Nodule special screening data processing method based on artificial intelligence and blood examination

Abstract

The invention relates to a nodule special screening data processing method based on artificial intelligence and blood examination. The nodule special screening data processing method based on artificial intelligence and blood examination comprises the following steps: acquiring an initial data set, wherein the initial data set comprises physical examination information of a high-risk target crowd and physical examination information of a healthy contrast crowd; performing data processing on the initial data set, and determining incorporation factors; according to the incorporated factors, constructing a first basic model for prompting that nodule special screening needs to be carried out; constructing a second basic model used for prompting whether malignant nodules possibly occur in a long term or not; performing parameter adjustment on the first basic model parameter and the second basic model parameter by adopting the verification set until the first basic model parameter and the second basic model parameter converge; and acquiring to-be-processed data, and respectively inputting the to-be-processed data into the first basic model and the second basic model to obtain a prompt result. According to the method and the device, the problem of inaccurate nodule prompt of indexes under the condition of lack of physical examination or lifestyle survey is solved.

Description

Technical Field

[0001] The present application relates to the technical field of data processing, and in particular to a method for processing data for special screening of nodules based on artificial intelligence and blood tests. Background Art

[0002] Malignant tumors, as a major global public health issue, have seriously threatened human health. In 2020, the International Agency for Research on Cancer (IARC) estimated that there would be approximately 19.3 million new cancer cases and nearly 10 million cancer deaths worldwide based on the GLOBOCAN project's estimates of cancer incidence and mortality.

[0003] In recent years, with the development of social economy, the continuous improvement of people's living standards and the change of lifestyle, the incidence and mortality of malignant tumors have also shown an increasing trend year by year, causing a serious disease burden. It is worth noting that the occurrence and development of most malignant tumors usually take several years. Nodules, as possible precursors of malignant tumors, are receiving more and more attention. Common nodules include lung nodules, thyroid nodules and breast nodules. In recent years, many technologies have been developed to determine whether nodules have the risk of deterioration, such as low-dose spiral CT of the lungs. In addition, many studies have also constructed predictive models for the risk of tumor occurrence.

[0004] However, the prediction models for tumor risk are all based on patients who are confirmed to have cancer as the target population, and each model selects different risk factors, so there are two problems. First, the identification of benign and malignant nodules requires the discovery of nodules first, and the detection methods are mainly imaging examinations or ultrasound examinations. According to the requirements of the health examination guidelines, the intervals of this special examination are not fixed for different risk groups. Second, the effectiveness of the model depends on the selected factors. When the examinee cannot provide the data required by the model, the model cannot make accurate predictions. Summary of the invention

[0005] In order to help remind physical examinees who have not undergone special nodule examinations whether they need to undergo relevant additional examinations, this application provides a special nodule screening data processing method based on artificial intelligence and blood tests.

[0006] The present invention provides a method for processing data for special screening of nodules based on artificial intelligence and blood examination, comprising:

[0007] S1: Get the initial data set;

[0008] S2: Data processing of the initial data set to determine the included factors;

[0009] S3: Based on the included factors, the first basic model is constructed to indicate the need for special screening for nodules;

[0010] S4: Construct a second basic model to indicate whether malignant nodules may occur in the long term;

[0011] S5: Obtain a validation set, use the validation set to respectively validate the first basic model and the second basic model and obtain validation results, and adjust parameters of the first basic model and the second basic model based on the validation results until the parameters of the first basic model and the parameters of the second basic model converge;

[0012] S6. Obtain the data to be processed, input them into the first basic model and the second basic model respectively, and obtain the prompt results of the nodule-specific screening and the possible prompt results of the long-term malignant nodule occurrence.

[0013] Furthermore, the initial data set includes physical examination information of a high-risk target population and physical examination information of a healthy control population; the physical examination information includes gender, age, blood test indicators and imaging examination data.

[0014] Further, the S2 includes:

[0015] S21: Based on the imaging examination data, the initial data set was divided into three nodule groups: a lung nodule group with a possible risk of malignancy, a thyroid nodule group with a possible risk of malignancy, and a breast nodule group with a possible risk of malignancy;

[0016] S22: Based on the physical examination information of each nodule group, the propensity score calculated based on logistic regression was used to perform overlapping weighting on each sample in the initial data set, and the generalized estimating equation was used to screen the factors of each sample after overlapping weighting to obtain the included factors.

[0017] Further, the S22 includes:

[0018] S221: Define the confounding factor effect estimation regression model used to divide the physical examination information of each nodule group into a high-risk group and a healthy control group:

[0019] A~+x′1,…,+x′ k

[0020] Where A represents the processing factor of the physical examination information. When A=1, the physical examination information is classified as a high-risk group, and when A=0, the physical examination information is classified as a healthy control group; x′ k Represents the kth confounding factor of the physical examination information, including age and gender; k represents the confounding factor number;

[0021] S222: Calculate the propensity score of the physical examination information using the processing factors of the physical examination information:

[0022] E(X i )=P(A i =1|X i )

[0023] Among them, E(X i ) represents the propensity score of the i-th physical examination information; P(A i =1|X i ) represents the probability that the i-th physical examination information is assigned to the high-risk group under the conditions of statistical age and gender; A i represents the processing factor of the i-th physical examination information; X i represents the confounding factors of the ith physical examination information;

[0024] S223: Define a logistic regression model for estimating the propensity score of physical examination information:

[0025]

[0026] Among them, P(A = 1|X) represents the probability of the physical examination information being assigned to the high-risk group under the condition of statistical age and gender; X represents the set of confounding factors of the physical examination information; β0 represents the constant term of the logistic regression model; β1 and β k All represent the regression coefficients of the Logistic regression model;

[0027] S224: Based on the Logistic regression model for estimating the propensity score of the physical examination information, the propensity prediction score of each physical examination information is calculated:

[0028]

[0029] Where E′(X) represents the propensity prediction score of physical examination information; P(A=1|X i ) represents the probability that the i-th physical examination information is assigned to the high-risk group under the conditions of statistical age and gender;

[0030] S225: According to the tendency prediction score of the physical examination information, weights are assigned to the physical examination information to obtain the weights of each physical examination information:

[0031]

[0032] h(X i )=E′(X i )(1-E′(X i ))

[0033] Among them, W i represents the weight of the i-th physical examination information; h(X i ) is E′(X i) of the binomial function; E′(X i ) represents the propensity prediction score of the i-th physical examination information;

[0034] S226: According to the weight of the physical examination information, a weighted generalized linear model is defined. The weighted generalized linear model includes a random part, a systematic part, and a link function part:

[0035]

[0036] a(φ)=φ / W i

[0037]

[0038] g(μ)=ln(μ)

