Cervical cancer assessment method based on high-risk HPV genotyping data

Abstract

The invention relates to the technical field of machine learning, in particular to a cervical cancer assessment method based on high-risk HPV genotyping data. The method comprises the following steps: collecting cervical cancer high-risk HPV genetic typing sample data; performing data preprocessing on the sample data; segmenting the sample data into model training data, model calibration data and model verification data by adopting an equal-proportion layered random sampling method, and constructing a cervical cancer positive probability prediction model according to the model training data; and predicting to-be-detected sample data according to the cervical cancer positive probability prediction model. According to the technical scheme, the cervical cancer positive probability prediction model is constructed based on the high-risk HPV genotyping sample data and the machine learning method, the prediction accuracy is improved by optimizing the model, the method can be applied to early screening of cervical cancer, and valuable auxiliary reference information is provided for clinicians.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and in particular to a method for evaluating cervical cancer based on high-risk HPV genotyping data. Background Art

[0002] Cervical cancer is one of the most common gynecological malignancies in my country. Its incidence rate is second only to breast cancer, ranking second among female malignancies, which seriously affects the physical and mental health of women. With the increasing number of cases and deaths of cervical cancer every year, and the younger age of the onset population, the screening and prevention of cervical cancer is extremely important. In the process of cervical cancer screening, high-risk HPV genotyping data is of great significance for understanding HPV infection, preventing cervical cancer and related diseases. High-risk HPV refers to those HPV types that are closely related to serious diseases such as cervical cancer and cervical precancerous lesions. According to the classification of the International Agency for Research on Cancer, high-risk HPV mainly includes the following types: type 16, type 18, type 31, type 33, etc. Persistent infection with high-risk HPV is the main pathogenic factor of cervical cancer and cervical precancerous lesions. Clinical studies have shown that more than 95% of cervical cancer cases are related to high-risk HPV infection. Therefore, the detection and monitoring of high-risk HPV is of great significance for the prevention of cervical cancer.

[0003] With the rapid development of artificial intelligence technology, machine learning technology has played a vital role in the auxiliary screening and diagnosis of various diseases. However, in the field of cervical cancer, there has not been a good application of high-risk HPV genotyping data for finding high-risk populations, and it is impossible to conduct efficient disease screening and positioning for high-risk populations, thereby achieving early prevention and treatment and reducing the incidence rate.

[0004] Therefore, how to use the existing high-risk HPV genotyping data, pathological data resources and machine learning technology to realize the prediction and evaluation of cervical cancer in patients is an urgent problem that needs to be solved. Summary of the invention

[0005] One of the purposes of the present invention is to provide a cervical cancer assessment method based on high-risk HPV genotyping data, to construct a cervical cancer positive probability prediction model based on high-risk HPV genotyping sample data and machine learning methods, to improve the accuracy of prediction by optimizing the model, and to apply it to early screening of cervical cancer, providing valuable auxiliary reference information for clinicians.

[0006] In order to solve the above technical problems, the present invention provides a cervical cancer assessment method based on high-risk HPV genotyping data, comprising the following steps:

[0007] S1. Collect high-risk HPV genotyping sample data for cervical cancer;

[0008] S2, performing data preprocessing on the sample data;

[0009] S3, using the equal proportion stratified random sampling method to divide the sample data into model training data, model calibration data and model verification data, the equal proportion stratified random sampling formula is: Among them, h is the number of layers into which the sample data is divided, n h is the sample size drawn at the hth level, N h represents the population size of the hth layer, N represents the population capacity, and n represents the total sample size;

[0010] S4, constructing a cervical cancer positive probability prediction model according to the model training data;

[0011] S5. Predict the sample data to be tested according to the cervical cancer positive probability prediction model.

[0012] Furthermore, the data preprocessing of the sample data comprises the following steps:

[0013] S21, confirming that the sample data is logically consistent with the commonly used medical reference value range;

[0014] S22. Set a default missing value for each data feature;

[0015] S23. Perform one-hot encoding on the sample data with multi-classification features.

[0016] Furthermore, the distribution ratio of the model training data, the model calibration data and the model verification data is 8:1:1.

