Construction method, system and equipment for esophageal and gastric varicosity prediction model

Abstract

The invention discloses a construction method, system and equipment for an esophageal and gastric varicosity prediction model. The invention establishes a prediction model for esophageal and gastric varicose veins of a subject, which can predict the occurrence of esophageal and gastric varicose veins of patients with cirrhosis and has high prediction ability. In addition, a clinical decision curve proves the feasibility of the prediction model for clinically evaluating the esophageal gastric varicosity of the liver cirrhosis patient. The model constructed by the invention has high accuracy, and can be used as a non-invasive tool for evaluating esophageal and gastric varicose veins of patients with cirrhosis. The prediction model disclosed by the invention is expected to play an important role in evaluating the degree of esophageal and gastric varicose veins of patients with cirrhosis, and can help clinicians to make clinical decisions.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence disease prediction, and in particular to a method, system and device for constructing an esophageal and gastric varices prediction model. Background Art

[0002] Cirrhosis is a common chronic liver disease caused by viral infection, alcoholism, non-alcoholic fatty liver, biliary tract disease, autoimmune disease or genetic factors. Due to the extensive necrosis of liver cells under the long-term action of different factors, diffuse proliferation of liver fibrosis occurs on this basis, forming nodules and pseudolobules, which further destroy the normal structure and blood supply of the liver. Cirrhosis is the 11th most common cause of death in the world. About 1.16 million people die from cirrhosis each year, of which 1 million die from complications of cirrhosis.

[0003] Portal hypertension is a major consequence of liver cirrhosis and the cause of most complications, including ascites, esophageal and gastric varices and their rupture and bleeding, and encephalopathy. Gastroesophageal varices (GOV) are mostly caused by increased portal vein pressure in liver cirrhosis, which leads to the anastomosis of the gastric coronary vein with the esophageal vein and azygos vein of the vena cava system, causing retrograde blood flow and dilation of the esophageal and gastric veins. Esophagogastric variceal bleeding (EVB) is one of the most common fatal complications of GOV in liver cirrhosis. Nearly 50% of patients with liver cirrhosis may develop GOV, which is proportional to the degree of liver damage. About 40% of patients with Child-Pugh score A and 85% of patients with grade C liver cirrhosis are accompanied by GOV, and the degree of varicose veins will increase at a rate of 10%-15% per year. The mortality rate of the first bleeding caused by GOV is 30%-50%. Clinically, early detection and treatment of esophageal and gastric varices can reduce the risk of esophageal varicose vein rupture and bleeding by 5%-15%. Therefore, early identification and early detection are of great significance to improving the quality of life of patients with cirrhosis.

[0004] At present, gastroscopy is the gold standard for assessing the presence of GOV and bleeding in patients with cirrhosis. The guidelines recommend that patients with cirrhosis without GOV undergo gastroscopy every 2 years, and patients with mild varicose veins undergo gastroscopy once a year. However, gastroscopy is an invasive procedure that can easily induce bleeding risk, cause discomfort to patients, and lead to poor compliance, and has high associated costs. Therefore, seeking non-invasive, accurate, and feasible non-invasive methods to predict the risk of GOV is a current challenge. With the advancement of technology, detection methods are also gradually developing. The length, width, and thickness of the spleen can be measured by ultrasound. A meta-analysis showed that the platelet count to spleen diameter ratio (PSR) can be used as a non-invasive diagnosis of GOV with high diagnostic efficiency. The AUC values ​​of this model for diagnosing varicose veins and high-risk varicose veins were 0.872 and 0.813, with sensitivities of 84.0% and 78.0%, and specificities of 78.0% and 67.0%, respectively. CT is also a widely used non-invasive examination method. Yu-Jen Tseng et al. showed through meta-analysis that the AUC value of CT diagnosis of GOV was 0.860, with a sensitivity and specificity of 89.6% and 72.3%, respectively. Non-invasive predictive indicators such as liver stiffness and spleen stiffness measured by transient elastography have high predictive value. Pons, Monica MD et al. found that liver stiffness <15 kPa and platelet count ≥150×10 9 / L can exclude clinically significant portal hypertension from most causes. The Baveno VII consensus states that patients with compensated progressive chronic liver disease (cACLD) should have a liver stiffness of less than 20 kPa and a platelet count of more than 150×10 9 / L can exempt the inner diameter screening for GOV, or the spleen hardness ≤40 kPa can identify the risk of high-risk GOV. Although there are many indicators to predict GOV, there is no unified non-invasive diagnostic standard.

[0005] Machine learning (ML) is to select algorithms based on existing data, and to build models based on algorithms and data to predict the occurrence of events. More and more studies are applying ML technology to promote the prediction of GOV, other complications of cirrhosis, and other disease progression. For example, Tien S Dong et al. established an EVendo scoring model using random forests to screen high-risk GOV. The EVendo scoring model mainly includes international normalized ratio, aspartate aminotransferase, platelets, urea nitrogen, hemoglobin, and the presence or absence of ascites. The AUC values ​​of this model for mild GOV screening in the training and validation cohorts were 0.84 and 0.82, respectively, and the AUC values ​​for high-risk GOV screening in the two cohorts were 0.74 and 0.75, respectively. This makes ML-based models a valuable tool for implementation in clinical practice.

