A fracture risk prediction model based on multi-level feature mining of big data

Abstract

The present invention provides a fracture risk prediction model based on multi-level feature mining of big data, characterized in that: the fracture risk prediction model is a hybrid model of XGBoost and SVM. Based on the model, fracture risk can be predicted.

Description

Technical Field

[0001] The present invention provides an auxiliary medical model, and in particular, relates to a fracture risk prediction model for elderly osteoporosis patients based on big data multi-level feature mining. Background Art

[0002] As the aging population increases, the burden on the health system caused by fractures is becoming more and more serious. Fragility fractures are the most serious complication of osteoporosis, and the incidence rate is very high among elderly osteoporosis patients. Surveys show that the incidence rate of osteoporosis in the elderly over 60 years old exceeds 36%, of which the incidence rate in men is 23% and the incidence rate in women is 49%. The number of people with fractures caused by osteoporosis is huge.

[0003] For patients, fractures may cause pain, reduce the quality of life, and shorten the life expectancy of patients; for society, a large number of fracture patients will also bring huge medical burden. In contrast to the high incidence and serious harm of fractures, the diagnosis and treatment rate of fracture population is very low. Since bone density loss usually occurs slowly and there are no obvious clinical symptoms in the early stage, it is difficult for patients to prevent fractures caused by osteoporosis. Some patients only discover the problem of bone density loss when they are examined in the hospital due to other diseases.

[0004] Therefore, if we can predict the risk of fracture in elderly patients with osteoporosis, it will be of great help to patients and society. When patients visit the hospital, they will leave a large amount of clinical data, which is called electronic health records (EHR), which usually includes patient diagnosis information, medical records, past medical history, physical signs, etc. Patients may not go for orthopedic examinations specifically, but when conducting blood and urine tests, some of the indicators may have an indicative effect on predicting fractures.

[0005] Currently, the fracture risk assessment tool FRAX developed by the World Health Organization calculates the probability of osteoporotic fractures in the next 10 years from some relatively easy-to-obtain clinical risk factors (CRFs), including age, gender, body mass index (BMI), history of fragility fractures, history of hip fractures in parents, current smoking, long-term use of glucocorticoids, rheumatoid arthritis, heavy drinking, etc.; the prediction effect can also be enhanced by inputting bone density (BMD).

[0006] Although FRAX is considered one of the most accurate tools for predicting fracture risk, it still has many limitations. First, it only includes a part of the risk factors, and does not include common and easily available blood, urine and other test data in clinical data, which contain factors that play an important role in prediction. In addition, FRAX can only predict the risk of fracture in people aged 40-90 years old, but it is also necessary to predict the risk of fracture in the elderly over 90 years old. Summary of the invention

[0007] The present invention aims to overcome the above-mentioned defects and establish a machine learning model based on real-world data sets to predict the fracture risk of elderly osteoporotic patients. At the same time, the risk factors of fractures are mined from electronic medical records using data mining methods, which plays a certain guiding role in clinical practice, thereby promoting fracture prevention in high-risk groups and moving the threshold of diagnosis and treatment forward.

[0008] The present invention provides a fracture risk prediction model based on big data multi-level feature mining, which is characterized by:

[0009] The fracture risk prediction model is a hybrid model of XGBoost and SVM.

[0010] Furthermore, the present invention provides a fracture risk prediction model based on multi-level feature mining of big data, which is also characterized in that:

[0011] The hybrid model of XGBoost and SVM first uses XGBoost to transform features, and then uses support vector machine SVM as a classifier to predict fracture risk.

[0012] Furthermore, the present invention provides a fracture risk prediction model based on multi-level feature mining of big data, which is also characterized in that:

[0013] The model parameters of the XGBoost are set to: Max number of trees = 300;

[0014] The model parameter settings of the SVM are: bf Gaussian kernel, C=1.0, C=0.008.

[0015] Furthermore, the present invention provides a fracture risk prediction model based on multi-level feature mining of big data, which is also characterized by:

[0016] The characteristics that determine the prediction results are: blood calcium, apolipoprotein A / B ratio, whether the weight is reduced, C-reactive protein, whether there is a history of vaccination, high-density lipoprotein, alpha2 globulin, alkaline phosphatase, whether there is heart disease, retinol-binding protein, whether there is a history of smoking, blood sodium, bilirubin, weight, alkaline granulocyte count, lumbar vertebra 1 bone density, urine leukocyte esterase, blood insulin 2 hours after meal, age, and the difference between lumbar vertebra 1 bone density and lumbar vertebra 2 bone density.

