Training methods, prediction methods, equipment, and media for preoperative PTC risk prediction models

Abstract

The present application discloses a training method, prediction method, device and medium for a preoperative PTC risk prediction model. According to preset sample screening criteria, preoperative clinical information and postoperative PTC risk stratification results of preoperative FNA specimens of samples are obtained; predictive features are determined from the preoperative clinical information; the original BRAFV600E mutation abundance of each FNA specimen obtained by ddPCR technology is obtained, the tumor purity of each FNA specimen is detected based on a preset tumor purity detection method, and the original BRAFV600E mutation abundance is divided by the tumor purity to obtain a standard BRAFV600E mutation abundance, so as to determine a training preoperative FNA specimen from the preoperative FNA specimens of each sample; the predictive features and the standard BRAFV600E mutation abundance of each training preoperative FNA specimen are used as input features of the model, the postoperative PTC risk stratification results of each training preoperative FNA specimen are used as labels of the input features, and low risk of preoperative PTC or medium and high risk of preoperative PTC is used as the prediction result of the model, and a preoperative PTC risk prediction model is obtained based on training of the XGBoost algorithm.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular, to a training method, prediction method, device, and medium for a preoperative PTC risk prediction model. Background Art

[0002] Thyroid cancer (TC) is an endocrine solid tumor with a steadily increasing incidence. Approximately 90% of thyroid tumors are pathologically classified as papillary thyroid carcinoma (PTC). Although most PTCs have a favorable overall prognosis, with a 10-year survival rate exceeding 90%, some tumors exhibit highly aggressive biological behaviors in the early stages, such as cervical lymph node metastasis (LNM), invasion of extraglandular soft tissues or the trachea and esophagus, and distant metastasis. These highly invasive characteristics often indicate recurrence and metastasis. Once PTC progresses to the late stage, all treatment options are unlikely to benefit the patient, and survival rates are significantly reduced.

[0003] Currently, fine needle aspiration cytology (FNAB) combined with imaging studies has a high diagnostic efficacy for PTC. However, the accuracy of these tests in assessing biological characteristics such as cervical lymph node metastasis and tumor invasion is only approximately 60%. Studies have shown that among patients diagnosed preoperatively as cN0 ("c" represents clinical stage, "N" represents regional lymph nodes, and "0" indicates no evidence of regional lymph node metastasis), 15.3% to 40.9% of postoperative pathology reports indicate lymph node metastasis (pN1). Other approaches have also attempted to construct clinical prediction models based on postoperative clinical characteristics, but retrospective studies have limited practical value.

[0004] Therefore, a more accurate method to predict the risk of PTC before surgery is urgently needed. Summary of the Invention

[0005] The present application aims to solve one of the technical problems in the related art to a certain extent. To this end, the present application provides a training method, prediction method, device and medium for a preoperative PTC risk prediction model.

[0006] As a first aspect of the present application, a method for training a preoperative papillary thyroid carcinoma (PTC) risk prediction model is provided, wherein the method comprises:

[0007] According to the preset sample screening criteria, preoperative clinical information of fine needle aspiration (FNA) specimens and postoperative PTC risk stratification results were obtained;

[0008] determining a predictive signature from the preoperative clinical information;

[0009] Obtaining the original BRAFV600E mutation abundance of each FNA specimen detected by droplet digital polymerase chain reaction (ddPCR) technology, detecting the tumor purity of each FNA specimen based on a preset tumor purity detection method, and dividing the original BRAFV600E mutation abundance by the tumor purity to obtain a standard BRAFV600E mutation abundance of the corresponding FNA specimen;

[0010] According to the standard BRAFV600E mutation abundance, a training preoperative FNA specimen is determined from the preoperative FNA specimens of each sample;

[0011] The predictive features of each training preoperative FNA specimen and the standard BRAFV600E mutation abundance are used as the input features of the model, the postoperative PTC risk stratification results of each training preoperative FNA specimen are used as the labels of the corresponding input features, and the preoperative PTC low risk or preoperative PTC medium-high risk are used as the prediction results of the model. A preoperative PTC risk prediction model is obtained based on the extreme gradient boosting XGBoost algorithm.

[0012] Optionally, the predictive features include gender, age and ultrasound features, and the ultrasound features include ultrasound measurement of tumor diameter, ultrasound display of multifocality, ultrasound assessment of extramembranous invasion, and ultrasound combined with serum indicators suggesting Hashimoto's thyroiditis.

[0013] Optionally, determining a predictive feature from the preoperative clinical information includes:

[0014] The postoperative PTC risk stratification results of the preoperative FNA specimens of each of the samples are used as prediction targets, and the preoperative clinical information of the preoperative FNA specimens of each of the samples is used as candidate features. The predictive features are selected from the candidate features based on the LASSO algorithm; wherein, the postoperative PTC risk stratification results include low risk of postoperative PTC and medium and high risk of postoperative PTC.

[0015] Optionally, determining a training preoperative FNA specimen from each of the preoperative FNA specimens according to the standard BRAFV600E mutation abundance comprises:

[0016] Preoperative FNA specimens with standard BRAFV600E mutation abundance exceeding the preset abundance threshold are identified as BRAFV600E mutation-positive specimens, and preoperative FNA specimens with standard BRAFV600E mutation abundance not exceeding the preset abundance threshold are identified as BRAFV600E mutation-negative specimens;

[0017] From the preoperative FNA specimens of each sample, the specimens positive for BRAFV600E mutation were determined as training preoperative FNA specimens.