[0006] The information in the background technology is only for illustrating the general background of the present invention and should not be regarded as admitting or suggesting in any form that such information constitutes the prior art known to a person skilled in the art. Summary of the invention

[0007] The present invention aims to establish a diagnostic model based on machine learning for the evaluation of GOV in patients with liver cirrhosis, and to provide a simple and convenient non-invasive prediction method for GOV patients. Specifically, the present invention includes the following contents.

[0008] A first aspect of the present invention provides a method for constructing a prediction model based on machine learning, which comprises the following steps: (1) providing a training sample set comprising a plurality of samples, wherein each of the samples comprises basic personal information and test information data and a label of a subject, wherein the subject comprises a patient with liver cirrhosis, the basic personal information and the test data are obtained by screening by introducing a penalty coefficient into a regression model, and the label is esophageal and gastric varices disease; (2) Inputting the personal basic information and the test data into the XGB machine learning model for training to obtain a prediction model for esophageal and gastric varices in patients with liver cirrhosis.

[0009] In certain embodiments, according to the method for constructing a prediction model based on machine learning described in the present invention, the number of samples in the training sample set is greater than 50.

[0010] In certain embodiments, the method for constructing a prediction model based on machine learning according to the present invention further comprises providing a validation sample set, and performing validation using the validation sample set, thereby adjusting or optimizing the prediction model according to the validation results.

[0011] In certain embodiments, the method for constructing a prediction model based on machine learning according to the present invention further comprises providing a test sample set, and using the test sample set to perform testing, thereby evaluating the prediction model.

[0012] In certain embodiments, according to the method for constructing a prediction model based on machine learning according to the present invention, the personal basic information and test data include parameter data selected from age, spleen thickness, platelets, albumin, prealbumin and cholinesterase.

[0013] A second aspect of the present invention provides a system for building a prediction model based on machine learning, comprising: A data acquisition unit, configured to acquire data from a training sample set and an optional validation sample set and a test sample set, wherein the training sample set comprises a training sample set of multiple samples, and each sample comprises personal basic information and test data and a label of a subject, the subject comprises a patient with liver cirrhosis, the personal basic information and test data are obtained by screening by introducing a penalty coefficient into a regression model, and the label is esophageal and gastric varices disease; A construction unit includes an XGB machine learning model and is configured to input the personal basic information and detection data in a training sample set into the XGB machine learning model for training. Optionally, it is further configured to use the verification sample set to verify the trained prediction model, thereby adjusting or optimizing the prediction model according to the verification result. Optionally, it further includes testing with a test sample set to evaluate the prediction model.

[0014] The third aspect of the present invention provides a method for predicting esophageal and gastric varices in a subject, wherein the method includes the step of obtaining the subject's basic personal information and test data and inputting them into a prediction model to obtain a prediction probability, wherein the subject is a patient with cirrhosis, and the basic personal information and test data include parameter data of age, spleen thickness, platelets, albumin, prealbumin and cholinesterase, and the prediction model is obtained according to the method of the present invention.

[0015] A fourth aspect of the present invention provides a system for predicting esophageal and gastric varices in a subject, comprising: a data acquisition unit configured to acquire basic personal information and test data of a subject, wherein the subject includes a patient with liver cirrhosis, and the basic personal information and test data include parameter data of age, spleen thickness, platelets, albumin, prealbumin, and cholinesterase; and The prediction unit is configured to input the data acquired by the data acquisition unit into a prediction model and output a prediction result, wherein the prediction model is obtained according to the method of the present invention.

[0016] A fifth aspect of the present invention provides a device, wherein the device comprises: a memory, a processor, and a program stored in the memory and capable of running on the processor, wherein the program is configured to implement the method for constructing a prediction model or the steps of the prediction method according to the present invention.

[0017] According to a sixth aspect of the present invention, a computer-readable storage medium is provided, wherein a program is stored on the computer-readable storage medium, and when the program is executed by a processor, the method for constructing a prediction model or the steps of the prediction method described in the present invention is implemented.

[0018] The present invention establishes a new prediction model for esophageal and gastric varices in subjects, which can predict the occurrence of GOV in patients with cirrhosis and has a high predictive ability. The present invention uses a clinical decision curve (DCA) to confirm the feasibility of the prediction model in clinically evaluating GOV. The model constructed by the present invention has high accuracy and can be used as a non-invasive tool for evaluating GOV in patients with cirrhosis. Therefore, the prediction model of the present invention is expected to play an important role in the assessment of the degree of esophageal and gastric varices in patients with cirrhosis, and can help clinicians make clinical decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Data processing workflow for patients with cirrhosis.

[0020] Figure 2 Distribution of causes of liver cirrhosis.

[0021] Figure 3 is the regression coefficient path diagram.

[0022] Figure 4 is the Lasso regression cross validation curve.

[0023] Figure 5 Comparison of variables selected by Lasso regression in different GOV grades. A. Comparison of Age in different GOV grades; B. Comparison of ST in different GOV grades; C. Comparison of PLT in different GOV grades; D. Comparison of ALB in different GOV grades; E. Comparison of PA in different GOV grades; F. Comparison of CHE in different GOV grades. Age: age; ST: spleen thickness; PLT: platelet; ALB: albumin; PA: prealbumin; CHE: cholinesterase.

[0024] Figure 6 ROC curves of different machine learning models on the test set. A: training set; B test set; C: external validation. RF: random forest; SVM: support vector machine; XGB: XGBoost; GNB: Gaussian naive Bayes; KNN: K-nearest neighbor algorithm.

