Early-stage deep venous thrombosis prediction model establishment strategy, prediction method and device

Abstract

The invention relates to the field of deep venous thrombosis medical diagnosis, in particular to an early-stage deep venous thrombosis prediction model establishment strategy and prediction method and device. Screening out DVT-related features from the electronic medical record data of the fracture patient by adopting a filtering type, a wrapping type, an embedded type, a model method and a causal relationship feature selection method; putting the screened features as input ends into Logistic regression, random forest, Bagging and LightGBM machine learning models, evaluating the optimal strategy of each model for predicting the DVT onset risk of the fracture patient, and obtaining an early-stage deep venous thrombosis prediction model; and a DVT prediction visualization convenient interaction program is constructed to assist a clinician in making a medical decision. According to the method, the clinical electronic medical record data of the fracture patient is collected, the fracture patient DVT risk monitoring model is built by innovatively using a causal relationship feature selection method in combination with machine learning, and a new attempt is provided for improving the detection rate of DVT and reducing the death rate of the fracture patient.

Description

Technical Field

[0001] The present invention relates to the medical field of early deep vein thrombosis, and in particular to an early deep vein thrombosis prediction model establishment strategy, prediction method and device. Background Art

[0002] Fractures are often complicated by thrombosis, which is a risk factor for deep vein thrombosis (DVT). After fractures, the blood is in a hypercoagulable state due to damage to the vascular endothelium, and the condition may progress to deep vein thrombosis. The process of embolus shedding and displacement is hidden, which can easily cause pulmonary embolism, which is the main cause of death in fracture patients. Therefore, early assessment of the risk of DVT in fracture patients and timely medical intervention are key measures to reduce the mortality rate of fracture patients. Despite my country's efforts to allocate and invest a lot of resources to strengthen the healthcare system, it is clear that the number of clinicians is still far from enough to meet the dynamic monitoring needs of patients' disease development.

[0003] The rational use of electronic medical records is essential for extracting a large amount of valuable information from routine clinical tests. For patients with fractures, the electronic medical record information of hospital admission examination plays a key role in predicting the potential risk of DVT. The challenges faced by clinicians are two-fold. First, the medical resources are insufficient. The number of professional clinicians makes it impossible for them to allocate their energy to monitor every fracture patient in real time and prevent the occurrence of DVT without assessing the risk of death of the patient. On the other hand, clinicians are unable to quantitatively evaluate the diagnostic value of these complex and high-dimensional DVT data, which will lead to the failure of timely medical intervention for patients with potential high-risk DVT. In the era of digital medicine pandemic, facing the above problems, the application of artificial intelligence in early disease risk prediction can assist clinicians in disease prevention and optimize medical resources. Therefore, the automatic diagnosis system that combines machine learning algorithms with electronic medical record data can not only improve the detection rate of DVT in patients with fractures, but also greatly save the relevant economic and medical costs through efficient and rapid detection methods.

[0004] At present, machine learning algorithms have been widely used in the detection and prevention of different diseases such as DVT. However, traditional machine learning methods usually optimize the quantity and quality of features based on the correlation between the target variable and the features. The situation that has statistical correlation but does not conform to the actual causal law is called pseudo-correlation. Due to the instability of pseudo-correlation itself, the trained prediction model will have insufficient generalization when applied to other data sets. In addition, the feature selection method based on correlation cannot explain the influence mechanism between DVT and features, resulting in insufficient interpretability of the prediction results and lack of persuasiveness. In recent years, machine learning methods based on causal feature selection have attracted widespread attention in the field of biology. Machine learning based on causal relationships introduces the research method of causal inference into feature selection, which not only considers the correlation between the target variable and the features, but also considers the causal relationship between them. Applying this method to the early prevention of diseases can help us identify clinical test indicators with strong causal relationships with DVT and improve the generalization ability and interpretability of DVT prediction models. Summary of the invention

[0005] In order to improve the generalization ability and interpretability of the DVT prediction model, the present invention provides a strategy for establishing an early deep vein thrombosis prediction model, a prediction method and a device.

