Method and device for constructing female osteoporosis prediction model based on physical examination data

Abstract

The invention provides a method and a device for constructing a female osteoporosis prediction model based on physical examination data. The construction method comprises the steps of obtaining a potential prediction variable set of female osteoporosis based on literature query; screening a sample set from the physical examination data; dividing the sample set into a training set and a verification set, and based on the preprocessed first training set, obtaining an intersection of variables screened by using a Boruta algorithm, an LASSO algorithm, a varSelRF algorithm and an SVM-RFE algorithm, and taking the intersection as a feature variable set of the osteoporosis prediction model; obtaining a second training set and optimizing the plurality of constructed initial prediction models by using the second training set to obtain a plurality of parameter tuning prediction models; and obtaining a test set, evaluating each parameter tuning prediction model by using the test set, and determining a female osteoporosis prediction model based on an evaluation result. The female osteoporosis can be predicted on the basis of eight conventional physical examination data, and the method has the advantages of high accuracy and medium sensitivity.

Description

Technical Field

[0001] The present invention belongs to the technical field of osteoporosis prediction, and in particular relates to a method and device for constructing a female osteoporosis prediction model based on physical examination data. Background Art

[0002] Hip fractures are associated with decreased mobility, morbidity, mortality and high medical costs, and approximately 90% of hip fractures in the elderly are related to osteoporosis. Therefore, screening and early intervention for hip osteoporosis are particularly important in the population.

[0003] Dual-Energy X-ray Absorptiometry (DXA) is the gold standard method for measuring bone mineral density (BMD) and subsequently diagnosing osteoporosis. However, the use of DXA is limited. DXA requires a room with infrastructure capable of handling the equipment, which is large, expensive, fixed, and must be operated by qualified technicians. Therefore, it may not be suitable for widespread osteoporosis screening, especially in large rural areas.

[0004] Koh et al. developed the Osteoporosis Self-assessment Tool for Asians (OSTA) to screen high-risk patients based on the individual's age and weight. However, the area under the curve (AUC) of OSTA was between 0.65 and 0.70, which limited the use of the tool due to its limited accuracy. Most studies using machine learning (ML) to diagnose osteoporosis used imaging examinations - x-rays, CT scans, or ultrasound, which are not suitable for grassroots and community screening due to the required equipment and instruments. A few studies predicted osteoporosis based on clinical data, but they all had problems; for example, some models showed significant overfitting, the AUC was close to perfect (AUC ≥ 0.99), some studies had small sample sizes, and some lacked external validation of the model. Suh et al. reported an interpretable deep learning (DL) model that used clinical features to screen for osteoporosis risk. However, most of the variables in the model were based on questions about medical history, which may be affected by recall bias. Summary of the invention

[0005] The purpose of the present invention is to provide a prediction model and construction method for predicting osteoporosis based on age, BMI, systolic blood pressure, red blood cell count, glycosylated hemoglobin, alanine aminotransferase, aspartate aminotransferase and uric acid in physical examination data, which has the advantages of convenient acquisition of prediction variables, high prediction accuracy and medium prediction sensitivity.

[0006] To achieve the above-mentioned object, the present invention provides a method for constructing a female osteoporosis prediction model based on physical examination data, the construction method comprising the following steps: step S10, based on literature query, obtaining a potential prediction variable set for female osteoporosis, the potential prediction variable set comprising a plurality of demographic variables, a plurality of physical examination variables, a plurality of laboratory examination variables, a body mass index variable and an estimated glomerular filtration rate variable, wherein the plurality of demographic variables comprises age, gender, height and weight; the plurality of physical examination variables comprises: pulse, systolic blood pressure and diastolic blood pressure; the plurality of laboratory examination variables comprises: white blood cell count, hematocrit, red blood cell count, lymphocyte percentage, mean corpuscular hemoglobin concentration, mean corpuscular volume The present invention relates to a method for screening the blood samples of the present invention, comprising: screening the blood samples of the present invention with respect to the blood, the blood volume, the percentage of monocytes, the mean platelet volume, the percentage of segmented granulocytes, the platelet count, the neutrophil count, the lymphocyte count, the monocyte count, the hemoglobin, the number of eosinophils, the percentage of eosinophils, the fasting blood glucose, the glycosylated hemoglobin, the high-density lipoprotein cholesterol, the low-density lipoprotein cholesterol, the triglycerides, the total cholesterol, the free triiodothyronine, the free thyroxine, the thyroid stimulating hormone, the 25-hydroxyvitamin D, the parathyroid hormone, the creatinine, the blood urea nitrogen, the alkaline phosphatase, the alanine aminotransferase, the aspartate aminotransferase, the serum albumin, the uric acid, the calcium, the phosphorus, the magnesium, the sodium, the potassium and the chloride; and step S20, screening the sample set from the physical examination data according to the inclusion criteria and the exclusion criteria, wherein the sample set is screened from the physical examination data according to the inclusion criteria and the exclusion criteria. The samples in the sample set are women aged ≥50 years or postmenopausal and have bone density data; step S30, dividing the sample set into a training set and a validation set, and obtaining a first training set based on the training set, wherein the first training set includes data corresponding to the potential predictive variable set in the training set and bone density data; step S40, preprocessing the first training set, and based on the preprocessed first training set, obtaining an intersection of variables screened by using the Boruta algorithm, the LASSO algorithm, the varSelRF algorithm, and the SVM-RFE algorithm, respectively, and using the intersection as a feature variable set of an osteoporosis prediction model, wherein the feature variable set includes 8 feature variables, namely age, BMI, systolic blood pressure, red blood cell count, number, glycosylated hemoglobin, alanine aminotransferase, aspartate aminotransferase and uric acid; step S50, based on the feature variable set and the training set, a second training set is obtained, and the plurality of initial prediction models constructed are optimized by using the second training set to obtain a plurality of parameter-tuned prediction models, wherein the second training set includes data corresponding to the feature variable set in the training set, and bone density data; step S60, based on the feature variable set and the validation set, a test set is obtained, and each parameter-tuned prediction model is evaluated by using the test set, and a female osteoporosis prediction model is determined in the plurality of parameter-tuned prediction models based on the evaluation result, wherein the test set includes data corresponding to the feature variable set in the validation set, and bone density data.