[0039] Among them, f(·) represents the random part, which indicates the probability distribution of the dependent variable; y g represents the value of the dependent variable corresponding to the g-th observation; g represents the observation value number of the examination index in the physical examination information; θ g represents a natural parameter related to the mean of the observations; φ represents a discrete parameter related to the variance of the observations; a(φ) represents a weight function related to the variance of the observations; b(θ g ) represents the weight function related to the mean of the observations; c(y g , φ) represents the normalization constant used to ensure that the integral of the entire probability distribution function is 1; η represents the system part, which is a linear value; g(μ) represents the link function part; μ represents the link function value; m represents the number of inspection indicators; j is the number of inspection indicators; Y j represents the regression coefficient corresponding to the jth inspection index; x j represents the jth inspection index;

[0040] S227: Use the logarithmic maximum likelihood estimation method to estimate the parameters of the weighted generalized linear model and establish the logarithmic likelihood function L:

[0041]

[0042] Among them, L g represents the log-likelihood function of the g-th observation value; n represents the total number of observation values ​​of the examination indicators in the physical examination information;

[0043] S228: Use the Newton-Raphson iterative method to maximize the log-likelihood function L and obtain γ j The maximum likelihood estimate b j ;

[0044] S229: Maximum likelihood estimate b j A likelihood ratio test was performed to obtain the included factors.

[0045] Further, the S229 includes:

[0046] S2291: The maximum likelihood estimate b j Substitute into the weighted generalized linear model, and fit the weighted generalized linear model without the inspection index, and find the log-likelihood function lnL0 corresponding to the weighted generalized linear model without the inspection index;

[0047] S2292: adding the inspection indicators to the generalized linear model for fitting in turn to obtain a new log-likelihood function value lnL1;

[0048] S2293: Calculate the likelihood ratio statistic G based on lnL0 and lnL1:

[0049] G=2(lnL1-lnL0)

[0050] S2294: Determine whether the likelihood ratio statistic G is greater than or equal to the preset statistical test standard value. If so, set the inspection indicator corresponding to lnL1 as the inclusion factor, and return to S2292 to screen the factors of the next inspection indicator. Otherwise, directly return to S2292 to screen the factors of the next inspection indicator.

[0051] Further, the S3 includes:

[0052] S31: According to the inclusion factors, irrelevant data are removed from the initial data set to obtain a processed initial data set, and each nodule group in the processed initial data set is randomly divided into corresponding training sets and test sets;

[0053] S32: using the training set corresponding to the group of pulmonary nodules that may have a risk of malignancy, respectively training a random forest classifier, an extremely random tree classifier, and an adaptive boosting classifier based on a decision tree, to obtain a trained random forest classifier, a trained extremely random tree classifier, and a trained adaptive boosting classifier based on a decision tree;

[0054] S33: import the test set corresponding to the group of pulmonary nodules with possible malignant risks into the trained random forest classifier, the trained extreme random tree classifier and the trained decision tree-based adaptive boosting classifier respectively, obtain the AUC value of the trained random forest classifier, the AUC value of the trained extreme random tree classifier and the AUC value of the trained decision tree-based adaptive boosting classifier, and select the classifier with the largest AUC value as the pulmonary nodule special screening prompt model;

[0055] S34, determining whether the pulmonary nodule special screening prompt model meets the accuracy requirement according to the AUC value of the pulmonary nodule special screening prompt model, if so, directly proceed to S35, otherwise, return to S32;

[0056] S35: Based on the training set and test set corresponding to the group of thyroid nodules with possible malignant risk, a special screening prompt model for thyroid nodules was constructed using the methods proposed in S32-S34;

[0057] S36: Based on the training set and test set corresponding to the group of breast nodules with possible malignant risks, a breast nodule special screening prompt model was constructed using the methods proposed in S32-S34;

[0058] S37: Integrate the lung nodule special screening prompt model, the thyroid nodule special screening prompt model and the breast nodule special screening prompt model to obtain the first basic model for prompting the need for nodule special screening.

[0059] Furthermore, the step S32 includes:

[0060] S321: construct a basic decision tree model using the training set D corresponding to the group of pulmonary nodules that may have a risk of malignancy; the z-th branch node of the basic decision tree model includes all the examination indicators θ in the training set D with a value of θ z Samples of

[0061] S322: The training set D in the basic decision tree model is sampled using the bagging method to obtain a trained random forest classifier:

[0062] RF={h(x,θ k′ ), k′=1,2,…,t}

[0063] Among them, RF represents the trained random forest classifier; h represents the sampling set of the bagging method; θ k′ represents the k′th random vector that follows an independent distribution; t represents the number of decision trees generated, and x represents the samples in the training set;

[0064] S323: using the training set D to train the extremely randomized tree classifier to obtain a trained extremely randomized tree classifier;

[0065] S324: Using the training set D to train the decision tree-based adaptive boosting classifier, to obtain a trained decision tree-based adaptive boosting classifier.

[0066] Furthermore, the step S324 includes:

[0067] S3241: Calculate the weight W of the samples in the training set D n :

[0068]

[0069] Among them, w N Represents the weight of the Nth sample; N represents the total number of samples in the training set D;

[0070] S3242: Select a weak classifier as the tth base classifier h t , and according to the sample weight W n , calculate the base classifier h t The fitting error rate is:

[0071]

[0072] Among them, e t represents the fitting error rate of the tth base classifier; n′ represents the sample number in the training set D; w tn′ represents the weight of the n′th sample in the tth base classifier; I(·) is used to judge h t (x n′ ) and y n′ Whether it is unequal or not; h t (x n ′) represents the sample x n′ The t-th base classifier of x n′ represents the n′th sample in the training set D; y n′ represents the observed value of the n′th sample, that is, whether a nodule occurs;

[0073] S3243: According to the base classifier h t The fitting error rate of the base classifier h t The parameter α required to update the weights in the final classifier t :

[0074]

[0075] Among them, ∈ t Represents the classifier h t (x n ′)’s error weight rate;

[0076] S3244: According to parameter α t , update the sample weight W t+1 :

[0077] W t+1 =W t exp{α t I(h t (x n′ )≠y n′ )}

[0078] Among them, W t+1Represents the updated sample weight; W t Represents the sample weight before updating;

[0079] S3245: The weak classifier h t With parameter α t Combined into a strong classifier H t :

[0080] H t =H t-1 +α t h t

[0081] Among them, H t represents the strong classifier of the tth iteration; H t-1 represents the strong classifier of the t-1th iteration;

[0082] S3246: Use symbolic functions to compare the effects of weak classifiers and strong classifiers:

[0083] C=sign(∑α t h t )

[0084] Among them, sign(·) represents the sign function, if ∑α t h t If it is less than 0, -1 is returned, indicating that the strong classifier H formed after this iteration t The effect is not as good as the weak classifier h t , if ∑α t h t If it is greater than 0, it returns 1, indicating that the strong classifier H formed after this iteration t The effect is better than the weak classifier h t ;

[0085] S3247: Repeat steps S3241-S3246 for multiple iterations, and select the best strong classifier as the trained decision tree-based adaptive boosting classifier H.

