Risk prediction model for severe fever with thrombocytopenia syndrome as well as construction method and application of risk prediction model

Abstract

The invention provides a severe fever with thrombocytopenia syndrome risk prediction model and a construction method and application thereof, first 10 relatively important risk factors are respectively selected through a machine learning algorithm, and coincident risk factors of the two algorithms are Virus, MYO, UREA, AGE, DD and BNP. According to the method, an extreme gradient boosted tree (XGBOOST) is adopted to score a variable set, and when the first five variables are selected, the model score is the highest, so that five prediction indexes Virus, MYO, UREA, AGE and DD are selected to construct an SFTS onset risk prediction model. The model has a good prediction value for SFTS prognosis, or the model can be applied to SFTS prognosis evaluation.

Description

Technical Field

[0001] The present invention relates to the field of biomedical detection technology, and in particular to a risk prediction model for fever with thrombocytopenia syndrome and a construction method and application thereof. Background Art

[0002] Severe fever with thrombocytopenia syndrome (SFTS) is an acute epidemic infectious disease caused by Dabie Banda virus (DBV, also known as new Bunyavirus). Its clinical manifestations are mainly fever and thrombocytopenia, which may be accompanied by symptoms such as nausea, vomiting, fatigue, poor appetite, diarrhea and muscle aches. Patients with severe conditions may experience skin ecchymosis, pulmonary hemorrhage, gastrointestinal bleeding, impaired consciousness and even multiple organ failure. In the early stage of the disease, the mortality rate was as high as 30%. With the advancement of medical technology, the domestic mortality rate of SFTS dropped from 10.58% to 5.07% from 2011 to 2021. However, the mortality rate of severe patients remains high. The mortality rate of SFTS patients with central nervous system involvement and multiple organ failure can be as high as 44.7%.

[0003] In the existing technology, the prognosis evaluation of fever with thrombocytopenia syndrome (SFTS) mainly relies on clinical observation and some traditional biochemical indicators. These methods often lack unified standards and focus on different indicators. In addition, there are few existing prediction models, and most of them are based on small sample studies and use traditional analysis methods, which limits the accuracy and generalization ability of the models.

[0004] Therefore, establishing a prognostic risk warning model for SFTS patients can help clinicians identify potential critically ill patients at an early stage, detect changes in their condition in a timely manner, and take proactive measures and nursing interventions, which is of great significance for improving the success rate of patient treatment and reducing the mortality rate. Summary of the invention

[0005] The purpose of the present invention is to provide a risk prediction model for fever with thrombocytopenia syndrome and a method for constructing the same. The Logistic regression model of the present invention exhibits a high AUC value and good stability in the training set and the validation set, and is selected as the best prediction model, which is of great significance in the risk prediction of fever with thrombocytopenia syndrome.

[0006] The present invention adopts the following technical solution:

[0007] In a first aspect of the present invention, a method for constructing a risk prediction model for fever with thrombocytopenia syndrome is provided, the method comprising:

[0008] Obtain laboratory test indicators for patients with fever and thrombocytopenia syndrome;

[0009] The laboratory test indicators were preliminarily screened by baseline analysis and correlation analysis, and the preliminarily screened variables were obtained and randomly divided into training sets and test sets;

[0010] In the training set, the variables after the preliminary screening are ranked by importance using a machine learning method, and the variable set is scored using an extreme gradient boosting tree to obtain the final screened variables; the final screened variables are input using multiple initial prediction models to respectively construct multiple trained prediction models;

[0011] The multiple trained prediction models are evaluated using a model evaluation index, and the best trained prediction model after the evaluation is determined as the fever with thrombocytopenia syndrome risk prediction model.

[0012] Furthermore, the laboratory test indicators are obtained by collecting blood samples from patients with fever and thrombocytopenia syndrome within 48 hours of admission, and include:

[0013] SEX, AGE, Virus, WBC, NEUT, LYM, MONO, RBC, HGB, PLT, PT, FIB, APTT, TT, DD, PCT, BNP, CTnI, MYO, CK-MB, ALT, AST, TP, ALB, TB, DB, ALP, GGT, CHOL, TG, CK, LDH, K, NA, CL, CA, UREA, CREA, UA, LIP, AMY-P, CRP and FER.

[0014] Furthermore, the baseline analysis includes: screening for significant variables with P < 0.05 and filling missing values ​​using the random forest method;

[0015] The correlation analysis includes: removing strongly correlated variables with partial correlation coefficients greater than 0.7, and performing variance homogeneity analysis to remove variables with variance inflation factors greater than 10.

[0016] Furthermore, the method also includes: using the test set to verify the prediction performance of the prediction model, and randomly using 80% of the data of the variables after the preliminary screening as a training set and 20% of the data as a test set.

[0017] Furthermore, the multiple initial prediction models include: extreme gradient boosting tree, logistic regression algorithm, lightweight gradient boosting machine, K-nearest neighbor algorithm, complementary naive Bayes classification algorithm, and support vector machine algorithm model in machine learning.