[0007] In a specific embodiment, the method further includes evaluating each parameter-tuned prediction model using Nhanes data, and verifying the accuracy of the evaluation results of each parameter-tuned prediction model based on the test set using the evaluation results of each parameter-tuned prediction model based on the Nhanes data.

[0008] In a specific embodiment, in step S20, the inclusion criteria are women with bone density data; the exclusion criteria include: (1) samples of the second or more bone density tests, (2) women under 50 years old or who have not yet menopausal, (3) samples lacking laboratory test data, and (4) sample groups with a large number of missing variables in one of the multiple variables in the potential predictive variable set.

[0009] In a specific implementation, in step S40, the preprocessing includes removing zero variance variables and interpolating missing data, and the missing data is interpolated using a bagged tree method.

[0010] In a specific embodiment, in step S50, the decision tree, k-nearest neighbor, LightGBM, random forest, extreme gradient boosting, support vector machine and multilayer perceptron in the multiple initial prediction models try all possible hyperparameter combinations through grid search, and use 5-fold cross validation to evaluate the performance of each hyperparameter combination to determine the best hyperparameter combination.

[0011] In a specific implementation, in step S50, the decision tree, k-nearest neighbor, LightGBM, random forest, extreme gradient boosting, support vector machine and multi-layer perceptron in the multiple initial prediction models are optimized using five-fold cross validation to optimize the hyperparameters of the model.

[0012] In a specific embodiment, in step S60, the female osteoporosis prediction model determined based on the evaluation result is an extreme gradient boosting model.

[0013] In a specific embodiment, the evaluation result includes multiple groups of model evaluation indicators corresponding to each parameter-tuned prediction model, each group of model evaluation indicators includes the area under the ROC curve AUC and sensitivity; the use of the test set to evaluate each parameter-tuned prediction model, and determining the female osteoporosis prediction model in multiple parameter-tuned prediction models based on the model evaluation indicators includes: step (1), based on the test set, obtaining the probability value predicted by each parameter-tuned prediction model for each sample in the test set; step (2), according to the probability value predicted by each parameter-tuned prediction model in the test set, establishing multiple test set operating characteristic curves ROC; step (3), calculating the area under the ROC curve AUC and sensitivity of each parameter-tuned prediction model, and taking the prediction model with the area under the ROC curve AUC greater than or equal to the AUC preset value as a candidate model; step (4), when the number of the candidate models is one, taking the candidate model as the female osteoporosis prediction model, and when the number of the candidate models is greater than one, taking the candidate model with the highest sensitivity as the female osteoporosis prediction model.

[0014] In a specific embodiment, the female osteoporosis prediction model is a female hip osteoporosis prediction model.

[0015] The present invention also provides a device for constructing a female osteoporosis prediction model based on physical examination data, the construction device comprising a sample set screening module, a first acquisition module, a second acquisition module, a characteristic variable screening module, a model optimization module and a model determination module. The sample set screening module is used to screen a sample set from the physical examination data according to the inclusion criteria and the exclusion criteria, and the samples in the sample set are women aged ≥50 years or postmenopausal and have bone density data. The first acquisition module is used to obtain a potential predictive variable set for osteoporosis based on literature, and the potential predictive variable set includes multiple demographic variables, multiple physical examination variables, multiple laboratory examination variables, body mass index variables and estimated glomerular filtration rate variables, wherein the multiple demographic variables include age, gender, height and weight; the multiple physical examination variables include: pulse, systolic blood pressure and diastolic blood pressure; the multiple laboratory examination variables include: white blood cell count, hematocrit, red blood cell count, lymphocyte percentage, mean corpuscular hemoglobin concentration, mean corpuscular volume, monocyte percentage The second acquisition module is used to divide the sample set into a training set and a validation set, and to acquire a first training set based on the training set, wherein the first training set includes data corresponding to the prediction variable set in the training set, and bone density data. The characteristic variable screening module is used to preprocess the first training set, and based on the preprocessed first training set, obtain the intersection of variables screened by Boruta algorithm, LASSO algorithm, varSelRF algorithm and SVM-RFE algorithm respectively, and use the intersection as the characteristic variable set of the osteoporosis prediction model, the characteristic variable set includes 8 characteristic variables, namely age, BMI, systolic blood pressure, red blood cell count, glycosylated hemoglobin, alanine aminotransferase, aspartate aminotransferase and uric acid. The model optimization module is used to obtain a second training set based on the characteristic variable set and the training set, and optimize the constructed multiple initial prediction models using the second training set to obtain multiple parameter-tuned prediction models, wherein the second training set includes data corresponding to the characteristic variable set in the training set, and bone density data.The model determination module is used to obtain a test set based on a feature variable set and a validation set, and use the test set to evaluate each parameter-tuned prediction model, and determine a female osteoporosis prediction model from multiple parameter-tuned prediction models based on model evaluation indicators, wherein the test set includes data in the validation set corresponding to the feature variable set and bone density data.