[0018] Optionally, the preset sample screening criteria include inclusion criteria and exclusion criteria; the inclusion criteria include: the age of the patient corresponding to the sample is within the preset age range, the patient corresponding to the sample is confirmed to have typical PTC by postoperative pathology, the patient corresponding to the sample has undergone unilateral thyroid lobectomy or total thyroidectomy and central lymph node dissection, the patient corresponding to the sample has undergone therapeutic lateral neck lymph node dissection in the case of lateral neck lymph node metastasis, and the preoperative clinical information and postoperative clinical information of the patient corresponding to the sample are complete; the exclusion criteria include: the patient corresponding to the sample has a history of neck trauma, the patient corresponding to the sample has other tumors, the postoperative pathological diagnosis of the patient corresponding to the sample is PTC subtype or non-PTC, and the preoperative clinical information and postoperative clinical information of the patient corresponding to the sample are incomplete.

[0019] Optionally, the step of using the predictive features of each of the training preoperative FNA specimens and the standard BRAFV600E mutation abundance as input features of the model, using the postoperative PTC risk stratification results of each of the training preoperative FNA specimens as labels of the corresponding input features, and using preoperative PTC low risk or preoperative PTC medium-high risk as the prediction results of the model, and training the preoperative PTC risk prediction model based on the extreme gradient boosting XGBoost algorithm, includes:

[0020] The predictive features of each training preoperative FNA specimen and the standard BRAFV600E mutation abundance are divided into a training set and a validation set according to a preset ratio. The training set is used for the XGBoost algorithm to construct a model, and the validation set is used for the XGBoost algorithm to verify the constructed model.

[0021] Optionally, the XGBoost algorithm optimizes the hyperparameters of the model based on a five-fold cross-validation algorithm and a Bayesian algorithm until the area under the receiver operating characteristic curve (AUC) of the model on the validation set reaches a maximum;

[0022] The XGBoost algorithm optimizes the parameters of the model based on an objective function, which includes a training loss and a regularization term between the label of the input feature and the prediction result corresponding to the input feature.

[0023] As a second aspect of the present application, a preoperative PTC risk prediction method is provided, wherein the method comprises:

[0024] To obtain the predictive characteristics of preoperative FNA specimens in patients with preoperative diagnosis of PTC;

[0025] Obtaining the original BRAFV600E mutation abundance of the predicted preoperative FNA specimen detected based on ddPCR technology, detecting the tumor purity of the predicted preoperative FNA specimen based on a preset tumor purity detection method, and dividing the original BRAFV600E mutation abundance by the tumor purity to obtain a standard BRAFV600E mutation abundance of the predicted preoperative FNA specimen;

[0026] The predictive features of the preoperative FNA specimens and the standard BRAFV600E mutation abundance are used as input features of the model and input into the preoperative PTC risk prediction model trained by the training method of the preoperative PTC risk prediction model according to any one of claims 1-7 to obtain the prediction results output by the preoperative PTC risk prediction model.

[0027] As a third aspect of the present application, an electronic device is provided, wherein the electronic device includes:

[0028] one or more processors;

[0029] A memory having one or more computer programs stored thereon, which, when executed by the one or more processors, causes the one or more processors to implement any of the following:

[0030] The first aspect of this application provides a training method for a preoperative PTC risk prediction model;

[0031] The second aspect of this application provides a preoperative PTC risk prediction method.

[0032] As a fourth aspect of the present application, a computer-readable medium is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, any of the following is implemented:

[0033] The first aspect of this application provides a training method for a preoperative PTC risk prediction model;

[0034] The second aspect of this application provides a preoperative PTC risk prediction method.

[0035] The training method of the preoperative PTC risk prediction model provided in the embodiment of the present application obtains the preoperative clinical information and postoperative PTC risk stratification results of the preoperative fine needle aspiration FNA specimen according to the preset sample screening criteria, determines the predictive features from the preoperative clinical information, obtains the original BRAFV600E mutation abundance of each FNA specimen detected by droplet digital polymerase chain reaction ddPCR technology, detects the tumor purity of each FNA specimen based on a preset tumor purity detection method, and divides the original BRAFV600E mutation abundance by the tumor purity to obtain the corresponding FNA The method uses the standard BRAFV600E mutation abundance of the sample to determine the training preoperative FNA specimens from the preoperative FNA specimens of each sample based on the standard BRAFV600E mutation abundance. The predictive features of each training preoperative FNA specimen and the standard BRAFV600E mutation abundance are used as the input features of the model. The postoperative PTC risk stratification results of each training preoperative FNA specimen are used as the labels of the corresponding input features. The low risk of preoperative PTC or the medium-high risk of preoperative PTC is used as the prediction result of the model. The preoperative PTC risk prediction model is obtained by training based on the extreme gradient boosting (XGBoost) algorithm. The method can combine the quantitative results of the BRAFV600E mutation, i.e., the standard BRAFV600E mutation abundance, with preoperative clinical information to construct a preoperative PTC risk prediction model, improve the accuracy of preoperative PTC risk prediction, and provide strong data support for clinical diagnosis and treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The present application will be further described below with reference to the accompanying drawings:

[0037] Figure 1 This is a flowchart of an implementation method of a preoperative PTC risk prediction model training method provided in an embodiment of the present application;

[0038] Figure 2 This is a flowchart of another embodiment of the method for training a preoperative PTC risk prediction model provided in the examples of the present application;

[0039] Figure 3 This is a flowchart of another embodiment of the method for training a preoperative PTC risk prediction model provided in the examples of the present application;

[0040] Figure 4 This is a flowchart of an implementation method of the preoperative PTC risk prediction method provided in the examples of the present application;

[0041] Figure 5 This is a module diagram of an implementation of an electronic device provided in an embodiment of the present application;

[0042] Figure 6It is a schematic diagram of the computer-readable medium provided in an embodiment of the present application.

[0043] Description of Reference Numerals

[0044] 101: Processor 102: Memory

[0045] 103: I / O interface 104: bus DETAILED DESCRIPTION

[0046] The embodiments of the present application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described in the embodiments are intended to be used to explain the present application and are not to be construed as limiting the present application.

[0047] References in this specification to "one embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with the embodiment itself can be included in at least one embodiment disclosed herein. The appearances of the phrase "in one embodiment" in various places in the specification are not necessarily all referring to the same embodiment.