[0018] Furthermore, the preset model evaluation indicators include: drawing the receiver operating characteristic curves of the 6 prediction models, evaluating the effectiveness of the prediction model by the area under the ROC curve AUC, and evaluating the performance of the machine learning algorithm by 6 indicators: area under the curve, sensitivity, accuracy, specificity, negative predictive value, positive predictive value and F1 score.

[0019] Furthermore, the method further comprises:

[0020] Inputting the preliminarily screened variables into the risk prediction model for fever with thrombocytopenia syndrome to obtain prediction results, and constructing a SHAP graph based on the prediction results;

[0021] The predicted probabilities of the variables after the preliminary screening in the SHAP diagram are calculated and added together to obtain the prognostic risk probability of the fever with thrombocytopenia syndrome risk prediction model.

[0022] In the second aspect of the present invention, a risk prediction model for fever with thrombocytopenia syndrome obtained by the above method is provided.

[0023] In a third aspect of the present invention, a system for constructing a risk prediction model for fever with thrombocytopenia syndrome is provided, the system comprising:

[0024] A data acquisition module, used to obtain laboratory test indicators of patients with fever and thrombocytopenia syndrome;

[0025] The data initial screening module performs baseline analysis and correlation analysis on the laboratory test indicators for initial screening, obtains the variables after initial screening, and randomly divides them into training set and test set;

[0026] A model training module is used to use a machine learning method to sort the importance of the variables after the preliminary screening in the training set, and use an extreme gradient boosting tree to score the variable set to obtain the final screened variables; input the final screened variables and use multiple initial prediction models to construct multiple trained prediction models respectively;

[0027] The model evaluation module is used to evaluate the multiple trained prediction models using preset model evaluation indicators, and determine the best trained prediction model after evaluation as the retinopathy risk prediction model.

[0028] In a fourth aspect of the present invention, a risk prediction system for fever with thrombocytopenia syndrome is provided, the system comprising:

[0029] A processor and a memory, the memory being coupled to the processor, the memory storing instructions which, when executed by the processor, perform the steps of the method.

[0030] One or more technical solutions in the embodiments of the present invention have at least the following technical effects or advantages:

[0031] The risk prediction model for fever with thrombocytopenia syndrome and the construction method thereof provided by the present invention have the following advantages compared with the prior art solutions:

[0032] (1) High accuracy: The logistic regression model showed high AUC values ​​in both the training set and the validation set, indicating high prediction accuracy.

[0033] (2) Generalization ability: The model also shows good prediction performance on the independent test set, indicating that the model has good generalization ability.

[0034] (3) Potential for clinical application: The model uses variables that are easily obtained clinically, making it easy to apply in clinical settings.

[0035] (4) Interpretability: The model is visually explained through the SHAP method, which increases the transparency and credibility of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0037] Figure 1 This was a correlation analysis between the initial screening variables.

[0038] Figure 2 Flow chart of the SFTS prognosis prediction model study.

[0039] Figure 3 Ranking variable importance (random forest algorithm).

[0040] Figure 4 Rank variable importance (extreme gradient boosting tree algorithm).

[0041] Figure 5 Rank variable importance (extreme gradient boosting tree algorithm).

[0042] Figure 6 Venn diagram of the final variants for three machine learning algorithms, with annotations: XGBoost, extreme gradient boosted trees; RF, random forests; AdaBoost, adaptive boosting

[0043] Figure 7 Modeling performance vs. number of variables.

[0044] Figure 8 ROC curves for the six machine learning models in the training set.

[0045] Fig. 9 The ROC curves of the six machine learning models in the validation set. Notes: Validation ROC Curve, validation set receiver operating characteristic curve; sensitivity, sensitivity; Specificity, specificity; extreme gradient boosting tree (XGBoost), logistic regression algorithm (logisitc), lightweight gradient boosting machine (LightGBM), K-nearest neighbor algorithm (KNN), complementary naive Bayes classification algorithm (CNB), support vector machine (SVM) algorithm.

[0046] Fig.10 Evaluation of the logistic prediction model for the training set.

[0047] Fig.11 Evaluation of the logistic prediction model on the validation set.

[0048] Fig.12 Evaluation of the logistic prediction model for the test set.

[0049] Fig.13 Learning curves for the logistic regression model for the training set and validation set.

[0050] Fig.14 This is a SHAP summary diagram, annotations and explanations: In the figure, blue to red represent the prediction factor values ​​from low to high. Each point represents a prediction factor and the Shapley value of an instance. The position on the vertical axis is determined by the prediction factor, and the position on the horizontal axis is determined by the Shapley value. The higher the Shapley value, the greater the probability of the predicted event.