[0016] The beneficial effects of the present invention include at least:

[0017] 1. The present invention provides a method for constructing a female osteoporosis prediction model based on physical examination data, the construction method comprising the following steps: step S10, based on literature query, obtaining a potential prediction variable set for female osteoporosis, the potential prediction variable set comprising multiple demographic variables, multiple physical examination variables, multiple laboratory examination variables, a body mass index variable and an estimated glomerular filtration rate variable; step S20, screening a sample set from the physical examination data according to inclusion criteria and exclusion criteria, the samples in the sample set being females aged ≥50 years or postmenopausal and having bone density data; step S30, dividing the sample set into a training set and a validation set, and obtaining a first training set based on the training set; step S40, preprocessing the first training set, and based on the preprocessed first training set, obtaining an intersection of variables screened respectively using the Boruta algorithm, the LASSO algorithm, the varSelRF algorithm and the SVM-RFE algorithm, and using the intersection as a feature variable set for the osteoporosis prediction model, the feature variable set comprising 8 feature The variables are age, BMI, systolic blood pressure, red blood cell count, glycated hemoglobin, alanine aminotransferase, aspartate aminotransferase and uric acid; step S50, based on the feature variable set and the training set, a second training set is obtained, and the second training set is used to optimize the constructed multiple initial prediction models to obtain multiple parameter-tuned prediction models; step S60, based on the feature variable set and the validation set, a test set is obtained, and the test set is used to evaluate each parameter-tuned prediction model, and based on the evaluation result, a female osteoporosis prediction model is determined in the multiple parameter-tuned prediction models; the prediction model constructed by the construction method provided by the present invention is an XGBoost model, and the feature variables corresponding to the model are age, BMI, systolic blood pressure, red blood cell count, glycated hemoglobin, alanine aminotransferase, aspartate aminotransferase and uric acid. The data of the above 8 feature variables can be easily collected and applied in most medical institutions, and are suitable for use in physical examination institutions, community institutions and rural areas, and do not require additional labor costs, and have good application prospects.

[0018] 2. The female osteoporosis prediction model (XGBoost model) constructed by the construction method of the present invention has an area under the ROC curve AUC value greater than 0.8 and a sensitivity greater than 0.7 in the test set and the external validation set, indicating that the model has high accuracy and moderate sensitivity when used to screen osteoporosis patients.

[0019] 3. The female osteoporosis prediction model (XGBoost model) constructed by the present invention is displayed in a web app, and a web application is constructed. The test subjects can obtain the pre-test results based on the web application, and decide whether to give priority to bone density examination according to the prediction results; this helps to screen people at high risk of osteoporosis and further examine bone density to achieve early detection and diagnosis of osteoporosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A flowchart of the steps of the method for constructing a female osteoporosis prediction model based on physical examination data provided by the present invention;

[0021] Figure 2 A flow chart of constructing a female osteoporosis prediction model for the present invention;

[0022] Figure 3 The ROC curves corresponding to the prediction models Logistic, DT, KNN, LightGBM, RF, XGBoost, SVM and MLP established by the present invention based on the test set respectively;

[0023] Figure 4 The ROC curves corresponding to the prediction models Logistic, DT, KNN, LightGBM, RF, XGBoost, SVM and MLP established by the present invention based on the external validation set respectively;

[0024] Figure 5 A screenshot of a web page corresponding to input data when making a prediction based on the female osteoporosis prediction model provided by the present invention;

[0025] Figure 6 Screenshot of the webpage corresponding to the output result of the female osteoporosis prediction model provided by the present invention based on the input data;

[0026] Figure 7 This is a module diagram of a device for constructing a female osteoporosis prediction model based on physical examination data provided by the present invention. DETAILED DESCRIPTION

[0027] The present invention is described in detail below with reference to the accompanying drawings and embodiments, but the present invention can be implemented in many different ways as limited and covered by the claims.

[0028] In view of the shortcomings of inconvenient and high cost of obtaining bone density data, the present invention provides a model for predicting female osteoporosis without the need for bone density data. The prediction model can predict female hip osteoporosis based on age, BMI, systolic blood pressure, red blood cell count, glycosylated hemoglobin, alanine aminotransferase, aspartate aminotransferase and uric acid in physical examination data, and has high prediction accuracy (AUC greater than 0.8) and medium sensitivity (sensitivity 0.7). It can be used in physical examination institutions and community institutions without increasing additional manpower costs; at the same time, it can also be applied to individuals with a previous medical history or physical examination records, as long as these 8 variable data are available, which is very efficient.

[0029] See also Figure 1 The present invention provides a method for constructing a female osteoporosis prediction model based on physical examination data, the construction method comprising the following steps:

[0030] Step S10: Based on literature query, a potential predictive variable set for osteoporosis in women is obtained, wherein the potential predictive variable set includes multiple demographic variables, multiple physical examination variables, multiple laboratory examination variables, a body mass index variable, and an estimated glomerular filtration rate variable, wherein the multiple demographic variables include age, gender, height, and weight; the multiple physical examination variables include: pulse, systolic blood pressure, and diastolic blood pressure; the multiple laboratory examination variables include: white blood cell count, hematocrit, red blood cell count, lymphocyte percentage, mean corpuscular hemoglobin concentration, mean corpuscular volume, monocyte percentage ratio, mean platelet volume, percentage of segmented granulocytes, platelet count, neutrophil count, lymphocyte count, monocyte count, hemoglobin, eosinophil count, eosinophil percentage, fasting blood glucose, glycosylated hemoglobin, high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, triglycerides, total cholesterol, free triiodothyronine, free thyroxine, thyroid-stimulating hormone, 25-hydroxyvitamin D, parathyroid hormone, creatinine, blood urea nitrogen, alkaline phosphatase, alanine aminotransferase, aspartate aminotransferase, serum albumin, uric acid, calcium, phosphorus, magnesium, sodium, potassium, and chloride.

[0031] Among them, body mass index (BMI) (kg / m2) is calculated by dividing weight (kg) by height squared (m2).

[0032] Estimated glomerular filtration rate (eGFR) was calculated from serum creatinine using the CKD-EPI equation.

