A method for constructing a hidden obesity prediction model, a hidden obesity prediction method, and related devices

Abstract

The application provides a hidden obesity prediction model construction method, a hidden obesity prediction method and related equipment, and relates to the technical field of electric digital data processing and the technical field of medical health informatics. The application constructs a hidden obesity prediction model based on non-invasive features, combines a clinical minimum feature set for efficient prediction, and avoids the pain and risk of invasive examination. The method has high accuracy, sensitivity and real-time response capability, can quickly process data in a low computing power environment, improve prediction efficiency, and the model has strong regularization capability, reduces overfitting, and ensures the stability and reliability of the prediction process.

Description

Technical Field

[0001] This invention relates to the fields of electronic digital data processing technology and healthcare informatics technology, and more specifically, to a method for constructing a hidden obesity prediction model, a method for predicting hidden obesity, and related equipment. Background Technology

[0002] In current technology, with the accelerating aging of the global population, the burden of metabolic diseases is increasing. Metabolic syndrome (MetS), as a complex group of metabolic disorders, is an independent risk factor for cardiovascular disease (CVD) and type 2 diabetes mellitus (T2DM). A large-scale prospective study confirmed that MetS significantly increases all-cause mortality and cardiovascular mortality, and this risk association remains robust across different age strata.

[0003] In traditional clinical understanding and public health strategies, the Metabolic Obesity Study (MetS) is usually closely linked to overweight or obesity (high BMI), leading screening strategies to often use BMI as the first line of defense. However, this strategy masks a clinically significant but often overlooked subgroup: the metabolically obese normal-weight (MONW) population. This concept refers to individuals with a BMI within the normal range (18.5–24 kg / m²) but with severe metabolic disorders. Recent data from the Chinese Chronic Disease Prospective Study (CKB) show that the MONW phenotype is prevalent among Chinese adults, with their cardiovascular disease risk even higher than that of metabolically healthy obese individuals.

[0004] Furthermore, current diagnostic methods for MONW rely on invasive biochemical tests such as blood glucose and blood lipids, which are costly and difficult to conduct frequently in resource-constrained large-scale community screenings. MONW is particularly prominent and insidious in middle-aged and elderly populations. The aging process is accompanied by sarcopenic obesity, characterized by a decrease in muscle mass and the infiltration of visceral fat. This muscle loss masks the pathological changes of increased fat, rendering BMI, a total weight-based indicator, completely insensitive in the elderly. Middle-aged and elderly individuals with MONW face a serious risk of diagnostic delay: on the one hand, they are easily missed by primary healthcare screening systems due to normal BMI; on the other hand, their metabolic risk is not reduced despite normal weight. Although blood biochemical testing is the gold standard for diagnosis, invasive blood collection is costly and difficult to conduct frequently in resource-constrained community screenings.

[0005] Although existing studies have used machine learning to predict MetS, screening models for the specific subgroup MONW still face the following bottlenecks: Lack of targeted non-invasive indicator system: Existing models mostly rely on traditional waist circumference (WC) or BMI, which are difficult to capture the pathological characteristics of visceral fat accumulation and muscle loss unique to the elderly.

[0006] Eigencollinearity interference: There is strong multicollinearity among existing non-invasive anthropometric indicators (such as waist circumference WC, body roundness index BRI, and body shape index ABSI), which cannot be effectively decoupled by traditional linear models.

[0007] Class imbalance leading to missed diagnoses: In people of normal weight, MONW is a relatively low-frequency event. Conventional algorithms tend to sacrifice recall, resulting in a large number of hidden patients being missed. Summary of the Invention

[0008] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a method for constructing a hidden obesity prediction model, a method for predicting hidden obesity, and related equipment, so as to overcome the shortcomings of the existing technology.

[0009] The above-mentioned technical objective of the present invention is achieved through the following technical solution: Firstly, a method for constructing a hidden obesity prediction model, comprising the following steps: A clinical dataset of middle-aged and elderly people is obtained, which includes several samples, each of which includes several non-invasive features and their corresponding feature values; according to a predetermined classification standard, labels are added to each sample in the clinical dataset of middle-aged and elderly people, and the labels are: patients with hidden obesity or healthy people; Calculate the positive sample weight factor, and use the positive sample weight factor to optimize the original loss function of each preset machine learning model to obtain the cost-sensitive loss function corresponding to each machine learning model. The aforementioned clinical dataset of middle-aged and elderly individuals is cleaned, and the non-invasive features after data cleaning are screened based on a three-stage feature optimization cascade algorithm to obtain a minimal clinical feature set. The three-stage feature optimization cascade algorithm includes: calculating the variance inflation factor to remove highly collinear features, using the Boruta algorithm to remove low-relevance features, and using the RFE algorithm to remove low-importance features. Specifically, the RFE algorithm uses the cost-sensitive loss function or the original loss function to train various preset machine learning models. The trained machine learning models are evaluated using a pre-defined independent validation set. The model with the best evaluation performance is selected for optimization to obtain a hidden obesity prediction model.

[0010] In some embodiments, the model with the best performance evaluation is specifically the CatBoost model.

[0011] In some embodiments, the cost-sensitive loss function of the CatBoost model is specifically: ; in, This represents the loss function value of the model; Indicates the number of samples; Indicates the weighting factor for positive samples; Indicates the first The true labels of each sample; Indicates the first The predicted probability of each sample; the positive sample weighting factor is specifically: ;in, Indicates the weighting factor for positive samples; Indicates the number of positive samples; This indicates the number of negative samples.

[0012] In some embodiments, the minimum clinical feature set includes the following non-invasive features: BRI, ABSI, relative grip strength, PEF, pulse pressure, pulse, age, sex, alcohol consumption, self-reported health, and history of heart disease.

[0013] In some embodiments, calculating the variance inflation factor to remove highly collinear features specifically includes: calculating the first... A linear regression equation is established between the first non-invasive feature and the remaining non-invasive features; based on the linear regression equation, the first non-invasive feature is calculated. The coefficient of determination for each non-invasive feature Based on the aforementioned determination coefficient Calculate the first The variance inflation factors corresponding to each non-invasive feature include: Based on the variance inflation factor corresponding to each non-invasive feature, an iterative elimination strategy is executed to generate a low collinearity feature set.