[0006] The present invention is achieved through the following technical solution: a strategy for establishing an early deep vein thrombosis prediction model, comprising the following steps:

[0007] 1) Collect electronic medical record data of fracture patients;

[0008] 2) The electronic medical record data of fracture patients were used to screen out features related to DVT using feature selection methods, and the electronic medical record data were divided into training and test sets;

[0009] 3) The features selected from the training set were put into the machine learning model as input to fit the model; the test set was used to detect the generalization ability and robustness of each model; the accuracy, recall, precision and F1 score indicators were used to evaluate the optimal strategy of each model to predict the risk of DVT in fracture patients, and an early deep vein thrombosis prediction model was obtained.

[0010] As a further improvement of the technical solution of the present invention, in step 3), different causal structure learning algorithms are used to perform causal relationship analysis on electronic medical record data, modify and eliminate obviously erroneous causal relationships, obtain causal relationships that are as close to the actual ones as possible, so as to ensure the reliability of the screened indicators, screen out features that are causally related to DVT, and put the features screened by the causal feature selection method in the training set as input into the machine learning model for fitting the model; the test set is used to detect the generalization ability and robustness of each model; the optimal strategy for each model to predict the risk of DVT in fracture patients is evaluated by accuracy, recall, precision and F1 score indicators, and an early deep vein thrombosis prediction model based on causal relationship features is obtained.

[0011] As a further improvement of the technical solution of the present invention, there are 9 characteristics that are causally related to DVT, specifically: age, D-dimer, albumin, fibrinogen, red blood cell count, white blood cell count, glucose, chloride ions and sodium ions.

[0012] As a further improvement of the technical solution of the present invention, in order to ensure the reliability of the screened indicators, the reliability of the screened causal relationship is tested, and the propensity score estimator is used to calculate the causal effect of the characteristics related to DVT. Then, the placebo effect provided by the propensity score estimator is used as a refutation method to test the robustness of the model.

[0013] As a further improvement of the technical solution of the present invention, the placebo effect is to replace the actual causal effect with a randomly generated placebo variable. If the causal relationship is established, the placebo effect should be smaller than the actual causal effect; otherwise, the causal relationship is not established.

[0014] The present invention further provides a prediction method for an early deep vein thrombosis prediction model, the method comprising:

[0015] Obtain electronic medical record data of fracture patients; input the electronic medical record data into an early deep vein thrombosis prediction model to identify people at high risk of DVT among fracture patients; the early deep vein thrombosis prediction model is obtained through the early deep vein thrombosis prediction model establishment strategy as described above.

[0016] The present invention also provides a visualization prediction device for an early deep vein thrombosis prediction model, comprising:

[0017] Input module, analysis module and prediction result output module;

[0018] The input module includes an electronic medical record data input unit for fracture patients;

[0019] The analysis module inputs the information variables input by the input module into the early deep vein thrombosis prediction model to calculate and obtain the DVT risk prediction value of the fracture patient;

[0020] The output module is used to output the predicted value of DVT incidence risk of fracture patients.

[0021] The early deep vein thrombosis prediction model establishment strategy, prediction method, device and equipment provided by the present invention have the following advantages compared with the prior art:

[0022] The present invention collects clinical electronic medical record data of fracture patients, and evaluates the prediction efficiency of DVT based on five feature selection methods, namely filtering, wrapping, embedded, model method and causal relationship, combined with machine learning. The results show that causal relationship machine learning can accurately predict potential DVT patients and improve the detection rate of DVT. On this basis, page design and model encapsulation are performed based on Python.tkinter, and visual production and operation software are performed. A convenient interactive program for DVT prediction visualization is constructed to assist clinicians in making medical decisions. The present invention collects clinical electronic medical record data of fracture patients, and innovatively uses the causal relationship feature selection method combined with machine learning to construct a DVT risk monitoring model for fracture patients, which provides a new attempt to improve the detection rate of DVT and reduce the mortality rate of fracture patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0025] Figure 1 A causal relationship structure diagram between features.

[0026] Figure 2 The following is a bar chart showing the prediction performance of each model.

[0027] Figure 3 Demonstration diagram of the interactive program for visualizing DVT prediction.

[0028] Figure 4 The present invention is an overall construction mode diagram of the method.

[0029] Figure 5This is a flow chart of an early deep vein thrombosis prevention system based on causal machine learning described in the present invention. DETAILED DESCRIPTION

[0030] In order to more clearly understand the above-mentioned objectives, features and advantages of the present invention, the scheme of the present invention will be further described below. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0031] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein; it is obvious that the embodiments in the specification are only part of the embodiments of the present invention, rather than all of the embodiments.

[0032] The specific embodiments of the present invention are described in detail below.

