validation method is adopted, and the simplified training set is randomly divided into five non-overlapping subsets. Four of these subsets are used as training data and the remaining subset is used as validation data for iterative training and validation. Based on the cross-validation process, different parameter combinations are systematically traversed in the preset hyperparameter space using grid search technology, and the hyperparameter configuration that optimizes model performance is found using the average validation set performance as an indicator. Using the optimal hyperparameter configuration, the model is retrained on the complete simplified training set to obtain the optimized sarcopenia assessment model.

[0059] Specifically, the model training and hyperparameter optimization processes employed a systematic and rigorous technical approach to ensure the acquisition of a sarcopenia assessment model with optimal performance. After selecting the model architecture, the model building and training module 130 first precisely extracts feature columns corresponding to the optimal feature subset from the preprocessed complete training set, constructing a streamlined training set containing only key features. This step effectively eliminates the interference of redundant features, significantly reducing model complexity and computational overhead, laying the foundation for subsequent efficient training.

[0060] The system employs a 5-fold cross-validation method for model training and validation: the streamlined training set is randomly and uniformly divided into five non-overlapping subsets (i.e., 5 folds). During iterative training, four subsets are used sequentially as training data for model parameter learning, while the remaining subset is used as validation data to evaluate model performance. This process is repeated five times until each subset has been used as a validation set at least once. The core advantage of this cross-validation mechanism is its ability to fully utilize limited data resources. By training and validating the model multiple times, a more robust and reliable estimate of its generalization ability is obtained, effectively avoiding evaluation bias caused by the randomness of a single data partitioning.

[0061] Based on the aforementioned cross-validation framework, the system further implements hyperparameter optimization using grid search technology: within a pre-defined hyperparameter space (such as the penalty coefficient of a support vector machine). With kernel function parameters The algorithm systematically iterates through all possible parameter combinations (such as the regularization strength of logistic regression). For each hyperparameter configuration, its average validation set performance (e.g., average AUC) is calculated using 5-fold cross-validation, and this metric is used as the evaluation standard. Finally, the hyperparameter configuration that achieves the best average performance is selected as the optimal parameter setting for the model. Although this exhaustive search strategy is computationally intensive, it ensures that no potential optimal solution is missed, thereby maximizing the potential of the model.

[0062] After hyperparameter optimization, the system uses the determined optimal hyperparameter configuration to retrain the model on the complete, simplified training set, resulting in the final optimized sarcopenia assessment model. Compared to a model trained only on a subset of data, retraining with all training data enables the model to learn a more comprehensive and stable feature-label mapping relationship, further improving its generalization performance and robustness in clinical applications.

[0063] This systematic training and optimization scheme has yielded significant technical results. By combining cross-validation and grid search, it has achieved objectivity and optimization of hyperparameter selection, avoiding the subjectivity and limitations of manual parameter tuning. This process maximizes the use of existing data, effectively improving the stability and reliability of the model in small-sample clinical data environments. Most importantly, the sarcopenia assessment model optimized through this rigorous process not only demonstrates high accuracy, high AUC (area under the ROC curve) value, and high sensitivity on the test set, but also possesses strong generalization ability, enabling it to accurately identify unknown patient data and provide clinicians with a truly reliable and trustworthy auxiliary diagnostic tool.

[0064] Furthermore, when evaluating the generalization performance of the model, the model evaluation module 140 includes: The test set is input into the optimized sarcopenia assessment model to obtain the model's prediction results for each test sample; The prediction results are compared with the corresponding real labels in the test set, and multiple performance indicators, including accuracy, sensitivity, specificity, area under the receiver operating characteristic curve, and F1 score, are calculated. Multiple sample subsets are generated by random sampling with replacement from the test set, and the performance metrics are recalculated on each sample subset. Based on the performance index distribution obtained from multiple resampling, the confidence interval of each performance index is calculated to quantify the statistical reliability of the evaluation results.

