Early noninvasive endometrial cancer assessment method based on machine learning

Abstract

The invention discloses an early noninvasive endometrial cancer assessment method based on machine learning, and the method comprises the following steps: S1, obtaining the index data of a potential endometrial lesion patient in advance, and carrying out the data integration based on the index data according to a preset standard to obtain a feature set; s2, performing standardized preprocessing on the feature set in the step S1 to obtain a training set and a test set of standardized endometrial lesion data; s3, constructing a disease prediction model based on a preset multi-modal machine learning model and used for predicting an endometrial cancer result on the basis of the training set and the test set containing the endometrial lesion data obtained in the step S2; and S4, inputting endometrial lesion index data of a patient to be evaluated into the disease prediction model to implement early noninvasive endometrial cancer prediction evaluation so as to obtain an endometrial cancer prediction result of the patient. The method has the advantages that non-invasive screening can be met at the same time, the prediction accuracy is high, and the economic burden is small.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical treatment, and more particularly to an early non-invasive endometrial cancer evaluation method based on machine learning. BACKGROUND

[0002] In clinical practice, for women, endometrial cancer (EC) is a gynecological malignant tumor that affects women's health. In recent decades, with the increase in obesity rate and the delay in childbearing age, the incidence of EC has shown an upward trend in countries around the world. According to clinical data, about 20% of EC patients are difficult to be detected early due to hidden symptoms, and most patients are diagnosed and treated due to symptoms such as abnormal vaginal bleeding. Among EC patients diagnosed due to abnormal vaginal bleeding, about 70% of patients are in the early stage of EC, and nearly 30% of patients are in the late stage of EC, which leads to a significantly poorer clinical outcome. Therefore, early detection of EC is of great significance to improve patient prognosis and survival rate.

[0003] With the increasing importance of early EC diagnosis, there is currently no reliable non-invasive screening method that can meet the basic needs of high prediction accuracy, simple operation and economic affordability. The most commonly used screening methods currently available include hysteroscopy and diagnostic curettage, which can provide relatively accurate diagnostic information, but the above methods have significant limitations such as invasiveness, complexity of operation and high medical costs, which not only cause physical trauma to patients, but also increase their economic burden; and the above methods also have a certain risk of infection, which also limits their widespread application. In addition, there is a minimally invasive screening method, i.e. fine needle aspiration, which has less trauma, but its diagnostic performance is relatively low, and it can only assess less than 50% of the uterine cavity, and the amount of tissue collected is usually insufficient for effective histopathological evaluation. And the screening method of independent imaging method, the independent imaging technology may have some value in some cases, but due to its limited sensitivity, it often cannot meet the screening needs. At the same time, although some biomarkers such as CA-125 and HE-4 are closely related to the clinical characteristics and prognosis of endometrial cancer, the diagnostic accuracy and specificity of these biomarkers alone for screening are still insufficient, with a false negative rate of more than 40%, which is far beyond the clinically acceptable range, and cannot meet the clinical screening needs.

[0004] How to solve the above problems has become a technical problem to be solved. SUMMARY

[0005] The technical problem to be solved by the present application is to provide an early non-invasive endometrial cancer evaluation method based on machine learning that can simultaneously meet the requirements of non-invasive screening, high prediction accuracy and low economic burden.

[0006] To achieve the above object, the technical scheme adopted by the present application is as follows:

[0007] The present application provides an early non-invasive endometrial cancer evaluation method based on machine learning, which comprises the following steps:

[0008] S1, pre-acquire the index data of potential endometrial lesion patients, and implement data integration based on the index data according to a preset standard to obtain a feature set;

[0009] S2, standardize the feature set in step S1 to obtain a training set and a test set of standardized endometrial lesion data;

[0010] S3, based on the training set and the test set containing endometrial lesion data obtained in step S2, a disease prediction model for predicting endometrial cancer results based on a preset multi-modal machine learning model is constructed;

[0011] S4, input the endometrial lesion index data of the patient to be evaluated into the disease prediction model to implement early non-invasive endometrial cancer prediction evaluation to obtain the endometrial cancer prediction result of the patient.