[0017] Furthermore, the present invention provides a fracture risk prediction model based on multi-level feature mining of big data, which is also characterized by:

[0018] The method for obtaining the features that determine the prediction results is:

[0019] comprehensive feature importance score=α×Weight+β×gain+ γ×cover+θ×SHAP values

[0020] Among them, the weight coefficient of cover is 0.3;

[0021] The weight coefficient of gain is 0.3;

[0022] The weight coefficient of weight is 0.2;

[0023] The weight coefficient of shap is 0.2.

[0024] Furthermore, the present invention provides a fracture risk prediction model based on multi-level feature mining of big data, which is also characterized by:

[0025] The method for establishing the fracture risk prediction model is as follows:

[0026] S1. Get the dataset;

[0027] S2. Perform descriptive statistics and significance tests on the data set to establish statistically significant differences between the fracture group and the non-fracture group;

[0028] S3. Process missing values ​​and unbalanced labels of the dataset;

[0029] S4. Evaluate the classification effects of various models based on accuracy, precision, recall, and F-score to obtain the optimal model.

[0030] Furthermore, the present invention provides a fracture risk prediction model based on multi-level feature mining of big data, which is also characterized in that:

[0031] The missing value filling method for missing data is the random forest regression method.

[0032] Furthermore, the present invention provides a fracture risk prediction model based on multi-level feature mining of big data, which is also characterized in that:

[0033] The method for processing the label imbalance is as follows:

[0034] S3.2.1. Randomly divide the entire data set into training set and test set;

[0035] S3.2.2. Perform a sampling method that combines oversampling SMOTE and undersampling Tomek links on the training set. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 . The specific working method of the model of this embodiment;

[0037] Figure 2 .Data extraction process;

[0038] Figure 3 .Random forest regression filling effect;

[0039] Figure 4 .Comparison of various sampling methods;

[0040] Figure 5 .The structure of the hybrid model;

[0041] Figure 6 .Comparison between XGBoost model and hybrid model;

[0042] Figure 7 .The 20 most important features obtained based on Weight, Gain, and Cover;

[0043] Figure 8 The 20 most important features obtained using SHAP;

[0044] Fig. 9 . Based on the comprehensive feature importance score, sort and get the 20 features with the highest scores;

[0045] Fig.10 Comparison of the results of inputting only 20 features and the results of inputting all features;

[0046] Fig.11 .Prediction results based on blood calcium;

[0047] Fig.12 .Prediction results based on CPR;

[0048] Fig.13 .Prediction results based on ALP; DETAILED DESCRIPTION

[0049] Example 1. A fracture risk prediction model based on multi-level feature mining of big data

[0050] The fracture risk prediction model is a hybrid model of XGBoost and SVM.

[0051] Specific work flow: Figure 1 shown.

[0052] Embodiment 2. The specific research method is as follows:

[0053] 2.1 Data Collection: The process of data extraction is as follows Figure 2 shown.

[0054] Inclusion criteria: Information on outpatients and inpatients from 2012 to 2017 was collected from the electronic medical record system of our hospital. Including patient information, medical records, test information, vital signs, symptoms, medication, doctor's orders, medical records, diagnosis results, etc. This database contains a total of 14,418 patients. Elderly osteoporosis patients were selected as the research target, and 8,530 patients were obtained.

[0055] Exclusion criteria: Abnormal data due to recording errors were excluded; patients whose fractures occurred before osteoporosis and patients whose fractures were caused by trauma were excluded.

[0056] Data collection: In the diagnostic data, the patient's condition was extracted according to the ICD-10 code of the disease and expressed in binary form; in the medical record data, the patient's age, gender and medical history information was obtained, such as smoking history, drinking history, previous fracture history, vaccination history, etc.; in the test data, blood and urine test data were obtained, such as blood sugar, blood calcium, urine protein and other parameters.