[0065] Specifically, the model evaluation module 140 executes a rigorous, multi-level generalization performance evaluation process to ensure the reliability of the final sarcopenia assessment model in real clinical settings. This module first inputs completely independent test sets into the optimized sarcopenia assessment model to obtain the model's prediction results (including classification labels or probability values) for each test sample. Subsequently, the system precisely compares these prediction results with the real labels in the test set based on the gold standard, and calculates a comprehensive set of performance metrics: accuracy measures the overall proportion of correct classifications by the model; sensitivity assesses the model's ability to identify real sarcopenia patients, which is crucial for screening applications; specificity measures the model's ability to exclude non-sarcopenia patients; the subject AUC value comprehensively reflects the model's classification and discrimination ability at different diagnostic thresholds; and the F1 score accurately balances the performance of sensitivity and accuracy.

[0066] To overcome the potential for randomness in single performance metrics and provide evidence of statistical reliability, the model evaluation module 140 further employs an advanced statistical analysis method based on Bootstrap resampling. This method generates a large subset of samples comparable in size to the original test set by performing multiple random samplings with replacement (e.g., 1000 times) from the original test set. On each Bootstrap sample subset, the system recalculates all the aforementioned performance metrics to obtain the empirical distribution for each metric. Based on these distributions, the module uses the percentile method to calculate the 95% confidence interval for each performance metric.

[0067] This systematic evaluation scheme yielded significant technical advantages and beneficial effects: First, by using completely independent test sets for evaluation, it provided an unbiased estimate of the model's generalization ability on unseen data, effectively avoiding optimism bias in model evaluation. Second, the multi-indicator comprehensive evaluation system comprehensively characterized the model's diagnostic performance from different clinical perspectives. High sensitivity ensured an extremely low false negative rate, while high specificity effectively controlled the risk of misdiagnosis, and the AUC value demonstrated the model's excellent comprehensive discriminative ability. Most importantly, by calculating confidence intervals through Bootstrap resampling, it provided, for the first time in the field, a statistical reliability quantification of model performance, enabling clinical users to clearly understand the fluctuation range of performance indicators. This greatly enhanced the credibility of the evaluation results and their clinical decision-making reference value, providing a solid scientific basis for the final deployment and application of the model.

[0068] Furthermore, the practical application module 150 realizes the complete process of deploying the trained sarcopenia assessment model into clinical use. The system serializes the best-performing model and its parameters after training, optimization, and validation, and solidifies them into a model file (preferably in .pkl format) that can be loaded and called by a computer. This file is then persistently stored in a non-volatile storage medium on a server or terminal device, thus completing the deployment preparation of the model.

[0069] This module initiates the application process when a new MHD patient needs to be assessed for sarcopenia. During the data acquisition phase, specific values ​​for four key characteristics corresponding to the optimal feature subset for the patient to be assessed are obtained through routine clinical procedures and BIA device measurements: gender, weight, Re, and Ri.

[0070] Furthermore, the system executes model invocation and calculation. The computer program automatically loads the pre-defined model file into memory, uses the collected four feature values ​​as input vectors, and transmits them to the loaded sarcopenia assessment model. The model performs forward calculations based on its internally defined algorithm logic and optimal parameters.

[0071] In the results output and presentation phase, the model generates an assessment of the patient. The results can be presented in two forms: first, a direct classification result clearly showing the conclusion of "diagnosed with sarcopenia" or "not diagnosed with sarcopenia"; second, a more refined probability result, displaying a specific value between 0 and 1, representing the predicted probability that the patient has sarcopenia. This assessment result is ultimately presented intuitively to medical staff on a graphical user interface.

[0072] The practical application module 150 constructs an end-to-end automated evaluation process from feature acquisition and model calculation to result output. Its core benefits are: it transforms complex machine learning models into convenient clinical tools. By acquiring only four readily available key features, it can generate objective and quantitative evaluation results instantly, which greatly improves screening efficiency and provides immediate and reliable auxiliary evidence for medical staff's clinical decision-making. Ultimately, it realizes the rapid, routine and standardized early screening of sarcopenia.