[0025] Figure 7 ROC curve comparison of XGB with LSM, PLT and the combination of the two. A. Test set; B. External validation. XGBtest: XGB model in the test set; XGBval: XGB model in the external validation set; LSM: liver stiffness measurement > 20 kPa; PLT: platelet count > 150 × 10^9; LP: liver stiffness measurement > 20 kPa and platelet count > 150 × 10^9.

[0026] Figure 8 Clinical decision curve of XGB model. A: test set; B: external validation. DETAILED DESCRIPTION

[0027] Various exemplary embodiments of the present invention will now be described in detail. This detailed description should not be considered as limiting the present invention, but should be understood as a more detailed description of certain aspects, features, and embodiments of the present invention.

[0028] It should be understood that the terms described in the present invention are only for describing special embodiments and are not intended to limit the present invention. In addition, for the numerical range in the present invention, it should be understood that the upper and lower limits of the scope and each intermediate value therebetween are specifically disclosed. Each smaller range between the intermediate value in any stated value or stated range and any other stated value or intermediate value in the described range is also included in the present invention. The upper and lower limits of these smaller ranges can be independently included or excluded in the scope.

[0029] Unless otherwise indicated, all technical and scientific terms used herein have the same meanings as commonly understood by those skilled in the art to which the invention belongs. Although the present invention describes only preferred methods and materials, any methods and materials similar or equivalent to those described herein may also be used in the implementation or testing of the present invention. All documents mentioned in this specification are incorporated by reference to disclose and describe the methods and / or materials related to the documents. In the event of a conflict with any incorporated document, the content of this specification shall prevail.

[0030] Methods for building prediction models based on machine learning One aspect of the present invention provides a method for constructing a prediction model based on machine learning, comprising the following steps: (1) providing a training sample set comprising a plurality of samples, wherein each of the samples comprises basic personal information and test data of a subject and a label, wherein the subject comprises a patient with liver cirrhosis, the basic personal information and test data are obtained by screening by introducing a penalty coefficient into a regression model, and the label is esophageal and gastric varices disease; (2) Inputting the personal basic information and the test data into the XGB machine learning model for training to obtain a prediction model for esophageal and gastric varices in patients with liver cirrhosis.

[0031] In step (1) of the present invention, a training sample set containing multiple samples is first provided. The sample type is not particularly limited and may be a fluid sample (blood sample) or a tissue sample from a subject. The subject includes a patient with liver cirrhosis, and the personal basic information and test data are obtained by screening by introducing a penalty coefficient in a regression model, and the label is esophageal gastric varices disease. In a preferred embodiment, the personal basic information and test data include or consist of the following variables: age (Age), spleen thickness (ST), platelets (PLT), albumin (ALB), prealbumin (PA), cholinesterase (CHE). It should be noted that the above variables are not selected arbitrarily. The inventors have conducted a large number of studies and used 30 influencing factors with inter-group differences in GOV as independent variables. Overfitting can be avoided by introducing a penalty coefficient λ, and the optimal penalty coefficient λ of lambda.1se is further selected using ten-fold cross validation. In a specific embodiment, λ is 0.004. In another specific embodiment, λ is 0.049.

[0032] In step (1) of the present invention, the number of samples in the training sample set is not particularly limited and can be as many as possible to ensure the predictive or diagnostic performance of the final prediction model. For example, the number of samples can be greater than 50, greater than 100, greater than 500, or greater than 1000.

[0033] In the present invention, the samples are patients with liver cirrhosis from the Chinese population. Therefore, in certain embodiments, the construction method of the present invention is a method for constructing an esophageal and gastric varices prediction model based on machine learning that is suitable for the Chinese subject population.

[0034] In step (2) of the present invention, the personal basic information and the test data are input into the XGB machine learning model for training to obtain a prediction model for esophageal and gastric varices in patients with liver cirrhosis. The machine learning model is not randomly selected in the construction method of the present invention. The XGB machine learning model finally selected is suitable for the 6 variables screened in step (1) and has significantly improved diagnostic or predictive efficacy compared to other commonly used machine learning models (including but not limited to Random forest (RF), Support Vector Machine (SVM), Gradient Boosting Classification (GBC), Gaussian Naive Bayes (GNB), K-Nearest Neighbor (KNN), etc.). Predictive efficacy related indicators include indicators such as AUC, sensitivity, and specificity of the constructed model.

[0035] The method of constructing a prediction model based on machine learning of the present invention further includes providing a verification sample set, and using the verification sample set for verification, so as to adjust or optimize the prediction model according to the verification results.

[0036] The method for constructing a prediction model based on machine learning of the present invention further includes providing a test sample set, and using the test sample set to perform testing, thereby evaluating the prediction model.

[0037] A system for building predictive models based on machine learning One aspect of the present invention provides a prediction model for esophageal and gastric varices disease constructed by the above method.

[0038] Another aspect of the present invention provides a system for building a prediction model based on machine learning (also referred to as a "device for building a prediction model based on machine learning"), comprising: A data acquisition unit, configured to acquire data from a training sample set and an optional validation sample set and a test sample set, wherein the training sample set comprises a training sample set of multiple samples, and each sample comprises basic personal information, test data and a label of a subject, the subject comprises a patient with liver cirrhosis, the basic personal information and test data are obtained by screening by introducing a penalty coefficient into a regression model, and the label is esophageal and gastric varices disease; A construction unit includes an XGB machine learning model and is configured to input the personal basic information and detection data in a training sample set into the XGB machine learning model for training. Optionally, it is further configured to use the verification sample set to verify the trained prediction model, thereby adjusting or optimizing the prediction model according to the verification result. Optionally, it further includes testing with a test sample set to evaluate the prediction model.