[0048] Thyroid cancer (TC) is an endocrine solid tumor with a steadily increasing incidence. Approximately 90% of thyroid tumors are pathologically classified as papillary thyroid carcinoma (PTC). Although most PTCs have a favorable overall prognosis, with a 10-year survival rate exceeding 90%, some tumors exhibit highly aggressive biological behaviors in the early stages, such as cervical lymph node metastasis (LNM), invasion of extraglandular soft tissues or the trachea and esophagus, and distant metastasis. These highly invasive characteristics often indicate recurrence and metastasis. Once PTC progresses to the late stage, all treatment options are unlikely to benefit the patient, and survival rates are significantly reduced.

[0049] Currently, fine needle aspiration cytology (FNAB) combined with imaging studies has a high diagnostic efficacy for PTC. However, the accuracy of these tests in assessing biological characteristics such as cervical lymph node metastasis and tumor invasion is only approximately 60%. Studies have shown that among patients diagnosed preoperatively as cN0 ("c" represents clinical stage, "N" represents regional lymph nodes, and "0" indicates no evidence of regional lymph node metastasis), 15.3% to 40.9% of postoperative pathology reports indicate lymph node metastasis (pN1). Other approaches have also attempted to construct clinical prediction models based on postoperative clinical characteristics, but retrospective studies have limited practical value.

[0050] In this regard, the applicant of this application proposes that when predicting the risk of PTC before surgery, the information value of prospective data (i.e., preoperative clinical information) is significantly higher than that of retrospective data (i.e., postoperative clinical information). In addition, since the mutation rate of the BRAFV600E gene in PTC is 28% to 87% and it is usually used as an important molecular marker for clinical auxiliary diagnosis of thyroid cancer, the BRAFV600E mutation status also has certain information value. However, as a qualitative result, the correlation between the BRAFV600E mutation status and the biological behavior of the tumor is ultimately limited. It is worth considering detecting the specific and standard BRAFV600E mutation abundance based on droplet digital polymerase chain reaction ddPCR technology and a preset tumor purity detection method, and then combining the quantitative results of the BRAFV600E mutation and preoperative clinical information to construct a preoperative PTC risk prediction model to accurately predict the preoperative risk of PTC.

[0051] As a first aspect of an embodiment of the present application, a method for training a preoperative papillary thyroid carcinoma (PTC) risk prediction model is provided, wherein: Figure 1 As shown, the method may include:

[0052] Step S110, obtaining preoperative clinical information of the preoperative fine needle aspiration (FNA) specimen and postoperative PTC risk stratification results according to preset sample screening criteria;

[0053] Step S120, determining predictive features from the preoperative clinical information;

[0054] Step S130, obtaining the original BRAFV600E mutation abundance of each FNA specimen detected by droplet digital polymerase chain reaction (ddPCR) technology, detecting the tumor purity of each FNA specimen based on a preset tumor purity detection method, and dividing the original BRAFV600E mutation abundance by the tumor purity to obtain a standard BRAFV600E mutation abundance of the corresponding FNA specimen;

[0055] Step S140, determining a training preoperative FNA specimen from each of the preoperative FNA specimens according to the standard BRAFV600E mutation abundance;

[0056] In step S150, the predictive features of each of the training preoperative FNA specimens and the standard BRAFV600E mutation abundance are used as input features of the model, the postoperative PTC risk stratification results of each of the training preoperative FNA specimens are used as labels of the corresponding input features, and the preoperative PTC low risk or preoperative PTC medium-high risk is used as the prediction result of the model, and a preoperative PTC risk prediction model is obtained based on the extreme gradient boosting XGBoost algorithm.

[0057] It should be noted that, in the embodiments of the present application, "preoperative PTC risk prediction" refers to the prediction of the patient's PTC risk before surgery, and "postoperative PTC risk stratification result" refers to the result of stratifying the patient's PTC risk after surgery.

[0058] Among them, the embodiment of the present application screens out sample preoperative FNA specimens from a large number of preoperative FNA specimens of PTC patients according to preset sample screening criteria. The sample preoperative FNA specimens refer to the preoperative FNA specimens of PTC patients used as samples, and the training preoperative FNA specimens refer to the preoperative FNA specimens of PTC patients used for training the model. It can be understood that step S140 is also a selection process, and the training preoperative FNA specimens are part of the sample preoperative FNA specimens.

[0059] Among them, the embodiments of the present application do not specifically limit how to obtain the postoperative PTC risk stratification results of preoperative FNA specimens. For example, a stratified assessment of tumor risk can be performed according to the differentiated thyroid cancer diagnosis and treatment guidelines issued by the American Thyroid Association (ATA), and the PTC risk can be stratified into low postoperative PTC risk or medium-high postoperative PTC risk after surgery.

[0060] Here, step S120 is used to determine features that are actually useful for preoperative PTC risk prediction from preoperative clinical information.

[0061] The embodiment of the present application does not specifically limit how step S130 is performed. For example, the tumor purity (i.e., the tumor cell ratio) of the preoperative FNA specimen is detected according to a preset tumor purity detection method (specifically, the original pathological image is segmented to obtain multiple sub-images to be detected; the multiple sub-images to be detected are input into a pre-trained cell counting model to obtain a detection density map of each sub-image to be detected output by the cell counting model; and finally, the tumor purity detection result is determined based on the detection density map of each sub-image to be detected. The detection density map includes a cancer cell density map and a non-cancerous cell density map; the cell counting model is obtained by optimizing the parameters and hyperparameters of the initial model based on the multiple training sub-images and their reference density maps; the reference density map is generated based on the training sub-images and their annotation results, and the annotation results include the correspondence between the cell center annotation position and color information). Deoxyribonucleic acid (DNA) is extracted from the preoperative FNA specimen for use in a QX-200 Droplet Digital PCR system (QX-200 Droplet Digital PCR system). Finally, the original BRAFV600E mutation abundance detected by ddPCR was divided by the tumor purity to obtain the final standard BRAFV600E mutation abundance.