[0017] Furthermore, the constructing of a cervical cancer positive probability prediction model according to the model training data comprises the following steps:

[0018] S41, performing category imbalance processing on the model training data;

[0019] S42, using the XGBoost method to train the cervical cancer positive probability prediction model using the model training data, the partial loss function formula of the XGBoost method is:

[0020]

[0021] Among them, x i is the i-th eigenvector, y i is the label corresponding to the i-th eigenvector, is the predicted value of the model, f t (x i ) is the predicted value of the tree model at step t, is the square of the predicted value of the tree model at step t, gi is the first-order derivative of the loss function, h i is the second-order derivative of the loss function;

[0022] S43, calibrating the cervical cancer positive probability prediction model using the rank-preserving regression method according to the model calibration data, the objective function of the calibration algorithm is: Where n represents the number of calibration samples, y i represents the label of the i-th sample, Represents the probability of the model output for the i-th sample;

[0023] S44, using a ten-fold cross validation method to tune the hyperparameters of the cervical cancer positive probability prediction model, the calculation formula of the ten-fold cross validation is: Where E is the total error, E i is the error on the i-th fold;

[0024] S45, using the Bayesian optimization method to optimize the hyperparameters of the cervical cancer positive probability prediction model, the Bayesian optimization acquisition functions are:

[0025] α EI (x) = E[max(0,f(x)-f(x) + ))],

[0026] α UCB (x) = μ(x) + κ·σ(x),

[0027]

[0028] Among them, α EI (x) is the expected improvement acquisition function, α UCB (x) is the confidence upper bound acquisition function, α PI (x) is the improved probability acquisition function, f(x) is the objective function, and f(x + ) is the current optimal value, κ is the parameter that controls the trade-off between exploration and exploitation, μ(x) is the mean, σ(x) is the standard deviation, E is the expectation, and Φ is the cumulative distribution function;

[0029] S46. Use the SHAP model interpretation method to evaluate the contribution of the characteristics of the sample data to the cervical cancer positive probability prediction model. The calculation formula of the Shapley value of the SHAP model interpretation method is:

[0030]

[0031] in, represents all subsets that do not contain the i-th feature, f(S) represents the predicted value of the model on the feature set S, |N| is the total number of features, |S| is the size of the feature set S, and f(S∪{i})-f(S) is the contribution of the i-th feature to the prediction;

[0032] S47, based on the evaluation results of the SHAP model interpretation method, select key features to simplify the model;

[0033] S48, using the model validation data to validate and evaluate the cervical cancer positive probability prediction model;

[0034] S49. Use a decision curve analysis method to evaluate the clinical practicality of the cervical cancer positive probability prediction model. The net benefit calculation formula of the decision curve analysis method is: Where N is the total number of samples, p t is the decision threshold for determining the probability of being positive, True Positives indicates the number of true positives predicted by the model at a given decision threshold, and False Positives indicates the number of false positives predicted by the model at a given decision threshold.

[0035] Furthermore, the class imbalance processing method includes oversampling, undersampling and class weight adjustment.

[0036] Different from the prior art, the technical solution of the present invention has the following beneficial effects:

[0037] 1. During the preprocessing of sample data, the technical solution of this application ensures the rationality and effectiveness of sample data through technical means such as logical range verification and one-hot encoding, which can improve the training accuracy of the cervical cancer positive probability prediction model.

[0038] 2. The cervical cancer positive probability prediction model of the present application adopts the XGBoost scalable distributed gradient boosting decision tree machine learning method, and uses the rank-preserving regression method for model calibration. The ten-fold cross-validation method and the Bayesian optimization method are used to tune the model's hyperparameters, perform category imbalance processing, and combine the SHAP model interpretation method to simplify and optimize the model. The decision curve analysis method is used to evaluate the practicality of the model to ensure that the model's cervical cancer positive prediction results are more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a flowchart of the steps of the cervical cancer assessment method based on high-risk HPV genotyping data of the present invention.

[0040] Figure 2 It is a flow chart of the steps of preprocessing sample data in the present invention.