[0039] In the system of the present invention for constructing a prediction model based on machine learning, the data acquisition unit can be any suitable device or instrument capable of collecting parameter data selected from age, spleen thickness, platelets, albumin, prealbumin and cholinesterase of the subject, such as an electronic information storage unit containing basic patient information, medical records, follow-up data, etc., or a kit or instrument capable of measuring or quantifying spleen thickness, platelets, albumin, prealbumin and cholinesterase. An exemplary data acquisition unit includes a multi-mode input port. The multi-mode input port can receive information data including the age, spleen thickness, platelets, albumin, prealbumin and amount of cholinesterase of the subject from different sources, including data from online and offline sources. Another exemplary data acquisition unit includes a port connected to a serological index measuring instrument (e.g., a chemiluminescent immunoassay analyzer).

[0040] In the system for constructing a prediction model based on machine learning of the present invention, the construction unit includes an XGB machine learning model and is configured to be able to input the personal basic information and detection data in the training sample set into the XGB machine learning model for training. Optionally, it is further configured to be able to use the verification sample set to verify the trained prediction model, so as to adjust or optimize the prediction model according to the verification result. Optionally, it further includes testing with a test sample set to evaluate the prediction model.

[0041] In the system for constructing a prediction model based on machine learning of the present invention, the form of the construction unit is not limited, as long as it includes the prediction model of the present invention. Exemplarily, the construction unit is a processor. Preferably, the construction unit is communicatively connected with the data acquisition unit, so that the data in the data acquisition unit can be retrieved and input into the construction model for calculation.

[0042] System for predicting esophageal and gastric varices in subjects The present invention further provides a system for predicting esophageal and gastric varices in a subject (or a device for predicting esophageal and gastric varices in a subject), comprising: A data acquisition unit, which is configured to acquire basic personal information and test data of a subject, wherein the subject includes a patient with liver cirrhosis, and the basic personal information and test data include parameter data of age, spleen thickness, platelets, albumin, prealbumin and cholinesterase; The prediction unit is configured to input the data acquired by the data acquisition unit into a prediction model and output a prediction result. The prediction model is obtained according to the method of the present invention.

[0043] In the system for predicting esophageal and gastric varices of a subject of the present invention, the data acquisition unit can be any suitable device or instrument capable of collecting parameter data selected from age, spleen thickness, platelets, albumin, prealbumin and cholinesterase of the subject, such as an electronic information storage unit containing basic patient information, medical records, follow-up data, etc., or a kit or instrument capable of measuring or quantifying spleen thickness, platelets, albumin, prealbumin and cholinesterase. An exemplary data acquisition unit includes a multi-mode input port. The multi-mode input port can receive information data including the age, spleen thickness, platelets, albumin, prealbumin and amount of cholinesterase of the subject from different sources, including data from online and offline sources. Another exemplary data acquisition unit includes a port connected to a serological index measuring instrument (e.g., a chemiluminescent immunoassay analyzer).

[0044] In the system for predicting esophageal and gastric varices of a subject of the present invention, the prediction unit comprises an XGB prediction model obtained by the construction method of the present invention, which is configured to input the data acquired by the data acquisition unit into the prediction model and output a prediction result.

[0045] In the system for predicting esophageal and gastric varices in a subject of the present invention, the form of the prediction unit is not limited, as long as it includes the prediction model of the present invention. Exemplarily, the prediction unit is a processor. Preferably, the prediction unit is communicatively connected with the data acquisition unit, so that the data in the data acquisition unit can be retrieved and input into the prediction model for calculation.

[0046] Device and computer readable storage medium The present invention further provides a device, wherein the device comprises: a memory, a processor, and a program stored in the memory and capable of running on the processor, wherein the program is configured to implement a method for constructing a prediction model or steps of a prediction method.

[0047] The present invention also provides a computer storage medium (readable storage medium) or cloud for use by an instruction execution system, device or equipment (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or equipment and execute instructions), or in combination with these instruction execution systems, devices or equipment. For the purposes of the present invention, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or equipment, or in combination with these instruction execution systems, devices or equipment. More specific examples of computer-readable media include the following: an electrical connection with one or more wirings (electronic devices), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or processing in other suitable ways if necessary, and then stored in a computer memory.

[0048] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0049] A person of ordinary skill in the art may understand that all or part of the steps in the method for implementing the above-mentioned implementation scheme may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one of the steps of the method embodiment or a combination thereof.

[0050] In addition, each functional unit in each embodiment of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into one module. The above-mentioned integrated module can be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. The above-mentioned storage medium can be a read-only memory, a disk or an optical disk, etc.

[0051] Example 1. Research subjects and methods 1.1 Research subjects This example included 578 patients with cirrhosis from 2006 to 2015 in the Fifth Medical Center of the PLA General Hospital, who met the requirements of the Guidelines for the Diagnosis and Treatment of Cirrhosis (2019) and had a diagnosis of GOV, and collected patients from the Fifth Medical Center from May to October 2024 for external validation.