[0012] Preferably, in step S3, the following steps are further included:

[0013] S3.1, based on a predetermined Python environment, a preset receiver operating characteristic curve (ROC) analysis method is used to implement an evaluation test on the test set according to a preset standard to obtain the classification threshold value changes of true positive (TP), true negative (TN), false positive (FP) and false negative (FN) and form a predetermined classification threshold value change curve;

[0014] S3.2, in the classification threshold value change curve of step S3.1, a predetermined classification threshold value is selected to implement performance evaluation according to a preset standard to obtain the performance and effectiveness of the disease prediction model.

[0015] Preferably, in step S3.2, the following steps are further included:

[0016] S3.2.1, in the Python environment, a preset Tree Explainer algorithm is used to obtain the index data of the classification threshold value according to a preset standard, and a preset Shap Python package is used to implement data processing on the index data to obtain SHAP values for explaining the prediction evaluation performance of the disease prediction model;

[0017] S3.2.2, the SHAP value is processed according to the preset standard by using the LSTM time recurrent neural network model to construct a predetermined patient-specific risk driving factor graph, and the importance measure and ranking of the index data of the characteristic are quantified and visualized by using the patient-specific risk driving factor graph, so as to obtain the local characteristic contribution degree of each characteristic to identify the influence of each characteristic on the disease prediction model, so as to publicly explain the prediction evaluation performance and effectiveness of the disease prediction model in the prediction process.

[0018] Preferably, in step S3.2, the corresponding classification threshold values of true positive (TP), true negative (TN), false positive (FP) and false negative (FN) are selected in the classification threshold change curve according to the preset standard to determine the accuracy, recall rate, precision, AUC-ROC and Brier score related indicators of the test set of the disease prediction model, wherein the larger the AUC-ROC and the higher the Brier score, the better the performance of the disease prediction model, so as to evaluate the performance and effectiveness of the disease prediction model.

[0019] Preferably, in step S1, the feature set includes clinical features, ultrasonic features and hematological features containing corresponding index data of the patient, wherein the index data contained in the clinical features includes the age, reproductive history, hypertension and diabetes history, body mass index (BMI) of the patient; the index data contained in the ultrasonic features includes the endometrial thickness; and the index data contained in the hematological features includes NLR, PLR, MLR, SIRI, SII, CA-125 and HE-4.

[0020] Preferably, in step S2, when there are missing values in the index data of each feature in the feature set, the missing values are replaced by the preset median of the feature to obtain the corresponding feature of the complete index data, and then the corresponding feature is standardized by using the preset Standard Scaler standard deviation standardization function method, so as to obtain the standard feature set with two different encoding category variables 0 and variable 1 to prevent potential bias caused by large data value range of the feature, and the data of the standard feature set is balanced by using the preset SMOTE minority class oversampling technology, and then the standard feature set is randomly divided into a training set and a test set according to the preset standard in a ratio of 8:2.

[0021] Preferably, in step S3, the method for constructing the disease prediction model comprises the following steps:

[0022] S7.1, according to the preset standard, the variable data in the endometrial lesion data of each feature in the training set and the test set are obtained and divided into feature variables and target variables;

[0023] S7.2, based on the feature variables and the target variables, constructing a disease prediction model based on the preset multi-modal machine learning model in the preset Python environment according to the preset standard.

[0024] Preferably, the multi-modal machine learning model is composed of base modules and meta modules interconnected in a cascading manner, wherein the base modules include 6 tree-based Bagging and Boosting base classifiers, the above base classifiers run in parallel and independently predict the features of the input training set according to the preset standard, and the predetermined five-fold cross-validation method is used to implement data validation and fitting processing according to the preset standard, so as to obtain the aggregated prediction results, and then the meta module composed of meta classifiers is used to obtain the aggregated prediction results and implement processing on the data of the aggregated prediction results by using the preset probability voting method based on the classification labels, thereby relieving the overfitting of the base classifiers to the feature data of the training set, and further obtaining the high-accuracy and high-robustness endometrial cancer prediction results.

