A method for constructing a lung cancer staging diagnosis model and a diagnosis model

Abstract

The present application belongs to the field of medical information technology, and relates to a method for constructing a lung cancer staging diagnosis model and a diagnostic model. The construction method comprises the following steps: S1: obtaining the lung cancer typing and TNM staging of the subject sample; S2: constructing a two-dimensional space based on the lung cancer typing and TNM staging of the subject sample, and integrating and correcting the subject samples in different TNM stages using the anchoring method; S3: forming a sample set with the subject samples uniformly distributed in the previous step and the corresponding feature variables, and performing machine learning using the random forest algorithm in the sample set to construct a lung cancer staging diagnosis model. The present application uses the two-dimensional space constructed by the lung cancer typing and TNM staging of the subject sample, and integrates and corrects the samples using the anchoring method, successfully making the samples in the sample set evenly distributed, so that the constructed data model is more reliable.

Description

Technical Field

[0001] The present application belongs to the field of medical information technology, and more specifically, relates to a method for constructing a lung cancer staging diagnosis model and a diagnosis model. Background Art

[0002] Lung cancer is a malignant tumor that originates in the bronchial mucosa or lung glands. Its incidence and mortality rates are rapidly increasing, making it one of the most threatening malignancies to human health and life. According to the American Journal of Clinicians in Cancer, the global incidence of lung cancer was 11.6% in 2018, with 31.5% occurring in men and 14.6% in women. 18.4% of deaths occurred, with 27.1% occurring in men and 11.2% occurring in women. According to the 2018 China Cancer Report, the cumulative risk of cancer for urban residents aged 0-85 has reached 35%. Approximately 75% of patients are diagnosed with advanced lung cancer, missing the optimal time for radical surgery. Early diagnosis can significantly improve prognosis and survival for lung cancer patients. This can be seen from the 5-year survival rate of patients with stage I lung cancer after surgery: the 5-year survival rate for stage I lung cancer is 77%-92%, while for stage IIA-IV lung cancer, it is only 10%-36%. Early diagnosis of lung cancer is crucial for improving its cure rate.

[0003] However, the early symptoms of lung cancer are very good at disguising themselves. Common lung cancer symptoms, such as cough, phlegm, chest tightness, chest pain, shortness of breath, and fever, mimic the clinical manifestations of many respiratory diseases and can be difficult to take seriously. Even when screened at a hospital, early-stage lung cancer is often overlooked because chest X-rays are ineffective in detecting it.

[0004] With the development of database and artificial intelligence (AI) technologies, data mining has become a hot emerging interdisciplinary and multi-disciplinary topic. Data mining technology has evolved from simple data queries to mining and discovering hidden information, and can help doctors diagnose lung cancer early. Therefore, providing a data mining-based analysis tool to assist doctors in early diagnosis of lung cancer is of great significance. Patent document CN113707327A discloses a multi-tumor marker tumor diagnosis model based on medical big data, which includes big data cleaning, big data organization, and big data classification. The model construction includes logistics regression, Lasso regression, ridge regression, Boosting and AdaBoost reinforcement learning, decision tree models, support vector models, elastic net regression models, naive Bayes, K-nearest neighbor algorithms, random forest algorithms, and neural network models to diagnose tumor markers, including lung cancer. Patent document CN113707298A discloses a predictive method for tumor diagnosis based on medical big data. Researchers collect markers from various medical records related to physical examinations and tumors, including age and gender, and use known tumor data to determine whether the tumor is malignant or benign. After iterative training, various machine learning algorithms are used to estimate the benign or malignant nature of the tumor. However, due to objective factors such as the lack of early diagnosis methods in economically underdeveloped areas and the death of late-stage patients, the number of samples from patients at different stages of lung cancer is often uneven. Furthermore, different medical institutions use different diagnostic procedures, and the parameters of the collected subject samples are sometimes missing, making it difficult to fully model the model, resulting in a lack of accuracy in the model predictions. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the purpose of this application is to improve the accuracy of constructing a lung cancer staging diagnosis model through a new method for constructing a lung cancer staging diagnosis model by integrating and correcting the subject samples, improving the uniformity of the samples.

[0006] To achieve the above objectives, the present application provides a method for constructing a lung cancer staging diagnosis model, the method comprising the following steps:

[0007] S1: Obtain the lung cancer classification and TNM stage of the subject samples and the corresponding characteristic variables; wherein the number of subject samples corresponding to each TNM stage is 2 or more;

[0008] S2: constructing a two-dimensional space based on the lung cancer classification and TNM stage (tumor node metastasis classification) of the subject samples, and integrating and correcting the subject samples in different TNM stages using an anchoring method so that the subject samples are evenly distributed in the two-dimensional space;

[0009] S3: The evenly distributed subject samples and the feature variables corresponding to the subject samples are combined into a sample set. A random forest algorithm is used to perform machine learning on the sample set to construct a lung cancer staging diagnosis model. The parameters of the random forest include the number of decision trees, the minimum number of leaf nodes, replacement sampling, and the number of samples randomly selected during node splitting.