[0073] Experiment 1: To objectively evaluate the overall performance of the sarcopenia assessment model constructed in this invention, this experiment established four machine learning models based on the optimal feature subset (gender, weight, Re, Ri) selected by LassoCV: SVM, logistic regression, random forest, and extreme gradient boosting tree (XGBoost). The performance of the models was comprehensively evaluated through a combination of internal cross-validation and independent test set validation.

[0074] The model training employed a 5-fold cross-validation method, recording the mean area under the curve (Mean AUC) and its standard deviation (Std AUC) during the cross-validation process to assess the robustness of the model training. Subsequently, the model was finally evaluated on a completely independent test set, calculating various performance metrics including accuracy, sensitivity, specificity, test set AUC (AUC(Test)), and F1 score. To further quantify the statistical reliability of the evaluation results, the 95% confidence interval (95% CI) of the test set AUC was calculated using the Bootstrap repeated sampling method. The comprehensive performance evaluation results of each model are shown in Table 1.

[0075] Table 1:

[0076] like Figure 8 As shown, the AUC values ​​of all models on the test set were higher than 0.91, indicating that the models built based on the four key features have a strong ability to distinguish whether patients have sarcopenia. Among them, the SVM and logistic regression models showed the best overall performance: both achieved a classification accuracy of 87.5%, with AUC values ​​as high as 0.947 and 0.949, respectively, and simultaneously possessed high sensitivity (91.7%) and high specificity (86.4%). This result proves that these two models achieved a good balance between accurately identifying sarcopenia patients (low false negative rate) and correctly excluding non-sarcopenia patients (low false positive rate). The XGBoost model also showed excellent performance, with an AUC value of 0.941. The random forest model had an AUC of 0.91, demonstrating effective discriminative ability, but its sensitivity was relatively low (66.7%), indicating a certain deficiency in identifying all positive samples.

[0077] The experimental results demonstrate that the machine learning models constructed based on the four key features (gender, weight, Re, Ri) selected in this invention, especially the SVM and logistic regression models, can achieve high-performance sarcopenia assessment. Their high accuracy, high AUC, and high sensitivity on the independent test set validate the effectiveness of the feature subset and the model's strong generalization ability. This provides a solid experimental basis for subsequently solidifying the optimized model into a practically usable assessment tool and integrating it into medical devices for clinical auxiliary diagnosis.

[0078] Example 2: The present invention also provides a sarcopenia assessment medical device, which includes the sarcopenia assessment system in Embodiment 1.

[0079] The sarcopenia assessment medical device provided by this invention simplifies the operation steps by deeply integrating software and hardware. The device realizes a closed loop of the entire process from measurement and calculation to diagnostic assistance, ensuring the integrity of data and the standardization of the assessment process, and effectively avoiding errors that may occur due to data transfer between different systems.

[0080] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0081] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0082] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0083] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0084] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A sarcopenia assessment system, characterized by, The method comprises the following modules: a data processing module, configured to split the collected multiple sets of clinical data of maintenance hemodialysis patients into a training set and a test set, and perform data preprocessing on the training set; a feature screening module, configured to perform feature screening on the preprocessed training data to determine an optimal feature subset from multiple features; a model construction and training module, configured to select a model architecture from multiple machine learning algorithms based on the optimal feature subset to construct a sarcopenia evaluation model, and train and optimize hyperparameters of the sarcopenia evaluation model to obtain an optimized sarcopenia evaluation model; a model evaluation module, configured to input the test set into the optimized sarcopenia evaluation model to evaluate the generalization performance of the model; an actual application module, configured to obtain feature data corresponding to the optimal feature subset of a patient to be evaluated in an application stage, and obtain an evaluation result by using the sarcopenia evaluation model.