[0062] Among them, step S140 is used to determine, from the preoperative FNA specimens, the specimens whose BRAFV600E mutation abundance is actually useful for preoperative PTC risk prediction.

[0063] It can be understood that the embodiment of the present application takes low risk of preoperative PTC or medium-high risk of preoperative PTC as the prediction result of the model. For each input feature (i.e., the predictive feature of each specimen and the standard BRAFV600E mutation abundance), the trained preoperative PTC risk prediction model will output one of low risk of preoperative PTC and medium-high risk of preoperative PTC.

[0064] The training method of the preoperative PTC risk prediction model provided in the embodiment of the present application obtains the preoperative clinical information and postoperative PTC risk stratification results of the preoperative fine needle aspiration FNA specimen according to the preset sample screening criteria, determines the predictive features from the preoperative clinical information, obtains the original BRAFV600E mutation abundance of each FNA specimen detected by droplet digital polymerase chain reaction ddPCR technology, detects the tumor purity of each FNA specimen based on a preset tumor purity detection method, and divides the original BRAFV600E mutation abundance by the tumor purity to obtain the corresponding FNA The method uses the standard BRAFV600E mutation abundance of the sample to determine the training preoperative FNA specimens from the preoperative FNA specimens of each sample based on the standard BRAFV600E mutation abundance. The predictive features of each training preoperative FNA specimen and the standard BRAFV600E mutation abundance are used as the input features of the model. The postoperative PTC risk stratification results of each training preoperative FNA specimen are used as the labels of the corresponding input features. The low risk of preoperative PTC or the medium-high risk of preoperative PTC is used as the prediction result of the model. The preoperative PTC risk prediction model is obtained by training based on the extreme gradient boosting (XGBoost) algorithm. The method can combine the quantitative results of the BRAFV600E mutation, i.e., the standard BRAFV600E mutation abundance, with preoperative clinical information to construct a preoperative PTC risk prediction model, improve the accuracy of preoperative PTC risk prediction, and provide strong data support for clinical diagnosis and treatment.

[0065] In some embodiments, the determined predictive features include gender, age, and ultrasound features, wherein the ultrasound features include ultrasound measurement of tumor diameter, ultrasound display of multifocality, ultrasound assessment of extrathyroidal extension (ETE), and ultrasound combined with serum indicators indicating Hashimoto's thyroiditis (HT).

[0066] It is understood that the predictive features are determined from preoperative clinical information, which at least includes the predictive features.

[0067] Among them, ultrasound combined with serum indicators to suggest Hashimoto's thyroiditis means that ultrasound combined with serological thyroid antibody positive indicators (thyroid peroxidase antibodies (TPO-Ab) and / or thyroglobulin antibodies (Tg-Ab)) suggest HT.

[0068] The applicant of the present application further proposes that predictive features can be selected based on the LASSO (Least Absolute Shrinkage and Selection Operator) algorithm. Accordingly, in some embodiments, such as Figure 2As shown, determining the predictive features from the preoperative clinical information (i.e., what is involved in step S120) may include:

[0069] In step S121, the postoperative PTC risk stratification results of the preoperative FNA specimens of each of the samples are used as prediction targets, and the preoperative clinical information of the preoperative FNA specimens of each of the samples is used as candidate features. The predictive features are selected from the candidate features based on the LASSO algorithm; the postoperative PTC risk stratification results include low risk of postoperative PTC and medium and high risk of postoperative PTC.

[0070] The applicant of the present application further proposes that a training preoperative FNA specimen can be screened from the sample preoperative FNA specimens according to a reasonable preset abundance threshold. Figure 3 As shown, determining a training preoperative FNA specimen from each of the preoperative FNA specimens according to the standard BRAFV600E mutation abundance (i.e., step S140 involved) may include:

[0071] Step S141, determining a preoperative FNA specimen whose standard BRAFV600E mutation abundance exceeds a preset abundance threshold as a BRAFV600E mutation-positive specimen, and determining a preoperative FNA specimen whose standard BRAFV600E mutation abundance does not exceed the preset abundance threshold as a BRAFV600E mutation-negative specimen;

[0072] Step S142: From the preoperative FNA specimens of each sample, a specimen positive for BRAFV600E mutation is determined as a training preoperative FNA specimen.

[0073] As an optional specific implementation, the preset abundance threshold may be 0.1%.

[0074] The applicant of this application further proposes that by presetting reasonable sample screening criteria, the accuracy of the model can be ensured. Accordingly, in some embodiments, the preset sample screening criteria include inclusion criteria and exclusion criteria; the inclusion criteria include: the age of the patient corresponding to the sample is within the preset age range, the patient corresponding to the sample is confirmed to be a typical PTC by postoperative pathology, the patient corresponding to the sample has undergone unilateral thyroid lobectomy or total thyroidectomy and central lymph node dissection, the patient corresponding to the sample has undergone therapeutic lateral neck lymph node dissection in the case of lateral cervical lymph node metastasis, and the preoperative clinical information and postoperative clinical information of the patient corresponding to the sample are complete; the exclusion criteria include: the patient corresponding to the sample has a history of neck trauma, the patient corresponding to the sample has other tumors, the patient corresponding to the sample is diagnosed as a PTC subtype or non-PTC by postoperative pathology, and the preoperative clinical information and postoperative clinical information of the patient corresponding to the sample are incomplete.

