A gradient-increment-based vte risk prediction method and system

Abstract

The application belongs to the technical field of intelligent medical big data disease risk prediction, and provides a VTE risk prediction method and system based on gradient increment, which comprises the following steps: collecting sample characteristic data groups through a hospital system; pre-processing the collected sample characteristic data; performing correlation analysis on the processed sample characteristic data; calculating the risk threshold of numerical sample characteristics; establishing a machine learning model by using a gradient increment algorithm; testing the prediction effect of the model; and performing VTE risk prediction by using the gradient increment model. The application combines the Pearson correlation coefficient and the gradient increment model feature importance sorting to double-filter the characteristic data related to VTE disease, simultaneously capture linear and nonlinear correlation characteristics, and solve the prediction deviation and insufficient interpretability caused by one-sided feature filtering and threshold setting deviating from clinical practice of existing machine learning models.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent medical big data disease risk prediction, and particularly relates to a VTE risk prediction method and system based on gradient increment. BACKGROUND

[0002] Venous thromboembolism (VTE) includes deep vein thrombosis (DVT) and pulmonary embolism (PE), and is the third leading cause of vascular death worldwide, often occurring in surgical patients, long-term bedridden patients and people with a certain lifestyle. Its early symptoms are occult, and if not timely intervention, it may lead to fatal complications such as pulmonary embolism, so accurate risk prediction is crucial for clinical decision-making. At present, the risk assessment scale widely used in clinical evaluation mainly relies on the subjective score of medical staff on patient symptoms and medical history, and assigns weights through preset rules and calculates risk levels. Although this method is simple to operate, the evaluation factors are difficult to cover complex clinical variables based on expert experience; and the weight allocation lacks data-driven optimization, resulting in significant differences in evaluation results between different medical institutions, which has obvious limitations.

[0003] The limitations of traditional evaluation methods have given rise to VTE prediction research based on machine learning. Although existing machine learning models can handle nonlinear relationships and integrate multidimensional data, most studies only rely on linear statistical methods such as Pearson correlation coefficient to screen features, which may miss key factors that are not linearly associated with VTE, and the feature selection is one-sided; at the same time, the risk threshold of numerical features such as D-dimer is calculated independently based on data distribution, without combining the diagnostic criteria in clinical guidelines, and its prediction results are out of line with the actual diagnosis and treatment process, so the existing machine learning models lack medical interpretability and have low clinical acceptance. SUMMARY

[0004] To solve the problem of one-sided feature selection and low clinical acceptance caused by insufficient model interpretability of existing machine learning-based VTE risk prediction methods, the application provides a VTE risk prediction method based on gradient increment.

[0005] In one scheme, a VTE risk prediction method based on gradient increment, the specific steps are as follows:

[0006] S1, collect sample feature data groups through hospital internal system export or API access;

[0007] S2, delete irrelevant features, process missing values, convert variable types, and normalize the collected sample feature data;

[0008] S3, perform VTE disease correlation analysis on the processed sample feature data using Pearson correlation coefficient;

[0009] S4, calculating the risk threshold of the numerical sample characteristics, using the Youden index method, ROC curve and kernel density estimation method to calculate the risk threshold of the numerical sample characteristics;

[0010] S5, using gradient incremental algorithm to establish machine learning model and train the model;

[0011] S6, testing the prediction effect of the model and verifying the clinical consistency;

[0012] S7, inputting the individual sample characteristic data of the user, using the gradient incremental model to predict the VTE risk of the user.

[0013] Specifically, in step S2, the basic sample characteristic data related to VTE prediction obtained after deleting irrelevant features has 15 items: age, plasma D-dimer, white blood cell, platelet, hemoglobin, neutrophil percentage, hypertension, coronary heart disease, tumor, gender, previous surgery history, diabetes, blood transfusion history, smoking history, and previous comorbidity.

[0014] Among them, age, white blood cell, hemoglobin, platelet, neutrophil percentage and plasma D-dimer are numerical sample characteristics.

[0015] Further, in step S3, after performing Pearson correlation coefficient calculation on each basic sample characteristic data, 8 sample characteristics with high correlation related to VTE are obtained: age, plasma D-dimer, platelet, hemoglobin, coronary heart disease, tumor, gender and previous comorbidity.

[0016] Further, in step S4, the risk threshold of each numerical sample characteristic obtained by the Youden index method is: age 80 years old, white blood cell 8.07x109 / L, hemoglobin 78g / L, platelet 307x109 / L, neutrophil percentage 72%, and plasma D-dimer 284ng / mL DDU.