[0033] Example 1

[0034] The strategy for establishing an early deep vein thrombosis prediction model includes the following steps:

[0035] 1) Collect electronic medical record data of fracture patients;

[0036] 2) The electronic medical record data of fracture patients were used to screen out features related to DVT using feature selection methods, and the electronic medical record data were divided into training and test sets;

[0037] 3) The features selected from the training set were put into the machine learning model as input to fit the model; the test set was used to detect the generalization ability and robustness of each model; the accuracy, recall, precision and F1 score indicators were used to evaluate the optimal strategy of each model to predict the risk of DVT in fracture patients, and an early deep vein thrombosis prediction model was obtained.

[0038] The above step 3) can be replaced by: using different causal structure learning algorithms to perform causal relationship analysis on electronic medical record data, modifying and eliminating obviously erroneous causal relationships, obtaining causal relationships that are as close to the actual ones as possible to ensure the reliability of the screened indicators, screening out features that are causally related to DVT, and using the features screened in the training set based on the causal relationship feature selection method as input into the machine learning model for fitting the model; the test set is used to detect the generalization ability and robustness of each model; the accuracy, recall rate, precision and F1 score indicators are used to evaluate the optimal strategy of each model to predict the risk of DVT in fracture patients, and obtain an early deep vein thrombosis prediction model based on causal relationship features.

[0039] There are 9 characteristics that are causally related to DVT, namely: age, D-dimer, albumin, fibrinogen, red blood cell count, white blood cell count, glucose, chloride ion, and sodium ion.

[0040] Furthermore, to ensure the reliability of the screened indicators, the reliability of the causal relationship of the screening was tested, and the propensity score estimator was used to calculate the causal effect of the characteristics related to DVT. Then, the placebo effect provided by the propensity score estimator was used as a refutation method to test the robustness of the model.

[0041] Then, the placebo effect is achieved by replacing the actual causal effect with a randomly generated placebo variable. If the causal relationship is established, the placebo effect should be smaller than the actual causal effect; otherwise, the causal relationship does not hold.

[0042] Furthermore, this embodiment also provides a prediction method for an early deep vein thrombosis prediction model, the method comprising:

[0043] Obtain electronic medical record data of fracture patients; input the electronic medical record data into an early deep vein thrombosis prediction model to identify people at high risk of DVT among fracture patients; the early deep vein thrombosis prediction model is obtained through the early deep vein thrombosis prediction model establishment strategy as described above.

[0044] Furthermore, this embodiment also provides a visualization prediction device for an early deep vein thrombosis prediction model, including:

[0045] Input module, analysis module and prediction result output module;

[0046] The input module includes an electronic medical record data input unit for fracture patients; the electronic medical record data input unit includes a personal basic information input unit and a clinical test information input unit for fracture patients;

[0047] The basic personal information input unit includes an age submodule; the clinical test information input unit includes a D-dimer (D-Dimer) submodule, an albumin (ALB) submodule, a fibrinogen (Fbg) submodule, a red blood cell count (RBC) submodule, a white blood cell count (WBC) submodule, a glucose (Glu) submodule, a chloride ion (Cl - ) submodule and sodium ion (Na + ) submodule;

[0048] The analysis module inputs the information variables input by the input module into the early deep vein thrombosis prediction model to calculate and obtain the DVT risk prediction value of the fracture patient;

[0049] The output module is used to output the predicted value of DVT incidence risk of fracture patients.

[0050] Example 2

[0051] 1. Data source:

[0052] Patients with fractures admitted to the Affiliated Hospital of Shanxi Medical University from January 2017 to October 2018 were collected. The inclusion and exclusion criteria were as follows: (1) venous ultrasound confirmed the presence of DVT; (2) the admission diagnosis included fracture and no thrombosis; (3) pregnancy and malignant tumors were excluded; (4) age >18 years. A total of 777 patients were screened for the construction of the subsequent DVT prediction system, including 405 fracture patients with DVT and 372 fracture patients without DVT. This study was approved by the Ethics Committee of Shanxi Medical University, approval number 2023GLL25.

[0053] 2. Collect electronic medical record data of fracture patients:

[0054] The electronic medical record data includes the patient's basic personal characteristics and clinical test data.