[0075] The applicant of this application further proposes that in order to further improve the accuracy of the model, the training data can be divided into a training set for building the model and a validation set for validating the model. Accordingly, in some embodiments, the step of using the predictive features of each of the training preoperative FNA specimens and the standard BRAFV600E mutation abundance as the input features of the model, using the postoperative PTC risk stratification results of each of the training preoperative FNA specimens as the labels of the corresponding input features, and using the low risk of preoperative PTC or the medium-high risk of preoperative PTC as the prediction results of the model, and training the preoperative PTC risk prediction model based on the extreme gradient boosting XGBoost algorithm may include:

[0076] The predictive features of each training preoperative FNA specimen and the standard BRAFV600E mutation abundance are divided into a training set and a validation set according to a preset ratio. The training set is used for the XGBoost algorithm to construct a model, and the validation set is used for the XGBoost algorithm to verify the constructed model.

[0077] As an optional specific implementation, the preset ratio may be 7:3.

[0078] The applicant of this application further proposed that the area under the receiver operating characteristic curve (AUC-ROC, AUC for short) of the model on the validation set can be used to evaluate the degree of hyperparameter optimization of the model, and the training loss and regularization term between the true value and the predicted value can be used to evaluate the degree of parameter optimization of the model.

[0079] Accordingly, in some embodiments, the XGBoost algorithm optimizes the hyperparameters of the model based on a five-fold cross-validation algorithm and a Bayesian algorithm until the area under the receiver operating characteristic curve (AUC) of the model on the validation set reaches a maximum;

[0080] The XGBoost algorithm optimizes the parameters of the model based on an objective function, which includes a training loss and a regularization term between the label of the input feature and the prediction result corresponding to the input feature.

[0081] XGBoost is a method for learning an ensemble of K classification and regression trees, where each additional tree ( ) are selected from the set of all possible regression trees and added to correct errors from previous learning iterations.

[0082] The model equation can be expressed as:

[0083] (1);

[0084] The objective function can be expressed as:

[0085] (2);

[0086] In (1) and (2), Indicates the model for the i-th input feature The predicted value of Indicates the model for the i-th input feature The true value of , F represents all function spaces in the regression forest, represents the weight of the i-th sample falling on the k-th leaf, and L represents the loss function used to estimate the predicted value and the true value The Ω term penalizes the complexity of the model to avoid overfitting. The XGBoost algorithm training model is relatively conventional and will not be described in detail here.

[0087] Among them, when the predicted value of the model is lower than the cutoff value, it can be classified as low risk of preoperative PTC; otherwise, it can be classified as medium or high risk of preoperative PTC. However, the embodiment of the present application does not specifically limit how to set the cutoff value of the model.

[0088] Furthermore, the applicants of this application propose that, in addition to AUC, model performance can also be evaluated based on metrics such as sensitivity (i.e., true positive rate (TPR)), specificity (i.e., true negative rate (TNR)), positive predictive value (PPV), negative predictive value (NPV), and accuracy. For example, a receiver operating characteristic (ROC) curve can be plotted, and the 95% confidence interval (CI) of the AUC values ​​can be estimated. AUC values ​​can then be compared using a two-sided DeLong test, with the optimal cutoff value determined by the maximum Youden index (Youden index = sensitivity + specificity - 1). The Hosmer-Lemeshow (HL) test can be used to assess the calibration and validity of the preoperative PTC risk prediction model. Finally, the SHapley Additive Interpretation (SHAP) algorithm, a well-known visual model interpretation tool, can be introduced to enhance the interpretability of the XGBoost model.

[0089] As a second aspect of the embodiment of the present application, a preoperative PTC risk prediction method is provided, wherein, Figure 4 As shown, the method may include:

[0090] Step S210, obtaining predictive features of preoperative FNA specimens for preoperative diagnosis of PTC patients;

[0091] Step S220, obtaining the original BRAFV600E mutation abundance of the predicted preoperative FNA specimen detected based on ddPCR technology, detecting the tumor purity of the predicted preoperative FNA specimen based on a preset tumor purity detection method, and dividing the original BRAFV600E mutation abundance by the tumor purity to obtain a standard BRAFV600E mutation abundance of the predicted preoperative FNA specimen;

[0092] In step S230, the predictive features of the predicted preoperative FNA specimen and the standard BRAFV600E mutation abundance are used as input features of the model and input into the preoperative PTC risk prediction model trained by the training method of the preoperative PTC risk prediction model provided according to the first aspect of the embodiment of the present application to obtain the prediction results output by the preoperative PTC risk prediction model.

[0093] It is understood that patients diagnosed with PTC preoperatively refer to patients who have not undergone surgery and have been clinically diagnosed with PTC based on preoperative FNA specimens. The present embodiment uses a pre-trained preoperative PTC risk prediction model to predict their PTC risk before surgery, obtaining risk stratification prediction results for low preoperative PTC risk or medium-to-high preoperative PTC risk.

[0094] It can be understood that, during the training of the model, predictive features have been determined from preoperative clinical information, and during the application of the model, predictive features of preoperative FNA specimens for preoperative diagnosis of PTC patients can be directly obtained.

[0095] Among them, it can be understood that compared with preoperatively diagnosed PTC patients whose standard BRAFV600E mutation abundance does not exceed the preset abundance threshold (i.e., they are determined to be BRAFV600E mutation negative), the prediction results given by the model will be more accurate for preoperatively diagnosed PTC patients whose standard BRAFV600E mutation abundance exceeds the preset abundance threshold (i.e., they are determined to be BRAFV600E mutation positive).