[0017] The risk threshold of each numerical sample characteristic obtained by the ROC curve is: age 80 years old, white blood cell 6.14x109 / L, hemoglobin 126g / L, platelet 222x109 / L, neutrophil percentage 68.3%, and plasma D-dimer 256ng / mL DDU.

[0018] The risk threshold of each numerical sample characteristic obtained by the kernel density estimation method is: age 80 years old, white blood cell 8.69x109 / L, hemoglobin 131.1g / L, platelet 197x109 / L, neutrophil percentage 69.8%, and plasma D-dimer 965ng / mL DDU.

[0019] Further, the risk threshold and accuracy of the numerical sample features obtained by the Youden index method, ROC curve and kernel density estimation method are compared, and the comprehensive risk threshold is obtained by combining the curve fluctuation of the kernel density curve in the corresponding numerical region: age 65 years, white blood cells 8.69x109 / L, hemoglobin 126g / L, platelets 197x109 / L, neutrophil percentage 72%, plasma D-dimer 965ng / mL DDU.

[0020] Further, in step S5, two different VTE risk prediction models are established by using XGBoost or LightGBM algorithm with 15 basic sample feature data and 8 high correlation sample feature data as input respectively;

[0021] Wherein, the initial prediction value is 0.5, and the initial probability of all testers is 50% by default, which is adjusted and accumulated by combining the sample feature data of the testers to predict the VTE disease probability of the testers.

[0022] In one scheme, the training process of the model is as follows:

[0023] Initialize the model, set the initial prediction score to 0.5 as the initial prediction value of all samples, i.e. the initial probability value, and specify to use decision tree as the base learner;

[0024] After the training starts, according to the parameter setting, no column sampling is performed, but all features are used for iteration. Each iteration calculates the gradient residual of the current model prediction value and the true label. The new tree fits these residuals by minimizing the objective function containing the first-order derivative-gradient and the second-order derivative-Hessian matrix information. Wherein, the tree node splitting is strictly controlled, and splitting is only performed when the loss reduction brought by splitting is at least 4.5; the sample weight of the child node is limited to 1 to prevent overfitting; the leaf node weight is smoothed according to the default L2 regularization of the algorithm;

[0025] After the tree construction is completed, the contribution of the new tree is scaled according to the learning rate 0.1 to prevent overfitting, and the limit change of the leaf node weight is further constrained to enhance the stability of the model;

[0026] Repeat the iteration until the performance of the validation set no longer improves, and stop the training;

[0027] Finally, the prediction results of all trees are weighted and combined to form the final model.

[0028] Further, in step S6, the prediction effect of the model is evaluated by using four evaluation indexes of ROC curve, AUC value, Matthew correlation coefficient and Kappa coefficient.

[0029] In addition, to achieve the above-mentioned purpose, the application also provides a gradient increment-based VTE risk prediction system, which is used to implement the gradient increment-based VTE risk prediction method, and the system comprises:

[0030] A data acquisition module is configured to acquire a group of basic sample feature data for VTE prediction through a hospital internal system or an API.

[0031] A data preprocessing module is configured to perform irrelevant feature deletion, missing value processing, variable type conversion and normalization processing.

[0032] A sample feature analysis module comprises a correlation analysis unit configured to quantitatively analyze the correlation between the basic sample feature data and VTE, and a risk threshold calculation unit configured to calculate a threshold value for a numerical sample feature.

[0033] A model construction and evaluation module is configured to construct a machine learning model based on a gradient increment algorithm and perform performance evaluation on the model based on the output data of the sample feature analysis module.

[0034] A risk prediction module is configured to receive individual feature data of a testee, and calculate a VTE risk probability by calling the gradient increment model.

[0035] Further, the data acquisition module is further configured with an anonymization processing unit and adopts an encryption transmission mechanism; when acquiring the basic sample feature data, direct identification data related to patient personal information is deleted or hashed, and quasi-identifiers related to indirect association identity features are desensitized.

[0036] The one or more technical solutions provided in the application have at least the following technical effects or advantages:

[0037] In the application, the Pearson correlation coefficient and the gradient increment model feature importance ranking are combined to double-filter the feature data related to VTE, and linear and nonlinear correlation features are captured; the Youden index method, the ROC curve and the kernel density estimation method are used to fuse the risk threshold calculated based on the clinical guideline standard, to realize dynamic calibration of numerical features, and to avoid single data-driven threshold from deviating from the diagnosis and treatment practice; the machine learning model constructed based on the gradient increment algorithm automatically learns the feature weight and adapts to the data distribution, the model decision logic is converted into interpretable medical evidence through feature importance analysis, to assist doctors in decision-making, and to provide intelligent prediction results with accuracy and clinical practicability for early screening of VTE. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0039] Figure 1 is a method flow chart of the VTE risk prediction method based on gradient increment in an embodiment of the present application;