[0025] Preferably, the 6 tree-based Bagging and Boosting base classifiers are decision tree (DT), random forest (RF), Extreme Gradient Boosting (XGBoost), Support Vector Mac (SVM), Gradient Boosting Machine Model (GBDT), Logistic Regression (LR) and neural network (MLP).

[0026] Preferably, the method for obtaining the aggregated prediction results comprises the following steps:

[0027] S9.1, dividing the training set into five different subsets of the same size according to the preset standard;

[0028] S9.2, obtaining each subset in step S9.1 by the 6 base classifiers and implementing iterative training according to the preset standard, so as to obtain five prediction result values;

[0029] S9.3, implementing data validation and aggregation processing on the prediction result values obtained in step S9.2 by the 6 base classifiers using the five-fold cross-validation method, so as to obtain the aggregated prediction results.

[0030] Preferably, the method for obtaining the endometrial cancer prediction results of the patient comprises the following steps:

[0031] S10.1, inputting the endometrial lesion index data of the patient to be tested into the disease prediction model, and implementing iterative training on the index data by the base classifiers, so as to obtain a plurality of corresponding endometrial cancer prediction result values;

[0032] S10.2, performing data verification and fitting processing on the endometrial carcinoma prediction result value in step S10.1 by the base classifier by using a five-fold cross-validation method to obtain an endometrial carcinoma aggregate prediction result;

[0033] S10.3, performing evaluation test on the endometrial carcinoma aggregate prediction result in step S10.2 according to a preset standard by using a preset receiver operating characteristic (ROC) analysis method to identify a classification threshold value of true positive (TP), true negative (TN), false positive (FP) and false negative (FN) of the endometrial carcinoma;

[0034] S10.4, performing cutoff value acquisition processing on the classification threshold value in step S10.3 according to a preset standard by using a preset Youden index, that is, acquiring the maximum value in each classification threshold value and taking the maximum value as the cutoff value, by using a preset Youden Index = Sensitivity + specificity - 1 function;

[0035] S10.5, inputting the cutoff value in step S10.4 into a meta-module composed of meta-classifiers, performing data processing on the cutoff value by the meta-module by using a classification label-based probability voting method, thus deciding to output the class with the most votes, which is the endometrial carcinoma prediction result, and the endometrial carcinoma prediction result is used to list patients into a corresponding low-risk group or high-risk group to realize early non-invasive screening of the patients and endometrial carcinoma prediction evaluation with high prediction accuracy and small economic burden.

[0036] Compared with the prior art, the present application has the following beneficial effects:

[0037] In the present application, the index data of potential endometrial lesion patients is acquired in advance, and the data integration is carried out based on the index data according to a preset standard to obtain a feature set, wherein the feature set of the patient includes 14 non-invasive index data such as clinical features including age, BMI, medical history, ultrasonic features including endometrial thickness, and hematological features including NLR, SII, CA-125 and the like, which can cover the full-dimensional representation of the disease. The standardization preprocessing is carried out on the above feature set, so as to obtain a standardized training set and a test set of endometrial lesion data. A disease prediction model for predicting the result of endometrial cancer based on a preset multi-modal machine learning model is constructed based on the acquired training set and test set containing endometrial lesion data. The endometrial lesion index data of the corresponding patient is input into the disease prediction model, and the multi-modal machine learning model is used to carry out nonlinear relationship sensitive capture iterative training and evaluation test on the endometrial lesion index data, so as to realize the non-invasive screening and prediction evaluation of endometrial cancer of the patient, and thus the prediction result of endometrial cancer of the patient is obtained. The present application has high prediction accuracy and small economic burden, and can make the high-risk population of endometrial lesions no longer repeatedly perform invasive endometrial biopsy, effectively improve the early detection accuracy of endometrial cancer, and assist doctors in personalized diagnosis, greatly facilitating medical staff and patients. Therefore, the present application has the advantages of simultaneously satisfying non-invasive screening, high prediction accuracy and small economic burden. BRIEF DESCRIPTION OF DRAWINGS

[0038] The present application will be further described below in combination with the drawings and examples.