[0051] Fig.15 This is the prediction result in Application Example 1. DETAILED DESCRIPTION

[0052] The present invention will be described in detail below in conjunction with specific implementations and examples, and the advantages and various effects of the present invention will be more clearly presented. It should be understood by those skilled in the art that these specific implementations and examples are used to illustrate the present invention, rather than to limit the present invention.

[0053] Throughout the specification, unless otherwise specifically stated, the terms used herein should be understood as meanings commonly used in the art. Therefore, unless otherwise defined, all technical and scientific terms used herein have the same meanings as those generally understood by those skilled in the art to which the present invention belongs. In the event of a conflict, the present specification takes precedence.

[0054] Unless otherwise specified, various raw materials, reagents, instruments and equipment used in the present invention can be purchased from the market or obtained by existing methods.

[0055] The general idea of ​​the present invention is as follows:

[0056] This application retrieved the SFTS data in the clinical research database of our hospital, and extracted the clinical data of 264 SFTS cases (165 cases in the survival group and 99 cases in the death group), and divided the samples into a training set of 168 cases, a validation set of 43 cases, and a test set of 53 cases.

[0057] Prediction models for SFTS prognosis were constructed by using machine learning-based extreme gradient boosting tree (XGBoost), logistic regression algorithm (logisitc), lightweight gradient boosting machine (LightGBM), K-nearest neighbor algorithm (KNN), complementary naive Bayes classification algorithm (CNB), and support vector machine (SVM) algorithm. The area under the receiver operating characteristic curve (AUC) was used to compare the predictive value of the models constructed by the above six algorithms for SFTS prognosis, and the model with the best predictive efficiency was screened out. The prediction results of the machine learning model were visualized and explained using the SHAP method.

[0058] Results: 1. Among the models constructed by the six machine learning algorithms, the AUC of the logistic regression model was higher in both the training set (0.913) and the validation set (0.926), and the model stability was better, with the best prediction efficiency. 2. Virus, MYO, UREA, AGE, and DD were candidate predictors for constructing the logistic model. 3. The SHAP summary plot analysis showed that Virus, MYO, UREA, AGE, and DD were risk factors for the prognosis of SFTS.

[0059] The logistic model constructed based on machine learning method in this application with Virus, MYO, UREA, AGE, and DD as predictive features has good predictive value for SFTS prognosis, or this model can be applied to SFTS prognosis assessment.

[0060] in,

[0061] Virus (new Bunyavirus) is a direct reflection of the virus infection status. The logarithmic value of its load can be used as an evaluation indicator, and the detection method is PCR (polymerase chain reaction) technology.

[0062] MYO (myoglobin) is a sensitive marker of myocardial cell damage. Changes in its concentration can reflect the extent of myocardial damage. The detection method is microparticle chemiluminescence.

[0063] UREA (urea) is one of the important indicators for evaluating renal function. An increase in its level often indicates renal damage. The detection method is enzyme kinetic colorimetry.

[0064] AGE (age) is a basic parameter of individual biological characteristics and is of great significance for disease risk assessment and treatment plan formulation. The data is obtained through the inpatient electronic medical record information system.

[0065] DD (D-dimer) is the product of thrombosis and activation of the fibrinolytic system, and the detection method is immunoturbidimetry.

[0066] SEX (full name is sex), AGE (full name is age), WBC (full name is white blood cell), NEUT (full name is neutrophil), LYM (full name is lymphocyte), MONO (full name is monocyte), RBC (full name is red blood cell), HGB (full name is hemoglobin), PLT (full name is platelet), PT (full name is prothrombin time), FIB (full name is fibrinogen), APTT (full name is activated partial prothrombin time), TT (full name is thrombin time), DD (full name is D dimer), PCT (full name is procalcitonin), BNP (full name is B-type natriuretic peptide), CTnI (full name is cardiac troponin I), MYO (full name is myoglobin), CK-MB (full name is creatine kinase isoenzyme MB), ALT (full name is alanine transferase Aminase), AST (full name aspartate aminotransferase), TP (full name total protein), ALB (full name albumin), TB (full name total bilirubin), DB (full name direct bilirubin), ALP (full name alkaline phosphatase), GGT (full name gamma-glutamyl transpeptidase), CHOL (full name cholesterol), TG (full name triglyceride), CK (full name creatine kinase), LDH (full name lactate dehydrogenase), K (full name potassium), NA (full name sodium), CL (full name chloride), CA (full name calcium), UREA (full name urea nitrogen), CREA (full name creatinine), UA (full name uric acid), LIP (full name lipase), AMY-P (full name pancreatic amylase), CRP (full name C-reactive protein), FER (full name ferritin).

[0067] The risk prediction model for fever with thrombocytopenia syndrome and its construction method of the present application will be described in detail below in combination with examples and experimental data.