[0040] Figure 2 is a heat map of the correlation analysis in an embodiment of the present application;

[0041] Figure 3 is a density curve chart of the age in an embodiment of the present application;

[0042] Figure 4 is a density curve chart of the white blood cells in an embodiment of the present application;

[0043] Figure 5 is a density curve chart of the hemoglobin in an embodiment of the present application;

[0044] Figure 6 is a density curve chart of the platelets in an embodiment of the present application;

[0045] Figure 7 is a density curve chart of the neutrophil percentage in an embodiment of the present application;

[0046] Figure 8 is a density curve chart of the plasma D-dimer in an embodiment of the present application;

[0047] Figure 9 is a model prediction performance ROC curve chart in an embodiment of the present application, inputting 15 basic sample characteristics;

[0048] Figure 10 is a model prediction performance ROC curve chart in an embodiment of the present application, inputting 8 high-correlation sample characteristics. DETAILED DESCRIPTION

[0049] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. The modules of the embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0050] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description of the application herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0051] In order to solve the problem that the existing machine learning-based VTE risk prediction method has one-sided feature screening and insufficient model interpretability, resulting in low clinical acceptance, the present application provides a gradient-increment-based VTE risk prediction method, please refer to Figure 1 , which comprises the following steps:

[0052] S1, collecting sample feature data groups through hospital internal system export or access API.

[0053] In step S1, the original sample feature data group with medical element information is obtained from the hospital by using the hospital internal system or the access API, which provides source data for subsequent correlation analysis and prediction model establishment.

[0054] S2, deleting irrelevant features, handling missing values, variable type conversion and normalization processing on the collected sample feature data.

[0055] In step S2, there are many sample features irrelevant to VTE prediction judgment in the original sample feature data obtained from the hospital, such as department, patient number, etc. Before correlation analysis, the original sample feature data is preprocessed, i.e. irrelevant sample feature items are deleted, so as to obtain basic sample feature data containing only 15 items related to VTE prediction, such as age, plasma D-dimer, white blood cells, platelets, hemoglobin, neutrophil percentage, hypertension, coronary heart disease, tumor, gender, previous surgery history, diabetes, blood transfusion history, smoking history and previous comorbidities. Among them, age, white blood cells, hemoglobin, platelets, neutrophil percentage and plasma D-dimer are numerical sample features.

[0056] In step S2, there are also incomplete random missing values in the original sample feature data. In order to avoid the prediction error of the model caused by these missing values, the missing values need to be processed. Common missing value processing methods include direct deletion method, mean filling method and mode filling method, etc. In order to improve the accuracy of model prediction, the mean filling method is selected to process the missing values. For example, the missing values in the sample feature "coronary heart disease" column are filled with the mean value 1.59396 of this column.

[0057] Further, the gender, previous comorbidity, previous surgery history, blood transfusion history and smoking history in the basic sample feature data also need to be converted into variable types. The application adopts binary variables for conversion, and the specific conversion results are shown in Table 1.

[0058] Table 1 Conversion results of basic sample feature variable types

[0059]

[0060] Further, the numerical sample feature data such as age and neutrophil percentage are normalized to convert the characteristic values of the data into a specific interval [0, 1] to eliminate the influence of the dimensions of different characteristics, so that the numerical sample feature data has numerical comparability, which helps to improve the performance and convergence speed of the machine learning model. Specifically, the minimum-maximum normalization processing method is adopted to linearly map the values of the numerical sample feature data to the range of [0, 1], while preserving the original distribution characteristics of the numerical sample feature data.

[0061] S3, Pearson correlation coefficient is used to analyze the correlation of VTE disease of the sample feature data processed.

[0062] In step S3, correlation is used to reflect the degree of association between each sample feature and VTE disease. In machine learning, it can be used to select features that have important influence on target variables. Sample features with high correlation may have a greater contribution to the prediction ability of the model, while sample features with low correlation can be removed to improve the performance and training efficiency of the model. Specifically, correlation is usually measured by correlation coefficient. The application uses sample Pearson correlation coefficient to measure the correlation of the basic sample features. The Pearson correlation coefficient is a commonly used statistical indicator for measuring the degree of linear association between two continuous variables. The correlation heat map of 15 basic sample features obtained by calculating the Pearson correlation coefficient matrix is shown in Figure 2 .