[0010] Preferably, in step S1, the number of subject samples is greater than 2000, and the number of characteristic variables is 15-25.

[0011] Preferably, the TNM staging in step S1 includes stage 0, stage 1, stage 2, stage 3 and stage 4.

[0012] Preferably, the lung cancer classification in step S1 includes small cell carcinoma, adenocarcinoma, squamous cell carcinoma, and large cell carcinoma.

[0013] Preferably, the characteristic variables are related to the TNM stage, including categorical variables and numerical variables, the categorical variables including age, gender, smoking, drinking, nodule density, nodule distribution and antibody test results; the antibody test results include GAGE7, MAGE A1 and GBU4-5.

[0014] As a further preferred method, the method for determining whether the characteristic variables are related to the TNM stage is specifically as follows: performing a normality test on the numerical variables, performing a Spearman correlation analysis on the numerical variables determined to be non-normally distributed, performing a Pearson correlation analysis on the numerical variables determined to be normally distributed, to obtain the numerical variables related to the TNM stage; performing a chi-square test on the categorical variables, to obtain the categorical variables related to the TNM stage.

[0015] As a further preferred embodiment, the feature variables are obtained from subject samples and corresponding incomplete feature variables, and the specific method is to obtain subject samples and corresponding incomplete feature variables; delete subject samples corresponding to feature variables whose missing amount of subject samples in the feature variables is less than a first threshold, and subject samples whose missing amount of feature variables of subject samples is greater than a second threshold, and use neural network generation technology to fill in other missing feature variables in the subject samples; the first threshold is 1%~3%, and the second threshold is 5%~10%.

[0016] As further preferred, the classification variables also include 7-AAB, CAGE, PGP9.5, SOX2 and P53.

[0017] As a further preference, the numerical variables are original numerical variables constructed through principal component analysis; the original numerical variables include nodule size, tumor marker levels, liver and kidney function test results, blood routine test results and electrolyte test results; the tumor marker levels include CEA and CA19-9.

[0018] As a further preferred embodiment, the principal component analysis is specifically as follows: according to the original numerical variable, the eigenvalue and the corresponding eigenvector are obtained; the eigenvectors are selected as principal components according to the eigenvalue from large to small, until the total variance explanation rate of the eigenvector is greater than the variance explanation rate threshold, and the principal component is the numerical variable; the variance explanation rate threshold is 60%~80%.

[0019] Preferably, step S2 specifically includes the following sub-steps:

[0020] S21: constructing a two-dimensional space based on the lung cancer classification and TNM stage of the subject sample, randomly selecting 2-5 anchor points in the two-dimensional space, and evaluating the uniformity of the subject sample using the nearest neighbor distance method. If the uniformity meets the requirements, proceed to step S3; otherwise, proceed to the next step;

[0021] S22: According to the number of subject samples in different TNM stages, oversampling and / or undersampling are performed on the corresponding subject samples so that the uniformity of the subject samples meets the requirements and the next step is entered; the oversampling method is the SMOTE method.

[0022] As a further preferred embodiment, in step S22, under-sampling is performed for subject samples whose number corresponding to the TNM stage is greater than a first threshold, and over-sampling is performed for subject samples whose number corresponding to the TNM stage is less than a second threshold.

[0023] As further preferred, the first threshold=the second threshold=the average value of the number of subject samples in different TNM stages.

[0024] Preferably, after step S3, the method further includes step S4: evaluating the lung cancer staging diagnosis model using the ROC curve and the AUC index, and if the evaluation criteria are not met, returning to S3.

[0025] As a further preference, the evaluation standard in step S4 is: the area below the ROC curve is 0.7-0.9, and the area below the ROC curve is greater than 0.9.

[0026] Another object of the present application is to provide a lung cancer staging diagnosis model obtained using the above-mentioned construction method.

[0027] Another object of the present application is to provide a diagnostic system including the above-mentioned lung cancer staging diagnostic model, further comprising:

[0028] A data input module for obtaining characteristic variables of the sample to be tested;

[0029] An output module for obtaining a diagnosis result based on the characteristic variables and the lung cancer staging diagnosis model.

[0030] Preferably, the diagnostic system further comprises a sample storage module, wherein the sample storage module is used to store subject samples with lung cancer typing and TNM staging, and characteristic variables corresponding to the subject samples.

[0031] Another object of the present application is to provide a computer program for executing the above-mentioned lung cancer staging diagnosis model or diagnosis system, and a non-transitory computer-readable storage medium storing the above-mentioned computer program.

[0032] The present application also provides a system for constructing a lung cancer staging diagnosis model and a computer program for executing the system, which are applied to the above-mentioned method for constructing a lung cancer staging diagnosis model. The system comprises:

[0033] a data acquisition module for obtaining subject samples;

[0034] A data processing module for performing integrated correction on the subject sample;

[0035] A model building module used to construct a lung cancer staging diagnosis model using subject samples to form a sample set and use the random forest algorithm for machine learning.