[0086] Furthermore, the step S4 comprises:

[0087] S41: Based on the initial data set, a target population with multiple physical examination information and whose first physical examination information clearly showed no malignancy was selected and defined as the change population;

[0088] S42: The changed population was divided into lung nodule group, thyroid nodule group and breast nodule group;

[0089] S43: The physical examination information of the changed population in the lung nodule group, thyroid nodule group, and breast nodule group was randomly divided into corresponding training sets and test sets;

[0090] S44: The training set corresponding to the population with changes in the pulmonary nodule group was imported into the long short-term memory network model, and a model for the possible occurrence of long-term malignant pulmonary nodules was constructed based on the included factors;

[0091] S45: import the test set corresponding to the changed population of the pulmonary nodule group into the model for possible indication of the occurrence of long-term malignant pulmonary nodules and output the true accuracy rate and false accuracy rate of the model for possible indication of the occurrence of long-term malignant pulmonary nodules, draw the ROC curve according to the true accuracy rate and false accuracy rate of the model for possible indication of the occurrence of long-term malignant pulmonary nodules, calculate the AUC value according to the ROC curve of the model for possible indication of the occurrence of long-term malignant pulmonary nodules, and determine whether the model for possible indication of the occurrence of long-term malignant pulmonary nodules meets the accuracy requirement according to the AUC value of the model for possible indication of the occurrence of long-term malignant pulmonary nodules, if so, directly enter S46, otherwise, return to S44;

[0092] S46: Based on the training set and test set corresponding to the population with changes in the thyroid nodule group, a model for the possible indication of the occurrence of long-term malignant thyroid nodules was constructed using the methods proposed in S44-S45;

[0093] S47: Based on the training set and test set corresponding to the population with changes in the breast nodule group, use the methods proposed in S44-S45 to construct a possible warning model for the occurrence of long-term malignant breast nodules;

[0094] S48: Integrate the possible indication model of the long-term malignant pulmonary nodule occurrence, the possible indication model of the long-term malignant thyroid nodule occurrence and the possible indication model of the long-term malignant breast nodule occurrence to obtain a second basic model for indicating whether malignant nodules may occur in the long term.

[0095] Furthermore, the step S44 is specifically as follows: irrelevant data is eliminated from the training set corresponding to the changed population of the lung nodule group based on the included factors to obtain a processed training set corresponding to the changed population of the lung nodule group, and a long short-term memory network model is trained based on the processed second training set corresponding to the changed population of the lung nodule group to obtain a possible prompt model for the occurrence of long-term malignant lung nodules.

[0096] In summary, the present application includes at least the following beneficial technical effects: the present method does not need to collect indicators that are easily missing, such as physical examination and lifestyle, and has a wider applicability. The method can also be used to recommend special examination items for nodules, to achieve early identification of benign and malignant nodules, to screen out more suitable potential risk groups, and then optimize the physical examination plan, save medical resources, and to achieve the purpose of prompting physical examination personnel who have not undergone special examinations for nodules whether they need to make up for the relevant examinations. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 This is a flowchart of an embodiment of the present application.

[0098] Figure 2This is a schematic diagram of the standardized mean difference of confounding factors before and after overlap weighting of the first lung nodule group in an embodiment of the present application.

[0099] Figure 3 This is a schematic diagram of the standardized mean difference of confounding factors before and after overlapping weighting of the first thyroid nodule group in the embodiment of the present application.

[0100] Figure 4 This is a schematic diagram of the standardized mean difference of confounding factors before and after overlapping weighting of the first breast nodule group in an embodiment of the present application. DETAILED DESCRIPTION

[0101] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0102] The terms used in the following embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to be used as limitations to the present application. As used in the specification and the appended claims of the present application, the singular expressions "one", "a kind of", "said", "above", "the" and "this" are intended to also include plural expressions, unless there is a clear contrary indication in its context. It should also be understood that the term "and / or" used in the present application refers to and includes any or all possible combinations of one or more listed items. The term "exemplary" means "used as an example, embodiment or illustrative", and any embodiment described as "exemplary" here does not need to be interpreted as being superior to or better than other embodiments. The terms "first" and "second" are only used for descriptive purposes, and cannot be understood as implying or implying relative importance or implicitly indicating the number of indicated technical features. Thus, the features defined as "first" and "second" can expressly or implicitly include one or more of these features, and in the description of the embodiments of the present application, unless otherwise specified, the meaning of "multiple" is two or more.

[0103] The embodiment of the present application provides a method for processing data for special screening of nodules based on artificial intelligence and blood examination.

[0104] refer to Figure 1 , a method for processing data for nodule-specific screening based on artificial intelligence and blood tests, comprising the following steps:

[0105] S1: Get the initial data set;

[0106] The initial data set includes physical examination information of high-risk target population and physical examination information of healthy control population; the physical examination information includes gender, age, blood test index and imaging examination data;

[0107] The selection of research subjects requires the determination of inclusion and exclusion criteria, which are the inclusion and exclusion criteria. The inclusion criteria are that the examinee undergoes routine physical examinations for renal function tests, liver function tests, blood lipid tests, blood routine tests, imaging tests (lung CT), and ultrasound tests (thyroid ultrasound or breast ultrasound); the exclusion criteria are that the overall missing rate of the examinee's physical examination indicators (inclusion criteria) reaches 30% or more.

[0108] S2: Data processing of the initial data set to determine the included factors;

[0109] The S2 includes:

[0110] S2 1: Based on the imaging examination data, the initial data set was divided into three nodule groups: a lung nodule group with possible malignancy risk, a thyroid nodule group with possible malignancy risk, and a breast nodule group with possible malignancy risk;

[0111] S22: Based on the physical examination information of each nodule group, the propensity score calculated based on logistic regression was used to perform overlapping weighting on each sample in the initial data set, and the generalized estimating equation was used to screen the factors of each sample after overlapping weighting to obtain the included factors.

[0112] refer to Figure 2 , Figure 3 and Figure 4 , the S22 comprises:

[0113] S221: Define the confounding factor effect estimation regression model used to divide the physical examination information of each nodule group into a high-risk group and a healthy control group:

[0114] A~+x′1,…,+x′ k

[0115] Where A represents the processing factor of the physical examination information. When A=1, the physical examination information is classified as a high-risk group, and when A=0, the physical examination information is classified as a healthy control group; x′ k Represents the kth confounding factor of the physical examination information, including age and gender; k represents the confounding factor number;

[0116] S222: Calculate the propensity score of the physical examination information using the processing factors of the physical examination information:

[0117] E(X i )=P(A i =1|X i )

[0118] Among them, E(X i ) represents the propensity score of the i-th physical examination information; P(A i =1|X i) represents the probability that the i-th physical examination information is assigned to the high-risk group under the conditions of statistical age and gender; A i represents the processing factor of the i-th physical examination information; X i represents the confounding factors of the ith physical examination information;

[0119] S223: Define a logistic regression model for estimating the propensity score of physical examination information:

[0120]