[0033] In the present invention, the estimated glomerular filtration rate is calculated using a formula provided by the prior art.

[0034] In the present invention, the number of demographic variables is 3, the number of physical examination variables is 3, and the number of laboratory examination variables is 40; the potential predictor variable set includes a total of 48 candidate variables.

[0035] Step S20: screening a sample set from the physical examination data according to the inclusion criteria and exclusion criteria, wherein each sample in the sample set is a female aged ≥50 years or postmenopausal and has bone density data.

[0036] Preferably, the inclusion criterion is women with bone density data.

[0037] Preferably, the exclusion criteria include: (1) samples that have undergone the second or more bone density tests, (2) women who are under 50 years old or have not yet reached menopause; (3) samples that lack laboratory test data, and (4) sample groups that have a relatively large number of missing variables in one of the multiple variables in the potential predictive variable set.

[0038] It should be noted that point (1) of the exclusion criteria can be understood as: when a sample has undergone multiple physical examinations and corresponds to multiple physical examination data, the second and subsequent physical examination data are deleted, and only the sample physical examination data corresponding to the first bone density test is retained.

[0039] The physical examination data may be physical examination data from a hospital or a physical examination center.

[0040] In the present invention, the physical examination data is the physical examination data of Xiangya Second Hospital, and the sample set is specifically: women who underwent bone density testing between September 2013 and November 2023, and the following samples are excluded: (1) second or more bone density tests; (2) under 50 years old or not yet menopausal; (3) samples without laboratory test data; (4) samples from 2017 to 2019 and 2013 to 2015 (because of the high proportion of vitamin D and electrolyte deficiencies).

[0041] Step S30: divide the sample set into a training set and a validation set, and obtain a first training set based on the training set, wherein the first training set includes data in the training set corresponding to the prediction variable set and bone density data.

[0042] In the present invention, the first training set includes multiple first data groups corresponding one by one to all samples in the training set, and each first data group includes the name of each candidate variable in the potential predictive variable set, the data corresponding to each candidate variable, the name of bone density, and the data corresponding to bone density.

[0043] It should be noted that the candidate variables are the age, gender, height and weight variables mentioned above, totaling 48.

[0044] In the present invention, the ratio of the training set to the validation set is 3:1, that is, the total number of samples in the training set is 3 / 4 of the total number of samples in the sample set, and the total number of samples in the validation set is 1 / 4 of the total number of samples in the sample set.

[0045] For ease of understanding, for example, assuming that the total number of samples in the sample set is 8039, then the number of samples in the training set is 6029, and the number of samples in the validation set is 2010. The first data set is composed of data corresponding to all variables of the 6029 samples in the training set, and all variables are the multiple demographic variables, multiple physical examination variables, multiple laboratory examination variables, body mass index variables and estimated glomerular filtration rate variables mentioned above.

[0046] Please refer to Figure 2 In the present invention, the total number of samples in the sample set is 8039.

[0047] Step S40, preprocessing the first training set, and based on the preprocessed first training set, obtaining the intersection of variables screened by the Boruta algorithm, LASSO algorithm, varSelRF algorithm and SVM-RFE algorithm respectively, and using the intersection as the characteristic variable set of the osteoporosis prediction model, the characteristic variable set includes 8 characteristic variables, namely age, BMI, systolic blood pressure, red blood cell count, glycated hemoglobin, alanine aminotransferase, aspartate aminotransferase and uric acid.

[0048] In this step, the preprocessing includes removing zero-variance variables and interpolating missing data.

[0049] Specifically, when the data of a certain variable corresponding to each sample are the same, the calculated variance is 0, then the data of the variable corresponding to each sample are all deleted. In the present invention, no zero variance variable is found; when the data of a certain variable corresponding to a certain sample is empty, the missing data is interpolated.

[0050] Preferably, in the present invention, the bagged tree method is used to interpolate the missing data.

[0051] Bagging is a method of interpolation that uses multiple decision trees to predict missing values. This method constructs multiple decision trees, each of which is trained independently on the data, and then the final result is obtained by voting or taking the average.

[0052] The bagging method includes: for numerical predictor variables containing missing values, 50 sub-samples with replacement are generated from the training data set; for each sub-sample, an independent decision tree model is trained; for each sample with missing values, all trained decision trees are used for prediction, and the prediction results of each tree are summarized; according to the summary rules (for numerical variables, the average value is used for interpolation, and missing values ​​are filled on categorical variables by majority voting), the final interpolation result is determined to replace the missing values ​​in the original data.

[0053] In the present invention, the variables obtained by the Boruta algorithm, the LASSO algorithm, the varSelRF algorithm and the SVM-RFE algorithm are shown in Table 1.

[0054] Table 1 Variables obtained by different algorithms

[0055]

[0056] It should be noted that the present invention does not improve the Boruta algorithm, LASSO algorithm, varSelRF algorithm and SVM-RFE algorithm.

[0057] The description of the four algorithms is as follows:

[0058] Boruta algorithm: First, create a corresponding shadow feature for each original feature. The shadow feature is trained by randomly shuffling the value of the original feature and taking the original feature and the shadow feature as input, and the target variable as output to train a random forest model. After the training is completed, the feature importance is determined by calculating the average accuracy of each feature (including the original feature and the shadow feature) in the random forest and dividing it by its standard deviation, which is the Z score. That is, for each feature, when predicting the model, first record the accuracy of the model. Then, shuffle the value of the feature, predict again, and record the new accuracy. The more important the feature, the more the performance of the model will decrease after shuffling. The Z-Score of each original feature is compared with the maximum Z-Score among shadow attributes (MZSA). Those original features with a Z score significantly lower than the shadow feature MZSA will be considered unimportant variables and will be deleted from the data set (Rejected). If the original feature Z score is significantly higher than the shadow feature MZSA, then the original feature is marked as "important" (Confirmed). For features whose importance cannot be determined (tentative variables), the original features are marked as "tentative".