[0036] The present application also provides a marker for lung cancer staging diagnosis and the use of the marker in a lung cancer staging diagnosis kit, wherein the marker includes age, gender, smoking status, drinking status, CT test results, antibody test results, CEA and CA19-9; wherein the CT test results include nodule size, nodule density and nodule distribution.

[0037] In general, the above technical solutions conceived by this application have the following technical advantages compared with the existing technologies:

[0038] 1. In step S2, in the two-dimensional space constructed by lung cancer classification and TNM staging, the nearest neighbor distance method is used in combination with the anchoring method to integrate and correct the subject samples so that the subject sample data meets the uniformity requirement. Furthermore, the nearest neighbor distance method is used to evaluate the uniformity of the subject samples. By accurately calculating the gap between the empirical nearest neighbor and the theoretical nearest neighbor of the anchor point, the distribution uniformity of the sample set is successfully quantified. It has been verified that in the evaluation of the embodiment, the method was used to discover the unevenness of the sample distribution, especially the significant difference in the number of nearest neighbors between the anchor points, which verifies the effectiveness of the method.

[0039] 2. For unevenly distributed subject sample data, this application adopts targeted data balancing measures (i.e., undersampling or oversampling the sample data of subjects at different TNM stages) for integration and correction. For example, in the embodiment, random undersampling of non-lung cancer and stage 1 lung cancer subject samples was performed, while the sample size of patients at other stages was expanded to ensure a balanced sample size across all stages. This improvement not only enhances the representativeness of the data, but also improves the accuracy and generalization ability of subsequent model training, providing more reliable data support for in-depth research.

[0040] 3. The present application preferably screened characteristic variables related to TNM staging through methods such as Spearman correlation analysis, Pearson correlation analysis, and chi-square test, and used the random forest method to construct a model. The diagnostic model finally constructed in the embodiment includes 21 independent characteristic variables related to CT test results, liver and kidney function test results, blood routine test results, antibody indicators, etc.: 11 categorical variables (sex, smoking, drinking, nodule distribution, nodule density, autoantibody levels related to cancer antigens (7-AAB), autoantibody CAGE, autoantibody GAGE ​​7, autoantibody MAGE A1, autoantibody PGP9.5, autoantibody SOX2), and 10 features obtained after principal component analysis; it was verified that the median accuracy of the multiple test in the embodiment of the present application was 0.881, and the maximum accuracy was greater than 0.93, which means that the model exhibits good predictive performance and indicates its practical application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A flowchart for constructing the diagnostic model for this application;

[0042] Figure 2 A schematic diagram of the system structure for this application;

[0043] Figure 3 A flowchart for constructing a prediction model for Example 1 of this application;

[0044] Figure 4 This is a diagram of the data dimensionality reduction process after steps S13 and S14 in Example 1 of the present application;

[0045] Figure 5 1 is the ROC curve and AUC index graph of the prediction model constructed in Example 1 of the present application;

[0046] Figure 6 It is the accuracy comparison of Example 1 and the comparative example. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0048] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, "plurality" means two or more, unless otherwise specifically defined.

[0049] Additionally, references throughout this specification to "one embodiment," "one embodiment," "an example," or similar language indicate that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present application. Thus, appearances of the phrase "in one embodiment," "in one embodiment," and similar language throughout this specification may, but do not necessarily, all refer to the same embodiment.

[0050] The present application provides a method for constructing a lung cancer staging diagnostic model, a construction system applied to the construction method, and a diagnostic model obtained by the construction method; the construction method generally includes four steps: obtaining a sample set, enhancing sample uniformity, constructing a diagnostic model, and evaluating accuracy. Figure 1 As shown, the details are as follows:

[0051] S1: Get sample set

[0052] Obtaining characteristic variables of the subject samples and corresponding lung cancer classification and TNM stage, wherein the TNM stage includes stage 0, stage 1, stage 2, stage 3, and stage 4, with stage 0 indicating no evidence of a primary tumor; the lung cancer classification includes small cell carcinoma, adenocarcinoma, squamous cell carcinoma, and large cell carcinoma; wherein the number of subject samples corresponding to each TNM stage is at least 2; for the accuracy of the model, in some embodiments, the number of subject samples is generally greater than 2000, and the number of characteristic variables is 15-25;