[0014] In some embodiments, the Boruta algorithm for removing low-relevance features specifically includes: constructing corresponding shadow features based on each non-invasive feature in the low-collinearity feature set; calculating the importance score of each shadow feature; and determining the most important shadow feature based on the importance score, with the score denoted as […]. ; calculate the importance score corresponding to each noninvasive feature in the low collinearity feature set, denoted as . ;Compare and ,like Greater than And if the preset significance condition is met, then it is determined that... The corresponding non-invasive features are significantly correlated and are retained; otherwise, the determination is made. The corresponding non-invasive features that are not significantly correlated are removed to obtain a set of highly correlated features.

[0015] In some embodiments, the removal of low-importance features using the RFE algorithm specifically includes: training each preset machine learning model using all features in the high-relevance feature set, wherein the training specifically employs several rounds of 10-fold cross-validation, and in each round, the contribution of each feature value to the model is evaluated, and the feature with the lowest contribution is removed to obtain the clinical minimum feature set.

[0016] In some embodiments, the optimization of the model with the best evaluation performance specifically includes: using a ranking boosting mechanism to process categorical features, encoding each categorical feature by introducing random permutation, wherein the encoded value is calculated based on the sample permutation order and prior information of the target variable, and smoothing the encoded value of the categorical feature using a smoothing parameter to reduce prediction bias; in the construction of the base model, a forgetting tree is used as a base learner, wherein each forgetting tree uses the same splitting feature and splitting threshold at each layer node, and the prediction mechanism of the forgetting tree includes using the splitting feature index and splitting threshold to perform binary judgment of the sample, dividing the sample to the corresponding leaf node, and accumulating the prediction results of each base learner through an additive model; the final prediction result is mapped to the probability of hidden obesity using a Sigmoid function.

[0017] Secondly, a method for predicting hidden obesity, the method comprising the following steps: Obtain the minimum clinical feature set of the target object, wherein all data in the minimum clinical feature set are non-invasive feature data; The clinical minimum feature set is input into the hidden obesity prediction model as described in any of the first aspects to make a prediction and obtain the probability of hidden obesity in the target subject.

[0018] Thirdly, a computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the method for constructing a hidden obesity prediction model as described in any one of the first aspects, or the method for predicting hidden obesity as described in the second aspect.

[0019] In summary, the present invention has the following beneficial effects: This application provides a method for constructing a hidden obesity prediction model and a method for predicting hidden obesity. Using the method of the present invention, a hidden obesity prediction model based on non-invasive features is constructed, and combined with a clinical minimum feature set for efficient prediction, avoiding the pain and risks of invasive examinations. This method has high accuracy, sensitivity, and real-time response capability, can quickly process data in low-computing-power environments, improves prediction efficiency, and the model has strong regularization capability, reducing overfitting and ensuring the stability and reliability of the prediction process. Attached Figure Description

[0020] Figure 1This is a flowchart of the method for constructing the hidden obesity prediction model of the present invention; Figure 2 This is a continuous variable heatmap of the present invention; Figure 3 This is a schematic diagram of the Boruta algorithm filtering results of the present invention; Figure 4 The results of recursive feature analysis and screening in this embodiment of the invention; Figure 5 This is a comparison chart of the ROC values ​​of ten models in the embodiments of the present invention; Figure 6 This is a global overview diagram of the importance of SHAP features in the embodiments of the present invention; Figure 7 This is a global feature influence distribution diagram in an embodiment of the present invention; Figure 8 This is a system output interface for risk assessment overview in an embodiment of the present invention; Figure 9 This is the system output interface for risk attribution analysis in this embodiment of the invention. Detailed Implementation

[0021] To make the objectives, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Several embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein.

[0022] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, and c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0023] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0024] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0025] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0026] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.

[0027] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0028] Example 1 To address the above problems, this invention provides a method for constructing a hidden obesity prediction model, such as... Figure 1 As shown, it includes the following steps: Obtain a clinical dataset for middle-aged and elderly people. The clinical dataset for middle-aged and elderly people consists of a number of samples. Each sample includes a number of non-invasive features and the corresponding feature values. Specifically, during the model construction process, it is first necessary to select a suitable target population. Currently, in the middle-aged and elderly population (age > 45 years), the MONW phenomenon is particularly prominent and concealed. The aging process is accompanied by sarcopenic obesity, that is, the coexistence of muscle mass reduction and visceral fat infiltration. This pathological change of muscle loss masking fat increase makes the BMI, an index based on total weight, completely lose its sensitivity in the elderly. Middle-aged and elderly people with MONW face a serious risk of diagnostic delay: on the one hand, middle-aged and elderly people are easily missed by the primary medical screening system due to normal BMI; on the other hand, their metabolic risk does not decrease due to normal weight. Although blood biochemical testing is the gold standard for diagnosis, in community screening with limited resources, invasive blood sampling is costly and difficult to carry out frequently. Therefore, to solve the problem that the MONW phenomenon in the middle-aged and elderly population is particularly prominent and concealed, the research object of this application is the middle-aged and elderly group. Specifically, the inclusion criteria are: age > 45 years; having complete fasting blood samples and anthropometric data; body mass index (BMI) strictly within the normal range (18.5 < BMI < 24.0). At the same time, it is also necessary to exclude some samples due to individual differences. The exclusion criteria are specifically: a history of malignant tumors; severe physical or intellectual disabilities; missing key outcome variables. Finally, 4622 people were included in the analysis cohort and randomly divided into a training set and a test set according to a 7:3 ratio. Among them, there are 3236 people in the training set and 1386 people in the test set. The data in this experiment comes from the China Health and Retirement Longitudinal Study (CHARLS) 2015 data.

[0029] Furthermore, in order to follow the principles of non-invasiveness and easy accessibility, this application mainly focuses on 39 indicators in five dimensions to construct a full-scale feature pool: The first type of features is demographic and basic physical characteristics, including: 1. Age; 2. Gender; 3. Type of residence; 4. Marital status; 5. Socioeconomic status; 6. Height; 7. Weight; 8. Waist circumference; 9. Body mass index (BMI); 10. Systolic blood pressure; 11. Diastolic blood pressure; 12. Pulse.

[0030] The second type of features is new anthropometric and hemodynamic-derived indicators, specifically including: 13. Body roundness index (BRI); 14. Body shape index (ABSI); 15. Waist-to-height ratio (WHtR); 16. Pulse pressure reflecting the degree of arteriosclerosis.

[0031] The third category of characteristics is a history of comorbid chronic diseases, specifically including: 17. heart disease; 18. stroke; 19. chronic lung disease; 20. liver disease; 21. kidney disease; 22. stomach disease; 23. arthritis; 24. asthma; 25. emotional / mental disorders; 26. memory-related disorders.