[0063] Further, the weight of each sample feature is measured based on the contribution of each basic sample feature to the impurity reduction of the decision tree. Each sample feature has an impurity reduction amount when it is split at each node of the decision tree. The reduction amount is weighted, and then the reduction amounts of all nodes are summed up. Finally, the weights of the basic sample features are obtained after normalization. Specifically, the correlation and weight of each basic sample feature and VTE prediction are obtained by analyzing and calculating 15 basic sample features. See Table 2 for details.

[0064] Table 2: Summary of correlation and weight of basic sample features

[0065]

[0066] Among them, the items marked with "*" represent sample features with high correlation with the result of whether the patient is a VTE patient. The judgment basis is the correlation coefficient > 0.05, that is, 8 sample features with high correlation with VTE are obtained: age, plasma D-dimer, platelet, hemoglobin, coronary heart disease, tumor, gender, and comorbidity.

[0067] S4, calculating the risk threshold of the numerical sample feature, and calculating the risk threshold of the numerical sample feature by using the Youden index method, ROC curve and kernel density estimation method.

[0068] In step S4, the risk threshold is a numerical limit to distinguish between normal and abnormal. For numerical sample features, determining the risk threshold can provide a reference for doctors and machine learning models to judge the risk of patients. The present application respectively uses the Youden index method, ROC curve and kernel density estimation method to calculate the risk threshold of 6 numerical sample features, and comprehensively analyzes and selects the most suitable risk threshold.

[0069] The Youden index is an important indicator for evaluating the accuracy of a binary classification diagnostic test in the fields of medicine and statistics, also known as the correct index. It measures the ability of a test to distinguish between "sick" and "healthy" populations by considering the sensitivity and specificity of the diagnostic test. The formula for calculating the Youden index is: Youden index = sensitivity + specificity - 1. The ROC curve is an important tool for evaluating the performance of a binary classification model in the fields of medical statistics and machine learning. It visualizes the relationship between the true positive rate (sensitivity) and the false positive rate (specificity) of a model at different thresholds, helping researchers choose the optimal decision threshold and compare the diagnostic abilities of different models. The ROC curve plots the false positive rate (FPR) on the x-axis and the true positive rate (TPR) on the y-axis. The point in the upper left corner represents the ideal classification state, where FPR is 0 (no false positives) and TPR is 1 (all positive samples are correctly classified). In the context of determining whether a person is sick, this means accurately identifying all sick individuals without misdiagnosing healthy people.

[0070] Specifically, first, the ROC curve of each basic sample feature is drawn. The sensitivity and specificity change with the threshold in the ROC curve. By calculating the Youden index at different thresholds, the threshold corresponding to the maximum Youden index is found, which can be used as a better risk threshold. After obtaining the risk threshold, the accuracy of the original sample data is verified. The results obtained by the Youden index method are shown in Table 3.

[0071] Table 3: Risk threshold results of Youden index method

[0072]

[0073] Further, since it is difficult to reach the ideal point in the upper left corner of the ROC curve, the threshold corresponding to the ideal point closest to the upper left corner of the ROC curve is selected as a better cutoff value, i.e., the risk threshold. The selection logic of this cutoff value is to maximize the correct identification rate (TPR) of sick people while minimizing the misdiagnosis rate (FPR) of healthy people to achieve a balance. Intuitively, the closer to the upper left corner in the ROC curve, the better the comprehensive performance of the sample feature value in distinguishing between sick and healthy people. The risk threshold results obtained by the ROC curve are shown in Table 4.

[0074] Table 4: Risk threshold results of ROC curve

[0075]

[0076] Further, kernel density estimation is a non-parametric statistical method used to estimate the probability density function of a random variable based on sample data. It constructs a continuous density curve by smoothing and weighting sample points using a "kernel function."

[0077] Specifically, refer to Figures 3 to 8 , the probability density of the sample characteristic value under different risk states is calculated by kernel density estimation method respectively, and the intersection point of the density curve or the point where the density difference changes obviously is taken as the risk threshold. This method is relatively more objective, can determine the threshold value by using the distribution information of the data, and the kernel density curve is drawn according to the real sample characteristic data, which has high reference value. The risk threshold value obtained by the kernel density estimation method is shown in Table 5.

[0078] Table 5 Risk threshold value of kernel density estimation method

[0079]

[0080] Further, by Figure 3 The density curve of age can be seen that the density difference of the curve changes obviously when the age is 60 to 65 years old, and the curve appears intersection point at 80 years old, so the accuracy rate of 60, 65 and 80 years old is calculated respectively, and the result is shown in Table 6.