[0121] Among them, P(A = 1|X) represents the probability of the physical examination information being assigned to the high-risk group under the condition of statistical age and gender; X represents the set of confounding factors of the physical examination information; β0 represents the constant term of the logistic regression model; β1 and β k All represent the regression coefficients of the Logistic regression model;

[0122] S224: Based on the Logistic regression model for estimating the propensity score of the physical examination information, the propensity prediction score of each physical examination information is calculated:

[0123]

[0124] Where E′(X) represents the propensity prediction score of physical examination information; P(A=1|X i ) represents the probability that the i-th physical examination information is assigned to the high-risk group under the conditions of statistical age and gender;

[0125] S225: According to the tendency prediction score of the physical examination information, weights are assigned to the physical examination information to obtain the weights of each physical examination information:

[0126]

[0127] h(X i )=E′(X i )(1-E′(X i ))

[0128] Where Wi represents the weight of the i-th physical examination information; h(X i ) is E′(X i ) of the binomial function; E′(X i ) represents the propensity prediction score of the i-th physical examination information;

[0129] S226: According to the weight of the physical examination information, a weighted generalized linear model is defined. The weighted generalized linear model includes a random part, a systematic part, and a link function part:

[0130]

[0131] a(φ)=φ / Wi

[0132]

[0133] g(μ)=ln(μ)

[0134] Among them, f(·) represents the random part, which indicates the probability distribution of the dependent variable; y g represents the value of the dependent variable corresponding to the g-th observation; g represents the observation value number of the examination index in the physical examination information; θ g represents a natural parameter related to the mean of the observations; φ represents a discrete parameter related to the variance of the observations; a(φ) represents a weight function related to the variance of the observations; b(θ g ) represents the weight function related to the mean of the observations; c(y g , φ) represents the normalization constant used to ensure that the integral of the entire probability distribution function is 1; η represents the system part, which is a linear value; g(μ) represents the link function part; μ represents the link function value; m represents the number of inspection indicators; j is the number of inspection indicators; Y j represents the regression coefficient corresponding to the jth inspection index; x j represents the jth inspection index;

[0135] S227: Use the logarithmic maximum likelihood estimation method to estimate the parameters of the weighted generalized linear model and establish the logarithmic likelihood function L:

[0136]

[0137] Among them, L g represents the log-likelihood function of the g-th observation value; n represents the total number of observation values ​​of the examination indicators in the physical examination information;

[0138] S228: Use the Newton-Raphson iterative method to maximize the log-likelihood function L and obtain γ j The maximum likelihood estimate b j ;

[0139] S229: Maximum likelihood estimate b j A likelihood ratio test was performed to obtain the included factors.

[0140] The S229 includes:

[0141] S2291: The maximum likelihood estimate b j Substitute into the weighted generalized linear model, and fit the weighted generalized linear model without the inspection index, and find the log-likelihood function lnL0 corresponding to the weighted generalized linear model without the inspection index;

[0142] S2292: adding the inspection indicators to the generalized linear model for fitting in turn to obtain a new log-likelihood function value lnL1;

[0143] S2293: Calculate the likelihood ratio statistic G based on lnL0 and lnL1:

[0144] G=2(lnL1-lnL0)

[0145] S2294: Determine whether the likelihood ratio statistic G is greater than or equal to the preset statistical test standard value. If so, set the inspection indicator corresponding to lnL1 as the inclusion factor, and return to S2292 to screen the factors of the next inspection indicator. Otherwise, directly return to S2292 to screen the factors of the next inspection indicator.

[0146] Table 1 is a schematic table of inspection indicators.

[0147] Table 1 Inspection index Chinese and English comparison table

[0148]

[0149]

[0150] Among them, the following Table 2 is the result of single factor analysis of the group of lung nodules that may have malignant risks. Table 3 is the result of single factor analysis of the group of thyroid nodules that may have malignant risks. Table 4 is the result of single factor analysis of the group of breast nodules that may have malignant risks.

[0151] Table 2 Results of univariate analysis of the group of pulmonary nodules with possible malignant risk

[0152]

[0153]

[0154] Table 3 Results of univariate analysis of thyroid nodules with possible malignant risk

[0155]

[0156]

[0157] Table 4 Results of univariate analysis of breast nodules with potential risk of malignancy

[0158]

[0159]

[0160]

[0161] According to the results in Table 2-4, the screening factors for each nodule group can be obtained:

[0162] The factors screened in the pulmonary nodule group were: total cholesterol, high-density lipoprotein cholesterol, hemoglobin, mean corpuscular hemoglobin, mean corpuscular hemoglobin concentration, platelet distribution width, mean platelet volume, neutrophil percentage, lymphocyte percentage, red blood cell distribution width, large platelet ratio, systemic immune inflammatory index, systemic inflammatory response index, neutrophil-to-lymphocyte ratio, derived neutrophil-to-lymphocyte ratio, platelet-to-lymphocyte ratio, monocyte-to-lymphocyte ratio, and neutrophil percentage-to-albumin ratio;

[0163] The factors screened in the thyroid nodule group were: urea, uric acid, alanine aminotransferase, triglycerides, high-density lipoprotein cholesterol, blood glucose, total white blood cell count, red blood cell count, hemoglobin, hematocrit, mean corpuscular volume, mean corpuscular hemoglobin, platelets, hematocrit, monocyte percentage, eosinophil percentage, absolute neutrophil value, absolute lymphocyte value, absolute monocyte value, absolute eosinophil value, absolute basophil value, red blood cell distribution width, triglyceride and glucose index, systemic immune inflammation index, systemic inflammatory response index, monocyte to high-density lipoprotein cholesterol ratio, and systemic inflammation comprehensive index;

[0164] The factors screened in the breast nodule group were: uric acid, albumin, total white blood cell count, red blood cell count, platelets, hematocrit, absolute neutrophil count, systemic immune inflammatory index, platelet-to-lymphocyte ratio, and neutrophil percentage-to-albumin ratio;

[0165] S3: Based on the included factors, the first basic model is constructed to indicate the need for special screening for nodules;

[0166] The S3 includes:

[0167] S31: According to the inclusion factors, irrelevant data are removed from the initial data set to obtain a processed initial data set, and each nodule group in the processed initial data set is randomly divided into corresponding training sets and test sets;

[0168] S32: using the training set corresponding to the group of pulmonary nodules that may have a risk of malignancy, respectively training a random forest classifier, an extremely random tree classifier, and an adaptive boosting classifier based on a decision tree, to obtain a trained random forest classifier, a trained extremely random tree classifier, and a trained adaptive boosting classifier based on a decision tree;

[0169] S33: import the test set corresponding to the group of pulmonary nodules with possible malignant risks into the trained random forest classifier, the trained extreme random tree classifier and the trained decision tree-based adaptive boosting classifier respectively, obtain the AUC value of the trained random forest classifier, the AUC value of the trained extreme random tree classifier and the AUC value of the trained decision tree-based adaptive boosting classifier, and select the classifier with the largest AUC value as the pulmonary nodule special screening prompt model;

[0170] S34, determining whether the pulmonary nodule special screening prompt model meets the accuracy requirement according to the AUC value of the pulmonary nodule special screening prompt model, if so, directly proceed to S35, otherwise, return to S32;

[0171] S35: Based on the training set and test set corresponding to the group of thyroid nodules with possible malignant risk, a special screening prompt model for thyroid nodules was constructed using the methods proposed in S32-S34;

[0172] S36: Based on the training set and test set corresponding to the group of breast nodules with possible malignant risks, a breast nodule special screening prompt model was constructed using the methods proposed in S32-S34;

[0173] S37: Integrate the lung nodule special screening prompt model, the thyroid nodule special screening prompt model and the breast nodule special screening prompt model to obtain the first basic model for prompting the need for nodule special screening.