[0052] Exclusion criteria: 1. Exclude cases without endoscopic examination: CT, MRI, blank results (endoscopy is the gold standard for GOV) (53 cases); 2. Exclude cases where varicose veins disappeared after treatment such as band ligation, tips, and sclerotherapy (6 cases); 3. Exclude duplicate cases: duplicate case numbers (5 cases); 4. Exclude cases with unclear definitions: esophageal varicose veins have a tendency to varicose veins and cannot be graded, and only gastric varices have no varicose veins and cannot be graded (46 cases). Finally, there were 468 patients with cirrhosis, including hepatitis B, hepatitis C, alcoholic, primary biliary, drug-induced, autoimmune cirrhosis, etc. The data processing flow chart and the distribution of causes of cirrhosis are shown in the figure. Figure 1 shown.

[0053] 1.2 Endoscopic examination GOV grading criteria: ① no varicose veins; ② mild: GOV is slightly tortuous and has no red sign; ③ moderate: GOV is slightly tortuous and has red sign; or serpentine; ④ severe: GOV is serpentine and has red sign; or is beaded, nodular or tumorous (with or without red sign).

[0054] GOV grouping: Patients with cirrhosis were divided into GOV group, including mild, moderate and severe GOV (360 cases) and non-GOV group (108 cases). There were 71 patients with GOV and 40 patients without GOV in the external validation group.

[0055] 1.3 Machine Learning Model Selection This embodiment first selects 6 models, including Random forest (RF), Support Vector Machine (SVM), Gradient Boosting Classification (GBC), eXtreme Gradient Boosting (XGB), Gaussian Naive Bayes (GNB), K-Nearest Neighbor (KNN), for predicting cirrhosis GOV. RF is an integrated learning algorithm that improves prediction accuracy by constructing multiple decision trees and combining them together. First, samples are randomly selected from the training set to form the training set of the tree using the self-service sampling method. After obtaining the random forest, a new sample is input, and each tree is judged separately and voted to determine the category to which the sample belongs. If a category is predicted by most trees, the category is designated as the final prediction category of the sample. The core idea of ​​SVM is to find the maximum distance between the line or hyperplane where the two closest points of variables of different categories are located, so as to separate different data points, and solve problems such as classification and regression. GBC is an ensemble learning method that builds a strong predictor by combining multiple weak predictors (usually decision trees). Each subsequent predictor attempts to correct the errors of the previous predictor, thereby gradually improving the accuracy and generalization ability of the model. XGB is a machine learning framework based on the gradient boosting decision tree algorithm, which aims to create a strong learner by integrating a large number of weak learners (such as decision trees). The objective function of XGB includes a loss function and a regularization term (L2 regularization). The loss function is used to measure the accuracy of the model's predictions, while the regularization term is used to control the complexity of the model and prevent overfitting. XGB uses Taylor series to expand the loss function to the second order in order to more accurately estimate the objective function and optimize more effectively. GNB is a machine learning model based on probability and Gaussian distribution. It assumes that the features are independent of each other and that the features also conform to the Gaussian distribution, that is, the probability density function of a continuous random variable obeys a normal distribution with a mean of μ and a variance of δ². GNB calculates the posterior probability by multiplying the prior probability and conditional probability of each feature, and obtains the probability of each category. Finally, the category with the largest probability is selected as the predicted category. The basic principle of KNN: When predicting a new variable, the category of the new variable is determined based on the category of the K points closest to it. In regression analysis, the KNN algorithm usually takes the average of the K nearest samples as the output result.

[0056] 1.4 Data Statistical Analysis Patient data were presented as continuous or categorical variables. The Kolmogorov-Smirnov test was used to assess whether the data followed a normal distribution. For normally distributed continuous variables, data were described as mean ± standard deviation and compared using the t-test. If continuous variables did not follow a normal distribution, the Mann-Whitney U test was used, and the results were expressed as median (interquartile range). Categorical data were presented as numbers and frequencies and compared using the chi-square test or Fisher's exact test. All statistical tests were two-sided, and p values ​​< 0.05 were considered statistically significant. SPSS software version 25.0 (IBM SPSS Statistics 25) was used for analysis.

[0057] 30 variables were selected as the total data set, with GOV as the dependent variable, and the independent variables included: general condition: gender, age, BMI; liver function indicators: alanine aminotransferase, aspartate aminotransferase, alkaline phosphatase, γ-glutamyl transpeptidase, total bile acid, cholinesterase, albumin, globulin, prealbumin, direct bilirubin, total bilirubin, Golgi protein, alpha-fetoprotein, alpha-fetoprotein isomers; coagulation indicators: prothrombin time, prothrombin time activity, prothrombin time international normalized ratio, fibrinogen; blood routine indicators: red blood cells, hemoglobin, white blood cells, neutrophils, lymphocytes, monocytes, platelets; imaging indicators: portal vein diameter, spleen thickness. First, LASSO regression was used to screen variables, and 20% of the screened data was randomly selected as the test set. The remaining data was put into the machine learning model, and the trained model was evaluated and adjusted to the appropriate parameters through 5-fold cross validation. Finally, the test set data was put into the adjusted model for testing. K-fold cross-validation splits the data K times, and each time a part is selected as the validation set, and the remaining part is the training set. In this way, the evaluation results of K validation sets are obtained, and each part can be used as a validation, which reduces the random error of the data set and improves the generalization ability of the model.

[0058] ROC curves were used to compare models, and sensitivity, specificity, positive predictive values ​​(PPV), negative predictive values ​​(NPV) were used to evaluate model effects. Decision Curve Analysis (DCA) was used to compare the benefits of different prediction models at different treatment decision thresholds.