[0032] The fourth category of characteristics is the assessment of physical function and cognitive ability, which is an objective indicator reflecting muscle strength and lung function. Specifically, it includes: 27. Relative grip strength (grip strength / body weight); 28. Maximum expiratory flow rate (PEF); 29. ​​Standing test time; 30. Instrumental activities of daily living (IADL); 31. History of falls; 32. Total cognitive function score.

[0033] The fifth category of characteristics is lifestyle and psychosocial factors, specifically including: 33. Smoking status; 34. Drinking status; 35. Nighttime sleep duration classification; 36. Nap duration classification; 37. Epidemiological investigation center depression scale score; 38. Self-reported health status; 39. Life satisfaction score.

[0034] That is, each middle-aged and elderly person corresponds to a sample. The middle-aged and elderly clinical dataset contains 4,622 samples. Each sample contains 39 non-invasive features, and each non-invasive feature corresponds to a non-invasive feature data.

[0035] Furthermore, in order to predict whether each sample has hidden obesity symptoms, a gold standard needs to be constructed to add a truth label to each sample. In this embodiment, the judgment criteria are as follows: (1) Waist circumference (WC) determination: male waist circumference ≥90cm, female waist circumference ≥85cm.

[0036] (2) Hyperglycemic state: fasting plasma glucose (FPG) ≥ 6.0 mmol / L, or those who have been diagnosed with type 2 diabetes and are receiving treatment.

[0037] (3) Hypertension status: systolic blood pressure (SBP) ≥130 mmHg or diastolic blood pressure (DBP) ≥85 mmHg, or those who have been diagnosed with hypertension and have received treatment.

[0038] (4) Hypertriglyceridemia: Fasting triglycerides (TG) ≥1.70 mmol / L.

[0039] (5) Low high-density lipoprotein cholesterol: fasting high-density lipoprotein cholesterol (HDL-C) <1.04 mmol / L.

[0040] Individuals meeting three or more of the above criteria were labeled as MONW (Monotrophic Obesity) patients, i.e., positive samples; those meeting only 0-2 of the above criteria were labeled as control group, i.e., negative samples. Each sample was tagged according to these screening criteria for use in subsequent training steps.

[0041] In this embodiment, to comprehensively evaluate the model's performance in the Monounseen Obesity (MONW) prediction task and ensure the selection of the optimal model, it's crucial to understand that different models may have varying strengths and limitations when processing data. By comparing the performance of multiple models, the most suitable model for the task can be identified, thereby improving the accuracy and stability of predictions. This is especially important when dealing with class imbalance problems, as different algorithms may have varying impacts on the identification ability of minority class samples. This embodiment selects ten machine learning models, including: RF model, CatBoost model, GLM model, KNN model, SVM model, NNET model, Decision Tree model, Naive Bayes model, XGBoost model, and LightGBM model. Comparing the performance of multiple models helps avoid the bias that may arise from using a single model, ensuring the reliability and practicality of the results.

[0042] In the labeled dataset, the number of patients with hidden obesity is significantly less than the number of normal patients. In some implementations, synthetic data is used to increase the number of positive samples, achieving a relative balance between negative and positive samples. However, this method also has the following drawbacks: synthetic samples are generated by interpolation between the original minority class samples, and these synthetic samples may not represent the true sample distribution. Especially in high-dimensional spaces, synthetic samples may cause the model to learn some unrealistic patterns, thus increasing noise. Generating synthetic samples also increases the size of the training data. When the sample size is already large, synthetic oversampling will further increase the size of the training set, leading to a significant increase in computational resources and time consumption. This application introduces a weighting factor, specifically a positive sample weighting factor, into the loss function, which allows the model to pay more attention to minority class samples during training, thereby improving the prediction ability of minority class samples and reducing the false negative rate. In this embodiment, the positive sample weighting factor... Defined as the number of negative samples and the number of positive samples The ratio, that is: Then, the loss function of the preset machine learning model is optimized using the sample weight factor to obtain the cost-sensitive loss function corresponding to each machine learning model, thereby increasing the penalty intensity of the model when learning positive samples, making the model pay more attention to minority class samples.

[0043] For example, when the model is a CatBoost model, the original loss function is: ; After introducing a positive sample weighting factor, the cost-sensitive loss function of the model is: .

[0044] in, This represents the loss function value of the model; Indicates the number of samples; Indicates the first The true labels of each sample; Indicates the first The probability of the predicted value for each sample.

[0045] Furthermore, this embodiment provides an example of a second cost-sensitive loss function: when the model is an XGBoost model, its original binary cross-entropy (Log-Loss) loss function is: ; After introducing a positive sample weight factor, the model pays more attention to the gradient update of the minority class when splitting training nodes, and its cost-sensitive loss function is transformed into weighted cross-entropy: .

[0046] Furthermore, this embodiment provides an example of a third cost-sensitive loss function: when the model is an SVM model, its original objective function based on hinge loss is: ; After introducing a positive sample weighting factor, in order to impose a greater penalty on misclassification of the minority class (hidden obesity patients), different penalty coefficients are assigned to different classes, and the cost-sensitive loss function is adjusted as follows: ; Where, when the true label of the i-th sample When the sample is positive, the penalty weight When the sample is negative, ; For regularization parameters, It is the normal vector. This is for predicting the output.

[0047] Furthermore, since the Naive Bayes algorithm is based on Bayes' theorem and belongs to the generative probabilistic model, its training process directly obtains the prior probability and conditional probability through maximum likelihood estimation of sample frequency, rather than relying on gradient descent algorithm to iteratively optimize the target loss function. Therefore, introducing [the necessary parameters] into the loss function is crucial. The weighted iteration mechanism is not applicable to it. Therefore, this application constructs a cost-sensitive loss function mechanism for the other nine discriminative models, while retaining the original probability calculation logic for the Naive Bayes model, thus ensuring the rigor and applicability of the underlying algorithm.

[0048] In practical applications, datasets often contain multiple highly correlated features (such as multiple measurement indicators reflecting the same phenomenon). These features provide repetitive information, which not only fails to increase model effectiveness but also leads to overfitting. In machine learning, the more features there are, the more computational resources are required for model training and prediction. This is especially true when the dataset has very high feature dimensionality, significantly increasing training time and storage requirements. Feature simplification can reduce computational load during training, making the training process more efficient, which is particularly important when dealing with large amounts of data. It is essential to filter and simplify data by reducing redundant features, retaining only the features most valuable for predicting the target variable, thereby improving model performance, reducing computational complexity, and enhancing the model's generalization ability. To filter important features from the aforementioned full feature pool and remove redundant features, this application proposes a three-stage feature optimization cascade algorithm, specifically including the following steps: Phase 1: Collinearity assessment and redundancy removal based on variance inflation factor.