[0055] Basic personal characteristics include age, height, weight, body mass index (BMI), smoking history, drinking history, pulse, respiratory rate, systolic blood pressure, diastolic blood pressure, history of activity limitation, history of hypertension, history of diabetes, history of blood transfusion, history of coagulation dysfunction, history of cerebrovascular disease, history of cognitive impairment, and history of cardiac conduction system disease.

[0056] Clinical laboratory data included creatinine (Crea), red blood cell count (RBC), hemoglobin (Hb), platelet count (PLT), white blood cell count (WBC), albumin (ALB), glucose (Glu), low-density lipoprotein cholesterol (LDL-C), prothrombin time (PT), fibrinogen (Fbg), creatine kinase (CK), D-dimer (D-Dimer), antithrombin activity (AT:A), total cholesterol (TC), prothrombin international normalized ratio (PT-INR), chloride ion (Cl – ), sodium ion (Na + ), potassium ion (K + ) and magnesium ions (Mg 2+ ).

[0057] This embodiment uses the synthetic minority oversampling technique (SMOTE) algorithm to solve the potential data imbalance phenomenon (DVT: 405, no DVT: 372) in order to improve the effectiveness of the DVT prediction results. The above data is divided into a training set and a test set in a ratio of 8:2. The training set is used to fit the model, and the test set data is used to detect the generalization ability and robustness of each model. Four models, Logistic regression, random forest, Bagging, and LightGBM, are used for DVT prediction. Four evaluation indicators, accuracy, recall, precision, and F1 score, are used to evaluate the prediction performance of DVT.

[0058] 3. Building a DVT prediction model based on the parcel feature selection method

[0059] A total of 24 features were selected based on the parcel variance selection feature method, namely age, respiratory rate, pulse, diastolic blood pressure, systolic blood pressure, height, weight, BMI, creatinine (Crea), red blood cell count (RBC), hemoglobin (Hb), platelet count (PLT), white blood cell count (WBC), albumin (ALB), glucose (Glu), low-density lipoprotein cholesterol (LDL-C), prothrombin time (PT), fibrinogen (Fbg), creatine kinase (CK), D-dimer (D-Dimer), antithrombin activity (AT:A), chloride ion (Cl - ), sodium ion (Na + ), total cholesterol (TC). The above features after filtering are used as input into the machine learning model. The results are shown in Table 1. Although the recall rate of the Bagging model is the highest, reaching 86.4%, the accuracy rate is only 48.7%, indicating that the model mistakenly misclassifies most patients without DVT as DVT. The random forest performs best, with an accuracy rate of 78.8%, a recall rate of 79.0%, and all evaluation indicators are above 78%, with good predictive efficiency.

[0060] Table 1: Comparison of prediction performance of various models based on filtering feature selection

[0061]

[0062] 4. Building a DVT prediction model based on the parcel feature selection method

[0063] Based on the wrap-around recursive feature elimination method, a total of 8 features were selected, namely age, creatinine (Crea), red blood cell count (RBC), platelet count (PLT), glucose (Glu), D-dimer (D-Dimer), antithrombin activity (AT:A) and magnesium ion (Mg 2+ ). Similarly, the above features after being filtered by the wrapping method are put into the machine learning model as input. The results are shown in Table 2. Unlike the filtering method, the bagging model based on the wrapping feature selection method has the best discrimination efficiency for DVT, with a recall rate of 85.1%, an accuracy rate of 82.0%, and all evaluation indicators above 80.0%. Random forest performed second, with a recall rate of more than 80.0%, indicating that most patients who may have DVT were correctly identified. LightGBM has the highest accuracy, reaching 83.5%, but the accuracy, recall rate and F1 score are all less than 80.0%.

[0064] Table 2: Comparison of prediction performance of various models based on package feature selection

[0065]

[0066] 5. Building a DVT prediction model based on embedded feature selection method

[0067] Based on the embedded Lasso regression feature selection method, a total of 13 features were selected, namely age, pulse, diastolic blood pressure, systolic blood pressure, weight, creatinine (Crea), hemoglobin (Hb), platelet count (PLT), albumin (ALB), glucose (Glu), D-dimer (D-Dimer), antithrombin activity (AT:A), chloride ion (Cl - ). Compared with the package feature selection method, the embedded method screens more basic features of the patient. The above features selected by the embedded method are put into the machine learning model as input. The results are shown in Table 3. The random forest model has the best DVT prediction efficiency, with an accuracy of 79.4%, and the remaining three indicators are all 80.2%. Although the embedded method screens more features than the package method, the overall prediction efficiency of the embedded method is slightly lower than that of the package method.