[0053] S11: Obtain the characteristic information of the subject sample and the corresponding lung cancer classification and TNM stage. The characteristic information includes numerical variables and categorical variables, including basic clinical information, test information, etc. The categorical variables include gender, age, smoking, drinking, lung CT results and cancer antigen-related autoantibody levels (7-aab, CAGE, GAGE7, GBU4-5, MAGE A1, P53, PGP9.5, SOX2); Numerical variables included tumor marker levels (CA19-9, ProGRP, NSE, SCC, CEA), routine blood test results (white blood cell count, monocyte percentage, neutrophil count, neutrophil percentage, lymphocyte count, lymphocyte percentage, basophil count, basophil percentage, eosinophil count, eosinophil percentage, platelet count, mean platelet volume, platelet hematocrit, platelet distribution width, large platelet ratio, red blood cell count, coefficient of variation of red blood cell distribution width, red blood cell distribution width, mean corpuscular hemoglobin, mean corpuscular hemoglobin), Protein concentration, hematocrit, hemoglobin), liver and kidney function test results (indirect bilirubin, direct bilirubin, total bilirubin, creatinine, uric acid, urea, albumin, globulin, albumin / globulin ratio, cholinesterase, total protein, alanine aminotransferase, aspartate aminotransferase, gamma-glutamyl transpeptidase, bicarbonate, estimated glomerular filtration rate, alkaline phosphatase) and electrolyte examination (sodium, potassium, chloride, calcium, calibrated calcium); specifically, the number and type of characteristic information collected can be adjusted according to the quality of the subject's sample itself, but it is best to at least include age, sex, smoking, drinking, nodule size, nodule density, nodule distribution, CEA, CA19-9, GAGE7, MAGE A1 and gbu4-5 are the key feature information for lung cancer diagnosis, and the number should be greater than or equal to 15. Among them, the lung CT results are not the lung CT images used directly, but the lung CT results are standardized to obtain information on nodule size, nodule density, and nodule distribution. Among them, nodule size is a numerical variable, and nodule density (divided into partially solid nodules, solid nodules, and ground glass density nodules) and nodule distribution (divided into multiple nodules and solitary nodules) are both categorical variables.

[0054] S12. In the actual process of collecting samples, it is very common that the characteristic information of the subject samples is incomplete. If the subject samples are sufficient, the subject samples with missing features can be directly discarded, but in practice, the data volume is usually insufficient. In some embodiments, the subject samples corresponding to the feature variables whose missing amount in the feature variables is less than the first threshold and the subject samples whose missing amount in the feature variables is greater than the second threshold are deleted, and the neural network generation technology is used to fill in the other missing feature variables in the subject samples to improve the sample set. The first threshold is 1%~3%, and the second threshold is 5%~10%. The neural network generation technology is used to process high-proportion missing values ​​(missing values>1%~3%), and the strategy of directly eliminating small-proportion missing values ​​(missing values<5%~10%) ensures the integrity and accuracy of the data.

[0055] S13. Determine the correlation between characteristic variables and TNM staging: Perform a normality test (KS test) on numerical variables, perform Spearman correlation analysis on numerical variables determined to be non-normally distributed, and perform Pearson correlation analysis on numerical variables determined to be normally distributed. This reveals the relationship between characteristic variables and TNM staging and obtains numerical variables related to TNM staging. Perform a chi-square test on categorical variables to obtain categorical variables related to TNM staging (using P value < 0.001 as the judgment criterion).

[0056] In some embodiments, it has been verified that the categorical variables most correlated with TNM staging include age, gender, smoking status, alcohol consumption, nodule density, nodule distribution, and antibody test results; the antibody test results include GAGE7, MAGEA1, and GBU4-5, etc.; the numerical variables most correlated with TNM staging include nodule size, CEA, and CA19-9; therefore, the above-mentioned characteristic variables can be used as markers for lung cancer staging diagnosis and for constructing detection markers for a lung cancer staging diagnostic kit.

[0057] S14. After step S13, the number of numerical variables is still large. Principal component analysis (PCA) can be used to reduce the dimensionality of the numerical variables. Specifically, the Kaiser-Meyer-Olkin (KMO) sampling adequacy test and Bartlett's sphericity test are used to determine whether the current sample is suitable for PCA.

[0058] If principal component analysis is not suitable, it usually means that the original dataset does not meet certain key conditions or assumptions required for this statistical method. There may be problems with the original dataset in terms of correlation, dimensionality reduction applicability, and sample size. At this time, you can temporarily proceed to the next step of modeling. If the model effect is not good, you should re-evaluate the data integrity and preprocessing process, try to collect more data to increase diversity and representativeness, or consider using alternative statistical methods. If principal component analysis is suitable, the original numerical variables are condensed to obtain eigenvalues ​​and corresponding eigenvectors. Eigenvectors are selected as principal components from large to small eigenvalues ​​until the total variance explained by the eigenvector is greater than the variance explained threshold. The principal component is the numerical variable; the variance explained threshold is usually set to 60% to 80%.

[0059] The data was further reduced in dimensionality through principal component analysis, which significantly reduced the number of features while retaining key information, and a streamlined sample set was constructed, demonstrating the significant advantages of this application in the field of data processing and analysis.