[0068] Example 1: A risk prediction model for fever with thrombocytopenia syndrome and its construction method

[0069] 1. Research subjects

[0070] This study is a retrospective analysis. A total of 264 SFTS patients admitted to the Department of Infectious Diseases and the Intensive Care Unit of this hospital from April 1, 2023 to July 18, 2024 were selected as the research subjects, including 123 males and 141 females. The average age was (64.170±10.226) years old. Inclusion criteria: ① Age ≥ 18 years old; ② Meet the diagnostic criteria for SFTS. That is, the case specimens tested positive for the new Bunyavirus nucleic acid. Exclusion criteria: ① Patients with incomplete laboratory-related data; ② Patients diagnosed in the outpatient clinic but not admitted to the hospital; ③ Patients with malignant tumors and blood system diseases. This study was reviewed and approved by the hospital ethics committee.

[0071] 2. General information collection

[0072] The relevant risk factor data were collected by reviewing the electronic medical records of the selected SFTS during their hospitalization. The clinical data of SFTS included: body temperature, days from onset to diagnosis, days of hospitalization, chills, dizziness, cough, sputum, chest tightness, loss of appetite, abdominal pain, hepatosplenomegaly, rash, poor appetite, fatigue and decreased physical strength, headache, coma, bleeding gums, diarrhea before consultation, skin and mucous membrane ecchymosis, lymphadenopathy, impaired consciousness, muscle aches, nausea, vomiting, changes in mental status, muscle tremors, diabetes, hypertension, coronary heart disease, cerebral infarction, smoking history, and drinking history.

[0073] 3. Laboratory test indicators

[0074] All subjects had blood samples collected within 48 hours of admission and sent for blood routine, biochemical, coagulation function and other indicators. This study used the first examination results after the patient was admitted to the hospital. This study selected SEX, AGE, Virus, WBC, NEUT, LYM, MONO, RBC, HGB, PLT, PT, FIB, APTT, TT, DD, PCT, BNP, CTnI, MYO, CK-MB, ALT, AST, TP, ALB, TB, DB, ALP, GGT, CHOL, TG, CK, LDH, K, NA, CL, CA, UREA, CREA, UA, LIP, AMY-P, CRP, FER and other indicators as observation objects. Japanese SYSMES fully automatic analyzers XE2100 five-category and XS800i five-category were used to detect blood routine related indicators; Swiss Roche C8000 biochemical analyzer was used to check blood biochemistry and other related indicators (ALT, AST, GGT, creatinine, etc.); French STAGO coagulation analyzer was used to check blood coagulation and other indicators (APTT, D-dimer, etc.). The American Abbott fully automatic biochemical immunoassay was used to detect three indicators of myocardial infarction (CTnI, MYO, CK-MB), the Swiss Roche cobas E602 electrochemiluminescence immunoassay was used to detect (PCT, BNP), and the Xi'an Tianlong PANA9600s extractor and GENTIER 96E amplifier were used to detect virus load, and the amplification reagents were produced by Sun Yat-sen University Da'an Gene Co., Ltd.

[0075] 4. Data preprocessing and variable screening

[0076] 4.1 First, we conducted a baseline analysis and preliminarily selected 33 significant (P < 0.05) variables from the baseline analysis table (see Table 1). We used the random forest method to fill in the missing values.

[0077] Table 1 SFTS baseline characteristics (characteristic variables with P>0.05 have been eliminated)

[0078]

[0079]

[0080]

[0081] 4.2 In order to reduce the model error caused by the mutual influence between variables, correlation analysis is performed (see Figure 1), and some strongly correlated variables (correlation coefficient>0.7) were eliminated, including PCT, ALT, DB, CREA, LIP, TT, CK, TP, FERR, and GGT. At the same time, in order to reduce the impact of homoscedasticity on the model, variance homogeneity analysis was performed, and it was found that there were no variables with variance inflation factor (VIF)>10, so there was no need to eliminate the variables.

[0082] 4.3 Use machine learning methods to rank the importance of the variables after preliminary screening, so as to further extract important features. We used three mainstream machine learning methods: Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Adaptive Boosting (AdaBoost) to rank the importance of features. The top 10 sets of variables of the three models have common features: Virus, MYO, UREA, AGE, DD, and BNP.

[0083] Extreme gradient boosting tree (XGBOOST) was used to score the variable set. When the first five variables were selected, the model score was the highest, so five predictive indicators Virus, MYO, UREA, AGE, and DD were selected as the final screened variables (i.e., the input variables of the model).

[0084] 5. Machine learning algorithms build predictive models

[0085] 5.1 Using screening variables to build a prediction model

[0086] Taking the final variables described in 4.3 as input variables, six prediction models were constructed using machine learning-based extreme gradient boosting tree (XGBoost), logistic regression algorithm (logisitc), lightweight gradient boosting machine (LightGBM), K-nearest neighbor algorithm (KNN), complementary naive Bayes classification algorithm (CNB), and support vector machine (SVM) algorithm. 80% of the data was used as the training set, 20% as the validation set, and a five-fold cross validation was performed to observe the performance in different models.