[0057] If a patient has multiple test records, the record closest to the diagnosis result is selected; in the vital signs data, the patient's height, weight and other information are obtained; in the doctor's orders and medication data, medication information is extracted, such as whether insulin, heparin, warfarin and other drugs that may affect bone metabolism are taken; in the symptom data, whether the patient has symptoms such as joint pain and weight loss is obtained. In the bone density data, the bone density information of the patient's lumbar vertebrae 1 to lumbar vertebrae 4 is obtained.

[0058] First, we included all possible features, and then excluded those with too many missing values. In this way, a total of 146 features were obtained, which can be divided into 6 categories: demographic information, medical history information, test results, disease conditions, medication conditions, and clinical symptoms.

[0059] 2.2 Statistical analysis

[0060] Descriptive statistics and significance tests were performed on the data set to see whether there were significant differences in the indicators between the fracture group and the non-fracture group from a statistical point of view. Features are divided into two categories: categorical variables and numerical variables. Categorical variables include gender, medical history and other characteristics, and numerical variables include age, height, weight, and various test results. If the distribution of the feature meets the criteria for parametric testing, then the chi-square test and t-test are used respectively, otherwise the non-parametric test is used. Some indicators are shown in the following table:

[0061] Table 2-1 Statistical analysis results of some features

[0062]

[0063]

[0064] 2.3 Data Preprocessing

[0065] The original data set has undergone a series of data preprocessing, including missing value processing and unbalanced label processing.

[0066] 2.3.1 Determining the criteria for removing missing values

[0067] Missing data is common in clinical research. Missing data occurs when the values ​​of the variables of interest for all subjects in the sample are not measured or recorded. This dataset has a high missing rate. If all missing values ​​are deleted directly, the dataset will become very small; if a large amount of missing data is retained and filled, the results will be greatly disturbed.

[0068] Therefore, the standard for retaining data is first determined. When the missing rate of each feature of each sample reaches a certain value, the sample is deleted from the data set.

[0069] The XGBoost model was used for preliminary experiments because it can automatically process real data. The results of the experiment are shown in the following table:

[0070] Table 2-2 Preliminary experimental results

[0071]

[0072]

[0073] From the above table, it can be found that when the missing rate of each sample of a record exceeds 55%, deleting the record from the data set can achieve the best effect, thus retaining 1913 samples.

[0074] 2.3.2 Missing Value Filling

[0075] The following methods of filling missing values ​​are used:

[0076] Stratified mean filling: fill the height and weight with mean according to gender, and fill the test data with mean according to age group.

[0077] missMDA: This is an R package developed by Hussonhe Josse that performs principal component analysis on data with missing values. It can interpolate categorical and continuous variables.

[0078] Random Forest Regression Filling: Random Forest (RF) is becoming increasingly popular for handling missing data, especially in biomedical research. It does not assume normality and does not require specifying a parametric model. For a set of data with n features, if feature T has missing values, select T as the label and select the other n-1 features and labels as the new feature matrix. Then, the non-missing part of feature T can be used as training data to predict the missing part. If other features also have missing values, traverse all features and start filling in the missing parts from the least missing part.

[0079] XGBoost is used to compare the effects of the above filling methods, and AUC (area under the ROC curve) is used as an indicator to measure the filling effect. Figure 3 As shown in the figure, the random forest regression filling method performs best on this dataset.

[0080] 2.3.2 Dealing with label imbalance

[0081] The labels in our dataset are very unbalanced (fracture: non-fracture = 1090:7440), which results in poor prediction performance of the model. An effective solution is to resample before training the machine learning classification model. The entire dataset is first randomly divided into a training set and a test set, and then various sampling methods are performed only on the training set. Usually, a combination of oversampling and undersampling of the majority class can achieve better classifier performance. Use the XGBoost model to evaluate the effect of different sampling methods.

[0082] like Figure 4 As shown, the combination of SMOTE (oversampling) and Tomek links (undersampling) is the best.

[0083] 2.4 Model

[0084] Input and output of the model: The factors extracted from EHR can be represented by the matrix x = {x1, x2, ..., xd} T as the input of the model. The predicted fracture risk can be represented by y, which is the output of the model and is a binary classification result.