[0081] Table 6 Accuracy rate calculation result of age 60, 65 and 80 years old

[0082]

[0083] Further, combined with the accuracy of each threshold value and the screening range, the preferred comprehensive risk threshold of the age sample characteristic is finally determined as 65 years old, and the preferred comprehensive risk threshold of the remaining numerical sample characteristics is obtained by the same analysis and comparison, as shown in Table 7.

[0084] Table 7 Comprehensive risk threshold of numerical sample characteristics

[0085]

[0086] S5, a machine learning model is established by using gradient increasing algorithm, and the model is trained.

[0087] In step S5, after completing the data preprocessing, two machine learning models are respectively established with 15 basic sample feature data and 8 high correlation sample feature data as inputs. Gradient boosting algorithm is used to establish the model. The core principle of the gradient boosting model is to iteratively train multiple weak learners and combine the prediction results of these weak learners in an accumulative manner to form a strong learner, so as to improve the overall prediction performance. It is mainly based on the gradient ascent algorithm. In machine learning, for a differentiable objective function, the gradient ascent algorithm updates the parameters along the gradient direction of the objective function iteratively, so that the value of the objective function is constantly increasing. Therefore, the gradient boosting model can effectively use the gradient information of the objective function to find the optimal parameter value. Compared with random search and other methods, it can more quickly and accurately find the parameter value that maximizes the objective function to improve the model performance; at the same time, for complex objective functions, especially in high-dimensional parameter space, the gradient boosting algorithm can gradually optimize the objective function by constantly updating the parameters along the gradient direction, and has good performance in machine learning and prediction.

[0088] For example, the XGBoost algorithm is used to establish the gradient boosting model, and the main parameter settings are as follows:

[0089] (1) base_score = 0.5, the initial prediction score of all samples

[0090] (2) booster = 'gbtree', the type of base learner is selected as gradient boosting decision tree

[0091] (3) colsample_bylevel = 1, the sampling ratio of columns (level) in the tree construction process

[0092] (4) colsample_bynode = 1, the sampling ratio of columns (node) in the tree construction process

[0093] (5) gamma = 4.5, the minimum loss reduction required for tree node splitting

[0094] (6) learning_rate = 0.1, learning rate, controls the step size of each iteration

[0095] (7) max_delta_step = 3, limits the maximum step size of weight update

[0096] (8) min_child_weight = 1, the minimum weight sum required for child nodes

[0097] (9) objective = 'binary:logistic', set the objective function type to binary logistic regression objective function

[0098] The core of the model is to correct errors step by step with multiple small decision trees. In the initial stage, it is assumed that all people have a 50% probability of suffering from VTE, and then the mission of each tree is to correct the errors of the previous step of prediction. For example, the first tree finds that patients with "plasma D-dimer > risk threshold" are more likely to suffer from the disease, so the prediction probability of the patient is increased from 0.5 to 0.6. The second tree may find that patients with "age > 60 years old and smoking" are at higher risk, and further adjust the prediction value. The final prediction result is the cumulative adjustment value of all trees, which is converted to a probability between 0 and 1 through a logical function, i.e. the VTE disease probability of the testee.

[0099] Further, the training process of the model is as follows:

[0100] Initialize the model, set the initial prediction score to 0.5 as the initial prediction value of all samples, i.e. the initial probability value, and specify the use of decision trees as the base learner;

[0101] After the training starts, according to the parameter setting, no column sampling is performed and all features are used for iteration. Each round of iteration calculates the gradient residual of the current model prediction value and the true label, and the new tree fits these residuals by minimizing the objective function containing the first-order derivative-gradient and the second-order derivative-Hessian matrix information. Among them, the tree node splitting is strictly controlled, and splitting is only performed when the loss reduction brought by splitting is at least 4.5; the sample weight of the child node is limited to 1 to prevent overfitting; the leaf node weight is smoothed according to the default L2 regularization of the algorithm;

[0102] After the tree construction is completed, the contribution of the new tree is scaled according to the learning rate 0.1 to prevent overfitting, and the limit change of the leaf node weight is further constrained to enhance the stability of the model;

[0103] Repeat the iteration until the performance of the validation set no longer improves, and stop the training;

[0104] Finally, the prediction results of all trees are combined to form the final model.

[0105] S6, test the prediction effect of the model and perform clinical consistency verification.

[0106] In step S6, in order to evaluate the prediction accuracy of the prediction model, evaluation indicators need to be introduced to analyze the accuracy and performance of the prediction model. In this embodiment, the prediction error is generally the difference between the model prediction value and the actual value of the data set. Through the evaluation of the prediction result, it can be reasonably judged whether the prediction model has high accuracy and reliability. The present application uses four evaluation indicators, ROC curve, AUC value, Matthew correlation coefficient and Kappa coefficient, to evaluate the prediction result of the model.