[0059] 2. Results 2.1 Comparison of basic conditions of patients Among the 468 patients with liver cirrhosis, 360 patients had GOV (including 118 patients with mild GOV, 118 patients with moderate GOV, and 124 patients with severe GOV), including 257 males (79.10%) and 103 females (72.00%). There were 108 patients without GOV, including 68 males (20.90%) and 40 females (28.00%). The age of patients with GOV was 52 years old, while that of patients without GOV was 48 years old, with a statistically significant difference (P < 0.05). The results of univariate analysis showed that there was no statistically significant difference in gender, BMI, AFPL3, MONO, and ALT between the two groups of patients (P>0.05). There were statistically significant differences in age, AST, ALP, GGT, TBA, CHE, ALB, GLB, PA, DBiL, TBiL, GP73, AFP, PT, PR, INR, FIB, RBC, HGB, WBC, NEUT, LY, PLT, PVD, and ST between the two groups (P<0.05), as shown in Table 1.

[0060] Table 1 Comparison of basic conditions of patients Note: *The difference is statistically significant at the α=0.05 level.

[0061] ALT: alanine aminotransferase; AST: aspartate aminotransferase; ALP: alkaline phosphatase; GGT: gamma-glutamyl transpeptidase; TBA: total bile acid; CHE: cholinesterase; ALB: albumin; GLB: globulin; PA: prealbumin; DBiL: direct bilirubin; TBiL: total bilirubin; PT: prothrombin time; PR: prothrombin time activity; INR: prothrombin time international normalized ratio; FIB: fibrinogen; RBC: red blood cell; HGB: hemoglobin; WBC: white blood cell; NEUT: neutrophil; LY: lymphocyte; MONO: monocyte; PLT: platelet; GP73: Golgi protein; AFP: alpha-fetoprotein; AFP-L3: alpha-fetoprotein variant; PVD: portal vein diameter; ST: spleen thickness.

[0062] 2.2 LASSO regression variable selection With GOV as the dependent variable, a total of 30 independent variables were included. LASSO regression was performed on all independent variables. Figure 3 It can be seen that as the penalty coefficient λ increases, the coefficients of the included independent variables are compressed, and some independent variable coefficients are compressed to 0, avoiding overfitting. Use ten-fold cross validation to select the best penalty coefficient λ. lambda.min refers to the lambda value with the smallest deviation (0.004), and lambda.1se is the lambda value (0.049) that obtains the simplest model within a variance range of lambda.min ( Figure 4 ). In this embodiment, lambda.1se value is selected, and the variables finally screened are age (Age), spleen thickness (ST), platelets (PLT), albumin (ALB), prealbumin (PA), and cholinesterase (CHE).

[0063] 2.3 Comparison of the selected variables with and without GOV and in four different degrees of GOV In the study, the age (52 years) and ST (52.00 mm) of patients with GOV were higher than those without GOV (48 years and 40.00 mm). The PLT (60.00 × 10^9 / L), ALB (31.00 g / L), PA (91.50 mg / L), and CHE (3157.00 U / L) of patients with GOV were lower than those without GOV (101.00 × 10^9 / L, 40.00 g / L, 167.00 mg / L, and 6125.00 U / L, respectively) (Table 1). The age of patients without GOV (48 years) was lower than that of patients with mild (52 years), moderate (51 years), and severe (52 years) GOV (P < 0.05) ( Figure 5 Figure A). The ST of the four groups was significantly different (P < 0.05). The ST of patients without GOV (40.00 mm) was the smallest. The ST increased from the no GOV group to the severe GOV group (40.00 mm, 48.00 mm, 53.00 mm, and 54.50 mm, respectively) ( Figure 5 Figure B). The PLT level of patients without GOV (101.00×10^9 / L) was higher than that of patients with mild (67.50×10^9 / L), moderate (59.00×10^9 / L), and severe (54.00×10^9 / L) GOV. The PLT level of patients with mild GOV was higher than that of patients with severe GOV (P<0.05) ( Figure 5 The levels of ALB (40.00 g / L), PA (167.00 mg / L), and CHE (6125.00 U / L) in patients without GOV were higher than those in the other three groups (mild GOV: 31.00 g / L, 99.00 mg / L, 3306.00 U / L, moderate GOV: 32.00 g / L, 86.50 mg / L, 2999.00 U / L, severe GOV: 31.00 g / L, 92.00 mg / L, 3194.50 U / L) (P < 0.05) ( Figure 5 Figures D, E, and F).

[0064] 2.4 Comparison of Machine Learning Results The above 6 variables were put into 6 machine learning models. The AUC values ​​of the 6 models in the training set ranged from 0.8791 to 0.9776 ( Figure 6The sensitivity ranged from 58.70% to 96.09%, the specificity ranged from 75.71% to 100.00%, the positive predictive value ranged from 92.50% to 100.00%, and the negative predictive value ranged from 42.40% to 87.70% (Table 2). The AUC values ​​of the six models in the test set ranged from 0.7414 to 0.9423 ( Figure 6 The sensitivity range was 74.29%-98.57%, the specificity range was 45.83%-100.00%, the positive predictive value range was 84.10%-100.00%, and the negative predictive value range was 56.10%-91.70% (Table 3). In the external validation, the AUC values ​​of the six models ranged from 0.5000 to 0.8384, and the XGB model performed better ( Figure 6 (Figure C), sensitivity range: 35.21%-100.00%, specificity range: 0.00%-92.50%, positive predictive value range: 75.80%-91.40%, negative predictive value range: 36.00%-75.00% (Table 4). In summary, the XGB model has good predictive efficiency.