[0059] If there are still features in the "pending" state and the number of iterations has not exceeded 100, repeat all the above steps. Until all features are marked as "important", "unimportant" or 100 iterations are reached. In each iteration, the random forest model is retrained and the feature importance is recalculated. After multiple iterations, the important features finally determined are those marked as "important" (Confirmed). These features are considered to be important for the prediction of the target variable and are used for subsequent modeling and analysis.

[0060] Lasso algorithm: Generate 100 different regularization parameter Lambda values, divide the training data set into 10 parts, and take turns using 9 of them as training sets and 1 part as validation set. Repeat this 10 times to ensure that each part has a chance to be used as a validation set. In each cross-validation, use different Lambda values ​​to optimize and adjust the model. For each Lambda value, calculate the AUC value of the model on the validation set. In this way, the performance of the model under different Lambda values ​​can be obtained. Finally, by comparing the AUC values ​​of the model under different Lambda values, find the regularization parameter combination that enables the model to perform best under a given evaluation indicator (AUC), that is, the Lambda value that maximizes the AUC value.

[0061] varSelRF algorithm

[0062] 1) Eliminate irrelevant variables: The data was optimized using a random forest with 500 trees. When building each tree, 20 features were randomly selected at each node for splitting. During the training process, the average number of splits for each feature in all decision trees was calculated as an indicator of the importance of the feature. Based on the importance score, variables whose importance was less than half of the average importance score were removed.

[0063] 2) Select relevant explanatory variables: A random forest with 100 trees was constructed. Ten features were randomly selected at each node for splitting. The importance of each variable was evaluated again by comparing the importance score of each variable with the mean importance score plus 1 standard deviation. If its importance score was greater than the mean plus 1 standard deviation, it was retained.

[0064] 3) Use random forest with 100 trees and 10 features randomly selected at each node. Evaluate the correlation between variables. If the correlation between two variables is more than 1, then one of the variables is considered redundant and remove those variables that are considered redundant.

[0065] SVM-RFE algorithm

[0066] RFE starts with retaining 2 features and gradually increases to retaining all features. Use the current feature subset and the specified support vector machine and cross-validation method to train the model. Evaluate the performance of the current feature subset based on the evaluation indicator ROC. Gradually remove the least important features, then train the model again and evaluate the performance until the model performance no longer improves. In each iteration, record the current feature subset and the corresponding performance indicator. Compare the performance of different feature subsets, and finally determine the feature subset that maximizes the evaluation indicator, which is the best feature subset.

[0067] Step S50: Based on the feature variable set and the training set, a second training set is obtained, and the multiple initial prediction models constructed are optimized using the second training set to obtain multiple parameter-tuned prediction models, wherein the second training set includes data in the training set corresponding to the feature variable set and bone density data.

[0068] In the present invention, after the initial prediction model is optimized by the second training set, the obtained parameter-tuned prediction models all have the best prediction rate, and each parameter-tuned prediction model has the best parameter combination or the best hyperparameter combination.

[0069] In the present invention, the second training set includes multiple second data groups corresponding one by one to all samples in the training set, and each of the second data groups includes the name of each feature variable in the feature variable set, the data corresponding to each feature variable, the name of bone density, and the data corresponding to bone density.

[0070] In the present invention, each prediction model has a different algorithm.

[0071] Preferably, the number of the initial prediction models is 8, namely logistic regression (LR), decision tree (DT), k nearest neighbor (KNN), LightGBM, random forest (RF), extreme gradient boosting (XGBoost), support vector machine (SVM) and multi-layer perceptron (MLP).

[0072] Preferably, the decision tree (DT), k-nearest neighbor (KNN), LightGBM, random forest (RF), extreme gradient boosting (XGBoost), support vector machine (SVM) and multi-layer perceptron (MLP) in the multiple initial prediction models try all possible hyperparameter combinations through grid search, and use 5-fold cross validation to evaluate the performance of each combination to determine the optimal hyperparameter combination.

[0073] Preferably, the parameters of the logistic regression model in the multiple initial prediction models are determined based on the maximum likelihood method.

[0074] Step S60: Based on the feature variable set and the validation set, a test set is obtained, and each parameter-tuned prediction model is evaluated using the test set, and a female osteoporosis prediction model is determined in multiple parameter-tuned prediction models based on the evaluation results, wherein the test set includes data in the validation set corresponding to the feature variable set and bone density data.

[0075] In the present invention, the test set includes multiple third data groups corresponding one by one to all the samples in the validation set, and each of the third data groups includes the name of each feature variable in the feature variable set, the data corresponding to each feature variable, the name of the bone density, and the data corresponding to the bone density.

[0076] Preferably, the evaluation result includes multiple groups of model evaluation indicators corresponding to each parameter-tuned prediction model, and each group of model evaluation indicators includes the area under the ROC curve AUC and sensitivity.

[0077] Preferably, each group of model evaluation indicators may also include accuracy, specificity and Brier score.

[0078] Preferably, the female osteoporosis prediction model is a female hip osteoporosis prediction model.

[0079] Preferably, the female osteoporosis prediction model is suitable for women who are 50 years old or older or who have reached menopause.

[0080] The step of evaluating each parameter-optimized prediction model using the test set and determining a female osteoporosis prediction model from among multiple parameter-optimized prediction models based on the evaluation results includes:

[0081] Step (1): based on the test set, obtain the probability value predicted by each parameter-tuned prediction model.

[0082] Step (2), establishing multiple test set operating characteristic curves (ROC) according to the probability values ​​predicted by the prediction model with each parameter tuning.