[0068] Table 3: Comparison of prediction performance of various models based on embedded feature selection

[0069]

[0070] 6. Constructing DVT prediction model based on model-based feature selection

[0071] The decision tree feature selection method based on the model method screened out a total of 8 features, namely age, red blood cell count (RBC), glucose (Glu), prothrombin time (PT), D-dimer (D-Dimer), total cholesterol (TC), potassium ion (K + ), magnesium ion (Mg2+). Similar to the parcel feature selection method, the model method screens 8 features and 5 of them are the same as the parcel method. The above features screened by the model method are put into the machine learning model as input. The results are shown in Table 4. Unlike the above method, the model-based LightGBM has the best prediction performance for DVT, with an accuracy of 80.1% and a recall rate of 82.7. Although the Bagging model has the highest recall rate, which is the same as the filtering method, the other indicators show that the model performance is extremely poor. Compared with filtering and embedded, the model-based feature selection is slightly better for the prediction of DVT, but the parcel-based feature selection method uses the least features and has the best prediction performance for DVT.

[0072] Table 4: Comparison of prediction performance of each model based on model-based feature selection

[0073]

[0074] 7. Construction of DVT prediction model based on causal feature selection

[0075] Machine learning based on causal feature selection can identify and understand the causal relationship between variables and improve prediction performance. The main steps are: first, it is necessary to clarify the research objectives and potential variable types, then visualize and describe the potential causal relationship between variables based on domain knowledge and structural equation models; then use causal inference methods such as propensity score matching to calculate the causal effect to verify the causal relationship between the feature and the target variable, and use the placebo effect to test the reliability of the causal relationship; finally, use the selected features to build a machine learning model to achieve the purpose of prediction.

[0076] We used different causal structure learning algorithms (Tetrad toolkit, Centre for CausalDiscovery) to perform causal relationship analysis on all clinical data, and the different causal graph structures generated ensured that the causal relationship was as close to the actual one as possible. In addition, we manually modified and removed the causal relationships with obvious errors in the algorithm output. On this basis, we used the model averaging method (Bayesys toolkit, Constantinou) to summarize the results of the above causal structure algorithms. Figure 1As shown in Figure 2, the causal relationship diagrams finally output by the model averaging method contain a total of 57 types, where each line represents the number of times the causal relationship appears in the results of the above different causal structure algorithms, and the more times it appears, the wider the line is ( Figure 1 ). And the Chinese translation of each English in the figure can be found in Table 5.

[0077] Table 5 English and Chinese comparison table of cause and effect diagram

[0078]

[0079]

[0080] In addition, we removed the causal relationships that appeared less than 10 times in the causal structure algorithm results to ensure the reliability of the screened indicators. The results showed that there were 9 indicators with causal relationship with DVT, including age, D-dimer (D-Dimer), albumin (ALB), fibrinogen (Fbg), red blood cell count (RBC), white blood cell count (WBC), glucose (Glu), chloride ion (Cl - ) and sodium ions (Na + ).

[0081] Afterwards, we conducted a reliability test on the causal relationship screened above. The first choice was to calculate the causal effect of 9 indicators related to DVT based on the propensity score estimator (DoWhy toolkit, Sharma and Kiciman). In addition, the placebo effect provided by DoWhy was used as a refutation method to test the robustness of the model. The placebo effect is to replace the actual causal effect with a randomly generated placebo variable. Therefore, if the causal relationship is established, the placebo effect should be smaller than the actual causal effect. We used the p-value to evaluate the results. If the P value is less than 0.05, the causal effect estimate is considered to be unsuccessful in the refutation test (Table 6). Then, the features selected based on the causal feature selection method were put into the machine learning model as input. The results showed (Table 7) that the random forest had the best prediction performance, with a recall rate of 91.3% and an accuracy of 85.2%. In addition to the Logistics model, the other models also showed excellent performance. Compared with other feature selection methods, causal machine learning significantly improved the detection rate of DVT ( Figure 2 ).