[0065] Table 2 Performance evaluation of machine learning models on the training set Table 3 Performance evaluation of different machine learning models on the test set Table 4 Effectiveness evaluation of different machine learning models in external validation Note: RF: random forest; SVM: support vector machine; XGB: XGBoost; GNB: Gaussian naive Bayes; KNN: K-nearest neighbor algorithm; PPV: positive predictive value; NPV: negative predictive value.

[0066] 2.5 Comparison of the XGB model with liver stiffness <20 kPa (LSM), platelet count >150×10^9 (PLT), and the combination of the two (LP) The performance of the XGB model in the test set (AUC=0.9423) is higher than that of LSM (AUC=0.7609), PLT (AUC=0.6005) and LP (AUC=0.7860) ( Figure 7 Figure A), the performance of the XGB model in external validation was also good (AUC=0.8384) ( Figure 7Figure B). The AUC for predicting GOV efficacy of LSM was 0.7609 (95%CI: 0.6681-0.8536, sensitivity: 75.25%, specificity: 76.92%, positive predictive value: 92.70%, negative predictive value: 44.40%). The AUC for predicting GOV efficacy of PLT was 0.6005 (95%CI: 0.5163-0.6848, sensitivity: 97.03%, specificity: 23.08%, positive predictive value: 83.10%, negative predictive value: 66.70%). The AUC for predicting GOV efficacy of the two combined was 0.7860 (95%CI: 0.6873-0.8847, sensitivity: 75.25%, specificity: 76.92%, positive predictive value: 92.70%, negative predictive value: 44.40%) (Table 5).

[0067] Table 5 Comparison of XGB model with LSM, PLT and LP XGBtest: XGB model in the test set; XGBval: XGB model in the external validation set; LSM: liver stiffness measurement >20 kPa; PLT: platelet count >150×10^9; LP: liver stiffness measurement >20 kPa and platelet count >150×10^9; PPV: positive predictive value; NPV: negative predictive value.

[0068] 2.6 Evaluation of the performance of GOV prediction in the XGB model by clinical decision curve analysis The reference line "All" in DCA refers to the net benefit obtained by intervening on all people at different risk thresholds. As the risk threshold increases, the net benefit decreases. The reference line "None" means no intervention for all people, and no benefit. The intersection of the two reference lines is the actual GOV prevalence rate of the data in this study. In the test set, the actual prevalence rate is 76.92%. When the risk threshold is higher than 0.2, the net benefit of intervention is higher than the reference line ( Figure 8 Figure A), the actual prevalence in external validation was 63.96%, and the net benefit of intervention was higher than the reference line when the risk threshold was higher than 0.4 ( Figure 8 (Figure B), the threshold interval is relatively large, and both have good clinical benefits.

[0069] 3. Discussion GOV is the most common complication in patients with cirrhosis. Early and accurate judgment of whether there is GOV can prevent the disease from developing in a worse direction. However, the current gold standard for GOV assessment is invasive, which can easily lead to poor patient compliance. Therefore, a non-invasive, convenient, reliable, accurate and safe method is needed for early detection. With the advancement of machine learning, clinicians can now build feasible models with large amounts of data to improve the performance of disease prediction. More and more studies have established ML models to predict the risk of GOV and EVB. For example, Azadeh Bayani et al. found that the prediction efficiency of ML for each level of GOV was higher than that of the regression method, and the RF model had better prediction results for each level (AUC value 0.990), which was better than the SVM model (AUC value 0.860) and the ANN model (AUC value 0.800). Chul-Min Lee et al. obtained the ratio of spleen volume and platelets based on deep learning model analysis to help detect high-risk esophageal varices, and the model had high sensitivity (69.4%) and specificity (78.5%). Zhang Qun et al. developed a neural network model to construct a prediction model for rebleeding in EVB patients. The AUC value of the prediction model was 0.782, which was significantly better than the COX regression model (AUC value 0.672), Childpugh score (AUC value 0.557) and MELD score (AUC value 0.562).

[0070] Compared with traditional methods, ML models show good performance in disease prediction. In this embodiment, a ML-based model is established to predict the occurrence of GOV in patients with cirrhosis. The training set is cross-validated to adjust the parameters. Multiple adjustments in this process will cause more and more information to be leaked, and a better performing model will be obtained, but the generalization ability is weak. Therefore, a test set is needed to evaluate the effect of the model to improve the generalization ability of the model. In the test set, the XGB model showed good predictive ability (AUC=0.9423, 95%CI: 0.8980-0.9865, sensitivity: 81.43%, specificity: 100.00%, positive predictive value: 100.00%, negative predictive value: 65.00%).

[0071] Baveno VII consensus states that: cACLD patients with LSM < 20 kPa and PLT > 150 × 10 9 / L can be exempted from internal diameter screening GOV, but according to the "Chinese Ultrasound Elastic Imaging Technology Diagnosis of Portal Hypertension in Cirrhosis Expert Consensus (2023 Edition)", this standard can avoid gastroscopy screening for cACLD patients, but it is not suitable for patients with cirrhosis of all causes. It is still necessary to explore a prediction model suitable for patients with cirrhosis. In the data of this example, the causes of cirrhosis patients include hepatitis B, hepatitis C, alcoholic, primary biliary, drug-induced, autoimmune cirrhosis, etc. 127 patients had liver stiffness measured, of which LSM <20kPa and PLT>150×10 9 Only 6 patients (4.72%) with 95% CI / 2 / L had no GOV. The combined predictive efficacy of the two for GOV in patients with cirrhosis was AUC 0.7860. The AUC of the XGB model was 0.9423, and the specificity was the highest, at 100.00%. In addition, in the external validation data, the test efficiency of the XGB model was also good (AUC = 0.8384, 95% CI: 0.7627-0.9141, sensitivity: 81.69%, specificity: 80.00%, positive predictive value: 87.90%, negative predictive value: 71.10%), and the DCA curve of the XGB model in the test set and external validation was above the curve of the reference strategy. The model had a high clinical benefit, which helped clinicians to detect GOV early, quickly and conveniently, so as to facilitate further examination and auxiliary treatment decisions.