[0039] Figure 1 It is a flowchart of an early non-invasive endometrial cancer prediction and evaluation method based on machine learning according to the present application. DETAILED DESCRIPTION

[0040] Referring to Figure 1 The present embodiment provides an early non-invasive endometrial cancer prediction and evaluation method based on machine learning, which comprises the following steps:

[0041] S1, acquiring index data of potential endometrial lesion patients in advance, and carrying out data integration based on the index data according to a preset standard to obtain a feature set;

[0042] S2, carrying out standardization preprocessing on the feature set in step S1, so as to obtain a standardized training set and a test set of endometrial lesion data;

[0043] S3, constructing a disease prediction model for predicting the result of endometrial cancer based on a preset multi-modal machine learning model based on the acquired training set and test set containing endometrial lesion data in step S2;

[0044] S4, input the endometrial lesion index data of the patient to be evaluated into the disease prediction model to implement early non-invasive endometrial cancer prediction evaluation to obtain the endometrial cancer prediction result of the patient.

[0045] In use, the index data of a potential endometrial lesion patient is acquired in advance, and based on the index data, data integration is implemented according to a preset standard to obtain a feature set, wherein the feature set of the patient includes 14 non-invasive index data such as clinical features including age, BMI, medical history, ultrasonic features including endometrial thickness, and hematological features including NLR, SII, CA-125, etc., which can cover the full-dimensional representation of the disease to ensure the accuracy of subsequent endometrial cancer prediction. The index data of the feature set is subjected to standardization preprocessing to obtain a standardized training set and test set of endometrial lesion data; based on the acquired training set and test set containing endometrial lesion data, a disease prediction model for predicting the result of endometrial cancer is constructed based on a preset multi-modal machine learning model; the endometrial lesion index data of the corresponding patient is input into the disease prediction model, and the multi-modal machine learning model is used to implement nonlinear relationship sensitive capture iterative training and evaluation test on the endometrial lesion index data to realize non-invasive screening and prediction evaluation of endometrial cancer of the patient, thereby obtaining the endometrial cancer prediction result of the patient. The prediction accuracy is high, the economic burden is small, and the high-risk population of endometrial lesions no longer needs to repeatedly perform invasive endometrial biopsy, which can effectively improve the early detection accuracy of endometrial cancer and assist doctors in personalized diagnosis, greatly facilitating medical staff and patients.

[0046] In this embodiment, in step S3, the following steps are further included:

[0047] S3.1, based on a predetermined Python environment, a preset receiver operating characteristic curve (ROC) analysis method is used to implement evaluation test on the test set according to a preset standard to obtain classification threshold value changes of true positive (TP), true negative (TN), false positive (FP) and false negative (FN) and form a predetermined classification threshold value change curve;

[0048] S3.2, in the classification threshold value change curve of step S3.1, a predetermined classification threshold value is selected to implement performance evaluation according to a preset standard to obtain the performance and effectiveness of the disease prediction model.

[0049] In this embodiment, in step S3.2, the following steps are further included:

[0050] S3.2.1, in the Python environment, adopting a preset Tree Explainer algorithm to obtain index data of a classification threshold according to a preset standard, and then using a preset Shap Python package to process the index data to obtain SHAP values for explaining the prediction evaluation performance of the disease prediction model;

[0051] S3.2.2, using an LSTM time recurrent neural network model to process the SHAP values according to a preset standard to construct a predetermined patient-specific risk driver factor graph, and using the patient-specific risk driver factor graph to quantitatively and visually explain the importance measure and ranking of the index data of the features, so as to obtain the local feature contribution degree of each feature to identify the predetermined marker of each feature affecting the disease prediction model, thereby being able to publicly explain the prediction evaluation performance and effectiveness of the disease prediction model in the prediction process.

[0052] In this embodiment, in step S3.2, the corresponding classification threshold of true positive (TP), true negative (TN), false positive (FP) and false negative (FN) is selected in the classification threshold change curve according to the preset standard to determine the accuracy, recall rate, precision, AUC-ROC and Brier score related indicators of the test set of the disease prediction model, wherein the larger the AUC-ROC and the higher the Brier score, the better the performance of the disease prediction model, thereby evaluating the performance and effectiveness of the disease prediction model.