[0174] The step S32 comprises:

[0175] S321: construct a basic decision tree model using the training set D corresponding to the group of pulmonary nodules that may have a risk of malignancy; the z-th branch node of the basic decision tree model includes all the examination indicators θ in the training set D with a value of θ z Samples of

[0176] S322: The training set D in the basic decision tree model is sampled using the bagging method to obtain a trained random forest classifier:

[0177] RF={h(x,θ k′ ), k′=1,2,…,t}

[0178] Among them, RF represents the trained random forest classifier; h represents the sampling set of the bagging method; θ k′ represents the k′th random vector that follows an independent distribution; t represents the number of decision trees generated, and x represents the samples in the training set;

[0179] S323: using the training set D to train the extremely randomized tree classifier to obtain a trained extremely randomized tree classifier;

[0180] S324: Using the training set D to train the decision tree-based adaptive boosting classifier, to obtain a trained decision tree-based adaptive boosting classifier.

[0181] The step S324 includes:

[0182] S3241: Calculate the weight W of the samples in the training set D n :

[0183]

[0184] Among them, w N Represents the weight of the Nth sample; N represents the total number of samples in the training set D;

[0185] S3242: Select a weak classifier as the tth base classifier h t , and calculate the base classifier h according to the sample weight Wn t The fitting error rate is:

[0186]

[0187] Among them, e t represents the fitting error rate of the tth base classifier; n′ represents the sample number in the training set D; w tn′ represents the weight of the n′th sample in the tth base classifier; I(·) is used to judge h t (X n′ ) and y n′ Whether it is unequal or not; h t (x n′ ) represents the sample x n′ The t-th base classifier of x n′ represents the n′th sample in the training set D; y n′ represents the observed value of the n′th sample, that is, whether a nodule occurs;

[0188] S3243: According to the base classifier h t The fitting error rate of the base classifier h t The parameter α required to update the weights in the final classifier t :

[0189]

[0190] Among them, ∈ t Represents the classifier h t (x n′ )’s error weight rate;

[0191] S3244: According to parameter α t , update the sample weight W t+1:

[0192] W t+1 =W t exp{α t I(h t (x n ′)≠y n′ )}

[0193] Among them, W t+1 Represents the updated sample weight; W t Represents the sample weight before updating;

[0194] S3245: The weak classifier h t With parameter α t Combined into a strong classifier H t :

[0195] H t =H t-1 +α t h t

[0196] Among them, H t represents the strong classifier of the tth iteration; H t-1 represents the strong classifier of the t-1th iteration;

[0197] S3246: Use symbolic functions to compare the effects of weak classifiers and strong classifiers:

[0198] C=sign(∑α t h t )

[0199] Among them, sign(·) represents the sign function, if ∑α t h t If it is less than 0, -1 is returned, indicating that the strong classifier H formed after this iteration t The effect is not as good as the weak classifier h t , if ∑α t h t If it is greater than 0, it returns 1, indicating that the strong classifier H formed after this iteration t The effect is better than the weak classifier h t ;

[0200] S3247: Repeat steps S3241-S3246 for multiple iterations, and select the best strong classifier as the trained decision tree-based adaptive boosting classifier H.

[0201] Step S33 includes:

[0202] For any one of the first basic model that indicates the need for a special screening prompt model for lung nodules, a special screening prompt model for thyroid nodules, and a special screening prompt model for breast nodules, the following steps are required to determine the first model parameters:

[0203] Construct the first confusion matrix, as shown in Table 5:

[0204] Table 5 The first confusion matrix table

[0205]

[0206] The first confusion matrix is ​​divided into four categories: true positive TP1, true negative TN1, false positive FP1, and false negative FN1;

[0207] Among them, the true positive TP1 is the healthy group sample correctly marked by the first basic model, that is, it is actually in the high-risk group and is predicted to be the high-risk group; the true negative TN1 is the healthy group sample correctly marked by the first basic model, that is, it is actually in the healthy group and is predicted to be the healthy group; the false positive FP1 is the healthy group sample incorrectly marked by the first basic model, that is, it is actually in the healthy group and is predicted to be the high-risk group; the false negative FN1 is the healthy group sample incorrectly marked by the first basic model, that is, it is actually in the high-risk group and is predicted to be the healthy group.

[0208] Calculate the model evaluation parameters: accuracy value, specificity value, F1 score value and AUC value;

[0209] The accuracy value is calculated as:

[0210]

[0211] Among them, Accuracy1 represents the accuracy value, which is the percentage of correctly classified samples in the first basic model to all samples, reflecting the recognition ability of the first basic model for various types of samples.

[0212] The specificity value is calculated as:

[0213]

[0214] Among them, Specifity1 represents the specificity value, which is the percentage of the number of healthy people correctly identified in the prediction results of the first basic model to the total number of healthy people, reflecting the recognition ability of the first basic model for healthy people.

[0215] The F1 score value calculation formula is:

[0216]

[0217] Among them, F11 represents the F1 score value, Precision1 represents the ratio of the number of correctly identified high-risk groups in the prediction results of the first basic model to the sum of the number of correctly identified high-risk groups and the number of incorrectly identified healthy groups, that is, the precision rate, which reflects the credibility of the first basic model in identifying high-risk groups. Recall1 represents the percentage of the number of correctly identified high-risk groups in the prediction results of the first basic model to the total number of high-risk groups, that is, the recall rate, which reflects the first basic model's ability to identify high-risk groups. The F1 score value is the harmonic mean of the precision and recall rates in the first basic model, reflecting the ratio of high precision and high recall rates.

[0218] in,

[0219] Calculate the AUC value;

[0220] A first ROC curve is drawn according to the true accuracy rate and the false accuracy rate in the first basic model, and the area under the first curve is calculated, and the value corresponding to the area under the first curve is taken as the AUC value.

[0221] in,

[0222] When the AUC value is between 0.7 and 0.85, it can be considered that the first basic model has average discrimination ability, and when it is greater than 0.85, it can be considered that the first basic model has good discrimination ability.

[0223] After obtaining the accuracy value, the specificity value, the F1 score value, and the AUC value, the first model parameters of the first basic model can be determined.