[0083] In the present invention, the number of prediction models is 8, and accordingly, the number of ROC curves of the test set is 8, see Figure 3 .

[0084] Step (3), calculating the area under the ROC curve (AUC) and sensitivity of each parameter-tuned prediction model, and taking the prediction model whose area under the ROC curve (AUC) is greater than or equal to the AUC preset value as the candidate model.

[0085] Preferably, the AUC preset value is 0.8.

[0086] The model evaluation results of each parameter-tuned prediction model are detailed in Table 2.

[0087] Table 2 Model evaluation results of different prediction models

[0088] The name of the prediction model 95% CI AUC value Sensitivity Specificity Accuracy Cutoff value Brier score Logistic 0.755-0.810 0.782 66.9％ 76.2％ 0.748 0.159 0.107 DT 0.733-0.791 0.762 67.9％ 74.6％ 0.736 0.167 0.107 KNN 0.677-0.744 0.711 49.5％ 81.4％ 0.768 0.214 0.115 Lightgbm 0.782～0.834 0.808 66.9％ 78.3％ 0.767 0.182 0.102 RF 0.766-0.821 0.793 68.6％ 75.6％ 0.746 0.106 0.105 Xgboost 0.779-0.831 0.805 70.6％ 74.4％ 0.739 0.161 0.102 SVM 0.641～0.714 0.677 44.4％ 84.7％ 0.789 0.145 0.124 MLP 0.759-0.814 0.786 67.2％ 75.1％ 0.739 0.338 0.146

[0089] It can be seen from Table 2 that the area under the ROC curve AUC values ​​of the Lightgbm model and the Xgboost model are both greater than 0.8. In the present invention, the models to be selected include the Lightgbm model and the Xgboost model.

[0090] It should be noted that the sensitivity data in Table 2 were determined based on the optimal threshold for the classification of osteoporosis in women determined by the Youden Index.

[0091] Step (4): when the number of candidate models is one, the candidate model is used as the female osteoporosis prediction model; when the number of candidate models is greater than one, the candidate model with the greatest sensitivity is used as the female osteoporosis prediction model.

[0092] Comparing the sensitivities of the Lightgbm model and the Xgboost model, it can be seen that the sensitivity of the Xgboost model is greater than that of the Lightgbm model. The present invention uses the Xgboost model as a prediction model for osteoporosis in women.

[0093] In the present invention, based on the optimal threshold for classification of osteoporosis in women determined by the Youden index, the sensitivity and specificity of the XGBoost model in the test samples were 70.6%, and the sensitivity of the Lightgbm model was 66.9%, which was lower than the sensitivity of the XGBoost model.

[0094] The method also includes evaluating each parameter-tuned prediction model using Nhanes data, and verifying the accuracy of the evaluation results of each parameter-tuned prediction model based on the test set using the evaluation results of each parameter-tuned prediction model based on the Nhanes data.

[0095] It is understandable that when evaluating the prediction model of each parameter tuning, the data corresponding to the feature variable set in the Nhanes data is applied, that is, Figure 2 and Figure 4 The external validation set in .

[0096] If the absolute value of the difference between the evaluation result of each parameter-tuned prediction model based on Nhanes data and the area under the ROC curve AUC value of the evaluation result of each parameter-tuned prediction model based on the test set is within the preset difference range or the preset error rate, it means that the diagnostic accuracy of the prediction model evaluated by the test set is high, and the female osteoporosis prediction model determined based on the evaluation result is appropriate.

[0097] The multiple Nhanes data operating characteristic curves ROC established by the present invention based on Nhanes data are detailed in Figure 4 .

[0098] Will Figure 3 and Figure 4 From the comparison, it can be seen that the diagnostic accuracy of the prediction model evaluated by the test set is high, and the Xgboost model can be used as a prediction model for osteoporosis in women.

[0099] In the present invention, XGBoost also showed moderate sensitivity (77.5%) and specificity (70.1%) in the external validation dataset based on the optimal threshold for osteoporosis classification in women determined by Youden Index.

[0100] The present invention also provides a method for screening osteoporosis in women, the screening method comprising:

[0101] Display the parameter-tuned XGBoost model in a web app and build a web application.

[0102] The subjects' age, BMI, systolic blood pressure, red blood cell count, glycated hemoglobin, alanine aminotransferase, aspartate aminotransferase, and uric acid were input to obtain the osteoporosis prediction results.

[0103] When the osteoporosis prediction result is greater than or equal to a preset probability, the subject is diagnosed as being a high-risk group for osteoporosis, and a recommendation is given to complete the bone density examination as soon as possible.

[0104] Preferably, the preset probability is 16.1%.

[0105] In the present invention, the URL of the web application is https: / / xy2osteo.shinyapps.io / shiny / , and the instructions for obtaining osteoporosis prediction results based on the URL are as follows.

[0106] The subject's multiple characteristic variables were: age 58 years, systolic blood pressure 128 mmHg, red blood cell count 4.46x10^9 / L, glycosylated hemoglobin 5.8%, alanine aminotransferase 16.8 U / L, aspartate aminotransferase 20.5 U / L, uric acid 283.2 umol / L, and BMI 17.4 kg / m 2 .

[0107] Enter the above data into the URL, see Figure 5 After entering, click Osteoporosis Prediction. The prediction results displayed on the web page are as follows: Figure 6 .

[0108] from Figure 6It can be seen that the predicted result of the subject is 29.29%, which is greater than 16.1%, and the subject is considered to be a high-risk group. The predicted result displayed on the webpage can be: Your predicted result is greater than 16.1%, which means you are in the high-risk group. It is recommended that you complete the bone density test as soon as possible.