[0053] In this embodiment, in step S1, the feature set includes clinical features, ultrasonic features and hematological features containing corresponding index data of the patient, wherein the index data contained in the clinical features includes the age, reproductive history, hypertension, diabetes history and body mass index (BMI) of the patient; the index data contained in the ultrasonic features includes the endometrial thickness; and the index data contained in the hematological features includes NLR, PLR, MLR, SIRI, SII, CA-125 and HE-4.

[0054] In this embodiment, in step S2, when there are missing values in the index data of each feature in the feature set, the missing values are replaced by a preset median of the feature according to a preset standard to obtain corresponding features of complete index data, and then a preset Standard Scaler standard deviation standardization function method is used to implement standardization processing on the corresponding features, so as to obtain a standard feature set with two different encoding category variables 0 and variable 1 to prevent potential bias caused by a large data value range of the features. After balancing the data of the standard feature set by a preset SMOTE minority class oversampling technology, the standard feature set is randomly divided into a training set and a test set according to a preset standard in a ratio of 8:2.

[0055] In the present embodiment, in step S3, the method steps of constructing the disease prediction model are as follows:

[0056] S7.1, obtaining variable data in the endometrial lesion data of each feature in the training set and the test set according to the preset standard and dividing into feature variables and target variables;

[0057] S7.2, based on the feature variables and the target variables, constructing a disease prediction model based on the preset multi-modal machine learning model in the preset Python environment according to the preset standard.

[0058] In the present embodiment, the multi-modal machine learning model is composed of basic modules and meta-modules interconnected in a cascading manner, wherein the basic modules include six tree-based Bagging and Boosting basic classifiers, the above basic classifiers run in parallel and independently predict the features of the input training set according to the preset standard, and a predetermined five-fold cross-validation method is used to implement data validation and fitting processing according to the preset standard, so as to obtain the aggregated prediction result, and then the meta-module composed of meta-classifiers is used to obtain the aggregated prediction result and implement processing on the data of the aggregated prediction result by using the preset probability voting method based on the classification label, thereby relieving the overfitting of the basic classifiers to the feature data of the training set, and further obtaining high-accuracy and high-robustness endometrial cancer prediction results.

[0059] In the present embodiment, the six tree-based Bagging and Boosting basic classifiers are decision tree (DT), random forest (RF), extreme gradient boosting (XGBoost), support vector machine (SVM), gradient boosting machine model (GBDT), logistic regression (LR) and neural network (MLP).

[0060] In the present embodiment, the method for obtaining the aggregated prediction result includes the following steps:

[0061] S9.1, randomly dividing the training set into five different subsets of the same size according to the preset standard;

[0062] S9.2, obtaining each subset in step S9.1 by the six basic classifiers and implementing iterative training according to the preset standard, so as to obtain five prediction result values;

[0063] S9.3, implementing data validation and aggregation processing on the prediction result values obtained in step S9.2 by the six basic classifiers using the five-fold cross-validation method, so as to obtain the aggregated prediction result.

[0064] In the present embodiment, the method for obtaining the endometrial cancer prediction result of the patient comprises the following steps:

[0065] S10.1, input the endometrial lesion index data of the patient to be tested into the disease prediction model, and implement iterative training on the index data by the base classifier respectively, so as to obtain a plurality of corresponding endometrial cancer prediction result values;

[0066] S10.2, implement data verification and fitting processing on the endometrial cancer prediction result values in step S10.1 by the base classifier using the five-fold cross-validation method to obtain an endometrial cancer aggregate prediction result;

[0067] S10.3, implement evaluation test on the endometrial cancer aggregate prediction result in step S10.2 according to a preset standard using a preset receiver operating characteristic curve (ROC) analysis method to identify the classification threshold value of true positive (TP), true negative (TN), false positive (FP) and false negative (FN) of the endometrial cancer;

[0068] S10.4, implement cutoff value acquisition processing on the classification threshold value in step S10.3 according to a preset standard using a preset Youden index through a preset Youden Index = Sensitivity + specificity - 1 function, that is, acquiring the maximum value in each classification threshold value and taking the maximum value as the cutoff value;

[0069] S10.5, input the cutoff value in step S10.4 into the meta-module composed of the meta-classifier, and implement data processing on the cutoff value by the meta-module using the probability voting method based on the classification label, so as to decide to output the category with the most votes, which is the endometrial cancer prediction result. The endometrial cancer prediction result is used to list the patient into the corresponding low-risk group or high-risk group to realize early non-invasive screening of the patient and endometrial cancer prediction evaluation with high prediction accuracy and small economic burden.