[0082] Table 6: Causal effect estimates and placebo test results

[0083]

[0084] Table 7: Comparison of prediction performance of each model based on causal feature selection

[0085]

[0086] On this basis, we built a visualization application based on causal feature selection and machine learning methods. The convenient operation software can help clinicians better use and understand the prediction process of DVT, and timely identify the high-risk group of DVT in fracture patients through a more simplified method to prevent the occurrence of DVT. The visualization interface provided in this embodiment is as follows Figure 3 As shown in the figure, the left side is the electronic medical record data input window for fracture patients, which includes age, D-dimer (D-Dimer), albumin (ALB), fibrinogen (Fbg), red blood cell count (RBC), white blood cell count (WBC), glucose (Glu), chloride ion (Cl - ) and sodium ions (Na + ); on the right are the Predict button and the DVT risk prediction window.

[0087] The DVT risk prediction window can be at least one of a data output window and a bar graph analysis window. The data output window includes whether the prediction result is a DVT risk population, the DVT risk percentage, and the non-DVT risk percentage; the bar graph analysis window includes whether the prediction result is a DVT risk population, a DVT risk bar graph, and a non-DVT risk bar graph.

[0088] In this embodiment, when the DVT risk percentage is greater than or equal to 50%, the person is identified as a DVT risk group; otherwise, the person is identified as a non-DVT risk group.

[0089] The above is only a specific implementation of the present invention, which enables those skilled in the art to understand or implement the present invention. Although detailed descriptions are given with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments, and they should all be covered by the protection scope of the claims.

Claims

1. A strategy for establishing an early deep vein thrombosis prediction model, characterized in that: The steps include: 1) Collect electronic medical record data containing individual characteristics and laboratory parameters of fracture patients; 2) The electronic medical record data of fracture patients were used to screen out features related to DVT using feature selection methods, and the electronic medical record data were divided into training and test sets; 3) The features selected from the training set were put into the machine learning model as input to fit the model; the test set was used to detect the generalization ability and robustness of each model; the accuracy, recall, precision and F1 score indicators were used to evaluate the optimal strategy of each model to predict the risk of DVT in fracture patients, and an early deep vein thrombosis prediction model was obtained.

2. The strategy for establishing an early deep vein thrombosis prediction model according to claim 1, characterized in that: In step 3), different causal structure learning algorithms are used to perform causal relationship analysis on the electronic medical record data, modify and eliminate obviously erroneous causal relationships, obtain causal relationships that are as close to the actual ones as possible, ensure the reliability of the screened indicators, screen out features that are causally related to DVT, and use the features selected in the training set based on the causal relationship feature selection method as input to the machine learning model for model fitting; the test set is used to detect the generalization ability and robustness of each model; the accuracy, recall rate, precision and F1 score indicators are used to evaluate the optimal strategy of each model for predicting the risk of DVT in fracture patients, and obtain an early deep vein thrombosis prediction model based on causal relationship features.

3. The strategy for establishing an early deep vein thrombosis prediction model according to claim 2, characterized in that: There are 9 characteristics that are causally related to DVT, namely: age, D-dimer, albumin, fibrinogen, red blood cell count, white blood cell count, glucose, chloride ion, and sodium ion.

4. The strategy for establishing an early deep vein thrombosis prediction model according to claim 2, characterized in that: To ensure the reliability of the screened indicators, the reliability of the causal relationship of the screening was tested, and the propensity score estimator was used to calculate the causal effect of the characteristics associated with DVT. Then, the placebo effect provided by the propensity score estimator was used as a refutation method to test the robustness of the model.

5. The strategy for establishing an early deep vein thrombosis prediction model according to claim 4, characterized in that: The placebo effect is achieved by replacing the actual causal effect with a randomly generated placebo variable. If the causal relationship is established, the placebo effect should be smaller than the actual causal effect; otherwise, the causal relationship does not hold.

6. A prediction method for an early deep vein thrombosis prediction model, characterized in that: The method comprises: Obtain electronic medical record data of fracture patients; input the electronic medical record data into an early deep vein thrombosis prediction model to identify people at high risk of DVT among fracture patients; the early deep vein thrombosis prediction model is obtained through the early deep vein thrombosis prediction model establishment strategy described in any one of claims 1-5.

7. A visual prediction device for an early deep vein thrombosis prediction model, characterized in that: include: Input module, analysis module and prediction result output module; The input module includes an electronic medical record data input unit for fracture patients; The analysis module inputs the information variables input by the input module into the early deep vein thrombosis prediction model according to any one of claims 1 to 5, and calculates and obtains the DVT risk prediction value of the fracture patient; The output module is used to output the predicted value of DVT incidence risk of fracture patients.

Images