[0109] See also Figure 7The present invention also provides a device 100 for constructing a female osteoporosis prediction model based on physical examination data, the device 100 comprising a sample set screening module 101, a first acquisition module 102, a second acquisition module 103, a characteristic variable screening module 104, a model optimization module 105 and a model determination module 106. The sample set screening module 101 is used to screen a sample set from the physical examination data according to the inclusion criteria and the exclusion criteria, wherein the samples in the sample set are females aged ≥50 years or postmenopausal and have bone density data. The first acquisition module 102 is used to acquire a potential predictive variable set for osteoporosis based on literature, wherein the potential predictive variable set includes multiple demographic variables, multiple physical examination variables, multiple laboratory examination variables, a body mass index variable, and an estimated glomerular filtration rate variable, wherein the multiple demographic variables include age, gender, height, and weight; the multiple physical examination variables include: pulse, systolic blood pressure, and diastolic blood pressure; the multiple laboratory examination variables include: white blood cell count, hematocrit, red blood cell count, lymphocyte percentage, mean corpuscular hemoglobin concentration, mean corpuscular volume, monocytes The second acquisition module 103 is used to divide the sample set into a training set and a validation set, and to acquire a first training set based on the training set, wherein the first training set includes data corresponding to the prediction variable set in the training set, and bone density data. The characteristic variable screening module 104 is used to preprocess the first training set, and based on the preprocessed first training set, obtain the intersection of variables screened by Boruta algorithm, LASSO algorithm, varSelRF algorithm and SVM-RFE algorithm, and use the intersection as the characteristic variable set of the osteoporosis prediction model, the characteristic variable set includes 8 characteristic variables, namely age, BMI, systolic blood pressure, red blood cell count, glycated hemoglobin, alanine aminotransferase, aspartate aminotransferase and uric acid. The model optimization module 105 is used to obtain a second training set based on the characteristic variable set and the training set, and optimize the constructed multiple initial prediction models using the second training set to obtain multiple parameter-tuned prediction models, wherein the second training set includes data corresponding to the characteristic variable set in the training set, and bone density data.The model determination module 106 is used to obtain a test set based on the feature variable set and the validation set, and use the test set to evaluate each parameter-tuned prediction model, and determine the female osteoporosis prediction model among multiple parameter-tuned prediction models based on the model evaluation index, wherein the test set includes data corresponding to the feature variable set in the validation set, and bone density data.

[0110] The construction device also includes a verification module, which is used to evaluate each parameter-tuned prediction model using Nhanes data, and use the evaluation results of each parameter-tuned prediction model based on Nhanes data to verify the accuracy of the evaluation results of each parameter-tuned prediction model based on the test set.

[0111] Those skilled in the art can clearly understand that the specific working process of the above-described devices and modules can be referred to the contents of the aforementioned method embodiments for the convenience and brevity of description, and will not be repeated here.

[0112] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions and substitutions can be made without departing from the concept of the present invention, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for constructing a female osteoporosis prediction model based on physical examination data, characterized in that: The construction method comprises the following steps: Step S10: Based on literature query, a potential predictive variable set for osteoporosis in women is obtained, wherein the potential predictive variable set includes multiple demographic variables, multiple physical examination variables, multiple laboratory examination variables, a body mass index variable, and an estimated glomerular filtration rate variable, wherein the multiple demographic variables include age, gender, height, and weight; the multiple physical examination variables include: pulse, systolic blood pressure, and diastolic blood pressure; the multiple laboratory examination variables include: white blood cell count, hematocrit, red blood cell count, lymphocyte percentage, mean corpuscular hemoglobin concentration, mean corpuscular volume, monocyte percentage , mean platelet volume, percentage of segmented granulocytes, platelet count, neutrophil count, lymphocyte count, monocyte count, hemoglobin, eosinophil count, eosinophil percentage, fasting blood glucose, glycosylated hemoglobin, high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, triglycerides, total cholesterol, free triiodothyronine, free thyroxine, thyroid-stimulating hormone, 25-hydroxyvitamin D, parathyroid hormone, creatinine, blood urea nitrogen, alkaline phosphatase, alanine aminotransferase, aspartate aminotransferase, serum albumin, uric acid, calcium, phosphorus, magnesium, sodium, potassium, and chloride; Step S20, screening a sample set from the physical examination data according to the inclusion criteria and exclusion criteria, wherein the samples in the sample set are women aged ≥50 years or postmenopausal women with bone density data; Step S30, dividing the sample set into a training set and a validation set, and acquiring a first training set based on the training set, wherein the first training set includes data corresponding to the potential prediction variable set in the training set and bone density data; Step S40, preprocessing the first training set, and based on the preprocessed first training set, obtaining the intersection of variables screened by the Boruta algorithm, the LASSO algorithm, the varSelRF algorithm, and the SVM-RFE algorithm, respectively, and using the intersection as a feature variable set of an osteoporosis prediction model, wherein the feature variable set includes 8 feature variables, namely, age, BMI, systolic blood pressure, red blood cell count, glycated hemoglobin, alanine aminotransferase, aspartate aminotransferase, and uric acid; Step S50: based on the characteristic variable set and the training set, a second training set is obtained, and the plurality of initial prediction models constructed are optimized using the second training set to obtain a plurality of parameter-optimized prediction models, wherein the second training set includes data corresponding to the characteristic variable set in the training set and bone density data; Step S60: Based on the feature variable set and the validation set, a test set is obtained, and each parameter-tuned prediction model is evaluated using the test set, and a female osteoporosis prediction model is determined in multiple parameter-tuned prediction models based on the evaluation results, wherein the test set includes data in the validation set corresponding to the feature variable set and bone density data.

2. The method for constructing a female osteoporosis prediction model based on physical examination data according to claim 1, characterized in that: The method also includes evaluating each parameter-tuned prediction model using Nhanes data, and verifying the accuracy of the evaluation results of each parameter-tuned prediction model based on the test set using the evaluation results of each parameter-tuned prediction model based on the Nhanes data.