[0070] In use, first, the index data of potential endometrial lesion patients is obtained in advance, and the index data is integrated according to a preset standard to obtain a feature set, wherein the feature set includes clinical features, ultrasonic features, and hematological features, the index data of the clinical features includes the age, fertility history, hypertension, diabetes history, and body mass index (BMI) of the patient, the index data of the ultrasonic features includes the endometrial thickness, and the index data of the hematological features includes NLR, PLR, MLR, SIRI, SII, CA-125, and HE-4; second, the index data of the feature set is standardized and preprocessed to obtain a training set and a test set of the standardized endometrial lesion data, when there are missing values in the index data of each feature in the feature set, the missing values are replaced with a preset feature median according to a preset standard to obtain a complete index data corresponding feature, and the corresponding feature is standardized by a preset Standard Scaler standard deviation standardization function method to obtain a standard feature set with two different encoding category variables 0 and variable 1 to prevent potential bias caused by a large data value range of the feature, the data of the standard feature set is balanced by a preset SMOTE minority class oversampling technology, and then the standard feature set is divided into the training set and the test set in a preset standard random ratio of 8:2; then, based on the obtained training set and test set containing endometrial lesion data, variable data in the endometrial lesion data of each feature in the training set and the test set is obtained according to a preset standard and is divided into a feature variable and a target variable; based on the feature variable and the target variable, a disease prediction model based on a preset multi-modal machine learning model is constructed in a preset Python environment according to a preset standard; based on the predetermined Python environment, a preset receiver operating characteristic curve (ROC) analysis method is used to evaluate the test set according to a preset standard to obtain the classification threshold value changes of true positive (TP), true negative (TN), false positive (FP), and false negative (FN) and form a predetermined classification threshold value change curve; in the classification threshold value change curve, a predetermined classification threshold is selected to perform performance evaluation according to a preset standard to obtain the performance and effectiveness of the disease prediction model, in the Python environment, a preset Tree Explainer algorithm is used to obtain the index data of the classification threshold according to a preset standard, and a preset Shap Python package is used to process the index data to obtain SHAP values for explaining the prediction and evaluation performance of the disease prediction model.The LSTM time recurrent neural network model is used to process the SHAP value data according to the preset standard to construct a predetermined patient-specific risk driving factor graph, the patient-specific risk driving factor graph is used to quantitatively and visually explain the importance measure and ranking of the index data of the feature, so as to obtain the local feature contribution degree of each feature to identify the influence of the predetermined marker of each feature on the disease prediction model, thereby being able to publicly explain the prediction evaluation performance and effectiveness of the disease prediction model in the prediction process; finally, the endometrial lesion index data of the to-be-tested patient is input into the disease prediction model, and the index data is iteratively trained by the basic classifier, thereby obtaining a plurality of corresponding endometrial cancer prediction result values; the basic classifier is used to perform data verification and fitting processing on the endometrial cancer prediction result value by using a five-fold cross-validation method to obtain an endometrial cancer aggregated prediction result; a preset receiver operating characteristic curve (ROC) analysis method is used to perform an evaluation test on the endometrial cancer aggregated prediction result according to a preset standard to identify a classification threshold value of true positive (TP), true negative (TN), false positive (FP) and false negative (FN) of the endometrial cancer; a preset Youden index is used to perform a cutoff value acquisition processing on the classification threshold value by using a preset Youden Index = Sensitivity + specificity - 1 function according to a preset standard, that is, the maximum value in each classification threshold value is obtained and the maximum value is used as the cutoff value; the cutoff value is input into the meta-module composed of the meta-classifier, and the cutoff value is processed by the meta-module by using a probability voting method based on the classification label, thereby deciding to output the class with the most votes, which is the endometrial cancer prediction result, and the endometrial cancer prediction result is used to classify the patient into a corresponding low-risk group or a high-risk group to realize early non-invasive screening of the patient, high prediction accuracy and low economic burden of the endometrial cancer prediction evaluation.