[0060] S2: Enhanced sample uniformity

[0061] A two-dimensional space is constructed according to the lung cancer classification and TNM stage of the subject samples, and the subject samples in different TNM stages are integrated and corrected using the anchoring method so that the subject samples are evenly distributed in the two-dimensional space. Specifically,

[0062] S21: Construct a two-dimensional space based on the lung cancer classification and TNM stage of the subject sample, randomly select 2-5 anchor points in the two-dimensional space, and use the nearest neighbor distance method to determine their error value, error If the uniformity is less than the threshold, it is considered to meet the requirements and the process goes to step S3; otherwise, the process goes to the next step.

[0063]

[0064] in, is the number of anchor points, is the actual number of nearest neighbors, is the theoretical nearest neighbor number.

[0065] In some embodiments, the threshold is set to 0.2.

[0066] S22: According to the number of subject samples in different TNM stages, oversampling and / or undersampling are performed respectively so that the uniformity of the subject samples meets the requirements, and the next step is entered; the oversampling method is the SMOTE method; in some embodiments, in order to simplify the calculation, the average value of the number of subject samples corresponding to different TNM stages can be directly calculated first, and the subject samples above the average value can be undersampled, and the subject samples below the average value can be oversampled by the SMOTE method; since the SMOTE method oversampling must ensure that there are more than two samples in this category, the number of subject samples corresponding to each TNM stage obtained in step S1 should be 2 or more.

[0067] S3: Diagnostic model construction

[0068] The subject samples uniformly distributed in the previous step and the corresponding characteristic variables are combined into a sample set. A random forest algorithm is used to perform machine learning on the sample set to construct a lung cancer staging diagnosis model. The characteristic variables of the samples to be tested are input into the lung cancer staging diagnosis model to obtain the TNM staging of the samples to be tested, thereby achieving lung cancer staging diagnosis. The parameters of the random forest include the number of decision trees, the minimum number of leaf nodes, replacement sampling, and the number of samples randomly selected during node splitting.

[0069] S4: Accuracy Assessment

[0070] The lung cancer staging diagnosis model is evaluated using the ROC curve and the AUC index. If the evaluation criteria are not met, the process returns to S3. In some embodiments, the evaluation criteria are: the area under the ROC curve is 0.7 to 0.9, and the area under the ROC curve is greater than 0.9.

[0071] A system for constructing a lung cancer staging diagnosis model can be executed on a computer program and applied to the above-mentioned construction method, wherein the construction system is as follows: Figure 2 Shown, including:

[0072] a data acquisition module for obtaining subject samples;

[0073] A data processing module for performing integrated correction on the subject sample;

[0074] A model building module used to construct a lung cancer staging diagnosis model using subject samples to form a sample set and use the random forest algorithm for machine learning.

[0075] The lung cancer staging diagnosis model can be used to construct a diagnosis system, which also includes a data input module and an output module;

[0076] The data input module is used to obtain characteristic variables of the sample to be tested;

[0077] The output module is used to obtain a diagnosis result based on the characteristic variables and the diagnosis model.

[0078] In some embodiments, the diagnostic system further includes a sample storage module to update and expand the parameters of the constructed diagnostic model, wherein the sample storage module is used to store subject samples with lung cancer classification and TNM staging, as well as characteristic variables corresponding to the subject samples.

[0079] The above-mentioned diagnostic model or diagnostic system may be implemented using a computer program as a medium and executed on a non-transitory computer-readable storage medium storing the computer program.

[0080] The application of the prediction model of this application is not limited to the routine clinical diagnosis of lung cancer, but can also include the following expanded applications:

[0081] 1) Early screening: This model can be used to conduct early screening of patients suspected of lung cancer, especially those who do not have obvious symptoms but have risk factors.

[0082] 2) As a medical supplement: Use this model to conduct early diagnosis and screening of lung cancer and expand the scope of diagnosis and treatment of lung cancer.

[0083] 3) Scientific research and education: This model can be used as a scientific research tool to identify factors related to the incidence and outcome of lung cancer, explore the pathogenesis of lung cancer, and the interactions between different factors.

[0084] It is particularly important to note that when using the model in this application, attention should be paid to data privacy protection and ethical review to ensure that all operations comply with relevant laws and regulations and medical ethics standards. Furthermore, because this model is constructed based on data from a specific population, differences in factors such as region and race should be taken into account when promoting and applying it, and appropriate adjustments and verification of the model may be required.

[0085] The following content is an embodiment

[0086] Example 1 Construction of lung cancer diagnosis model

[0087] The construction process of the diagnostic model in this embodiment is as follows Figure 3 As shown, specifically:

[0088] S1: Obtain the characteristic information of the subject samples and the corresponding lung cancer classification (stages 0-4) and TNM staging (divided into small cell lung cancer (SCLC) and non-small cell lung cancer (NSCLC, which is further subdivided into adenocarcinoma, squamous cell carcinoma, and large cell carcinoma). Since sample information collected from hospitals or research institutions usually has some missing information, and due to the large number of characteristic variables and the large amount of calculation, some preprocessing is required after collection; specifically, this step includes the following substeps:

[0089] S11: Obtaining characteristic information of the subject sample and the corresponding lung cancer classification (stages 0-4) and TNM staging; the characteristic information includes basic clinical information, laboratory information, etc.; the subject samples in this embodiment are from 3075 lung patients studied at Tongji Hospital, Tongji Medical College, and Huazhong University of Science and Technology from January 2021 to June 2022;

[0090] The basic clinical information included gender, age, smoking and drinking status; the test information included tumor marker levels (CA19-9, ProGRP, NSE, SCC, CEA), blood routine test results (white blood cell count, monocyte percentage, neutrophil count, neutrophil percentage, lymphocyte count, lymphocyte percentage, basophil count, basophil percentage, eosinophil count, eosinophil percentage, platelet count, mean platelet volume, platelet hematocrit, platelet distribution width, large platelet ratio, red blood cell count, red blood cell percentage, platelet count, platelet distribution width ... Table 1 shows the coefficient of variation of cell distribution width, red blood cell distribution width, mean corpuscular hemoglobin, mean corpuscular hemoglobin concentration, hematocrit, hemoglobin), liver and kidney function test results (indirect bilirubin, direct bilirubin, total bilirubin, creatinine, uric acid, urea, albumin, globulin, albumin / globulin ratio, cholinesterase, total protein, alanine aminotransferase, aspartate aminotransferase, γ-glutamyl transpeptidase, bicarbonate, estimated glomerular filtration rate, alkaline phosphatase), and levels of autoantibodies related to cancer antigens (7-aab, CAGE, GAGE7, GBU4-5, MAGE A1, P53, PGP9.5, SOX2), electrolyte tests (sodium, potassium, chloride, calcium, calibrated calcium), and lung CT results (nodule size, nodule density, and nodule distribution).

[0091] Table 1 Characteristic variables of basic clinical information and laboratory information

[0092]

[0093] S12: During the data inspection process, we found that 8 samples were not lung cancer cases but had lung cancer stage data, which were incorrect samples. We cleared the data in the sample feature of lung cancer stage and corrected it into the non-lung cancer group. In the first sample set collected, not all feature variables of each patient were complete. However, if as long as one feature variable was missing, the sample would be discarded, which would result in too little sample data and affect the quality of the prediction model. We first counted the subject samples corresponding to the feature values ​​whose missing sample data amount in all feature variables was less than 1% (i.e. 31 people), and discarded the subject samples as a whole. If the number of missing features of a patient was greater than 5% (i.e., the collected feature variables were less than 69), the sample was discarded, and the remaining missing feature variables were filled using neural network generation technology (i.e., the neural network was trained with subject samples without missing feature variables, and then the complete features in the missing samples were input to predict the missing features).

[0094] Through statistics, we found missing values ​​for 56 sample features, including nodule size, nodule density, nodule distribution, indirect bilirubin, and mean hematocrit. After using neural network generation technology to fill in the missing values, we finally obtained a second sample set constructed from 2,956 subject samples and 73 sample features.

[0095] S13: Correlation analysis was performed on the characteristic variables in the second sample set and the TNM stage, and data dimensionality reduction was performed to obtain the third sample set. Since the correlation analysis methods for numerical variables and categorical variables are different, all characteristic variables were first classified according to numerical variables and categorical variables, and then correlation analysis was performed, as follows:

[0096] Numerical variables were tested for normality (KS test). Spearman correlation analysis was performed for non-normally distributed numerical variables, and Pearson correlation analysis was performed for normally distributed numerical variables. If the data met the requirements, the chi-square test was used to assess the association between variables. If the requirements were not met, Fisher's exact test should be used to ensure the accuracy of the results. In this application, the data met the requirements, and the chi-square test was performed for categorical variables.

[0097] After Spearman correlation analysis and chi-square test, the probability that the following numerical and categorical variables were closely related to disease stage was 99.9%, that is, P value < 0.001:

[0098] Numerical variables included age, nodule size, carbohydrate antigen 19-9 (CA19-9), progastrin-releasing peptide (ProGRP), indirect bilirubin (IBIL), white blood cell count (WBC), albumin (ALB), monocyte count (Mono#), neutrophil count (Neut#), neutrophil percentage (Neut%), neuron-specific enolase (NSE), platelet count (PCT), platelet count (PLT), red blood cell volume distribution width coefficient of variation (RDWCV), red blood cell volume distribution width (RDW), squamous cell carcinoma antigen (SCC), total bilirubin (TBIL), albumin / globulin ratio (Alb / Glb), γ-glutamyl transpeptidase (γ-GT), globulin (Glb), estimated glomerular filtration rate (eGFR), alkaline phosphatase (ALP), lactate dehydrogenase (LDH), lymphocyte percentage (Lym%), and carcinoembryonic antigen (CEA).

[0099] Categorical variables included sex, smoking status, alcohol drinking status, nodule distribution, nodule density, levels of autoantibodies related to cancer antigens (7-AAB), autoantibody CAGE, autoantibody GAGE ​​7, autoantibody MAGE A1, autoantibody PGP9.5, and autoantibody SOX2.