[0049] Collinearity refers to a strong correlation between different features in a dataset. For example, features such as waist circumference, height, weight, and BMI may overlap, providing similar information. If the correlation between features is too strong, the model struggles to determine which features have the greatest impact on the prediction results, thus affecting the model's accuracy and stability. To identify which features exhibit severe collinearity, the variance inflation factor (VIF) can be used to quantify the degree of collinearity for each feature. VIF measures the correlation between a given feature and other features. The specific calculation method is as follows: for the first feature in the dataset... VIF uses one feature to perform a linear regression equation with all other features, including: ;in, Let be the target variable, representing the in the dataset. The value of each feature; The intercept of the linear regression; These are the regression coefficients in a regression model, representing the degree of influence of each independent variable (feature) on the target variable. The regression coefficients of each feature indicate that, with other factors remaining constant, the feature... unit change The impact; Let be the independent variable, representing the in the dataset. The value of each feature; It is the error term, representing the part that the model cannot explain, or the difference between the observed value and the predicted value of the regression model.

[0050] The coefficient of determination corresponding to the j-th noninvasive feature is calculated based on the linear regression equation. ,include: ; in, Indicates the first The true value of each observation, that is, the actual observed value of the target variable; Indicates the first The predicted value of each observation, i.e., the value predicted by the regression model. The coefficient of determination is the mean of the target variable, i.e., the average of all observations, where N is the sample size. The denominator represents the total variability of the target variable, which is the sum of squared deviations between the target variable and its mean. Coefficient of determination The numerator represents the sum of squares of the model prediction errors, i.e., the sum of squared differences between the actual and predicted values, and the coefficient of determination. The coefficient of determination calculates the proportion of variability explained by the regression model to the total variability. The value is between 0 and 1: This means the model can fully explain the variability of the data, i.e., a perfect fit. This indicates that the model cannot explain any variability, which is equivalent to the model making predictions solely based on the mean. Specifically, in this application, it means that the first... The coefficient of determination of the linear relationship between the first non-invasive feature and other non-invasive features; the larger the value, the stronger the linear relationship. The stronger the correlation between a non-invasive feature and other features.

[0051] Then based on the coefficient of determination Calculate the variance inflation factor ,include: Variance inflation factor (VIF) is an indicator that quantifies collinearity among features, measuring the degree of linear relationship between a feature and all other features. Specifically, the VIFs of continuous variables in the full feature pool are shown in Table 1.

[0052] Table 1. Results of collinearity diagnosis VIF

[0053] Analysis of the initial VIFs in the table reveals that some variables exhibit extreme collinearity, such as waist-to-height ratio (VIF=2058.78), ABSI (VIF=567.17), waist circumference (VIF=532.31), and BRI (VIF=445.31).

[0054] Furthermore, to confirm the correlation between any two variables, this embodiment further calculates the Pearson coefficient between any two non-invasive features, which is the correlation coefficient between any two non-invasive features, such as... Figure 2 As shown, the correlation heatmap of continuous variables in the full feature pool reveals a strong correlation between height and multiple features, such as waist-to-height ratio, waist circumference, BMI, height, and weight. Therefore, when eliminating collinear features, features strongly correlated with height should be eliminated first.

[0055] In the feature removal process, this embodiment employs an iterative removal strategy. After removing one highly collinear non-invasive feature (e.g., waist-to-height ratio), the VIF (Variable Interest Scale) is recalculated for all remaining features. Then, the non-invasive feature with the highest collinearity (e.g., ABSI) is removed, and the VIF among the remaining non-invasive features is calculated again. The final non-invasive feature removal results are shown in Table 1 above. Waist-to-height ratio, weight, height, BMI, and waist circumference were removed from the full feature pool.

[0056] Furthermore, analysis revealed that the non-invasive features BRI and ABSI had high VIF values ​​in the initial feature pool. However, because these two features effectively reflect visceral fat accumulation, they were reintroduced into the final feature set, and their VIF values ​​were recalculated. The final VIF for these two features was found to have decreased to [value missing]. , It possesses good independence. For example... Figure 2 As shown in the figure, this embodiment also found a strong correlation between systolic and diastolic blood pressure, with a correlation coefficient of 0.72, indicating a strong linear correlation between them. Retaining both simultaneously would lead to severe multicollinearity interference; retaining only one would inevitably result in the loss of blood pressure fluctuation information carried by the other feature. Pulse pressure, as the difference between systolic and diastolic blood pressure, not only perfectly resolves the multicollinearity problem but also achieves information fusion through feature derivation. Secondly, this invention focuses on the middle-aged and elderly population aged >45 years. In this population, the instantaneous absolute values ​​of systolic or diastolic blood pressure have limited predictive sensitivity for hidden obesity (MONW), while "pulse pressure difference" is a clinically recognized non-invasive and sensitive indicator reflecting the stiffness of large arteries and the degree of vascular aging. Finally, the SHAP interpretability analysis results (global feature influence distribution map) in Example 5 also confirm the correctness of this treatment. The results showed that pulse pressure had a decisive impact on the prediction output of MONW disease risk on a global scale. Therefore, in this embodiment, pulse pressure was used instead of diastolic and systolic blood pressure to reduce feature redundancy and correlation and improve the independence between features.

[0057] The feature set after removing collinear features is called the low collinearity feature set, which specifically includes: ABSI, BRI, pulse pressure, age, maximum expiratory flow rate, relative grip strength, total cognitive function score, standing time, depression score and pulse, as well as other non-continuous variable features in the full feature pool.

[0058] The second stage involves using the Boruta algorithm to remove low-relevance features, including the following steps: Based on each non-invasive feature in the aforementioned low-collinearity feature set, corresponding shadow features are constructed. Specifically, shadow features are pseudo-features generated by randomly permuting the original data. These pseudo-features do not contain any actual information; they are only used to compare with the real features to see which real features are more important than these pseudo-features. Taking waist circumference as an example, Boruta will randomly shuffle the data of these waist circumference features to generate a shadow waist circumference feature. The data value of the shadow waist circumference has no practical meaning. A corresponding shadow feature is generated for each feature.

[0059] Calculate the importance score for each shadow feature, and determine the most important shadow feature based on the importance score. The score of this most important shadow feature is denoted as [missing value]. Specifically, the importance index for calculation can be measured using metrics such as information gain and Gini impurity to assess the importance of shadow features.