[0085] Model structure: The structure of the mixed model is as follows Figure 5 As shown in Figure 2, XGBoost is first used to transform the features, and then the support vector machine (SVM) is used as a classifier to predict the risk of fracture.

[0086] This study used accuracy, precision, recall, and F-score to evaluate the classification effect of the model.

[0087] Accuracy: The ratio of the number of samples correctly classified by the classifier to the total number of samples

[0088]

[0089] Precision: The ratio of the number of samples classified as true positive examples to the number of samples classified as positive examples

[0090]

[0091] Recall: The ratio of the number of samples classified as true positive examples to the number of all true positive examples.

[0092]

[0093] F-score: An extension of accuracy, combining precision and recall

[0094]

[0095] Among them, TP=true positive, TN=true negatives, FP=false positive, FN=false negatives.

[0096] Example 3. Comparison of prediction results

[0097] 3-1. Comparison between different models

[0098] The dataset is divided into training set and test set according to a 4:1 ratio.

[0099] On the training set, SMOTE was used for oversampling, and Tomek links were used for undersampling to avoid overfitting. The experiment was repeated 100 times for each model to obtain the measurement results. The parameter settings of different models are shown in Table 3-1:

[0100] Table 3-1 Model parameter settings

[0101]

[0102] The prediction results of different models are shown in Table 3-2. It can be seen that the performance of the XGBoost and SVM hybrid model is better than other models.

[0103] Table 3-2 Model prediction results

[0104]

[0105] 3-2. The impact of different boosting trees on the model

[0106] In this study, the effect of the number of boosted trees in XGBoost on the recall rate was explored.

[0107] We are more concerned about recall here because it measures our ability to predict fracture risk among people at high risk of fracture. Figure 6 As can be seen in the figure, the performance of the hybrid model is always better than the single XGBoost model. When the number of trees reaches about 300, the prediction performance of the model basically reaches its upper limit.

[0108] 3-3. Feature mining results

[0109] In this study, a score to measure the importance of features was designed, taking into account the results of four aspects:

[0110] Weight: The number of times a feature is used to split nodes in all trees.

[0111] Gain: The total gain brought by a feature at each node split in all trees.

[0112] Cover: The number of all samples covered by a feature at each node split in all trees.

[0113] The 20 most important features obtained based on Weight, Gain, and Cover are as follows: Figure 7 shown.

[0114] SHAP (SHapley Additive exPlanations): A unified approach to explaining the output of any machine learning model, linking game theory with local explanations. In addition to being able to measure the importance of features, it can also explain the direction in which each feature contributes to the classification. The 20 most important features obtained using SHAP are as follows: Figure 8 As shown in the figure, one dot represents one sample. The redder the color, the larger the value of the feature itself, and the bluer the color, the smaller the value of the feature itself. The further to the left the dot is, the smaller the output result is, and the further to the right the dot is, the larger the output result is.

[0115] To obtain the comprehensive feature importance score, we first scale the score obtained by each method to between 0 and 1.

[0116]

[0117] We then linearly combine these feature scores.

[0118]

[0119] Among them, α, β, γ, and δ are the weight coefficients of each scoring method. In order to find the best coefficient, we tested different combinations of values. Under each combination, we obtained the comprehensive feature importance score based on this, sorted the 20 features with the highest scores, and input these 20 features into the model to obtain the accuracy, precision, recall, and F-score. The results are shown in Table 3-3.

[0120] Table 3-3 Effects of different weight combinations

[0121]

[0122]

[0123] When the weight coefficients of cover, gain, weight, and shap are 0.3, 0.3, 0.2, and 0.2 respectively, the top 20 features obtained based on this have the best prediction performance. The scores of these 20 features from high to low are: blood calcium, apolipoprotein A / B ratio, whether the weight is reduced, C-reactive protein, whether there is a history of vaccination, high-density lipoprotein, alpha2 globulin, alkaline phosphatase, whether there is heart disease, retinol binding protein, whether there is a history of smoking, blood sodium, bilirubin, weight, alkaline granulocyte count, lumbar vertebra 1 bone density, urine leukocyte esterase, blood insulin 2 hours after meal, age, and the difference between lumbar vertebra 1 bone density and lumbar vertebra 2 bone density.