[0096] The preoperative PTC risk prediction method provided in the embodiment of the present application obtains preoperative clinical information and postoperative PTC risk stratification results of preoperative fine needle aspiration FNA specimens according to preset sample screening criteria, determines predictive features from the preoperative clinical information, obtains the original BRAFV600E mutation abundance of each FNA specimen detected by droplet digital polymerase chain reaction (ddPCR) technology, detects the tumor purity of each FNA specimen based on a preset tumor purity detection method, and divides the original BRAFV600E mutation abundance by the tumor purity to obtain the corresponding FNA specimen. Standard BRAFV600E mutation abundance, based on the standard BRAFV600E mutation abundance, determine the training preoperative FNA specimens from each of the sample preoperative FNA specimens, use the predictive features of each training preoperative FNA specimen and the standard BRAFV600E mutation abundance as the input features of the model, use the postoperative PTC risk stratification results of each training preoperative FNA specimen as the label of the corresponding input feature, and use the preoperative PTC low risk or preoperative PTC medium-high risk as the prediction result of the model, and train the preoperative PTC risk prediction model based on the extreme gradient boosting XGBoost algorithm. It can combine the quantitative results of the BRAFV600E mutation, i.e., the standard BRAFV600E mutation abundance, and preoperative clinical information to construct a preoperative PTC risk prediction model for predicting the PTC risk of patients diagnosed with PTC preoperatively, thereby improving the accuracy of preoperative PTC risk prediction and providing strong data support for clinical diagnosis and treatment.

[0097] In order to verify the performance and accuracy of the preoperative PTC risk prediction model trained by the training method provided in the embodiment of the present application, the applicant of the present application also implemented the following verification process:

[0098] After obtaining consent from the enrolled patients and the hospital's research ethics committee, according to pre-set sample screening criteria, postoperative formalin-fixed paraffin-embedded (FFPE) specimens and their postoperative clinical information from 528 PTC patients were included in the retrospective data cohort; preoperative FNA specimens and their preoperative and postoperative clinical information from 167 PTC patients were included in the prospective data cohort; and preoperative FNA specimens and their preoperative and postoperative clinical information from 74 PTC patients from other sources were included in the external validation data cohort.

[0099] Postoperative FFPE specimens were serially sectioned into three 5-µm sections along the maximum diameter of the tumor surface. One section was stained with hematoxylin and eosin (HE) to assess tumor purity (ratio of tumor cells to total cells), and the remaining two sections were used for DNA extraction. Preoperative FNA specimens were sampled three times per nodule using a 23-gauge needle. The blood samples were evenly smeared on a glass slide, fixed with 95% ethanol, and stained with HE. Tumor purity was assessed first, followed by DNA extraction.

[0100] To detect tumor purity, to save time, postoperative FFPE slides and preoperative FNA slides were scanned into digital images using an automated high-resolution digital pathology scanner at a 20x magnification. The tumor region (ROI) on the digital image was visually inspected and manually delineated using Visiopharm digital pathology analysis software, and non-tumor cells were carefully excluded. The number of tumor cells within the ROI (a) and the total number of cells on the image (b) were counted using Visiopharm software. The tumor purity (R=a / b) was calculated, and the average of multiple calculations was taken as the final tumor purity.

[0101] Postoperative FFPE tissue specimens were dehydrogenated with xylene and rehydrated with 100% ethanol. DNA was extracted using the QIAamp DNA FFPE Tissue Kit. DNA was extracted from preoperative FNA biopsies using the QIAamp DNA Mini Kit. DNA was quantified on a NanoDrop 1000 micro-volume spectrophotometer, with a purity of 1.8 ≤ A260 / A280 ≤ 2.0 and a concentration greater than 10.0 μg / μl.

[0102] ddPCR was performed on the QX-200 Droplet Digital PCR System. BRAF V600E mutation detection primers and probes were prepared using the ddPCR BRAF Screening Kit. Two nucleotide difference probes targeting the mutation region were labeled with carboxyfluorescein (FAM) and VIC dye to detect mutant and wild-type BRAF V600E alleles, respectively. A final TaqMan PCR mixture (20 μL) was prepared using 50 ng of template DNA. Thermal cycling conditions for the BRAF V600E detection experiment included incubation at 95°C for 10 minutes, followed by incubation at 94°C for 15 seconds, 58°C for 1 minute, and then 4°C for 5 minutes. Results were analyzed using Quanta Soft software (version 1.3.2.0). Occasionally, single nonspecific droplets were observed in the positive region. According to the manufacturer's instructions, the presence of at least three droplets with FAM signal was defined as a positive signal for the mutation. Droplets with FAM or VIC signal passed through both the FAM and VIC channels and were automatically counted. The FAM channel counts droplets labeled with FAM (FAM+ / VIC- and FAM+ / VIC+), while the VIC channel counts droplets labeled with VIC (FAM- / VIC+ and FAM+ / VIC+). Double-negative droplets (FAM- / VIC-) are not counted. The software calculates the raw abundance of the BRAF V600E mutation based on a Poisson distribution using the following formula: (FAM-positive droplets + FAM / VIC double-positive droplets) / (FAM-positive droplets + VIC-positive droplets + FAM / VIC double-positive droplets). This yields the raw BRAF V600E mutation abundance for each specimen.

[0103] For each specimen, the raw BRAFV600E mutation abundance was divided by its tumor purity to obtain its standardized BRAFV600E mutation abundance. Specimens with a standardized BRAFV600E mutation abundance exceeding 0.1% were defined as BRAFV600E mutation-positive, and all other specimens were defined as BRAFV600E mutation-negative.

[0104] The model was trained using data from BRAF V600E mutation-positive patients, including 442 cases from the retrospective cohort and 140 cases from the prospective cohort. The retrospective and prospective cohorts were divided into training and validation sets in a 7:3 ratio. Predictive features were selected from preoperative clinical data using the LASSO algorithm: gender, age, and ultrasound features. Ultrasound features included ultrasound-measured tumor diameter, ultrasound-demonstrated multifocality, ultrasound-assessed endothelial growth factor (ETE), and ultrasound combined with serum thyroid antibody positivity (thyroid peroxidase antibodies (TPO-Ab) and / or thyroglobulin antibodies (Tg-Ab)) indicating HT. Similarly, gender, age, tumor diameter, multifocality, ETE, and HT were selected from postoperative clinical data as predictive features.