[0060] Calculate the importance score for each noninvasive feature in the low collinearity feature set, denoted as . Specifically, for each feature, the same steps are used to calculate the importance score corresponding to each feature. This information is then used for subsequent feature importance comparisons.

[0061] Importance score calculated using shadow features Importance scores for each feature Compare, if Significantly greater than Then determine The corresponding non-invasive features are significantly correlated and are retained; if No significantly greater than Then determine The corresponding non-invasive features were not significantly correlated and were removed, resulting in a set of highly correlated features. Specifically, significance was determined using a binomial distribution test: ,in For the number of iterations, for > The number of times.

[0062] like Figure 3 As shown, this step identified 16 relevant variables, including not only strongly correlated variables such as BRI, pulse pressure, and ABSI, but also weak signal variables such as "self-reported health status" and "depression score." Although these weak signal variables have low linear correlation with MONW, they provide unique predictive gains in the nonlinear space. Specifically, when constructing the full feature pool, this application divides non-invasive features into five dimensions. Among them, "self-reported health status" and "depression score" are specifically classified as "the fifth category of features: lifestyle and psychosocial factors." This ensures that the features, before entering the algorithm, already possess the clinically logical qualification to reflect physical functional compensation and metabolic psychological stress assessment.

[0063] Although self-reported health status and depression scores are weak signal variables and have low univariate linear correlation with hidden obesity (MONW), these two features successfully passed the Boruta algorithm's comparison with... The significance of the comparison was evaluated. This objectively proved that they were not random noise. Clinically, these two variables effectively captured the hidden psycho-metabolic stress and subjective signs of early sarcopenia unique to middle-aged and elderly people, providing an indispensable supplementary signal for the predictive model from a non-physical measurement dimension.

[0064] This invention does not rely on traditional linear regression equations, but instead utilizes a tree-based model-based nonlinear space to uncover complex interactions between features. Specifically, for the prediction gain of these weak signal variables in the nonlinear space, this invention employs the Tree Explainer algorithm within the SHAP interpretation framework for precise quantitative evaluation. This evaluation method is based on cooperative game theory, calculating the weighted average of the marginal contributions of features when included in all possible feature subset combinations (i.e., the SHAP value). In this way, the independent gain contribution of weak signal variables to the final predicted probability log-Odds when they interact nonlinearly with other features (such as age and body mass index) is objectively evaluated.

[0065] Phase 3: Recursive RFE Feature Elimination The core objective of this stage is to find the most important features using the Recursive Feature Elimination (RFE) method to improve the model's predictive performance. Specifically, by removing unimportant features, the number of input features is reduced, making the model more concise and efficient. First, the RFE algorithm uses the current feature set—the 16 features identified in the second stage—to train the model and calculate the model's performance, such as... Figure 4 As shown, this embodiment uses AUC (Area Under the Curve) as the evaluation metric. The higher the AUC value, the better the model's performance. The core step of RFE is to gradually eliminate features that contribute little to the model and have the lowest importance ranking.

[0066] In this way, the model becomes increasingly concise, removing features that have little impact on prediction. To ensure model stability and avoid overfitting, the RFE algorithm uses 10 rounds of 10-fold cross-validation to evaluate the model's performance after each feature removal. The algorithm repeatedly verifies whether the removal of each feature helps improve the model's generalization ability. Through iteration, RFE finds a balance point, that is, the optimal relationship between model performance (AUC) and the number of features.

[0067]

[0068] in, The balance point is the balance between model performance and the number of features, which is the final determined feature subset; Represents a feature subset. Represents the set of all features. Indicates from Select features from the set gather. The AUC value represents the calculated value on the feature subset S. It is usually used to evaluate the performance of classification models. The larger the AUC value, the better the model's performance. The term serves as a regularization term, penalizing the size of the feature subset and preventing the selection of too many features. Its purpose is to control the complexity of the model and prevent overfitting; This means selecting the subset that maximizes the objective function from all possible feature subsets S.

[0069] In this embodiment, the preset number of machine learning models is 10. Among them, the loss function of the Naive Bayes model is the original loss function, and the loss functions of the other 9 machine learning models are optimized using positive sample weight factors to obtain corresponding cost-sensitive loss functions, so that the machine learning models pay more attention to minority class samples.

[0070] As mentioned above, the sample size was 4622 people, randomly divided into training and test sets in a 7:3 ratio, with 3236 people in the training set and 1386 people in the test set. Furthermore, the 3236 people in the training set were used to perform 10-fold cross-validation on ten machine learning models. In each round of cross-validation, the importance of each feature to the model performance was evaluated. In some embodiments, the importance of the model can be evaluated by the degree of contribution of each feature to the AUC. By removing features with smaller contributions, the remaining features are used for retraining until the most useful features remain.

[0071] In this phase, the 16 features remaining from the second round of screening were further eliminated in the third phase, ultimately retaining 11 non-invasive features, including: BRI, ABSI, relative grip strength, PEF, pulse pressure, pulse, age, gender, alcohol consumption, self-reported health, and history of heart disease. Among these, BRI and ABSI are used to accurately quantify central obesity; relative grip strength and PEF are used to detect sarcopenia risk; pulse pressure and pulse reflect arteriosclerosis; and age, gender, alcohol consumption, self-reported health, and history of heart disease collectively reflect the patient's lifestyle across multiple dimensions.

[0072] Furthermore, after performing ten-fold cross-validation on the ten machine models, corresponding trained machine learning models are obtained. These ten trained machine learning models are then independently validated using the aforementioned validation set. Figure 5 As shown in Table 2, a balance of multiple metrics needs to be considered to comprehensively evaluate which model is optimal. Sensitivity and AUC are usually the most important metrics when dealing with class imbalance, as they directly affect the ability to identify minority class samples. As shown in Table 2, CatBoost has a high AUC (0.7791) and sensitivity (0.7382), meaning that CatBoost can distinguish between positive and negative samples well and has a high ability to identify minority class samples (MONW). XGBoost performs well in sensitivity (0.7435) and NPV (0.8755), especially in the accuracy of identifying minority class samples (hidden obesity) and negative class samples. From an overall performance perspective, CatBoost is a relatively balanced and effective choice, especially suitable for emphasizing minority class identification. Although XGBoost performs better in some aspects (such as NPV), overall, CatBoost's performance across multiple key metrics is more balanced, and therefore it can be considered the optimal model overall.

[0073] Table 2 Performance Comparison Chart of Models

[0074] After three stages of screening and evaluation, this technical solution selected the most suitable model, CatBoost, and 11 of the most representative non-invasive feature indicators.