[0124] We compare the results of inputting only these 20 features with the results of inputting all features. The results are as follows Fig.10 As shown in the figure, it can be seen that only inputting these 20 features can achieve better results than inputting all the features. This shows that these 20 features have good predictive performance and can be regarded as important risk factors for fractures.

[0125] Functions and effects of this embodiment:

[0126] In this study, a hybrid model was designed to predict the fracture risk in elderly patients with osteoporosis, and 20 important features were mined.

[0127] In this study, all variables were first used as inputs to the model. Then the data was processed for missing values ​​and unbalanced labels. Then we tested a variety of machine learning models, among which the hybrid model of XGBoost and SVM had the best prediction performance. Finally, a comprehensive feature importance score was obtained, which comprehensively considered the information of weight, cover, gain, and SHAP, and mined the 20 most important features. The scores were ranked from high to low: blood calcium, apolipoprotein A / B ratio, whether the weight was reduced, C-reactive protein, whether there was a history of vaccination, high-density lipoprotein, alpha2 globulin, alkaline phosphatase, whether there was heart disease, retinol binding protein, whether there was a history of smoking, blood sodium, bilirubin, weight, alkaline granulocyte count, lumbar vertebra 1 bone density, urine leukocyte esterase, blood insulin 2 hours after meal, age, and the difference between lumbar vertebra 1 bone density and lumbar vertebra 2 bone density.

[0128] Among them, some characteristics are well known to be related to bone metabolism, such as: whether there is a history of smoking, whether the weight is reduced, lumbar vertebra 1 bone density, weight, age, and the difference between lumbar vertebra 1 bone density and lumbar vertebra 2 bone density.

[0129] There are some characteristics that are often overlooked in clinical practice, but they are closely related to bone metabolism and we should pay special attention to them, such as indicators related to inflammation: blood calcium, C-reactive protein, and alkaline phosphatase.

[0130] It should be pointed out that based on the above research, the present invention has conducted in-depth research and analysis on blood calcium, C-reactive protein, and alkaline phosphatase:

[0131] Serum calcium is the amount of calcium in the blood. Calcium is one of the most important minerals in the body. 99% of the body's calcium is stored in the bones, and the remaining 1% circulates in the blood. The body cannot make its own calcium, and if it does not get enough calcium, the body will take calcium from the bones, and the bones will become weak. A normal calcium balance is important for maintaining a balanced bone metabolism. Abnormal calcium levels in the blood can be a sign of bone disease. In our model, if Fig.11 As shown, when the blood calcium level decreases (blue), the probability of predicted fracture increases.

[0132] C-reactive protein is involved in the immune response and is a non-specific, highly sensitive biomarker of inflammation. The immune system and the validation process are involved in a variety of health problems. Systemic chronic inflammation tends to increase with age. Studies have shown that the circulating levels of cytokines, chemokines, and acute phase proteins in the elderly are higher, and more genes involved in inflammation are expressed. In addition, systemic chronic inflammation is persistent and will eventually cause damage to tissues and organs over time, such as inducing osteoporosis. On the other hand, skeletal muscle produces and releases cytokines and actin into the blood, which can have a systemic effect of reducing inflammation. With increasing age, low physical activity leads to an increase in anabolic resistance and affects the levels of C-reactive protein and pro-inflammatory cytokines. These effects, in turn, promote some pathological changes related to inflammation, including osteoporosis.

[0133] The determination of alkaline phosphatase is a routine part of blood tests. It is a hydrolase responsible for removing phosphate groups from many types of molecules. It is most effective in an alkaline environment. Most of the alkaline phosphatase is produced in the liver, and the rest comes from the bones. There are several isoenzymes of alkaline phosphatase, which come from different tissues, such as: liver, bones, spleen, etc. In the bones, alkaline phosphatase is involved in the decomposition of pyrophosphate, which is a strong inhibitor of extracellular calcium phosphate. Therefore, alkaline phosphatase plays an important role in bone formation. The concentration of alkaline phosphatase in the blood varies with age, blood type, gender, etc. Abnormal levels of alkaline phosphatase usually indicate diseases of the bones, liver, kidneys, etc. For example, the automatic analyzer of Technicon is used to measure the alkaline phosphatase in the serum of osteoporotic patients with enteritis. Spinal bone density also shows a significant correlation with blood alkaline phosphatase. Studies have shown that serum alkaline phosphatase levels can be used as an indicator of osteoporosis in men. In addition, the level of alkaline phosphatase decreased after treatment with alendronate sodium, indicating that alkaline phosphatase can be used as an indicator to reflect the efficiency of osteoporosis drug treatment.