[0107] Wherein, the AUC value is an important index in machine learning and statistics for evaluating the performance of a binary classification model, the area under the whole curve is the area under the ROC curve, that is, the area under the receiver operating characteristic curve, which is essentially that a positive sample and a negative sample are randomly selected, and the probability of the model predicting the positive sample as positive is greater than the probability of the negative sample, the value range is between 0.5 and 1, and the closer the AUC value is to 1, the better the classification performance of the classifier.

[0108] The Matthew correlation coefficient is also a statistical index for evaluating the performance of a binary classification model, especially suitable for unbalanced sample scenarios. It comprehensively considers the four types of results, true positive (TP), true negative (TN), false positive (FP), and false negative (FN), provides a single value that can summarize the classification quality, and can more stably reflect the classification accuracy of the model.

[0109] The Kappa coefficient is a statistical index for measuring the consistency of classification results. It considers the proportion of consistent classification results and also corrects the influence of random consistency to more objectively evaluate the reliability of the classifier, judge or model. Its value range is between -1 and 1, wherein, 1 represents perfect classification, that is, the predicted result is completely consistent with the actual classification; 0 represents that the classification accuracy is the same as random guessing; and negative value represents that the classification accuracy is lower than random guessing.

[0110] Specifically, please refer to Figure 9 and Figure 10 The evaluation results of the prediction performance of the two groups of gradient incremental models with 15 basic sample feature data and 8 high correlation sample feature data as input are shown in Table 8.

[0111] Table 8 Model prediction performance evaluation results

[0112]

[0113] As can be seen from Table 8, both groups of prediction models can achieve good prediction results, and the accuracy is high. According to the Matthew correlation coefficient and the Kappa coefficient, the model has good classification ability and high consistency with the actual label. According to Figure 9 and Figure 10 The performance of the two groups of prediction models under different thresholds is excellent, and the prediction model with 15 basic sample features has better performance and can be used as the preferred model group.

[0114] S7, input user individual sample feature data, and use the gradient incremental model to predict the VTE risk of the user.

[0115] In step S7, the individual sample feature data of the test user is received by using the evaluated gradient incremental model, and the entered data is subjected to variable type conversion and normalization processing, the model is substituted into the processed data to perform a calculation process, and then personalized reports such as the VTE disease probability, risk level, key feature deviation degree, feature importance ranking, and the like of the feedback test user are correspondingly fed back to provide support for the clinical decision of the doctor.

[0116] Therefore, the present application combines the Pearson correlation coefficient and the gradient incremental model feature importance ranking to double-screen the feature data related to VTE disease, and simultaneously capture linear and nonlinear correlation features; the risk threshold is calculated by using the Youden index method, the ROC curve, and the kernel density estimation method to fuse the clinical guideline standard, the dynamic calibration of numerical features is realized, and the single data-driven threshold is avoided to deviate from the diagnosis and treatment practice; the machine learning model constructed based on the gradient incremental algorithm automatically learns the feature weight and adapts to the data distribution, the model decision logic is converted into interpretable medical evidence through feature importance analysis, which assists the doctor in decision-making, and provides intelligent prediction results with accuracy and clinical practicability for early screening of VTE.

[0117] As an implementation scheme, the present application also provides a gradient incremental-based VTE risk prediction system, which is used to implement the gradient incremental-based VTE risk prediction method in the foregoing embodiments, and the system comprises:

[0118] The data acquisition module is used to acquire a group of basic sample feature data for VTE prediction through a hospital internal system or an API, including patient age, plasma D-dimer, white blood cell count, platelet count, hemoglobin, neutrophil percentage, hypertension, coronary heart disease, tumor, gender, surgical history, diabetes, blood transfusion history, smoking history, and comorbidity information.

[0119] Further, the data acquisition module is further configured with an anonymization processing unit and adopts an encryption transmission mechanism; when the basic sample feature data is collected, the direct identifier data related to the personal identity information of the patient, such as the patient's name, ID number, and medical insurance number, is deleted or subjected to hash processing, and the quasi-identifier of the indirect association identity feature, such as the residence address and the treatment timestamp, is subjected to desensitization processing, so as to block the leakage risk of the patient's privacy data in the collection, transmission, and storage links.