[0087] 1. Extreme Gradient Boosting Tree (XGBoost)

[0088] XGBoost is a scalable machine learning system that implements the gradient boosting tree model internally and optimizes the algorithm in the model. For example, when dealing with missing values ​​in data, the XGBoost algorithm can automatically learn the split direction and handle missing values; it can simulate nonlinear effects and has high efficiency and accuracy when dealing with large data sets. At present, clinical prediction models built based on the XGBoost algorithm have been widely used in the field of medicine, which can significantly improve the decision-making efficiency in disease diagnosis and disease risk, prognosis and outcome, drug development and rational and safe drug use, and provide highly feasible solutions in these fields.

[0089] 2. Logistic Regression Algorithm:

[0090] Logistic regression is a widely used linear classification algorithm, mainly used for binary classification problems. It determines the class label by predicting the probability of an event. The logistic regression model estimates the parameters by maximizing the likelihood function, usually optimized using gradient descent or Newton's method. Although it is a linear model, in some cases, it can also handle nonlinear relationships through feature engineering and regularization.

[0091] 3. Lightweight Gradient Boosting Machine (LightGBM)

[0092] LightGBM is an improved version of XGBoost. Compared with the former, it adds many new methods to improve the model, such as parallel schemes, bundled exclusive features, and gradient-based unilateral detection. The central idea is to build the best model through iterative training of weak classifiers (decision trees). It has low memory utilization and fast running speed. It also supports parallel learning and avoids overfitting, which makes it further improve the gradient boosting tree and gradually becomes the mainstream algorithm for ensemble learning. At the same time, LightGBM abandons the layer-by-layer decision tree growth strategy used by most gradient boosting tree models, and uses a leaf growth strategy with depth restrictions for iteration, which greatly improves the model prediction efficiency, while ensuring high-precision predictions and preventing overfitting.

[0093] 4. K-Nearest Neighbor Algorithm (KNN):

[0094] KNN is an instance-based learning algorithm that makes predictions based on the concept of nearest neighbors. When making predictions, it finds the K samples in the training set that are closest to the new data point. The KNN algorithm is simple and intuitive, easy to implement, but computationally expensive, especially when dealing with high-dimensional data. It is a non-parametric method that does not require assumptions about the distribution of the data, so it performs well in some applications.

[0095] 5. Supplementary Naive Bayes Classification Algorithm (CNB)

[0096] The complementary naive Bayesian classification algorithm (CNB) is a simple Bayesian classifier

[45] . To some extent, it is an improvement on the standard multinomial naive Bayesian algorithm and is used to process continuous feature variables. It is based on the Bayesian theorem and has a relatively stable classification performance. It performs best when the distribution independence assumption holds. CNB can not only solve the problem of sample imbalance, but also ignore the complement of the naive assumption to a certain extent. Rennie et al. showed that the parameter estimation of CNB is more stable than that of ordinary multinomial naive Bayes, and it is very suitable for samples with imbalanced samples. Therefore, this study selected it as one of the machine learning algorithms for experiments.

[0097] 6. Support Vector Machine (SVM) Algorithm

[0098] Support vector machine is a binary classification model defined by the maximum margin in feature space. It uses sparse kernel machine algorithm and is a particularly powerful and flexible supervised learning model. It is mainly suitable for models with hyperplane separation of input data. It analyzes data for classification and regression. Its usual algorithmic complexity is polynomially related to the dimension of the data space and the number of data points. It only relies on a subset of data (support vector) to predict unknown class labels.

[0099] 5.2 Validation and performance evaluation of prediction models

[0100] In order to reduce the risk of overfitting and bias, this study will use 80% of the data as a training set and 20% of the data as a validation set. The performance in different models will be observed through five-fold cross-validation, and the receiver operating characteristic (ROC) curves of all models will be plotted. The area under the ROC curve (AUC) will be used to evaluate the performance of the prediction model. The performance of the machine learning algorithm will be evaluated through six indicators: area under the curve (AUC), sensitivity, accuracy, specificity, negative predictive value (NPV), positive predictive value (PPV) and F1 score. The prediction model with the best performance will be selected for the next step of research through the above evaluation indicators.

[0101] 5.3 Construction of SHAP Summary Graph

[0102] SHAP, as a popular machine learning interpretable tool, is a post hoc interpretation framework based on game theory and local interpretation. The SHAP summary diagram can combine the importance of predictors with the effects of predictors. Each point in the figure represents a predictor and the Shapley value of a corresponding instance. The position on the ordinate axis is determined by the predictor, and the position on the abscissa axis is determined by the Shapley value. Blue to red represent predictor values ​​from low to high. The distribution of the Shapley value of each predictor can be understood from the figure. The Shapley value estimated by SHAP can explain the prediction model globally and locally. Therefore, the SHAP summary diagram can be used to explain the contribution of each feature of each sample to the prediction result, increase the transparency of the model, and facilitate clinicians to analyze the reliability of the model.