3. The method for constructing a female osteoporosis prediction model based on physical examination data according to claim 1, characterized in that: In step S20, the inclusion criteria are women with bone density data; the exclusion criteria include: (1) samples of the second or more bone density tests, (2) women under 50 years old or who have not yet menopausal, (3) samples lacking laboratory test data, and (4) sample groups with a large number of missing variables in a plurality of variables in the potential predictive variable set.

4. The method for constructing a female osteoporosis prediction model based on physical examination data according to any one of claims 1 to 3, characterized in that: In step S40, the preprocessing includes removing zero-variance variables and interpolating missing data, and the missing data is interpolated using a bagged tree method.

5. The method for constructing a female osteoporosis prediction model based on physical examination data according to claim 1, characterized in that: In step S50, the multiple initial prediction models include: logistic regression, decision tree, k-nearest neighbor, LightGBM, random forest, extreme gradient boosting, support vector machine and multi-layer perceptron.

6. The method for constructing a female osteoporosis prediction model based on physical examination data according to claim 5, characterized in that: In step S50, the decision tree, k-nearest neighbor, LightGBM, random forest, extreme gradient boosting, support vector machine and multilayer perceptron in the multiple initial prediction models try all possible hyperparameter combinations through grid search, and use 5-fold cross validation to evaluate the performance of each hyperparameter combination to determine the best hyperparameter combination.

7. The method for constructing a female osteoporosis prediction model based on physical examination data according to claim 6, characterized in that: In step S60, the female osteoporosis prediction model determined based on the evaluation result is an extreme gradient boosting model.

8. The method for constructing a female osteoporosis prediction model based on physical examination data according to any one of claims 5 to 7, characterized in that: The evaluation results include multiple groups of model evaluation indicators corresponding to each parameter-tuned prediction model, each group of model evaluation indicators includes the area under the ROC curve AUC and sensitivity; The step of evaluating each parameter-tuned prediction model using the test set, and determining the female osteoporosis prediction model from a plurality of parameter-tuned prediction models based on the model evaluation index comprises: Step (1), based on the test set, obtaining the probability value predicted by each parameter-tuned prediction model for each sample in the test set; Step (2), according to the probability value predicted by each parameter-tuned prediction model for each sample in the test set, multiple test set operating characteristic curves (ROC) are established; Step (3), calculating the area under the ROC curve (AUC) and sensitivity of each parameter-tuned prediction model, and taking the prediction model whose area under the ROC curve (AUC) is greater than or equal to the AUC preset value as the candidate model; Step (4): when the number of the candidate models is one, the candidate model is used as the female osteoporosis prediction model; when the number of the candidate models is greater than one, the candidate model with the greatest sensitivity is used as the female osteoporosis prediction model.

9. The method for constructing a female osteoporosis prediction model based on physical examination data according to claim 1, characterized in that: The female osteoporosis prediction model is a female hip osteoporosis prediction model.

10. A device for constructing a female osteoporosis prediction model based on physical examination data, characterized in that: The construction device comprises: The first acquisition module is used to obtain a potential predictive variable set of osteoporosis based on literature, wherein the potential predictive variable set includes multiple demographic variables, multiple physical examination variables, multiple laboratory examination variables, a body mass index variable and an estimated glomerular filtration rate variable, wherein the multiple demographic variables include age, gender, height and weight; the multiple physical examination variables include: pulse, systolic blood pressure and diastolic blood pressure; the multiple laboratory examination variables include: white blood cell count, hematocrit, red blood cell count, lymphocyte percentage, mean corpuscular hemoglobin concentration, mean corpuscular volume, monocyte percentage ratio, mean platelet volume, percentage of segmented granulocytes, platelet count, neutrophil count, lymphocyte count, monocyte count, hemoglobin, eosinophil count, eosinophil percentage, fasting blood glucose, glycosylated hemoglobin, high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, triglycerides, total cholesterol, free triiodothyronine, free thyroxine, thyroid-stimulating hormone, 25-hydroxyvitamin D, parathyroid hormone, creatinine, blood urea nitrogen, alkaline phosphatase, alanine aminotransferase, aspartate aminotransferase, serum albumin, uric acid, calcium, phosphorus, magnesium, sodium, potassium, and chloride; A sample set screening module, used to screen a sample set from the physical examination data according to the inclusion criteria and exclusion criteria, wherein the samples in the sample set are women aged ≥50 years or postmenopausal and have bone density data; A second acquisition module is used to divide the sample set into a training set and a validation set, and acquire a first training set based on the training set, wherein the first training set includes data corresponding to the prediction variable set in the training set and bone density data; A feature variable screening module, used for preprocessing the first training set, and based on the preprocessed first training set, obtaining the intersection of variables screened by using the Boruta algorithm, the LASSO algorithm, the varSelRF algorithm and the SVM-RFE algorithm, respectively, and using the intersection as the feature variable set of the osteoporosis prediction model, wherein the feature variable set includes 8 feature variables, namely age, BMI, systolic blood pressure, red blood cell count, glycated hemoglobin, alanine aminotransferase, aspartate aminotransferase and uric acid; A model optimization module, for obtaining a second training set based on the characteristic variable set and the training set, and optimizing the multiple initial prediction models constructed using the second training set to obtain multiple parameter-tuned prediction models, wherein the second training set includes data corresponding to the characteristic variable set in the training set and bone density data; The model determination module is used to obtain a test set based on a feature variable set and a validation set, and use the test set to evaluate each parameter-tuned prediction model, and determine a female osteoporosis prediction model among multiple parameter-tuned prediction models based on model evaluation indicators, wherein the test set includes data corresponding to the feature variable set in the validation set, and bone density data.

Images