[0041] Figure 3It is a flow chart of the steps of constructing a cervical cancer positive probability prediction model based on model training data in the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] like Figure 1 FIG. 1 is a flowchart of the method for evaluating cervical cancer based on high-risk HPV genotyping data of the present invention, comprising the following steps:

[0044] S1. Collect high-risk HPV genotyping sample data for cervical cancer.

[0045] HPV genotyping data refers to indicators related to the genotype of the patient's HPV human papillomavirus. HPV infection with different genotypes can lead to different clinical lesions. Usually, HPV is divided into high-risk and low-risk types based on biological characteristics and carcinogenic potential. It has been confirmed that persistent infection with high-risk HPV is closely related to the occurrence of cervical cancer. High-risk HPV genotyping data is of great significance for understanding HPV infection, preventing cervical cancer and related diseases. According to the classification of the International Agency for Research on Cancer, high-risk HPV mainly includes the following types: type 16, type 18, type 31, type 33, type 35, etc. The high-risk HPV genotyping data described in this application refers to a certain number of individual data used to learn and construct a machine learning classification prediction model. The specific collection method can be data obtained from community cervical cancer screening or data from opportunistic cervical cancer screening in hospitals.

[0046] S2. Preprocess the sample data. Figure 2 FIG. 1 is a flowchart of the steps of preprocessing sample data according to the present invention, which includes the following steps:

[0047] S21. Confirm that the sample data is logically consistent with the commonly used medical reference value range.

[0048] In cervical cancer screening and prevention research, ensuring that the collected data is logical and within the commonly used medical reference value range is an important step in data quality control. By checking the data, its accuracy and consistency can be ensured, and analysis bias caused by data errors can be avoided.

[0049] S22. Set a default missing value for each data feature.

[0050] In cervical cancer screening and prevention research, processing missing values ​​in data is an essential step in data analysis. Processing missing values ​​can not only improve the accuracy of the model, but also avoid the bias caused by missing data. Setting missing value indicator variables helps to process missing data in subsequent statistical analysis and machine learning models, and evaluate the potential impact of missing data.

[0051] S23. Perform one-hot encoding on the sample data with multi-classification features.

[0052] In cervical cancer screening and prevention research, some data features (such as screening methods, regions, pathological diagnosis, etc.) may be multi-classified. This method can convert these features into a form suitable for machine learning model processing. One-Hot Encoding for multi-class features is a common encoding method for converting categorical variables into binary vectors. The one-hot encoding formula is: one_hot(C i )=[0,0,…,1,…,0], where C i Indicates the i-th feature category, where i is the feature category number. For the multi-class features in the sample data, one-hot encoding is performed. This method can convert these features into a form suitable for processing by machine learning models.

[0053] S3, using the equal proportion stratified random sampling method to divide the sample data into model training data, model calibration data and model verification data, the equal proportion stratified random sampling formula is: Among them, h is the number of layers into which the sample data is divided, n h is the sample size drawn at the hth level, N h represents the overall size of the hth layer, N represents the overall capacity, and n represents the total sampling sample size.

[0054] Stratified random sampling means that when dividing the sample data set into training, calibration, and validation sets, the proportion of endpoint events in each subset is consistent with the proportion in the entire data set. This method helps to avoid data imbalance problems caused by random division, especially when the sample size is small or the endpoint events are rare.

[0055] In a preferred embodiment, the distribution ratio of the model training data, model calibration data and model verification data is 8:1:1. That is, 80% of the sample data are randomly selected for model training, 10% of the sample data are used for model calibration, and the remaining 10% of the sample data are used for model internal verification.

[0056] S4. Construct a cervical cancer positive probability prediction model based on the model training data. Figure 3As shown, it is a flowchart of the steps of constructing a cervical cancer positive probability prediction model according to model training data in the present invention, including the following steps:

[0057] S41, performing category imbalance processing on the model training data.

[0058] Class imbalance is a common problem in machine learning and data mining, especially in classification tasks. Class imbalance occurs when the number of samples in one class is much larger than that in other classes. This problem can severely affect the accuracy and performance of the model, as the model tends to predict the majority class and ignore the minority class.