[0224] S4: Construct a second basic model to indicate whether malignant nodules may occur in the long term;

[0225] The step S4 comprises:

[0226] S41: Based on the initial data set, a target population with multiple physical examination information and whose first physical examination information clearly showed no malignancy was selected and defined as the change population;

[0227] S42: The changed population was divided into lung nodule group, thyroid nodule group and breast nodule group;

[0228] S43: The physical examination information of the changed population in the lung nodule group, thyroid nodule group, and breast nodule group was randomly divided into corresponding training sets and test sets;

[0229] S44: The training set corresponding to the population with changes in the pulmonary nodule group was imported into the long short-term memory network model, and a model for the possible occurrence of long-term malignant pulmonary nodules was constructed based on the included factors;

[0230] S45: import the test set corresponding to the changed population of the pulmonary nodule group into the model for possible indication of the occurrence of long-term malignant pulmonary nodules and output the true accuracy rate and false accuracy rate of the model for possible indication of the occurrence of long-term malignant pulmonary nodules, draw the ROC curve according to the true accuracy rate and false accuracy rate of the model for possible indication of the occurrence of long-term malignant pulmonary nodules, calculate the AUC value according to the ROC curve of the model for possible indication of the occurrence of long-term malignant pulmonary nodules, and determine whether the model for possible indication of the occurrence of long-term malignant pulmonary nodules meets the accuracy requirement according to the AUC value of the model for possible indication of the occurrence of long-term malignant pulmonary nodules, if so, directly enter S46, otherwise, return to S44;

[0231] S46: Based on the training set and test set corresponding to the population with changes in the thyroid nodule group, a model for the possible indication of the occurrence of long-term malignant thyroid nodules was constructed using the methods proposed in S44-S45;

[0232] S47: Based on the training set and test set corresponding to the population with changes in the breast nodule group, use the methods proposed in S44-S45 to construct a possible warning model for the occurrence of long-term malignant breast nodules;

[0233] S48: Integrate the possible indication model of the long-term malignant pulmonary nodule occurrence, the possible indication model of the long-term malignant thyroid nodule occurrence and the possible indication model of the long-term malignant breast nodule occurrence to obtain a second basic model for indicating whether malignant nodules may occur in the long term.

[0234] The step S44 is specifically as follows: irrelevant data is eliminated from the training set corresponding to the changed population of the lung nodule group based on the included factors to obtain a processed training set corresponding to the changed population of the lung nodule group, and a long short-term memory network model is trained based on the processed second training set corresponding to the changed population of the lung nodule group to obtain a possible prompt model for the occurrence of long-term malignant lung nodules.

[0235] Step S45 includes: constructing a second confusion matrix, as shown in Table 6:

[0236] Table 6 Second confusion matrix table

[0237]

[0238] The second confusion matrix is ​​divided into four categories: true positive TP2, true negative TN2, false positive FP2, and false negative FN2;

[0239] Among them, the true positive TP2 is the healthy group sample correctly marked by the second basic model, that is, it is actually in the high-risk group and is predicted to be the high-risk group; the true negative TN2 is the healthy group sample correctly marked by the second basic model, that is, it is actually in the healthy group and is predicted to be the healthy group; the false positive FP2 is the healthy group sample incorrectly marked by the second basic model, that is, it is actually in the healthy group and is predicted to be the high-risk group; the false negative FN2 is the healthy group sample incorrectly marked by the second basic model, that is, it is actually in the high-risk group and is predicted to be the healthy group.

[0240] Calculate the model evaluation parameters: accuracy value, specificity value, F1 score value and AUC value;

[0241] If the AUC value is between 0.7 and 0.85, it can be considered that the second basic model has average discrimination ability, and if it is greater than 0.85, it can be considered that the second basic model has good discrimination ability.

[0242] After obtaining the accuracy value, the specificity value, the F1 score value, and the AUC value, the second model parameters of the second basic model can be determined.

[0243] S5: Obtain a validation set, use the validation set to respectively validate the first basic model and the second basic model and obtain validation results, and adjust parameters of the first basic model and the second basic model based on the validation results until the parameters of the first basic model and the parameters of the second basic model converge;

[0244] Step S5 includes:

[0245] S51: 1,000 cases of the newly collected target population and control population are selected as the validation set, and then the validation set is processed. Among them, the selection criteria of the validation set are selected according to the inclusion and exclusion criteria, and the validation set includes the physical examination information of the newly collected target population and the control population. The physical examination information includes gender, age, blood test indicators, and imaging examination data or ultrasound examination data. For any physical examination indicator, if the missing rate reaches 20% or more, the physical examination indicator will not be included in the subsequent verification. The data processing of the validation set uses the validation set based on the imaging examination data or ultrasound examination data to classify the validation set, and then divide the validation set into three nodule groups including imaging examination data or ultrasound examination data, namely: a lung nodule validation group that may have a malignant risk, a thyroid nodule validation group that may have a malignant risk, and a breast nodule validation group that may have a malignant risk.

[0246] S52: Verify the model that prompts the need for special screening of lung nodules, the model that prompts the need for special screening of thyroid nodules, and the model that prompts the need for special screening of breast nodules.

[0247] The validation data corresponding to the lung nodule validation group is input into the model that prompts the need for special screening of lung nodules, and the first model result is output. The validation data corresponding to the thyroid nodule validation group is input into the model that prompts the need for special screening of thyroid nodules, and the second model result is output. The validation set data corresponding to the breast nodule validation group is input into the model that prompts the need for special screening of breast nodules, and the third model result is output. For any one of the first model result, the second model result, and the third model result, it includes: accuracy value, specificity value, F1 score value and AUC value.

[0248] Since the AUC values ​​of the model indicating the need for special screening for lung nodules, the model indicating the need for special screening for thyroid nodules, and the model indicating the need for special screening for breast nodules are respectively determined in step S33, the three AUC values ​​in the first model result, the second model result, and the third model result can be compared with the AUC value of the corresponding special screening prompt model. At this time, the AUC value of the special screening prompt model is used as the evaluation criterion to determine the situation of the above three models. If the standard of the first AUC value is not met, return to step S3 to adjust the parameters of the above three models until the first model parameters of the above three models converge.

[0249] S53: verifying the second basic model.

[0250] Select people who may have nodules and have multiple physical examination information and the first physical examination information clearly indicates that there is no possibility of malignancy from the validation set. Input the validation set data corresponding to the people who may have nodules into the second basic model, and output the fourth model result. The fourth model result includes an accuracy value, a specificity value, an F1 score value, and an AUC value. Since the AUC value of the second basic model is determined in step S4, the AUC value in the fourth model result can be compared with the AUC value of the second basic model. At this time, the AUC value of the second basic model is used as an evaluation index to determine the situation of the second basic model. If the standard of the second AUC value is not met, return to step S4 to adjust the parameters of the second basic model until the second model parameters in the second basic model converge.