[0075] Furthermore, traditional machine learning models may exhibit bias in handling categorical features, such as gender, history of heart disease, alcohol consumption status, and self-reported health status, especially when the distribution of categorical features is imbalanced. CatBoost's ranking boosting mechanism effectively reduces this bias and improves the modeling ability for categorical features. To ensure that the model can accurately identify the impact of different categorical features on the target variable during training, random permutation and smoothing of prior probabilities are introduced to avoid bias in the model regarding categorical features and to more fairly handle the relationships between different categorical features. Specifically, random permutation is introduced to avoid bias in the prediction process. For the first The first sample Category features Its encoded value By considering the arrangement The calculation was based on previous samples. The specific calculation method is as follows: ;in, It is an indicator function used to determine whether the current feature is consistent with the previous sample. It is the prior probability, representing the overall distribution of the category; It is a smoothing parameter, and satisfies This method helps avoid overfitting and smooth out the effects of noise during model training. It ensures that when processing categorical features, the model fully considers the relationship between categorical features and the target variable, while avoiding bias caused by categorical feature ranking issues.

[0076] Furthermore, in this embodiment, Oblivious Trees from the CatBoost algorithm are used as the base learner for model construction. Unlike the asymmetric growth strategy used by traditional decision tree algorithms (such as XGBoost or Random Forest), Oblivious Trees employ a fully symmetric decision tree (i.e., at each level of the tree, all nodes use the same split feature and split threshold). This structural design effectively improves the computational efficiency and regularization effect of the model. In the construction of the Oblivious Tree, nodes at each level must use the same split feature and split threshold. This forced symmetry structure not only gives the model stronger regularization capabilities during training, effectively preventing overfitting on small sample data, but also simplifies the tree's inference process, making its computation more efficient, thereby greatly improving the model's response speed in online prediction.

[0077] Specifically, the optimal prediction model of this invention It is by An additive model composed of forgetting trees. The prediction mechanism of the base learner: for a depth of... The A forgetting tree, the structure of which consists of a set of splitting rules. Definition, where For the first The layer splitting feature index corresponds to the 11 non-invasive indicators selected in this invention. This corresponds to the threshold. For the input sample vector... The index of the leaf node it falls into Calculated by accumulating binary bit weights: ;in The function is an indicator function (1 if the condition is met, 0 otherwise). This formula shows that the leaf node position of a sample depends only on the binary string composed of its binary judgment results on each layer of features.

[0078] The probability of developing hidden obesity By summing the weights of all base learners at the corresponding leaf nodes And obtained by mapping through the Sigmoid function: ; in For learning rate, Representing the Tree No. The output value of each leaf node.

[0079] In summary, forget trees, by enforcing the use of the same splitting features and thresholds at each level to form a symmetrical tree structure, naturally act as a regularization mechanism, effectively preventing overfitting when the model has limited data and high complexity. Regularization is particularly important for improving the model's generalization ability when dealing with small sample data (such as health data of the elderly). Forget trees simplify the inference process by using binary bitwise operations to calculate the leaf node where a sample falls. This allows the model to perform only simple comparisons and shift operations during the inference phase, avoiding the complex decision path calculations found in traditional decision tree models. Therefore, compared to traditional asymmetric tree structures (such as XGBoost), the inference speed is improved by approximately 8-10 times, enabling the model to achieve millisecond-level response even in low-computational environments (such as web-based prediction terminals), meeting the needs of real-time prediction. In performance validation, the model using forgetting trees not only maintained a high level of accuracy (Accuracy=0.7006), but also demonstrated significantly better identification of minority class samples than traditional models such as random forests (AUC=0.773) and decision trees (AUC=0.7293) through an improved AUC value (AUC=0.7791). This indicates that forgetting trees can more effectively capture patterns related to hidden obesity (MONW) in normal-weight elderly individuals, improving diagnostic accuracy.

[0080] After optimization, a hidden obesity prediction model was finally obtained, which can be used to predict the hidden obesity symptoms of target users.

[0081] Example 2 Furthermore, based on Embodiment 1, this application provides a method for predicting hidden obesity, the method comprising the following steps: S21. Obtain the minimum clinical feature set of the target object, wherein the data in the minimum clinical feature set are all non-invasive feature data; S22. Input the clinical minimum feature set into the hidden obesity prediction model described in Example 1 for prediction to obtain the probability of hidden obesity in the target object.

[0082] In this embodiment, the core of the hidden obesity prediction method lies in how to efficiently and non-invasively identify whether a target subject has hidden obesity (MONW) through clinical data. To achieve this goal, this invention designs a prediction method based on a minimal clinical feature set, aiming to accurately predict hidden obesity using available and non-invasive feature data.

[0083] First, in this method, the minimum clinical feature set of the target subject is obtained by extracting non-invasive feature data. The minimum clinical feature set refers to the minimal and necessary set of data required for predicting hidden obesity, including non-invasive features reflecting the target subject's physiological, functional, and health status. These features include, but are not limited to, body shape and fat distribution, muscle strength, hemodynamic parameters, medical history, physical function, and cognitive abilities. By collecting this non-invasive physiological and health data, sufficient information can be obtained to assess whether the target subject has hidden obesity without invasive examinations.

[0084] After obtaining the minimum clinical feature set, this data will be input into a hidden obesity prediction model for analysis. This model is constructed using the hidden obesity prediction algorithm described in the first embodiment of the present invention. Through the input feature data, the model assesses the hidden obesity risk of the target individual. Specifically, the model analyzes various non-invasive features in the data, combining its built-in algorithm and model training results to calculate whether the target individual has a potential risk of hidden obesity.

[0085] Through this process, the model generates a probability of hidden obesity for the target individual, representing the likelihood that the individual has hidden obesity. This probability value can be used not only for preliminary clinical diagnosis but also as a reference for subsequent, more detailed examinations. If the probability is high, further medical testing or intervention may be necessary for a more accurate assessment and treatment of the target individual's health condition.

[0086] The advantage of this predictive method lies in its non-invasive nature. Preliminary predictions of hidden obesity can be made using simple physiological characteristic data, eliminating the need for complex biochemical or imaging examinations. This significantly reduces the cost and time of prediction, while also avoiding the discomfort and risks associated with invasive procedures. Furthermore, the design of a minimal clinical feature set ensures that the prediction process covers key factors influencing hidden obesity, while avoiding information redundancy caused by excessive data, making the prediction process more efficient.

[0087] In summary, the method for predicting hidden obesity of the present invention can accurately and quickly assess the risk of hidden obesity in target subjects by collecting a minimal set of clinical features and inputting it into a hidden obesity prediction model, thereby achieving early detection and intervention. It has significant application value and clinical significance.