[0071] In summary, the method of the present application has the advantages of simultaneously meeting the requirements of non-invasive screening, high prediction accuracy and low economic burden.

[0072] For those skilled in the art, various corresponding changes and modifications can be made to the above-described technical solutions and concepts, and all such changes and modifications should be within the scope of protection of the claims of the present application.

Claims

1. A method for early non-invasive endometrial cancer assessment based on machine learning, characterized in that: The method comprises the following steps: S1. Pre-acquire indicator data of patients with potential endometrial lesions, and perform data integration based on the indicator data according to preset standards to obtain a feature set; S2. performing standardization preprocessing on the feature set in step S1, thereby obtaining a training set and a test set of standardized endometrial lesion data; S3. Building a disease prediction model for predicting endometrial cancer outcomes based on a preset multimodal machine learning model based on the training set and test set containing the endometrial lesion data obtained in step S2; S4. Inputting the endometrial lesion index data of the patient to be evaluated into the disease prediction model to perform early non-invasive endometrial cancer prediction evaluation to obtain the endometrial cancer prediction result of the patient.

2. The method for early non-invasive endometrial cancer assessment based on machine learning according to claim 1, characterized in that: In step S3, the following steps are also included: S3.

1. Based on a predefined Python environment, perform an evaluation test on the test set using a predefined receiver operating characteristic (ROC) analysis method according to predefined criteria to obtain the classification threshold changes for true positive (TP), true negative (TN), false positive (FP), and false negative (FN) and form a predefined classification threshold change curve. S3.

2. In the classification threshold change curve of step S3.1, a predetermined classification threshold is selected and a performance evaluation is performed according to a preset standard to obtain the performance and effectiveness of the disease prediction model.

3. The method for early non-invasive endometrial cancer assessment based on machine learning according to claim 2, characterized in that: In step S3.2, the following steps are also included: S3.2.

1. In the Python environment, use the pre-set Tree Explainer algorithm to obtain index data for classification thresholds according to pre-set criteria. Then, use the pre-set Shap Python package to process the index data to obtain SHAP values ​​for interpreting the predictive performance of the disease prediction model. S3.2.

2. Use the LSTM time recursive neural network model to perform data processing on the SHAP value according to preset standards to construct a predetermined patient-specific risk driver factor map. Use the patient-specific risk driver factor map to quantify and visualize the importance measurement and ranking of the indicator data of the feature, so as to obtain the local feature contribution of each feature to identify the impact of the predetermined marker of each feature on the disease prediction model, thereby being able to publicly explain the predictive evaluation performance and effectiveness of the disease prediction model in the prediction process.

4. The method for early non-invasive endometrial cancer assessment based on machine learning according to claim 2, characterized in that: In step S3.2, the corresponding classification thresholds of true positive (TP), true negative (TN), false positive (FP) and false negative (FN) are selected according to the preset standards in the classification threshold change curve to determine the accuracy, recall rate, precision, AUC-ROC, and Brier score related indicators of the test set of the disease prediction model. Among them, the larger the AUC-ROC and the higher the Brier score, the better the performance of the disease prediction model, thereby evaluating the performance and effectiveness of the disease prediction model.

5. The method for early non-invasive endometrial cancer assessment based on machine learning according to any one of claims 1 to 4, characterized in that: In step S1, the feature set includes clinical characteristics containing corresponding indicator data of the patient, as well as ultrasound characteristics and hematological characteristics, wherein the indicator data contained in the clinical characteristics include the patient's age, reproductive history, hypertension, diabetes history, and body mass index (BMI); the indicator data contained in the ultrasound characteristics include endometrial thickness; and the indicator data contained in the hematological characteristics include NLR, PLR, MLR, SIRI, SII, CA-125, and HE-4.