[0134] The relationship between abnormal lipid metabolism and osteoporosis deserves further discussion. For example, apolipoprotein A / B ratio, high-density lipoprotein cholesterol (HDL-C), alkaline leukocyte count. Ackert-Bicknell et al. demonstrated that bone metabolism and high-density lipoprotein have a genetic relationship and that high-density lipoprotein can directly interact with osteoblasts and osteoclasts.

[0135] Several indicators, such as alpha2 globulin, blood sodium, bilirubin (BIL), leukocyte lipase, and insulin 2 hours after a meal, have not been found to have a clear relationship with bone metabolism. Heart disease and immune history may play an important role in model prediction because they are binary variables with fewer missing values.

[0136] In the data set, several samples were randomly selected, which met the requirements of FRAX, and the model of this study was compared with the results obtained by FRAX to test the consistency of the two. The results showed that patients who were high-risk according to FRAX were also predicted to be high-risk by our model. However, FRAX cannot predict the elderly population over 90 years old, while the method of the present invention is not afraid of age barriers.

Claims

1. A fracture risk prediction model for elderly osteoporosis patients based on multi-level feature mining of big data, characterized by: The fracture risk prediction model is a hybrid model of XGBoost and SVM; The hybrid model of XGBoost and SVM first uses XGBoost to transform the features, and then uses SVM as a classifier to predict the risk of fracture; The model parameters of the XGBoost are set as: Maxnumberoftrees=300; The model parameter settings of the SVM are: bfGaussian kernel, C=1.0, γ=0.008; The method for obtaining the features that determine the prediction results is: comprehensivefeatureimportancescore=α×weight+β×gain+ Y×cover+θ×SHAPvalues Comprehensive feature importance score is the comprehensive feature importance score. In order to obtain the comprehensive feature importance score, the score obtained by each method is first scaled to between 0 and 1, and then these feature scores are linearly combined; α, β, γ, and θ are the weight coefficients of each scoring method. In order to find the best coefficient, different values ​​are combined and tested. Under each combination, the comprehensive feature importance score is obtained, and the 20 features with the highest scores are sorted; Among them, the weight coefficient γ of cover is 0.3; The weight coefficient β of gain is 0.3; The weight coefficient α of weight is 0.2; The weight coefficient θ of shap is 0.2; The characteristics that determine the prediction results are: blood calcium, apolipoprotein A / B ratio, whether the weight is reduced, C-reactive protein, whether there is a history of vaccination, high-density lipoprotein, alpha2 globulin, alkaline phosphatase, whether there is heart disease, retinol-binding protein, whether there is a history of smoking, blood sodium, bilirubin, weight, alkaline granulocyte count, lumbar vertebra 1 bone density, urine leukocyte esterase, blood insulin 2 hours after meal, age, and the difference between lumbar vertebra 1 bone density and lumbar vertebra 2 bone density.

2. A fracture risk prediction model for elderly osteoporosis patients based on big data multi-level feature mining as claimed in claim 1, characterized in that : The method for establishing the fracture risk prediction model is as follows: S1. Get the dataset; S2. Perform descriptive statistics and significance tests on the data set to establish statistically significant differences between the fracture group and the non-fracture group; S3. Process missing values ​​and unbalanced labels of the dataset; S4. Evaluate the classification effects of various models based on accuracy, precision, recall, and F-score to obtain the optimal model.

3. A fracture risk prediction model for elderly osteoporosis patients based on big data multi-level feature mining as claimed in claim 2, characterized in that : The missing value filling method for missing data is the random forest regression method.

4. A fracture risk prediction model for elderly osteoporosis patients based on big data multi-level feature mining as claimed in claim 2, characterized in that : The method for processing the unbalanced labels is as follows: S3.2.

1. Randomly divide the entire data set into training set and test set; S3.2.

2. Perform a sampling method that combines oversampling SMOTE and undersampling Tomek links on the training set.

Images