[0088] Example 3 A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for constructing a hidden obesity prediction model as described in Embodiment 1, or the method for predicting hidden obesity as described in Embodiment 2.

[0089] Example 4 In one embodiment, a computer device is provided, which may be a server. The computer device includes a processor, memory, a network interface, and a database connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. When the computer program is executed by the processor, it implements a method for constructing a hidden obesity prediction model, or a method for predicting hidden obesity.

[0090] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps: S11. Obtain a clinical dataset of middle-aged and elderly people, which includes several samples, each of which includes several non-invasive features and their corresponding feature values; add labels to each sample in the clinical dataset of middle-aged and elderly people according to a predetermined classification standard, wherein the labels are: patients with hidden obesity or healthy people; S12. Calculate the positive sample weight factor, and use the positive sample weight factor to optimize the original loss function of each preset machine learning model to obtain the cost-sensitive loss function corresponding to each machine learning model. S13. The middle-aged and elderly clinical dataset is cleaned, and the non-invasive features after data cleaning are screened based on a three-stage feature optimization cascade algorithm to obtain a clinical minimum feature set. The three-stage feature optimization cascade algorithm includes: calculating the variance inflation factor to remove highly collinear features, using the Boruta algorithm to remove low-correlation features, and using the RFE algorithm to remove low-importance features. The RFE algorithm specifically uses the cost-sensitive loss function or the original loss function to train each preset machine learning model. S14. Evaluate each trained machine learning model using a pre-set independent validation set, select the model with the best evaluation performance for optimization, and obtain the hidden obesity prediction model.

[0091] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps: S21. Obtain the minimum clinical feature set of the target object, wherein the data in the minimum clinical feature set are all non-invasive feature data; S22. Input the clinical minimum feature set into the hidden obesity prediction model as described in Example 1 for prediction to obtain the probability of hidden obesity in the target object.

[0092] Example 5 To address the challenge of understanding CatBoost ensemble models as black boxes, this invention introduces the SHapley Additive exPlanations (SHAP) framework. Specifically, it employs the TreeExplainer algorithm, optimized for tree models. Based on cooperative game theory, this algorithm treats the model's risk of MONW as a total payoff generated by the cooperation of all features, calculating the marginal contribution of each feature, such as BRI and pulse pressure, to the prediction result. This not only determines whether a feature increases or decreases risk but also captures the nonlinear interactions between features; for example, the contribution of BMI to risk varies across different age groups.

[0093] (1) Core Algorithm Formula: This invention achieves interpretable decomposition of risk through the following two core mathematical models: 1) Shapley Value Calculation Formula: For input samples , No. Features (e.g.) SHAP value Defined as the weighted average of the marginal contributions of the feature when added to all possible combinations of feature subsets: ; in: These are the 11 full feature sets selected by this invention; For features not included Any subset of features; For the model to know only a subset Conditional expected prediction value under the condition of feature information .

[0094] 2) Formula for additive attribution model based on 11-dimensional features: Based on the additive attributes of SHAP, the MONW prediction probability output by the CatBoost model. The log-odds ratio (Log-Odds) is precisely decomposed into the sum of the baseline value and the contributions of 11 key features. For the specific model of this invention, this formula is expressed as: ; in: This represents the average prediction baseline of the model across the entire dataset; to Each characteristic represents an independent contribution to the individual's disease risk (positive values ​​represent increased risk, and negative values ​​represent decreased risk).

[0095] (2) Implementation effect: By implementing the above mathematical model, the present invention achieves the following progressive technical effects: 1) such as Figure 6 , Figure 7 As shown, the visualization verification of the global pathological mechanism is performed using the SHAP global importance ranking chart ( Figure 6 ) and global feature influence distribution map ( Figure 7 The system reveals the pathogenic patterns of the MONW population from a macroscopic perspective. Experimental data show that BRI (Body Roundness Index), Pulse Pressure, and ABSI (Body Shape Index) have the greatest impact on the model output globally and show a significant positive correlation (i.e., the higher the value, the higher the risk). This mathematically and statistically verifies that central obesity and arteriosclerosis are the core pathological characteristics of this population, establishing the credibility of the model.

[0096] 2) such as Figure 9 The diagram illustrates the precise attribution of individual risk: for specific subjects, the system utilizes the SHAP attribution analysis graph to achieve precise profiling at the single-sample level. Figure 9 As shown, although the patient (female, 50 years old) did not exhibit obvious conventional obesity, her BRI=4.89 produced a positive SHAP contribution of +1.11, which was the dominant factor driving up the risk; at the same time, the chart shows that her pulse pressure of 45 produced a negative contribution of -0.381, which played a certain protective role.

[0097] This visual addition and subtraction logic allows doctors to directly explain to patients: "Although your blood pressure is well controlled (green bar), your body roundness index is too high (blue-green bar), which is the root cause of your assessment as high risk (69.0%)", thus enabling the development of targeted fat loss intervention plans.

[0098] 3) Establishing human-machine trust: The web terminal based on the Shiny framework transforms the complex Log-Odds logarithmic probability formula into an intuitive chart, realizing the leap from black-box prediction to white-box interpretation, which greatly improves the acceptance of AI prediction results by primary care physicians and elderly patients.

[0099] Web interactive terminal implementation Furthermore, the web terminal of this invention is designed based on a B / S (Browser / Server) architecture and is built using the R Shiny framework and the bslib modern theme library. The system backend logic layer contains three core processing units: a dynamic feature mapping unit, a risk stratification and quantification unit, and an interpretable waterfall visualization unit.

[0100] The dynamic feature mapping unit is used by the front-end interface to receive raw clinical parameters (height, weight, waist circumference, blood pressure, etc.) input by the user. Before entering the prediction model, the server first performs real-time feature engineering, mapping the low-dimensional raw data into high-dimensional model feature vectors. The system incorporates non-linear calculation formulas optimized for middle-aged and elderly users: Body Roundness Index (BRI) calculation: ; Body Shape Index (ABSI) calculation: ; Relative Grip Calculation: ; This unit ensures that the feature space input to the CatBoost model is completely consistent with that of the training phase, thereby guaranteeing the effectiveness of the prediction.

[0101] The risk stratification unit is used to call the model output MONW disease probability in the inference engine. Subsequently, instead of directly outputting numerical values, the system constructs a dynamic dashboard using the Echarts rendering engine. Simultaneously, the system incorporates a three-level risk warning logic based on preset threshold vectors. Automatic stratification of subjects: Low Risk The system identifies this as a green zone and suggests maintaining your current healthy lifestyle.