[0105] Training was performed using the XGBoost algorithm. Model A was constructed and validated based on sex, age, tumor diameter, multifocality, ETE, and HT in the retrospective data cohort. Model B was constructed and validated based on sex, age, tumor diameter, multifocality, ETE, and HT in the prospective data cohort. Model C was constructed and validated based on sex, age, ultrasound-measured tumor diameter, ultrasound-visual multifocality, ultrasound-assessed ETE, and ultrasound-combined thyroid antibody positivity (thyroid peroxidase antibodies (TPO-Ab) and / or thyroglobulin antibodies (Tg-Ab)) indicating HT. Model D was constructed and validated based on sex, age, ultrasound-measured tumor diameter, ultrasound-visual multifocality, ultrasound-assessed ETE, ultrasound-combined thyroid antibody positivity (thyroid peroxidase antibodies (TPO-Ab) and / or thyroglobulin antibodies (Tg-Ab)) indicating HT, and standard BRAFV600E mutation abundance. The area under the curves (AUCs) for models A, B, C, and D on the training and validation sets were obtained during training.

[0106] Using the external validation data queue, models B, C, and D were externally validated, and the AUCs of models B, C, and D during the external validation process were obtained.

[0107] Model A achieved an AUC of 0.93 (95% CI, 0.91-0.96) in the training set and 0.89 (95% CI, 0.83-0.94) in the validation set. Model B achieved an AUC of 0.91 (95% CI, 0.85-0.97) in the training set and 0.89 (95% CI, 0.83-0.99) in the validation set. Model C achieved an AUC of only 0.67 (95% CI, 0.56-0.78) in the training set and 0.65 (95% CI, 0.48-0.82) in the validation set. Model D achieved an AUC of 0.86 (95% CI, 0.79-0.94) in the training set and 0.86 (95% CI, 0.75-0.98) in the validation set. Furthermore, P values ​​for these models were all greater than 0.05, indicating good model calibration. During external validation, the AUCs of models B, C, and D were 0.78 (95% CI, 0.67-0.88), 0.61 (95% CI, 0.48-0.75), and 0.82 (95% CI, 0.71-0.93), respectively.

[0108] As can be seen, there was no significant difference in the AUC values ​​of models A, B, and D, indicating that model D had the same efficacy as models A and B in predicting preoperative risk stratification for PTC. Compared with model C, which was based solely on preoperative clinical information, the AUC values ​​of model D increased by 20% (P < 0.01), 22% (P = 0.04), and 21% (P = 0.02) in the training set, validation set, and external validation data cohort, respectively, confirming that the incorporation of preoperative clinical information and the standard BRAFV600E mutation abundance can significantly improve predictive efficacy.

[0109] Finally, to interpret each model, the SHAP value for each input feature was calculated for the corresponding model. For Model A, the top three features were ETE (SHAP value of 1.849), tumor diameter (SHAP value of 0.979), and age (SHAP value of 0.275). For Model B, the top three features were ETE (SHAP value of 0.851), tumor diameter (SHAP value of 0.448), and multifocality (SHAP value of 0.216). In Model C, ultrasound-measured tumor diameter (SHAP value of 0.333), sex (SHAP value of 0.333), and age (SHAP value of 0.256) were the top three features. For Model D, the top three features were standard BRAFV600E mutation abundance (FNA) (SHAP value of 1.321), sex (SHAP value of 0.280), and ultrasound-measured tumor diameter (SHAP value of 0.213). This once again confirmed that the standard BRAFV600E mutation abundance can be used as an important indicator for preoperative PTC risk prediction, and combined with preoperative clinical information, it can effectively and accurately predict preoperative PTC risk.

[0110] As a third aspect of the embodiments of the present application, an electronic device is provided, wherein, Figure 5 As shown, the electronic device includes:

[0111] One or more processors 101;

[0112] The memory 102 stores one or more computer programs. When the one or more computer programs are executed by the one or more processors 101, the one or more processors 101 implement any of the following:

[0113] The training method of the preoperative PTC risk prediction model provided in the first aspect of the embodiment of the present application;

[0114] The second aspect of the embodiments of the present application provides a preoperative PTC risk prediction method.

[0115] The electronic device may further include one or more I / O interfaces 103 connected between the processor 101 and the memory 102 and configured to implement information exchange between the processor 101 and the memory 102 .

[0116] Among them, the processor 101 is a device with data processing capabilities, including but not limited to a central processing unit (CPU); the memory 102 is a device with data storage capabilities, including but not limited to random access memory (RAM, more specifically such as SDRAM, DDR, etc.), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), and flash memory (FLASH); the I / O interface (read-write interface) is connected between the processor and the memory, and can realize information exchange between the processor and the memory, including but not limited to a data bus (Bus), etc.

[0117] In some embodiments, the processor 101 , the memory 102 , and the I / O interface 103 are connected to each other via a bus 104 , and further connected to other components of the computing device.

[0118] As a fourth aspect of the embodiment of the present application, Figure 6 As shown, a computer readable medium is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, any of the following is implemented:

[0119] The training method of the preoperative PTC risk prediction model provided in the first aspect of the embodiment of the present application;

[0120] The second aspect of the embodiments of the present application provides a preoperative PTC risk prediction method.

[0121] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program. Accordingly, the computer program can be stored in a non-volatile computer-readable storage medium, and when the computer program is executed, it can implement the method of any of the above-mentioned embodiments. Among them, any reference to memory, storage, database or other media used in the embodiments provided in the embodiments of the present application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0122] The above are only specific embodiments of the present application, but the scope of protection of the present application is not limited thereto. Those skilled in the art should understand that the present application includes but is not limited to the contents described in the drawings and the above specific embodiments. Any modifications that do not deviate from the functional and structural principles of the present application are included within the scope of the claims.