[0103] 6. Statistical methods

[0104] This study used R 3.6.3 software and Python 3.7 for statistical analysis and model construction based on machine learning and SHAP graph construction. Quantitative data were expressed as (x±s), and the independent sample t test was used for inter-group comparison; count data were expressed as number of cases and percentages, and the differences between groups were compared using the χ 2 P < 0.05 was considered statistically significant.

[0105] 3. Results

[0106] 3.1 Predictor Selection

[0107] A total of 264 SFTS (165 in the survival group and 99 in the death group) were included for screening predictive factors.

[0108] Random forest (RF) was used to analyze the importance of variables. The top 10 variables (from high to low) were: Virus, DD, MYO, LDH, UREA, AGE, BNP, CTnI, APTT, CA (see Figure 3 ).

[0109] The extreme gradient boosting tree (XGBoost) was used to analyze the feature importance. The 10 most important features (from high to low) were: Virus, UREA, MYO, AGE, DD, ALP, CTnI, BNP, AMYP, LDH (see Figure 4 ).

[0110] Adaptive boosting (AdaBoost) was used to analyze the feature importance. The top 10 features (from high to low) are: DD, AGE, FIB, UREA, BNP, ALP, AMYP, PT, MYO, Virus (see Figure 5 ).

[0111] The top 10 sets of variables of the three methods have common features: Virus, MYO, UREA, AGE, DD, BNP (see Figure 6 ).

[0112] The extreme gradient boosting tree (XGBOOST) was used to score the variable set. When the first five variables were selected, the model score was the highest (see Figure 7 ), so five predictive indicators, Virus, MYO, UREA, AGE, and DD, were selected to construct a risk prediction model for SFTS.

[0113] 3.2 Evaluation of Machine Learning Models

[0114] There are differences in performance between different models: the AUC of logistic is higher and the model is more stable in both the training set (0.913) and the validation set (0.926). The AUC of XGBoost is 1.000 in the training set and 0.909 in the validation set, and the AUC of LightGBM is 1.000 in the training set and 0.862 in the validation set. The model may be overfitting and has poor stability. The next best is the KNN model (training set 0.899, validation set 0.740), the SVM model (training set 0.810, validation set 0.725), and the AUC of the CNB algorithm (training set 0.725, validation set 0.784); see Figure 8 , Fig. 9 , Table 2, Table 3. In summary, compared with other models, the logistic algorithm is an effective prediction tool. Therefore, we will use the logistic algorithm to construct a SFTS prognostic risk prediction model in the next step.

[0115] Table 2 Evaluation of 6 prediction models of the training set

[0116]

[0117]

[0118] Table 3 Evaluation of 6 prediction models in the validation set

[0119]

[0120] 3.3Logistic algorithm is used to construct the SFTS prognostic risk prediction model

[0121] Logistic Regression machine learning method was used for classification. The classification variable was OUTCOME. The variables in the model included Virus, AGE, DD, MYO, and UREA. The model parameters were: C (regularization factor): 1.0; max_iter (number of iterations): 100; penalty (regularization type): l2; tol (convergence measure): 0.0001.

[0122] The logistic algorithm was used to construct the SFTS prognostic risk prediction model. Whether in the training set (0.938), validation set (0.936) or test set (0.927), the AUC value was high and the model stability was good. Fig.10 , 11 , 12. Given that the performance of the validation set under the AUC indicator does not exceed that of the test set or the excess ratio is less than 10%, it can be considered that the fit is successful and the logistic model can be used for the classification modeling task of this dataset. Fig.13 .

[0123] Table 4 Evaluation of logistic algorithm prediction model in three data sets

[0124]

[0125] 3.3 Interpreting Model Results—Construction and Visualization of SHAP Plots

[0126] In order to further clarify the positive / negative relationship of each predictor on the prediction results, we use the SHAP diagram to visualize the interpretation of the model results (see Fig.14 ). Based on the SHAP diagram, we can see that Virus has the greatest impact on the model. In addition, the larger the eigenvalue, the higher the Shapley value, that is, the greater the risk of death from SFTS. Age is the second most influential feature. The higher the eigenvalue, the higher the risk of death from SFTS. The same is true for UREA, MYO, and DD. The higher the eigenvalue, the higher the risk of death from SFTS. The above experimental results are consistent with relevant clinical and research results, further confirming the impact of Virus, MYO, UREA, AGE, and DD on the prognostic risk of SFTS. Although the SHAP summary diagram can briefly describe the positive or negative effects of the characteristics on the risk of SFTS, the amount of information is still limited. In order to better understand the interactive effects between the variables, further research is needed on the relationship between the predictive indicators.