[0072] This embodiment screened a total of 6 features through the screening process, including age, ST, PLT, ALB, PA, and CHE. The portal vein pressure of patients with cirrhosis increases, the blood return of the spleen is blocked, and blood stasis causes the spleen to enlarge. The spleen plays a major role in predicting GOV. The spleen thickness also plays an important role by measuring the spleen hardness, diameter, volume, etc. and combining other indicators for prediction. In this embodiment, the spleen thickness (ST) is displayed by the intercostal oblique section. ST is different in different GOV grades. The more severe the GOV, the larger the ST. Increased portal vein pressure in patients with cirrhosis is likely to lead to aggravation of GOV and ST enlargement. In the embodiment, the PLT of the non-GOV group is higher than that of the other three groups. In addition, the PLT of the mild GOV group is higher than that of the severe GOV group. A large amount of PLT is stored in the spleen, and the spleen is hyperfunctional after enlargement, resulting in a decrease in PLT. The variables screened in the study also include ALB, PA, and CHE, which may be affected by liver synthesis and catabolism. This example also found that the age of patients with and without GOV was different. The average age when GOV occurred was 52 years old, which was higher than that of patients without GOV.

[0073] In summary, the present invention has established a novel XGB model, which can predict the occurrence of GOV in liver cirrhosis and has a high predictive ability. The present invention uses DCA to confirm the feasibility of the XGB model in clinically evaluating GOV. The XGB model constructed by the present invention has a high accuracy and can be used as a non-invasive tool for predicting GOV in patients with liver cirrhosis.

[0074] Although the present invention has been described with reference to exemplary embodiments, it should be understood that the present invention is not limited to the disclosed exemplary embodiments. Various adjustments or changes may be made to the exemplary embodiments of the present specification without departing from the scope or spirit of the present invention. The scope of the claims should be based on the broadest interpretation to cover all modifications and equivalent structures and functions.

Claims

1. A method for constructing a prediction model based on machine learning, characterized in that: The following steps are involved: (1) providing a training sample set comprising a plurality of samples, wherein each of the samples comprises basic personal information and test data of a subject and a label, wherein the subject comprises a patient with liver cirrhosis, the basic personal information and test data are obtained by screening by introducing a penalty coefficient into a regression model, and the label is esophageal and gastric varices disease; (2) Inputting the personal basic information and the test data into the XGB machine learning model for training to obtain a prediction model for esophageal and gastric varices in patients with liver cirrhosis.

2. The method for constructing a prediction model based on machine learning according to claim 1, characterized in that: The number of samples in the training sample set is greater than 50.

3. The method for constructing a prediction model based on machine learning according to claim 1, characterized in that: It further includes providing a validation sample set, and using the validation sample set for validation, thereby adjusting or optimizing the prediction model according to the validation results.

4. The method for constructing a prediction model based on machine learning according to claim 1, characterized in that: The method further includes providing a test sample set, and using the test sample set to perform testing, thereby evaluating the prediction model.

5. The method for constructing a prediction model based on machine learning according to claim 1, characterized in that: The basic personal information and test data include parameter data of age, spleen thickness, platelets, albumin, prealbumin and cholinesterase.

6. A system for building a prediction model based on machine learning, characterized in that: include: A data acquisition unit, configured to acquire data from a training sample set and an optional validation sample set and a test sample set, wherein the training sample set comprises a training sample set of multiple samples, and each sample comprises basic personal information, test data and a label of a subject, the subject comprises a patient with liver cirrhosis, the basic personal information and test data are obtained by screening by introducing a penalty coefficient into a regression model, and the label is esophageal and gastric varices disease; A construction unit includes an XGB machine learning model and is configured to input the personal basic information and detection data in a training sample set into the XGB machine learning model for training.

7. The system for constructing a prediction model based on machine learning according to claim 6, characterized in that: The construction unit is further configured to use the verification sample set to verify the trained prediction model, thereby adjusting or optimizing the prediction model according to the verification result; the construction unit further uses the test sample set to test, thereby evaluating the prediction model.

8. A system for predicting esophageal and gastric varices in a subject, characterized in that: include: A data acquisition unit, which is configured to acquire basic personal information and test data of a subject, wherein the subject includes a patient with liver cirrhosis, and the basic personal information and test data include parameter data of age, spleen thickness, platelets, albumin, prealbumin, and cholinesterase; A prediction unit is configured to input the data acquired by the data acquisition unit into a prediction model and output a prediction result, wherein the prediction model is obtained according to the method according to any one of claims 1-5.

9. A device, characterized in that: The device comprises: a memory, a processor, and a program stored in the memory and capable of running on the processor, wherein the program is configured to implement the steps of the method according to any one of claims 1 to 5 or claim 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a program, and when the program is executed by a processor, the steps of the method according to any one of claims 1 to 5 or claim 8 are implemented.

Images