[0100] S14: After step S13, the number of numerical variables is still large. Principal component analysis can be used to reduce the dimensionality of the numerical variables in the third sample set to obtain the fourth sample set. Before performing principal component analysis, it is necessary to first check whether the data is suitable for principal component analysis. The main indicators include Kaiser-Meyer-Olkin (KMO) sampling adequacy measurement and Bartlett's sphericity test. If the test results KMO>0.5 and Bartlett's sphericity test significance level <0.05, principal component analysis can be performed.

[0101] The principal component analysis method is specifically as follows: the original numerical variables are concentrated to obtain eigenvalues ​​and corresponding eigenvectors, and eigenvectors are selected as principal components according to the eigenvalues ​​from large to small until the total variance explanation rate of the eigenvector is greater than the preset variance explanation rate threshold (75%). The principal components finally selected in this embodiment and their corresponding relationship with the original numerical variables are shown in Table 2.

[0102] Table 2 Principal components and their corresponding relationships with original numerical variables

[0103]

[0104] After steps S11-S14, the sample size of subjects was reduced from 3075 to 2956, and the characteristic variables were reduced from 73 to 21, including 11 categorical variables (gender, smoking, drinking, nodule type, multiple nodules, 7-AAB, CAGE, GAGE7, MAGE A1, PGP9.5, SOX2), and 10 principal components obtained after principal component analysis. The fourth sample set was constructed by the above 2956 subject samples and the corresponding 21 characteristic variables; the process of data dimensionality reduction in steps S13 and S14 is as follows: Figure 4 shown.

[0105] S2: If the amount of data for patients with different TNM stages in the sample set is different, it may affect the model construction in step S3. In this step, the sample uniformity of the fourth sample set is evaluated. If the sample uniformity is insufficient, the data is oversampled using the SMOTE method. At the same time, a uniform fifth sample set is obtained by undersampling. Specifically:

[0106] S21. This example utilizes the nearest neighbor distance method and incorporates the concept of anchor points to establish a method for evaluating the uniformity of discrete data. Lung cancer classification is used as the Y-axis (small cell carcinoma, adenocarcinoma, squamous cell carcinoma, and large cell carcinoma, in that order) and lung cancer stage is used as the X-axis (stage 0, stage 1, stage 2, stage 3, and stage 4, in that order). Each subject sample is represented as a discretely distributed point in a two-dimensional space. Because the distribution of these points is discrete and the number of points is limited, the concept of anchor points is introduced. Anchor points are defined as points that belong to the two-dimensional space and are not on the boundaries of the space. Boundary points are defined as points that have at most three adjacent points in different locations. The selection of anchor points is random, and their number can be adjusted based on actual conditions.

[0107] The fourth sample set was evaluated for sample uniformity, and 2-5 anchor points were selected for experiment to determine whether the error value;

[0108]

[0109] in is the number of anchor points, is the actual number of nearest neighbors, is the theoretical nearest neighbor number.

[0110] Test findings error >0.2, which means that the number of nearest neighbors of each anchor point is quite different from the theoretical number of nearest neighbors. At the same time, the differences between different tests are also large, indicating that the distribution of samples is indeed obviously uneven and the uniformity of samples needs to be enhanced.

[0111] S22. Undersampling and oversampling are performed as long as the subject samples meet the requirements of step S21. In this embodiment, to simplify calculations, we directly calculate the average number of subject samples corresponding to different TNM stages, and then undersample subject samples above the average (stage 0 lung cancer subject samples and stage 1 lung cancer subject samples). We also oversample subject samples below the average using the SMOTE method (stage 2-4 subject samples), thereby obtaining the fifth sample set.

[0112] SMOTE (Synthetic Minority Oversampling Technique) is a synthetic sampling technique for minority classes. It is an improvement on the random oversampling algorithm. The basic idea is to analyze samples from some classes and, based on the analysis results, add new samples to the features of the minority classes whose number is less than the average in the dataset. The main process of SMOTE is as follows:

[0113] a. For each minority sample i (i=0,1,2,3...), sort by Euclidean distance from nearest to farthest and select the first k nearest neighbors (k is 1% of the sample size, rounded up);

[0114] b. Randomly select a sample J from k samples;

[0115] c. Create composite samples in the feature space according to H=p*I+(1-p)*J to fill the sample set, where p is a randomly generated number between 0 and 1;

[0116] S3: Using the fifth sample set, we perform machine learning based on random forests to obtain a prediction model for early diagnosis of lung cancer:

[0117] Specifically, the fifth sample set was divided into 80% training and 20% validation. A random forest algorithm was used to train samples from the training set to generate an early prediction model for lung cancer diagnosis. The random forest algorithm used 120 decision trees with a minimum leaf size of 4. The trees were classified with replacement sampling, and 5 samples were randomly selected at each node during the split. Furthermore, 70 samples from each lung cancer stage were extracted from the original sample set (i.e., the samples from the second sample set constructed in step S12), for a total of 420 samples. 30% of these samples were randomly selected as a test set to verify the suitability of the trained model for the original samples.