[0127] Compared with models based on traditional regression analysis (such as logistic models), its advantages are mainly reflected in the following aspects: 1. Prediction accuracy: Machine learning models (especially logistic regression models) are usually able to handle complex nonlinear relationships and further optimize prediction performance through feature engineering. In contrast, traditional logistic models may be limited by their linear assumptions and have difficulty capturing complex interactions between variables. Therefore, machine learning models based on the five indicators of Virus, MYO, UREA, AGE, and DD may have higher prediction accuracy. 2. Model stability: The AUC values ​​of the logistic regression model on both the training set and the validation set were high (0.938 and 0.936, respectively), showing good model stability. This shows that the performance of the model on different data sets is relatively consistent and can stably predict the prognosis of SFTS patients. 3. Feature importance interpretation: In this study, we used the SHAP method proposed by Lundberg and Lee to interpret the logistic regression model. Because the SHAP diagram has the advantages of easy understanding, high accuracy, and visual interface, it can help clinicians make better diagnosis and treatment decisions, and has therefore been widely used in the prognosis and prediction analysis of clinical diseases. Therefore, we use the logistic regression algorithm to build a model based on the above five factors, and then use the SHAP diagram to visualize and explain the model. 4. Ability to process high-dimensional data: Machine learning models are usually able to process high-dimensional data and automatically select the features that have the greatest impact on the prediction results. In contrast, traditional regression analysis may face problems of computational complexity and overfitting when processing high-dimensional data. Therefore, machine learning models based on multiple indicators such as Virus, MYO, UREA, AGE, and DD can consider the patient's condition information more comprehensively and improve the accuracy of prediction. 5. Generalization ability: Machine learning models use methods such as cross-validation to reduce overfitting and improve the generalization ability of the model during training. This means that machine learning models based on indicators such as Virus, MYO, UREA, AGE, and DD not only perform well on the current data set, but may also maintain high prediction performance on other similar data sets.

[0128] In summary, the SFTS prognostic model based on the five indicators of Virus, MYO, UREA, AGE, and DD has advantages over the traditional regression analysis model in terms of higher prediction accuracy, model stability, feature importance interpretation ability, ability to process high-dimensional data, and generalization ability. These advantages make the model have higher practical value and promotion prospects in clinical applications.

[0129] The logistic regression model constructed based on machine learning with Virus, MYO, UREA, AGE, and DD as predictive features has a good predictive value for the prognosis of SFTS patients. The construction of the SHAP graph can help physicians further understand the principles behind the prediction results of machine learning methods, so as to make clinical decisions that are in the best interests of patients.

[0130] Example 2: Fever with thrombocytopenia syndrome risk prediction system

[0131] The embodiment of the present invention provides a fever with thrombocytopenia syndrome risk prediction system, the system comprising: a data acquisition module, used to obtain laboratory examination indicators of patients with fever with thrombocytopenia syndrome;

[0132] The data initial screening module performs baseline analysis and correlation analysis on the laboratory test indicators for initial screening, obtains the variables after initial screening, and randomly divides them into training set and test set;

[0133] A model training module is used to use a machine learning method to sort the importance of the variables after the preliminary screening in the training set, and use an extreme gradient boosting tree to score the variable set to obtain the final screened variables; input the final screened variables and use multiple initial prediction models to construct multiple trained prediction models respectively;

[0134] The model evaluation module is used to evaluate the multiple trained prediction models using preset model evaluation indicators, and determine the best trained prediction model after evaluation as the retinopathy risk prediction model.

[0135] Example 3: Fever with thrombocytopenia syndrome risk prediction system

[0136] An embodiment of the present invention provides a risk prediction system for fever with thrombocytopenia syndrome, the system comprising:

[0137] A processor and a memory, wherein the memory is coupled to the processor and stores instructions. When the instructions are executed by the processor, the steps described in Embodiment 1 or Embodiment 2 are used.

[0138] Embodiment 4: Computer readable storage medium

[0139] An embodiment of the present invention provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method of embodiment 1 and / or the steps of embodiment 2 are implemented.

[0140] Of course, the computer executable instructions of a storage medium including computer executable instructions provided by an embodiment of the present invention are not limited to the operations of the method described above, but can also execute related operations in the method provided by any embodiment of the present invention.

[0141] Through the above description of the implementation methods, the technicians in the relevant field can clearly understand that the present invention can be implemented by means of software and necessary general hardware, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for an electronic device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0142] It is worth noting that in the above embodiments, the various units and modules included are only divided according to functional logic, but are not limited to the above divisions, as long as the corresponding functions can be achieved; in addition, the specific names of the functional units are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention.

[0143] Application Example 1: Predicting the prognosis of patients with clinical fever and thrombocytopenia syndrome

[0144] The five indicators of Virus, MYO, UREA, AGE and DD of clinical patients with fever and thrombocytopenia syndrome are collected and input into the fever and thrombocytopenia syndrome risk prediction model or system obtained by the present invention, that is, the prognostic risk probability of the fever and thrombocytopenia syndrome risk prediction model.

[0145] Specifically, log in to the online prediction model website

[0146] http: / / 10.32.15.196:9999 / model / list / predict / model / html?mid=15376&symbol=2JX17231nc0649fg5pO7, you can see the interface as Fig.15 shown.