[0059] In a preferred embodiment, the class imbalance processing method described in the present application includes oversampling, undersampling and class weight adjustment. Among them, random oversampling is to simply copy minority class samples to increase their number, and balance the data set by increasing the number of minority class samples; undersampling is to randomly delete majority class samples to reduce their number, and balance the data set by reducing the number of majority class samples; class weight is to balance the data set by giving a larger weight to minority class samples in the loss function.

[0060] S42. Using the model training data, the XGBoost method is used to train the cervical cancer positive probability prediction model.

[0061] XGBoost, full name eXtreme Gradient Boosting, is an ensemble learning algorithm based on the Gradient Boosting Decision Tree (GBDT). It has made many improvements on the basis of GBDT, including the introduction of regularization terms, the use of second-order derivative information, and support for parallel processing, which significantly improves the performance and generalization ability of the model. Preferably, the partial loss function formula of the XGBoost method is:

[0062]

[0063] Among them, x i is the i-th eigenvector, y i is the label corresponding to the i-th eigenvector, is the predicted value of the model, f t (x i ) is the predicted value of the tree model at step t, is the square of the predicted value of the tree model at step t, g i is the first-order derivative of the loss function, h i is the second-order derivative of the loss function.

[0064] S43, calibrating the cervical cancer positive probability prediction model using the rank-preserving regression method according to the model calibration data, the objective function of the calibration algorithm is: Where n represents the number of calibration samples, y i represents the label of the i-th sample, Represents the probability of the model output for the i-th sample.

[0065] In model development and evaluation, isotonic regression is an effective calibration algorithm, especially for ensuring that model predictions meet certain monotonicity constraints. This method uses nonparametric regression techniques to adjust the predictions to make them more consistent with the monotonic trend of the actual observed data.

[0066] S44, using a ten-fold cross validation method to tune the hyperparameters of the cervical cancer positive probability prediction model, the calculation formula of the ten-fold cross validation is: Where E is the total error, E i is the error on the i-th fold.

[0067] Using 10-fold cross validation to tune the model's hyperparameters is a common and effective method in machine learning. Cross validation can help evaluate the performance of the model and select the best hyperparameter settings, thereby improving the generalization ability of the model. Ten-fold cross validation randomly divides the original sample data set into 10 non-overlapping, equal-sized subsets (or "folds"). Then, 10 independent training and testing processes are performed. In each iteration, one of the subsets is selected as the test set, and the remaining 9 subsets are used as training sets. In this way, each subset has the opportunity to be used as a test set, so that the model hyperparameters can be tuned more comprehensively.

[0068] S45. Use the Bayesian optimization method to optimize the hyperparameters of the cervical cancer positive probability prediction model.

[0069] Bayesian optimization is an effective method for optimizing hyperparameters, especially for optimizing models with high computational complexity. It establishes a probabilistic model of the objective function and uses this model to guide the search process to find the optimal hyperparameters. Preferably, the acquisition functions of the Bayesian optimization are:

[0070] α EI (x) = E[max(0,f(x)-f(x) + ))],

[0071] α UCB (x) = μ(x) + κ·σ(x),

[0072]

[0073] Among them, α EI (x) is the expected improvement acquisition function, α UCB (x) is the confidence upper bound acquisition function, α PI (x) is the improved probability acquisition function, f(x) is the objective function, and f(x + ) is the current optimal value, κ is the parameter that controls the trade-off between exploration and exploitation, μ(x) is the mean, σ(x) is the standard deviation, E is the expectation, and Φ is the cumulative distribution function.

[0074] S46. Use the SHAP model interpretation method to evaluate the contribution of the characteristics of the sample data to the cervical cancer positive probability prediction model.

[0075] SHAP (SHapley Additive exPlanations) is a game-theoretic method for explaining the predictions of machine learning models. It provides a unified framework that can explain the output of any machine learning model, including linear models, decision trees, random forests, gradient boosting machines, and deep learning models. SHAP values ​​are based on the concept of Shapley values ​​and are used to quantify the contribution of each feature to the model's prediction results. Using the SHAP method to evaluate the contribution of each candidate feature to predictive modeling is an effective means of understanding model output and feature importance. SHAP values ​​provide the specific contribution of each feature to the prediction results, helping to identify the most important features for a specific prediction (such as CIN2+ / CIN3+ occurrence).