2. The sarcopenia assessment system of claim 1, wherein: When the model construction and training module selects a model architecture from multiple machine learning algorithms to construct a sarcopenia evaluation model, the module comprises: selecting and instantiating a machine learning model from a preset model library, wherein the model library at least includes four classifiers of support vector machine, logistic regression, random forest and extreme gradient boosting tree; configuring corresponding initialization parameters for different model types, wherein the support vector machine is configured with a kernel function type and a penalty coefficient, the logistic regression is configured with a regularization strength and an optimization algorithm, and the random forest and the extreme gradient boosting tree are configured with a number and a depth of trees.

3. The sarcopenia assessment system according to claim 1 or 2, characterized by: When the model construction and training module trains and optimizes hyperparameters of the sarcopenia evaluation model to obtain an optimized sarcopenia evaluation model, the module comprises: extracting feature columns corresponding to the optimal feature subset from the preprocessed training set to construct a reduced training set; adopting a 5-fold cross-validation method, the reduced training set is randomly divided into 5 non-overlapping subsets, and 4 subsets are used as training data and the remaining 1 subset is used as validation data for cyclic training and validation; based on the cross-validation process, different parameter combinations are systematically traversed in a preset hyperparameter space by a grid search technique, and the optimal hyperparameter configuration is found by taking the average validation set performance as an indicator to optimize the model performance; retraining the optimal hyperparameter configuration on the complete reduced training set to obtain the optimized sarcopenia evaluation model.

4. The sarcopenia assessment system of claim 1, wherein: When the data processing module splits the collected multiple sets of clinical data of maintenance hemodialysis patients into a training set and a test set, the module comprises: dividing the clinical data into different levels, at least including two levels of sarcopenia patients and non-sarcopenia patients; in each level, a random sampling method is used to extract samples from each level according to a preset same sampling ratio; the samples extracted from each level are combined to form a training set, and the remaining samples in each level are combined to form a test set.

5. The sarcopenia assessment system of claim 1, wherein: The data processing module comprises the following steps when performing data preprocessing on the training set: detecting missing feature values in the training set, filling the missing feature values by iterative regression interpolation of multiple variables through a chain equation with Bayesian ridge regression as an estimator, and performing standardization or normalization processing on the filled continuous numerical feature values.

6. The sarcopenia assessment system of claim 1, wherein: The feature screening module, when screening features from the preprocessed training data to determine an optimal feature subset from multiple features, comprises: automatically finding an optimal regularization penalty coefficient through 5-fold cross-validation ; and under the condition of the optimal regularization penalty coefficient , compressing the feature coefficient with lower contribution to zero, automatically screening out features with non-zero regression coefficients to form the optimal feature subset.

7. The system for sarcopenia assessment according to claim 6, characterized in that: The optimal feature subset comprises gender, weight, extracellular fluid resistance, and intracellular fluid resistance.

8. The sarcopenia assessment system of claim 1, wherein: The model evaluation module comprises the following steps when evaluating the generalization performance of the model: inputting the test set into the optimized sarcopenia evaluation model to obtain a prediction result of the model for each test sample; comparing the prediction result with a corresponding true label in the test set to calculate multiple performance indexes including accuracy, sensitivity, specificity, area under the receiver operating characteristic curve, and F1 score; generating multiple sample subsets by multiple random sampling with replacement from the test set and recalculating the performance indexes on each sample subset; calculating a confidence interval of each performance index based on the performance index distribution obtained by multiple resampling to quantitatively evaluate the statistical reliability of the evaluation result.

9. The sarcopenia assessment system of claim 1, wherein: The sarcopenia evaluation system further comprises a model storage module configured to serialize the optimized sarcopenia evaluation model and its parameter into a model file that can be persistently stored, and store the model file in a non-volatile storage medium of a server or a terminal device.

10. A sarcopenia assessment medical device characterized by: The sarcopenia evaluation medical device comprises the sarcopenia evaluation system according to any one of claims 1-9.

Images