Claims

1. A training method for a preoperative papillary thyroid carcinoma (PTC) risk prediction model, characterized in that: The method comprises: According to the preset sample screening criteria, preoperative clinical information of fine needle aspiration (FNA) specimens and postoperative PTC risk stratification results were obtained; Determining predictive features from the preoperative clinical information; wherein the predictive features include gender, age, and ultrasound features, and the ultrasound features include ultrasound measurement of tumor diameter, ultrasound display of multifocality, ultrasound assessment of extramembranous invasion, and ultrasound combined with serum indicators indicating Hashimoto's thyroiditis; determining predictive features from the preoperative clinical information includes: using the postoperative PTC risk stratification results of the preoperative FNA specimens of each of the samples as a prediction target, using the preoperative clinical information of the preoperative FNA specimens of each of the samples as a candidate feature, and selecting the predictive features from the candidate features based on the LASSO algorithm; wherein the postoperative PTC risk stratification results include low risk of postoperative PTC and medium-high risk of postoperative PTC; Obtaining the original BRAFV600E mutation abundance of each FNA specimen detected by droplet digital polymerase chain reaction (ddPCR) technology, detecting the tumor purity of each FNA specimen based on a preset tumor purity detection method, and dividing the original BRAFV600E mutation abundance by the tumor purity to obtain a standard BRAFV600E mutation abundance of the corresponding FNA specimen; According to the standard BRAFV600E mutation abundance, a training preoperative FNA specimen is determined from the preoperative FNA specimens of each sample; The predictive features of each training preoperative FNA specimen and the standard BRAFV600E mutation abundance are used as the input features of the model, the postoperative PTC risk stratification results of each training preoperative FNA specimen are used as the labels of the corresponding input features, and the preoperative PTC low risk or preoperative PTC medium-high risk are used as the prediction results of the model. A preoperative PTC risk prediction model is obtained based on the extreme gradient boosting XGBoost algorithm.

2. The method according to claim 1, characterized in that Determining a training preoperative FNA specimen from each of the preoperative FNA specimens according to the standard BRAFV600E mutation abundance comprises: Preoperative FNA specimens with standard BRAFV600E mutation abundance exceeding the preset abundance threshold are identified as BRAFV600E mutation-positive specimens, and preoperative FNA specimens with standard BRAFV600E mutation abundance not exceeding the preset abundance threshold are identified as BRAFV600E mutation-negative specimens; From the preoperative FNA specimens of each sample, the specimens positive for BRAFV600E mutation were determined as training preoperative FNA specimens.

3. The method according to claim 1, characterized in that The preset sample screening criteria include inclusion criteria and exclusion criteria; the inclusion criteria include: the age of the patient corresponding to the sample is within the preset age range, the patient corresponding to the sample is confirmed to have typical PTC by postoperative pathology, the patient corresponding to the sample has undergone unilateral thyroid lobectomy or total thyroidectomy and central lymph node dissection, the patient corresponding to the sample has undergone therapeutic lateral neck lymph node dissection in the case of lateral cervical lymph node metastasis, and the preoperative clinical information and postoperative clinical information of the patient corresponding to the sample are complete; the exclusion criteria include: the patient corresponding to the sample has a history of neck trauma, the patient corresponding to the sample has other tumors, the patient corresponding to the sample is diagnosed as a PTC subtype or non-PTC by postoperative pathology, and the patient corresponding to the sample has incomplete preoperative clinical information and postoperative clinical information.

4. The method according to any one of claims 1 to 3, characterized in that The step of using the predictive features of each training preoperative FNA specimen and the standard BRAFV600E mutation abundance as input features of the model, using the postoperative PTC risk stratification results of each training preoperative FNA specimen as labels of the corresponding input features, and using preoperative PTC low risk or preoperative PTC medium-high risk as the prediction results of the model, and training the preoperative PTC risk prediction model based on the extreme gradient boosting (XGBoost) algorithm, includes: The predictive features of each training preoperative FNA specimen and the standard BRAFV600E mutation abundance are divided into a training set and a validation set according to a preset ratio. The training set is used for the XGBoost algorithm to construct a model, and the validation set is used for the XGBoost algorithm to verify the constructed model.

5. The method according to claim 4, characterized in that The XGBoost algorithm optimizes the hyperparameters of the model based on a five-fold cross-validation algorithm and a Bayesian algorithm until the area under the receiver operating characteristic curve (AUC) of the model on the validation set reaches a maximum; The XGBoost algorithm optimizes the parameters of the model based on an objective function, which includes a training loss and a regularization term between the label of the input feature and the prediction result corresponding to the input feature.

6. A method for predicting the risk of PTC before surgery, characterized in that: The method comprises: Obtaining predictive features of preoperative FNA specimens for preoperative diagnosis of PTC patients; wherein the predictive features include gender, age, and ultrasound features, including ultrasound measurement of tumor diameter, ultrasound display of multifocality, ultrasound assessment of extramembranous invasion, and ultrasound combined with serum indicators indicating Hashimoto's thyroiditis; Obtaining the original BRAFV600E mutation abundance of the predicted preoperative FNA specimen detected based on ddPCR technology, detecting the tumor purity of the predicted preoperative FNA specimen based on a preset tumor purity detection method, and dividing the original BRAFV600E mutation abundance by the tumor purity to obtain a standard BRAFV600E mutation abundance of the predicted preoperative FNA specimen; The predictive features of the preoperative FNA specimens and the standard BRAFV600E mutation abundance are used as input features of the model and input into the preoperative PTC risk prediction model trained by the training method of the preoperative PTC risk prediction model according to any one of claims 1-5 to obtain the prediction results output by the preoperative PTC risk prediction model.

7. An electronic device, characterized in that: The electronic device comprises: one or more processors; A memory having one or more computer programs stored thereon, which, when executed by the one or more processors, causes the one or more processors to implement any of the following: The training method of the preoperative PTC risk prediction model according to any one of claims 1 to 5; The preoperative PTC risk prediction method according to claim 6.

8. A computer-readable medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements any of the following: The training method of the preoperative PTC risk prediction model according to any one of claims 1 to 5; The preoperative PTC risk prediction method according to claim 6.

Images