[0251] S6. Obtain the data to be processed, input them into the first basic model and the second basic model respectively, and obtain the prompt results of the nodule special screening and the possible prompt results of the long-term malignant nodule occurrence.

[0252] As described above, the above embodiments are only used to introduce the technical solution of the present application in detail, but the description of the above embodiments is only used to help understand the method and core idea of ​​the present application and should not be understood as limiting the present application. Any changes or substitutions that can be easily thought of by technicians in this technical field within the technical scope disclosed in the present application should be included in the protection scope of the present application.

Claims

1. A method for processing data for special screening of nodules based on artificial intelligence and blood tests, characterized in that: include: S1: Get the initial data set; S2: Data processing of the initial data set to determine the included factors; S3: Based on the included factors, the first basic model is constructed to indicate the need for special screening for nodules; S4: Construct a second basic model to indicate whether malignant nodules may occur in the long term; S5: Obtain a validation set, use the validation set to respectively validate the first basic model and the second basic model and obtain validation results, and adjust parameters of the first basic model and the second basic model based on the validation results until the parameters of the first basic model and the parameters of the second basic model converge; S6. Obtain the data to be processed, input them into the first basic model and the second basic model respectively, and obtain the prompt results of the nodule special screening and the possible prompt results of the long-term malignant nodule occurrence.

2. The method for processing data for nodule-specific screening based on artificial intelligence and blood examination according to claim 1, characterized in that: The initial data set includes physical examination information of a high-risk target population and physical examination information of a healthy control population; the physical examination information includes gender, age, blood test indicators and imaging examination data.

3. The method for processing data for nodule-specific screening based on artificial intelligence and blood test according to claim 2, characterized in that: The S2 includes: S21: Based on the imaging examination data, the initial data set was divided into three nodule groups: a lung nodule group with a possible risk of malignancy, a thyroid nodule group with a possible risk of malignancy, and a breast nodule group with a possible risk of malignancy; S22: Based on the physical examination information of each nodule group, the propensity score calculated based on logistic regression was used to perform overlapping weighting on each sample in the initial data set, and the generalized estimating equation was used to screen the factors of each sample after overlapping weighting to obtain the included factors.

4. The method for processing data for nodule-specific screening based on artificial intelligence and blood test according to claim 3, characterized in that: The S22 includes: S221: Define the confounding factor effect estimation regression model used to divide the physical examination information of each nodule group into a high-risk group and a healthy control group: A~+x ′ 1,…,+x ′ k Where A represents the processing factor of the physical examination information. When A=1, the physical examination information is classified as a high-risk group, and when A=0, the physical examination information is classified as a healthy control group; x ′ k Represents the kth confounding factor of the physical examination information, including age and gender; k represents the confounding factor number; S222: Calculate the propensity score of the physical examination information using the processing factors of the physical examination information: E(X i )=P(A i =1|X i ) Among them, E(X i ) represents the propensity score of the i-th physical examination information; P(A i =1|X i ) represents the probability that the i-th physical examination information is assigned to the high-risk group under the conditions of statistical age and gender; A i represents the processing factor of the i-th physical examination information; X i represents the confounding factors of the ith physical examination information; S223: Define a logistic regression model for estimating the propensity score of physical examination information: Among them, P(A=1|X) represents the probability of the physical examination information being assigned to the high-risk group under the condition of statistical age and gender; X represents the set of confounding factors of the physical examination information; β0 represents the constant term of the logistic regression model; β1 and β k All represent the regression coefficients of the Logistic regression model; S224: Based on the Logistic regression model for estimating the propensity score of the physical examination information, the propensity prediction score of each physical examination information is calculated: Where E′(X) represents the propensity prediction score of physical examination information; P(A=1|X i ) represents the probability that the i-th physical examination information is assigned to the high-risk group under the conditions of statistical age and gender; S225: According to the tendency prediction score of the physical examination information, weights are assigned to the physical examination information to obtain the weights of each physical examination information: h(X i )=E′(X i )(1-E′(X i )) Among them, W i represents the weight of the i-th physical examination information; h(X i ) is E′(X i ) of the binomial function; E′(X i ) represents the propensity prediction score of the i-th physical examination information; S226: According to the weight of the physical examination information, a weighted generalized linear model is defined. The weighted generalized linear model includes a random part, a systematic part, and a link function part: a(φ)=φ / W i g(μ)=ln(μ) Among them, f(·) represents the random part, which indicates the probability distribution of the dependent variable; y g represents the value of the dependent variable corresponding to the g-th observation; g represents the observation value number of the examination index in the physical examination information; θ g represents a natural parameter related to the mean of the observations; φ represents a discrete parameter related to the variance of the observations; a(φ) represents a weight function related to the variance of the observations; b(θ g ) represents the weight function related to the mean of the observations; c(y g ,φ) represents the normalization constant used to ensure that the integral of the entire probability distribution function is 1; η represents the system part, which is a linear value; g(μ) represents the link function part; μ represents the link function value; m represents the number of inspection indicators; j is the number of inspection indicators; γ j represents the regression coefficient corresponding to the jth inspection indicator; x j represents the jth inspection index; S227: Use the logarithmic maximum likelihood estimation method to estimate the parameters of the weighted generalized linear model and establish the logarithmic likelihood function L: Among them, L g represents the log-likelihood function of the g-th observation value; n represents the total number of observation values ​​of the examination indicators in the physical examination information; S228: Use the Newton-Raphson iterative method to maximize the log-likelihood function L and obtain γ j The maximum likelihood estimate b j ; S229: Maximum likelihood estimate b j A likelihood ratio test was performed to obtain the included factors.

5. The method for processing data for nodule-specific screening based on artificial intelligence and blood test according to claim 4, characterized in that: The S229 includes: S2291: The maximum likelihood estimate b j Substitute into the weighted generalized linear model, and fit the weighted generalized linear model without the inspection index, and find the log-likelihood function lnL0 corresponding to the weighted generalized linear model without the inspection index; S2292: adding the inspection indicators to the generalized linear model for fitting in turn to obtain a new log-likelihood function value lnL1; S2293: Calculate the likelihood ratio statistic G based on lnL0 and lnL1: G=2(lnL1-lnL0) S2294: Determine whether the likelihood ratio statistic G is greater than or equal to the preset statistical test standard value. If so, set the inspection indicator corresponding to lnL1 as the inclusion factor, and return to S2292 to screen the factors of the next inspection indicator. Otherwise, directly return to S2292 to screen the factors of the next inspection indicator.