[0102] Medium Risk The system identified this area as a yellow zone, indicating a potential high-risk group for MONW (Montenegrown Obesity). The system suggested that the individual pay attention to abdominal obesity indicators and increase exercise appropriately.

[0103] High-risk tier The system identifies this as a red zone and outputs the following recommendation: a detailed medical examination and lifestyle intervention are recommended.

[0104] The interpretable waterfall visualization unit is used to visually demonstrate why a patient is classified as high-risk. This invention uses a SHAP waterfall chart for attribution visualization.

[0105] 1) Benchmark Anchoring: The chart uses the model's average risk value across the entire population. Starting point.

[0106] 2) Incremental accumulation: Accumulate the SHAP value of each feature. The data is rendered as cumulative bars. Blue bars (positive contribution) indicate that the indicator (such as a high BRI) increases the risk of disease; green bars (negative contribution) indicate that the indicator (such as not drinking alcohol) reduces the risk of disease.

[0107] 3) Final closed loop: The algebraic sum of all blocks ultimately points to the actual predicted probability of the current patient. .

[0108] Through this visualization method, clinicians can clearly see that although the elderly person's BMI is normal, their BRI (body roundness) and relative grip strength SHAP contribution values ​​are significantly positive, thus accurately pinpointing "sarcopenic obesity" as a hidden cause.

[0109] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0110] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0111] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for constructing a hidden obesity prediction model, characterized in that, Includes the following steps: Obtain a clinical dataset of middle-aged and elderly patients, which includes several samples, each of which includes several non-invasive features and their corresponding feature values; According to the predetermined classification criteria, labels are added to each sample in the middle-aged and elderly clinical dataset, and the labels are: hidden obesity patients or healthy people; Calculate the positive sample weight factor, and use the positive sample weight factor to optimize the original loss function of each preset machine learning model to obtain the cost-sensitive loss function corresponding to each machine learning model. The aforementioned clinical dataset of middle-aged and elderly individuals is cleaned, and the non-invasive features after data cleaning are screened based on a three-stage feature optimization cascade algorithm to obtain a minimal clinical feature set. The three-stage feature optimization cascade algorithm includes: calculating the variance inflation factor to remove highly collinear features, using the Boruta algorithm to remove low-relevance features, and using the RFE algorithm to remove low-importance features. Specifically, the RFE algorithm uses the cost-sensitive loss function or the original loss function to train various preset machine learning models. The trained machine learning models are evaluated using a pre-defined independent validation set. The model with the best evaluation performance is selected for optimization to obtain a hidden obesity prediction model.

2. The method for constructing a hidden obesity prediction model according to claim 1, characterized in that, The model with the best performance in the evaluation is specifically the CatBoost model.

3. The method for constructing a hidden obesity prediction model according to claim 2, characterized in that, The cost-sensitive loss function of the CatBoost model is as follows: ； in, This represents the loss function value of the model; Indicates the number of samples; Indicates the weighting factor for positive samples; Indicates the first The true labels of each sample; Indicates the first The probability of the predicted value for each sample; The specific positive sample weighting factor is: ; in, Indicates the weighting factor for positive samples; Indicates the number of positive samples; This indicates the number of negative samples.

4. The method for constructing a hidden obesity prediction model according to claim 1, characterized in that, The minimum clinical feature set includes the following non-invasive features: BRI, ABSI, relative grip strength, PEF, pulse pressure, pulse, age, sex, alcohol consumption, self-reported health, and history of heart disease.

5. The method for constructing a hidden obesity prediction model according to claim 1, characterized in that, The calculation of the variance inflation factor to remove highly collinear features specifically includes: Calculate the first The linear regression equation between one non-invasive feature and the remaining non-invasive features; Calculate the first based on the linear regression equation. The coefficient of determination for each non-invasive feature ; Based on the determination coefficient Calculate the first The variance inflation factors corresponding to each non-invasive feature include: ; Based on the variance inflation factor corresponding to each non-invasive feature, an iterative elimination strategy is executed to generate a low collinearity feature set.

6. The method for constructing a hidden obesity prediction model according to claim 5, characterized in that, The specific steps of using the Boruta algorithm to remove low-relevance features include: Based on each non-invasive feature in the aforementioned low collinearity feature set, corresponding shadow features are constructed respectively; Calculate the importance score for each shadow feature, and determine the most important shadow feature based on the importance score. The score of this most important shadow feature is denoted as [missing value]. ; Calculate the importance score for each noninvasive feature in the low collinearity feature set, denoted as . ; Compare and ,like Greater than And if the preset significance condition is met, then it is determined that... The corresponding non-invasive features are significantly correlated and are retained; otherwise, the condition is determined to be non-invasive. The corresponding non-invasive features that are not significantly correlated are removed to obtain a set of highly correlated features.

7. The method for constructing a hidden obesity prediction model according to claim 6, characterized in that, The removal of low-importance features using the RFE algorithm specifically includes: Using all features in the highly relevant feature set, each preset machine learning model is trained. The training specifically employs several rounds of 10-fold cross-validation. In each round, the contribution of each feature value to the model is evaluated, and the feature with the lowest contribution is removed to obtain the minimum clinical feature set.

8. The method for constructing a hidden obesity prediction model according to claim 2 or 3, characterized in that, The process of selecting and optimizing the model with the best performance specifically includes: A ranking boosting mechanism is used to process categorical features. Each categorical feature is encoded by introducing random permutation. The encoded value is calculated based on the order of the samples and the prior information of the target variable. A smoothing parameter is used to smooth the encoded value of the categorical features to reduce prediction bias. In the construction of the base model, a forgetting tree is used as the base learner. Each forgetting tree uses the same splitting feature and splitting threshold at each layer node. The prediction mechanism of the forgetting tree includes using the splitting feature index and splitting threshold to perform binary judgment of the sample, dividing the sample into the corresponding leaf node, and accumulating the prediction results of each base learner through an additive model. Finally, the prediction result is mapped to the Sigmoid function to obtain the probability of hidden obesity.

9. A method for predicting hidden obesity, characterized in that, The method includes the following steps: Obtain the minimum clinical feature set of the target object, wherein all data in the minimum clinical feature set are non-invasive feature data; The clinical minimum feature set is input into the hidden obesity prediction model as described in any one of claims 1-8 to make a prediction and obtain the probability of hidden obesity in the target subject.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method for constructing a hidden obesity prediction model as described in any one of claims 1-8, or the method for predicting hidden obesity as described in claim 9.

Images