[0147] In a certain sample, Virus (lg copies / ml) = 2.75, AGE (year) = 6.3, DD (μg / mL) = 2.62, MYO (ng / mL) = 109.2, and UREA (mmol / L) = 4.3.

[0148] The obtained disease occurrence probability is: 5.6%, the threshold for disease occurrence is 45.9%, and the disease probability is lower than the threshold, so the possibility of disease occurrence is considered low.

[0149] Finally, it should be noted that the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements that are inherent to such process, method, article, or apparatus.

[0150] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0151] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for constructing a risk prediction model for fever with thrombocytopenia syndrome, characterized in that: The method comprises: Obtain laboratory test indicators for patients with fever and thrombocytopenia syndrome; The laboratory test indicators were preliminarily screened by baseline analysis and correlation analysis, and the preliminarily screened variables were obtained and randomly divided into training sets and test sets; In the training set, the variables after the preliminary screening are ranked by importance using a machine learning method, and the variable set is scored using an extreme gradient boosting tree to obtain the final screened variables; the final screened variables are input using multiple initial prediction models to respectively construct multiple trained prediction models; The multiple trained prediction models are evaluated using a model evaluation index, and the best trained prediction model after the evaluation is determined as the fever with thrombocytopenia syndrome risk prediction model.

2. The method according to claim 1, characterized in that The laboratory examination indicators are obtained by collecting blood specimens from patients with fever and thrombocytopenia syndrome within 48 hours of admission to the hospital, including: SEX, AGE, Virus, WBC, NEUT, LYM, MONO, RBC, HGB, PLT, PT, FIB, APTT, TT, DD, PCT, BNP, CTnI, MYO, CK-MB, ALT, AST, TP, ALB, TB, DB, ALP, GGT, CHOL, TG, CK, LDH, K, NA, CL, CA, UREA, CREA, UA, LIP, AMY-P, CRP and FER.

3. The method according to claim 1, characterized in that The baseline analysis included: screening for significant variables with P < 0.05 and filling in missing values ​​using the random forest method; The correlation analysis includes: removing strongly correlated variables with partial correlation coefficients greater than 0.7, and performing variance homogeneity analysis to remove variables with variance inflation factors greater than 10.

4. The method according to claim 1, characterized in that: The method further comprises: using the test set to verify the prediction performance of the prediction model, wherein 80% of the data of the variables after the preliminary screening are randomly used as a training set and 20% of the data are used as a test set.

5. The method according to claim 1, characterized in that: The multiple initial prediction models include: extreme gradient boosting tree, logistic regression algorithm, lightweight gradient boosting machine, K-nearest neighbor algorithm, complementary naive Bayes classification algorithm, and support vector machine algorithm model in machine learning; The preset model evaluation indicators include: drawing the receiver operating characteristic curves of the 6 prediction models, evaluating the effectiveness of the prediction model by the area under the ROC curve AUC, and evaluating the performance of the machine learning algorithm by 6 indicators: area under the curve, sensitivity, accuracy, specificity, negative predictive value, positive predictive value and F1 score.

6. The method according to claim 1, characterized in that The method further comprises: Inputting the preliminarily screened variables into the risk prediction model for fever with thrombocytopenia syndrome to obtain prediction results, and constructing a SHAP graph based on the prediction results; The predicted probabilities of the variables after the preliminary screening in the SHAP diagram are calculated and added together to obtain the prognostic risk probability of the fever with thrombocytopenia syndrome risk prediction model.

7. A risk prediction model for fever with thrombocytopenia syndrome obtained by the method described in any one of claims 1 to 6.

8. A system for constructing a risk prediction model for fever with thrombocytopenia syndrome, characterized in that: The system comprises: A data acquisition module, used to obtain laboratory test indicators of patients with fever and thrombocytopenia syndrome; The data initial screening module performs baseline analysis and correlation analysis on the laboratory test indicators for initial screening, obtains the variables after initial screening, and randomly divides them into training set and test set; A model training module is used to use a machine learning method to sort the importance of the variables after the preliminary screening in the training set, and use an extreme gradient boosting tree to score the variable set to obtain the final screened variables; input the final screened variables and use multiple initial prediction models to construct multiple trained prediction models respectively; The model evaluation module is used to evaluate the multiple trained prediction models using preset model evaluation indicators, and determine the best trained prediction model after evaluation as the retinopathy risk prediction model.

9. A fever with thrombocytopenia syndrome risk prediction system, characterized in that: The system comprises: A processor and a memory, the memory being coupled to the processor, the memory storing instructions, and when the instructions are executed by the processor, the steps of any one of the methods 1-6 are used.

10. Use of the fever with thrombocytopenia syndrome risk prediction model according to claim 7 or the fever with thrombocytopenia syndrome risk prediction system according to claim 9 in the preparation of a fever with thrombocytopenia syndrome risk prediction product.

Images