[0120] A data preprocessing module is configured to perform irrelevant feature deletion, missing value processing, variable type conversion and normalization processing. The original sample feature data is preprocessed before correlation analysis, i.e. irrelevant sample feature items are deleted to obtain basic sample feature data related to VTE prediction, including age, plasma D-dimer, white blood cells, platelets, hemoglobin, neutrophil percentage, hypertension, coronary heart disease, tumor, gender, previous surgery history, diabetes, blood transfusion history, smoking history and 15 comorbidities; at the same time, the numerical value type features are subjected to mean value interpolation, the classification features are subjected to mode interpolation, and normalization processing is performed to eliminate the influence of the dimensions of different features.

[0121] A sample feature analysis module includes a correlation analysis unit configured to perform correlation analysis on the basic sample feature data based on a Pearson correlation coefficient, to further screen a high-correlation feature subset, and a risk threshold calculation unit configured to perform threshold calculation on numerical value type sample features, to generate dynamic risk thresholds by combining a Youden index method, ROC curve analysis and kernel density estimation method, and to fuse clinical diagnosis guideline standards.

[0122] A model construction and evaluation module is configured with a model construction unit and a model evaluation unit. The model construction unit takes the output data of the sample feature analysis module as input, and uses an XGBoost or LightGBM algorithm to construct a gradient incremental model. For a binary classification task, a logarithmic loss is used, and the optimization target is to minimize the deviation of the predicted probability and the true label. The residual learning logic is that in each iteration, the model calculates the residual of the current prediction result, adds a decision tree to fit the residual, adjusts the split nodes and leaf weights of the tree through gradient descent, and gradually approximates the true distribution. The model evaluation unit uses four evaluation indexes, including ROC curve, AUC value, Matthew correlation coefficient and Kappa coefficient, to test the performance of the model on the test set.

[0123] A risk prediction module is configured with an interactive input unit and an explainability report generation unit. The interactive input unit receives individual feature data of a testee and performs variable type conversion and normalization processing on the data. After the risk prediction module calls the gradient incremental model to calculate the VTE risk probability of the testee, the VTE disease probability, feature importance ranking, local interpretability and Shapley value analysis results are output through the explainability report generation unit to assist doctors in making clinical decisions.

[0124] In summary, the present application effectively solves the prediction deviation and lack of explainability of existing machine learning models caused by one-sided feature screening and threshold setting deviating from clinical practice by constructing a VTE risk prediction system with deep integration of data-driven and clinical knowledge. The system automatically captures the nonlinear correlation between features based on the gradient incremental algorithm, and combines a dynamic risk threshold generation mechanism to break through the limitations of traditional linear screening methods, ensuring that key risk factors such as the interactive effect of plasma D-dimer and surgical history are accurately identified. At the same time, through visual decision path and clinical consistency verification, the model prediction logic is converted into medical evidence that doctors can understand, improving the credibility and clinical acceptance of the prediction results. In addition, the system supports dual-mode input and privacy compliance processing, balancing prediction accuracy and feasibility, and through dynamic iterative optimization, it adapts to diverse medical scenarios, providing efficient, transparent and compliant intelligent decision support for early warning of VTE.

[0125] The above description of disclosed embodiments enables one of ordinary skill in the art to make or use the application. Various modifications to these embodiments will be readily apparent to those of ordinary skill in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Accordingly, the application will not be limited to the embodiments shown herein but is intended to accord with the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A gradient-increment-based VTE risk prediction method, characterized in that, The method comprises the following steps: S1, collecting sample characteristic data groups through hospital internal system export or API access; S2, deleting irrelevant features, processing missing values, variable type conversion and normalization processing on the collected sample characteristic data; S3, performing VTE disease correlation analysis on the processed sample characteristic data by using Pearson correlation coefficient; S4, calculating the risk threshold of the numerical sample characteristics, and comparing the risk threshold of the numerical sample characteristics by using the Youden index method, ROC curve and kernel density estimation method; S5, establishing a machine learning model by using gradient incremental algorithm, and training the model; S6, testing the prediction effect of the model and performing clinical consistency verification; S7, inputting user individual sample characteristic data, and predicting the VTE risk of the user by using the gradient incremental model.

2. The gradient-increasing based VTE risk prediction method of claim 1, wherein, In step S2, the basic sample characteristic data related to VTE prediction obtained after deleting irrelevant features has 15 items: age, plasma D dimer, white blood cell, platelet, hemoglobin, neutrophil percentage, hypertension, coronary heart disease, tumor, gender, previous surgery history, diabetes, blood transfusion history, smoking history, and previous comorbidity. Among them, age, white blood cell, hemoglobin, platelet, neutrophil percentage and plasma D dimer are numerical sample characteristics.

3. The gradient-increasing based VTE risk prediction method of claim 1, wherein, In step S3, 8 sample characteristics with high correlation related to VTE disease are obtained after calculating the Pearson correlation coefficient of each basic sample characteristic data: age, plasma D dimer, platelet, hemoglobin, coronary heart disease, tumor, gender and previous comorbidity.