[0076] In this application, the calculation formula of the Shapley value of the SHAP model interpretation method is:

[0077]

[0078] in, represents all subsets that do not contain the i-th feature, f(S) represents the predicted value of the model on the feature set S, |N| is the total number of features, |S| is the size of the feature set S, and f(S∪{i})-f(S) is the contribution of the i-th feature to the prediction.

[0079] S47. Based on the evaluation results of the SHAP model interpretation method, key features are selected to simplify the model.

[0080] Based on the SHAP algorithm, combined with the judgment of clinical practice experts, pathology experts, and statistics experts, the importance of the features is ranked comprehensively, and key features are selected to simplify the model: By combining the SHAP algorithm and expert judgment, key features can be scientifically selected and the model can be simplified. The SHAP value provides quantitative data on the importance of the features for prediction, while the clinical and scientific knowledge of the experts helps to confirm the practical significance of these features. This approach can improve the interpretability and clinical practicality of the model, while reducing the complexity of the model for easy implementation and interpretation.

[0081] Using different dimensions of evaluation indicators to evaluate the auxiliary diagnostic value of the simplified model for CIN2+ / CIN3+: Using multiple evaluation indicators to evaluate the performance of the simplified model can fully understand the auxiliary diagnostic value of the model for CIN2+ / CIN3+. Indicators such as accuracy, precision, recall, F1 score, ROC curve and AUC, PR curve and AUC provide a comprehensive evaluation of model performance, which helps to select the most appropriate model in practical applications.

[0082] S48. Use the model validation data to validate and evaluate the cervical cancer positive probability prediction model.

[0083] Use internal validation, time series validation, and external validation to evaluate the overall prediction performance of the model from different validation types, and at the same time, compare the pros and cons of different algorithm models: Through internal validation, time series validation, and external validation, the prediction performance of the model can be comprehensively evaluated from different perspectives. These validation methods help confirm the stability, time dependency, and generalization ability of the model. At the same time, by comparing the pros and cons of different algorithm models, the model that best suits a specific task can be selected and its performance can be optimized.

[0084] S49. Use the decision curve analysis method to evaluate the clinical practicality of the cervical cancer positive probability prediction model.

[0085] DCA, or decision curve analysis, is a statistical method for evaluating the clinical practical value of prediction models. It comprehensively considers the prediction accuracy of the model, the patient's preferences, and the decision results under different thresholds, thus providing a more comprehensive and intuitive evaluation method. DCA helps determine whether the model can provide valuable information within a specific threshold range and compare it with other possible strategies.

[0086] In this application, the net benefit calculation formula of the decision curve analysis method is:

[0087] Where N is the total number of samples, p tis the decision threshold for determining the probability of being positive, True Positives indicates the number of true positives predicted by the model at a given decision threshold, and False Positives indicates the number of false positives predicted by the model at a given decision threshold.

[0088] It should be noted that building a cervical cancer positive probability prediction model is a process of continuous tuning, and the above steps can be repeated multiple times according to the results of model verification until the prediction verification results achieve the expected effect.

[0089] S5. Predict the sample data to be tested according to the cervical cancer positive probability prediction model.

[0090] Finally, after the cervical cancer positive probability prediction model reaches the optimal value, the sample data of the patient to be tested can be collected according to the format of the sample data, and input into the cervical cancer positive probability prediction model, and the model predicts the probability value of the sample data to be tested being positive for cervical cancer. The cervical cancer positive probability prediction result of the present application can assist doctors in the early diagnosis and treatment of cervical cancer.

[0091] The above specific implementation modes are only detailed explanations of the technical solutions of the present invention. The present invention is not limited to the above embodiments. Any improvement or replacement based on the principles of the present invention should be within the protection scope of the present invention.