6. The method for processing data for nodule-specific screening based on artificial intelligence and blood test according to claim 3, characterized in that: The S3 includes: S31: According to the inclusion factors, irrelevant data are removed from the initial data set to obtain a processed initial data set, and each nodule group in the processed initial data set is randomly divided into corresponding training sets and test sets; S32: using the training set corresponding to the group of pulmonary nodules that may have a risk of malignancy, respectively training a random forest classifier, an extremely random tree classifier, and an adaptive boosting classifier based on a decision tree, to obtain a trained random forest classifier, a trained extremely random tree classifier, and a trained adaptive boosting classifier based on a decision tree; S33: import the test set corresponding to the group of pulmonary nodules with possible malignant risks into the trained random forest classifier, the trained extreme random tree classifier and the trained decision tree-based adaptive boosting classifier respectively, obtain the AUC value of the trained random forest classifier, the AUC value of the trained extreme random tree classifier and the AUC value of the trained decision tree-based adaptive boosting classifier, and select the classifier with the largest AUC value as the pulmonary nodule special screening prompt model; S34, determining whether the pulmonary nodule special screening prompt model meets the accuracy requirement according to the AUC value of the pulmonary nodule special screening prompt model, if so, directly proceed to S35, otherwise, return to S32; S35: Based on the training set and test set corresponding to the group of thyroid nodules with possible malignant risk, a special screening prompt model for thyroid nodules was constructed using the methods proposed in S32-S34; S36: Based on the training set and test set corresponding to the group of breast nodules with possible malignant risks, a breast nodule special screening prompt model was constructed using the methods proposed in S32-S34; S37: Integrate the lung nodule special screening prompt model, the thyroid nodule special screening prompt model and the breast nodule special screening prompt model to obtain the first basic model for prompting the need for nodule special screening.

7. The method for processing data for nodule-specific screening based on artificial intelligence and blood test according to claim 6, characterized in that: The step S32 comprises: S321: construct a basic decision tree model using the training set D corresponding to the group of pulmonary nodules that may have a risk of malignancy; the z-th branch node of the basic decision tree model includes all the examination indicators in the training set D The upper value is Samples of S322: The training set D in the basic decision tree model is sampled using the bagging method to obtain a trained random forest classifier: RF={h(x,θ k′ ),k′=1,2,…,t} Among them, RF represents the trained random forest classifier; h represents the sampling set of the bagging method; θ k′ represents the k′th random vector that follows an independent distribution; t represents the number of decision trees generated, and x represents the samples in the training set; S323: using the training set D to train the extremely randomized tree classifier to obtain a trained extremely randomized tree classifier; S324: Using the training set D to train the decision tree-based adaptive boosting classifier, to obtain a trained decision tree-based adaptive boosting classifier.

8. The method for processing data for nodule-specific screening based on artificial intelligence and blood test according to claim 7, characterized in that: The step S324 includes: S3241: Calculate the weight W of the samples in the training set D n : Among them, w N Represents the weight of the Nth sample; N represents the total number of samples in the training set D; S3242: Select a weak classifier as the tth base classifier h t , and according to the sample weight W n , calculate the base classifier h t The fitting error rate is: Among them, e t represents the fitting error rate of the tth base classifier; n′ represents the sample number in the training set D; w tn′ represents the weight of the n′th sample in the tth base classifier; I(·) is used to judge h t (x n′ ) and y n′ Whether it is unequal or not; h t (x n′ ) represents the sample x n′ The t-th base classifier of x n′ represents the n′th sample in the training set D; y n′ represents the observed value of the n′th sample, that is, whether a nodule occurs; S3243: According to the base classifier h t The fitting error rate of the base classifier h t The parameter α required to update the weights in the final classifier t : Among them, ∈ t Represents the classifier h t (x n′ )’s error weight rate; S3244: According to parameter α t , update the sample weight W t+1 : W t+1 =W t exp{α t I(h t (x n′ )≠y n′ )} Among them, W t+1 Represents the updated sample weight; W t Represents the sample weight before updating; S3245: The weak classifier h t With parameter α t Combined into a strong classifier H t : H t =H t-1 +α t h t Among them, H t represents the strong classifier of the tth iteration; H t-1 represents the strong classifier of the t-1th iteration; S3246: Use symbolic functions to compare the effects of weak classifiers and strong classifiers: C=sign(∑α t h t ) Among them, sign(·) represents the sign function, if ∑α t h t If it is less than 0, -1 is returned, indicating that the strong classifier H formed after this iteration t The effect is not as good as the weak classifier h t , if ∑α t h t If it is greater than 0, it returns 1, indicating that the strong classifier H formed after this iteration t The effect is better than the weak classifier h t ; S3247: Repeat steps S3241-S3246 for multiple iterations, and select the best strong classifier as the trained decision tree-based adaptive boosting classifier H.

9. The method for processing data for nodule-specific screening based on artificial intelligence and blood test according to claim 1, characterized in that: The step S4 comprises: S41: Based on the initial data set, a target population with multiple physical examination information and whose first physical examination information clearly showed no malignancy was selected and defined as the change population; S42: The changed population was divided into lung nodule group, thyroid nodule group and breast nodule group; S43: The physical examination information of the changed population in the lung nodule group, thyroid nodule group, and breast nodule group was randomly divided into corresponding training sets and test sets; S44: The training set corresponding to the population with changes in the pulmonary nodule group was imported into the long short-term memory network model, and a model for the possible occurrence of long-term malignant pulmonary nodules was constructed based on the included factors; S45: import the test set corresponding to the changed population of the pulmonary nodule group into the model for possible indication of the occurrence of long-term malignant pulmonary nodules and output the true accuracy rate and false accuracy rate of the model for possible indication of the occurrence of long-term malignant pulmonary nodules, draw the ROC curve according to the true accuracy rate and false accuracy rate of the model for possible indication of the occurrence of long-term malignant pulmonary nodules, calculate the AUC value according to the ROC curve of the model for possible indication of the occurrence of long-term malignant pulmonary nodules, and determine whether the model for possible indication of the occurrence of long-term malignant pulmonary nodules meets the accuracy requirement according to the AUC value of the model for possible indication of the occurrence of long-term malignant pulmonary nodules, if so, directly enter S46, otherwise, return to S44; S46: Based on the training set and test set corresponding to the population with changes in the thyroid nodule group, a model for the possible indication of the occurrence of long-term malignant thyroid nodules was constructed using the methods proposed in S44-S45; S47: Based on the training set and test set corresponding to the population with changes in the breast nodule group, use the methods proposed in S44-S45 to construct a possible warning model for the occurrence of long-term malignant breast nodules; S48: Integrate the possible indication model of the long-term malignant pulmonary nodule occurrence, the possible indication model of the long-term malignant thyroid nodule occurrence and the possible indication model of the long-term malignant breast nodule occurrence to obtain a second basic model for indicating whether malignant nodules may occur in the long term.

10. The method for processing data for nodule-specific screening based on artificial intelligence and blood test according to claim 9, characterized in that: The step S44 is specifically as follows: irrelevant data is eliminated from the training set corresponding to the changed population of the lung nodule group based on the included factors to obtain a processed training set corresponding to the changed population of the lung nodule group, and a long short-term memory network model is trained based on the processed second training set corresponding to the changed population of the lung nodule group to obtain a possible prompt model for the occurrence of long-term malignant lung nodules.

Images