4. The gradient-increasing based VTE risk prediction method of claim 1, wherein, In step S4, The risk threshold of each numerical sample feature obtained by the Youden index method is: age 80 years old, white blood cells 8.07*10 9 / L, hemoglobin 78 g / L, platelets 307*10 9 / L, neutrophil percentage 72%, plasma D-dimer 284 ng / mL DDU; The risk threshold of each numerical sample feature obtained by the ROC curve is: age 80 years old, white blood cells 6.14*10 9 / L, hemoglobin 126 g / L, platelets 222*10 9 / L, neutrophil percentage 68.3%, plasma D-dimer 256 ng / mL DDU; The risk threshold of each numerical sample feature obtained by the kernel density estimation method is: age 80 years old, white blood cells 8.69*10 9 / L, hemoglobin 131.1 g / L, platelets 197*10 9 / L, percentage of neutrophils 69.8%, plasma D-dimer 965 ng / mL DDU.

5. The gradient-increasing based VTE risk prediction method of claim 4, wherein, The risk threshold and accuracy of the numerical sample features obtained by the Youden index method, the ROC curve and the kernel density estimation method are compared, and the comprehensive risk threshold is obtained by combining the curve fluctuation of the kernel density curve in the corresponding numerical region: age 65 years old, white blood cell 8.69*10 9 / L, hemoglobin 126 g / L, platelet 197*10 9 / L, neutrophil percentage 72%, plasma D dimer 965 ng / mL DDU.

6. The gradient-increasing based VTE risk prediction method of claim 3, wherein, In step S5, 15 basic sample characteristic data and 8 high correlation sample characteristic data are used as input to establish two different VTE risk prediction models by using XGBoost or LightGBM algorithm. Among them, the initial prediction value is 0.5, and the initial disease probability of all testers is 50% by default, which is adjusted and accumulated by combining the sample characteristic data of the testers to predict the VTE disease probability of the testers.

7. The gradient-increasing-based VTE risk prediction method of claim 6, wherein, The training process of the model is as follows: Initialize the model, set the initial prediction score to 0.5 as the initial prediction value of all samples, i.e. the initial probability value, and specify to use decision tree as the base learner; After starting training, according to parameter setting, do not perform column sampling but use all features for iteration, calculate the gradient residual of the current model prediction value and the true label in each iteration, and fit these residuals by minimizing the objective function containing the first-order derivative-gradient and the second-order derivative-Hessian matrix information; wherein, strictly control the tree node splitting, only when the loss reduction amount brought by splitting is at least 4.5, the splitting is performed; limit the sample weight of the child node to 1 to prevent overfitting; according to the default L2 regularization smoothing leaf node weight of the algorithm; After the tree construction is completed, the contribution of the new tree is scaled according to the learning rate 0.1 to prevent overfitting, and the limit change of the leaf node weight is further constrained to enhance the stability of the model; Repeat the iteration until the performance of the validation set no longer improves, and stop the training; Finally, the prediction results of all trees are weighted and combined to form the final model.

8. The gradient-increasing based VTE risk prediction method of claim 1, wherein, In step S6, the prediction effect of the model is evaluated by using four evaluation indexes of ROC curve, AUC value, Matthew correlation coefficient and Kappa coefficient.

9. A gradient-increment-based VTE risk prediction system, comprising: A system for implementing the gradient-increasing based VTE risk prediction method according to any one of claims 1 to 8, the system comprising: a data collection module configured to obtain a group of basic sample feature data for VTE prediction from a hospital internal system or an API; a data preprocessing module configured to perform irrelevant feature deletion, missing value processing, variable type conversion and normalization processing; a sample feature analysis module comprising a correlation analysis unit configured to quantitatively analyze the correlation between the basic sample feature data and VTE disease based on a Pearson correlation coefficient, and a risk threshold calculation unit configured to calculate a threshold value for a numerical sample feature; a model construction and evaluation module configured to construct a machine learning model based on a gradient-increasing algorithm using the output data of the sample feature analysis module as input, and to perform performance evaluation of the model on a test set; a risk prediction module configured to receive individual feature data of a testee, and to calculate a VTE risk probability by calling the gradient-increasing model.

10. The gradient-increasing-based VTE risk prediction system of claim 9, wherein, The data collection module is further configured with an anonymization processing unit and uses an encryption transmission mechanism; when collecting the basic sample feature data, direct identification data involving patient personal information is deleted or hashed, and quasi-identifiers indirectly associated with identity features are desensitized.

Images