Claims

1. A cervical cancer assessment method based on high-risk HPV genotyping data, characterized in that: The following steps are involved: S1. Collect high-risk HPV genotyping sample data for cervical cancer; S2, performing data preprocessing on the sample data; S3, using the equal proportion stratified random sampling method to divide the sample data into model training data, model calibration data and model verification data, the equal proportion stratified random sampling formula is: Among them, h is the number of layers into which the sample data is divided, n h is the sample size drawn at the hth level, N h represents the population size of the hth layer, N represents the population capacity, and n represents the total sample size; S4, constructing a cervical cancer positive probability prediction model according to the model training data; S5. Predict the sample data to be tested according to the cervical cancer positive probability prediction model.

2. The cervical cancer assessment method based on high-risk HPV genotyping data according to claim 1, characterized in that: The data preprocessing of the sample data comprises the following steps: S21, confirming that the sample data is logically consistent with the commonly used medical reference value range; S22. Set a default missing value for each data feature; S23. Perform one-hot encoding on the sample data with multi-classification features.

3. The cervical cancer assessment method based on high-risk HPV genotyping data according to claim 1, characterized in that: The distribution ratio of the model training data, model calibration data and model verification data is 8:1:

1.

4. The cervical cancer assessment method based on high-risk HPV genotyping data according to claim 1, characterized in that: The method of constructing a cervical cancer positive probability prediction model according to the model training data comprises the following steps: S41, performing category imbalance processing on the model training data; S42, using the XGBoost method to train the cervical cancer positive probability prediction model using the model training data, the partial loss function formula of the XGBoost method is: Among them, x i is the i-th eigenvector, y i is the label corresponding to the i-th eigenvector, is the predicted value of the model, f t (x i ) is the predicted value of the tree model at step t, is the square of the predicted value of the tree model at step t, g i is the first-order derivative of the loss function, h i is the second-order derivative of the loss function; S43, calibrating the cervical cancer positive probability prediction model using the rank-preserving regression method according to the model calibration data, the objective function of the calibration algorithm is: Where n represents the number of calibration samples, y i represents the label of the i-th sample, Represents the probability of the model output for the i-th sample; S44, using a ten-fold cross validation method to tune the hyperparameters of the cervical cancer positive probability prediction model, the calculation formula of the ten-fold cross validation is: Where E is the total error, E i is the error on the i-th fold; S45, using the Bayesian optimization method to optimize the hyperparameters of the cervical cancer positive probability prediction model, the Bayesian optimization acquisition functions are: α EI (x)=E[max(0,f(x)-f(x + ))], a UCB (x)=μ(x)+κ·σ(x), Among them, α EI (x) is the expected improvement acquisition function, α UCB (x) is the confidence upper bound acquisition function, α PI (x) is the improved probability acquisition function, f(x) is the objective function, and f(x + ) is the current optimal value, κ is the parameter that controls the trade-off between exploration and exploitation, μ(x) is the mean, σ(x) is the standard deviation, E is the expectation, and Φ is the cumulative distribution function; S46. Use the SHAP model interpretation method to evaluate the contribution of the characteristics of the sample data to the cervical cancer positive probability prediction model. The calculation formula of the Shapley value of the SHAP model interpretation method is: in, represents all subsets that do not contain the i-th feature, f(S) represents the predicted value of the model on the feature set S, |N| is the total number of features, |S| is the size of the feature set S, and f(S∪{i})-f(S) is the contribution of the i-th feature to the prediction; S47, based on the evaluation results of the SHAP model interpretation method, select key features to simplify the model; S48, using the model validation data to validate and evaluate the cervical cancer positive probability prediction model; S49. Use a decision curve analysis method to evaluate the clinical practicality of the cervical cancer positive probability prediction model. The net benefit calculation formula of the decision curve analysis method is: Where N is the total number of samples, p t is the decision threshold for determining the probability of being positive, True Positives indicates the number of true positives predicted by the model at a given decision threshold, and False Positives indicates the number of false positives predicted by the model at a given decision threshold.

5. The cervical cancer assessment method based on high-risk HPV genotyping data according to claim 4, characterized in that: The class imbalance processing method includes oversampling, undersampling and class weight adjustment.

Images