6. The method for early non-invasive endometrial cancer assessment based on machine learning according to claim 5, characterized in that: In step S2, when there are missing values ​​in the indicator data of each feature in the feature set, the preset feature median is used to replace the missing value according to the preset standard to obtain the corresponding feature of the complete indicator data, and then the preset Standard Scaler standard deviation standardization function method is used to standardize the corresponding features, thereby obtaining a standard feature set with two different coding category variables 0 and variable 1 to prevent potential bias due to a large range of feature data values. After the data of the standard feature set is balanced by the preset SMOTE synthetic minority class oversampling technology, the standard feature set is randomly divided into a training set and a test set in a ratio of 8:2 according to the preset standard.

7. The method for early non-invasive endometrial cancer assessment based on machine learning according to claim 6, characterized in that: In step S3, the method steps for building a disease prediction model are as follows: S7.

1. Obtain variable data from the endometrial lesion data of each feature in the training set and the test set according to preset standards and divide them into feature variables and target variables; S7.

2. Based on the feature variables and target variables, a disease prediction model based on a preset multimodal machine learning model is constructed in a preset Python environment according to preset standards.

8. The method for early non-invasive endometrial cancer assessment based on machine learning according to claim 7, characterized in that: The multimodal machine learning model consists of basic modules and meta-modules interconnected in a cascade manner. The basic module includes six tree-based bagging and boosting basic classifiers. The above basic classifiers run in parallel and independently predict the features of the input training set according to preset standards. The predetermined five-fold cross-validation method is used to implement data verification and fitting processing according to preset standards to obtain aggregated prediction results. The aggregated prediction results are then obtained through the meta-module composed of meta-classifiers, and the data of the aggregated prediction results are processed using a preset probability voting method based on classification labels, thereby alleviating the overfitting of the basic classifiers to the feature data of the training set, and thus obtaining highly accurate and robust endometrial cancer prediction results.

9. The method for early non-invasive endometrial cancer assessment based on machine learning according to claim 8, characterized in that: The method for obtaining the aggregated prediction results includes the following steps: S9.

1. Randomly divide the training set into five different subsets of equal size according to the preset criteria; S9.

2. Using the six basic classifiers, each subset obtained in step S9.1 is trained iteratively according to a preset standard, thereby obtaining five prediction result values. S9.

3. The six basic classifiers use a five-fold cross-validation method to perform data verification and aggregation processing on the prediction result values ​​obtained in step S9.2, thereby obtaining an aggregated prediction result.

10. The method for predicting and evaluating early-stage non-invasive endometrial cancer based on machine learning according to claim 9, characterized in that: In step S4, the method for obtaining the prediction result of endometrial cancer of the patient to be tested includes the following steps: S10.

1. Inputting the endometrial lesion indicator data of the patient to be tested into the disease prediction model, and iteratively training the indicator data using a basic classifier to obtain a number of corresponding endometrial cancer prediction result values; S10.

2. Using a five-fold cross-validation method, the basic classifier performs data validation and fitting on the endometrial cancer prediction result value in step S10.1 to obtain an aggregated prediction result for endometrial cancer; S10.

3. Perform an evaluation test on the aggregated prediction results of endometrial cancer in step S10.2 using a pre-defined receiver operating characteristic (ROC) curve analysis method according to pre-defined criteria to identify classification thresholds for true positive (TP), true negative (TN), false positive (FP), and false negative (FN) of endometrial cancer; S10.

4. Using the preset Youden Index, perform a cutoff value acquisition process on the classification thresholds in step S10.3 according to preset standards using the preset Youden Index = Sensitivity + Specificity - 1 function. This process obtains the maximum value among the classification thresholds and uses this maximum value as the cutoff value. S10.

5. The cutoff value in step S10.4 is input into a meta-module composed of a meta-classifier. The meta-module performs data processing on the cutoff value using a probabilistic voting method based on classification labels, thereby outputting a decision to select a category with the most votes. This category is the endometrial cancer prediction result. Based on the endometrial cancer prediction result, the patient is classified into a corresponding low-risk group or high-risk group to achieve early non-invasive screening of the patient and a prediction evaluation of endometrial cancer with high prediction accuracy and low economic burden.

Images