[0118] S4: The model was validated by random forest and evaluated by ROC curve, AUC index and accuracy;

[0119] Receiver operating characteristic (ROC) curve analysis was performed to determine the diagnostic value of different indicators for predicting related tumors. The area under the ROC curve (AUC) was ideally 1. An AUC > 0.9 indicated a high degree of accuracy, while an AUC between 0.7 and 0.9 indicated a moderate degree of accuracy. An AUC between 0.5 and 0.7 indicated a low degree of accuracy, and an AUC < 0.5 indicated a valueless indicator.

[0120] like Figure 5 As shown in the figure, the AUC index of the established random forest model round test is greater than 0.9, indicating that the model has a high accuracy in predicting whether a patient is likely to have the disease.

[0121] Comparative Example

[0122] Example 1 is repeated with the same steps, except that step S11 directly proceeds to step S3, omitting the intermediate steps; the samples and parameters used in the steps that are not omitted are the same.

[0123] Verification Example

[0124] Accuracy is used to assess the degree of agreement between a patient's predicted stage and actual stage. The lung cancer staging diagnostic results for the original sample set were compared with the model prediction results from Example 1. The model achieved an accuracy rate of over 85% for all stages of the disease, with an overall accuracy rate exceeding 90%, demonstrating excellent predictive performance.

[0125] Table 3 Random Forest Model Accuracy

[0126]

[0127] Figure 6 This figure compares the accuracy of Example 1 (blue) and the comparative example (purple). The horizontal axis represents the number of experiments, and the vertical axis represents the accuracy of the random forest test set. As can be seen from the figure, the accuracy of the random forest test set improved by approximately 25% after processing the samples using the example method, demonstrating that the example method can effectively improve model performance.

[0128] This application provides an early prediction model and construction method for lung cancer diagnosis based on data mining. It collects 73 features such as basic clinical information of hospitalized patients, lung cancer diagnosis stage, CT results, antibody test results, liver and kidney function indicators and blood routine results. It uses correlation analysis and principal component analysis to reduce the dimension to 21 features, and performs sample uniformity assessment and data expansion. A machine model based on data mining is established with good predictive performance. The model predicts whether the subject is likely to have lung cancer based on the output value and diagnoses the stage of the patient's disease, which can accurately and objectively achieve early detection of lung cancer. The prediction model constructed in this application provides an important tool for the early diagnosis of lung cancer.

[0129] This example uses data from 3,075 lung disease patients collected from Tongji Hospital, Tongji Medical College, and Huazhong University of Science and Technology between January 2021 and June 2022, including 2,409 lung cancer patients and 666 benign lung nodules. This large sample size covers different types of lung cancer (such as small cell lung cancer, non-small cell lung cancer, and its subtypes such as adenocarcinoma, squamous cell carcinoma, and large cell lung cancer) and the corresponding collected feature variables, ensuring that the model training and validation process can fully cover all possible disease conditions. It also explores the relationship between parameters such as patient CT scan results, liver and kidney function test results, blood test results, antibody indicators, and the TNM stage of lung cancer. The samples used greatly enrich the diversity and representativeness of the sample set, allowing the model to learn more subtle differences and features, thereby improving the diagnostic accuracy of lung cancer and its subtypes. In addition, the large sample size helps reduce statistical errors, making the model results more stable and reliable.

[0130] It can be seen from the above examples that the present application has demonstrated excellent efficacy and potential value in the early diagnosis of lung cancer through large-scale and comprehensive clinical data collection, comprehensive and scientific data preprocessing procedures, innovative sample evaluation and homogenization strategies, and carefully constructed prediction models and rigorous verification; the results of the present application are expected to play an important role in clinical practice, providing strong support for the early detection, accurate classification and formulation of personalized treatment plans for lung cancer, thereby improving the treatment effect and survival rate of patients, and having far-reaching significance for improving the prognosis of lung cancer patients; in addition, the methods and strategies of the present application also provide a paradigm that can be used as a reference for the diagnosis and prediction research of other complex diseases, demonstrating its broad prospects in the fields of medical big data analysis and machine learning applications.

[0131] It is easy for those skilled in the art to understand that the above is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A method for constructing a lung cancer staging diagnosis model, characterized in that: The following steps are involved: S1: Obtain lung cancer classification and TNM stage of subject samples; wherein, the number of subject samples corresponding to each TNM stage is 2 or more; S2: constructing a two-dimensional space based on the lung cancer classification and TNM stage of the subject samples, and integrating and correcting the subject samples in different TNM stages using an anchoring method so that the subject samples are evenly distributed in the two-dimensional space; S3: The subject samples uniformly distributed in the previous step and the feature variables corresponding to the subject samples are combined into a sample set, and a random forest algorithm is used to perform machine learning on the sample set to construct a lung cancer staging diagnosis model, wherein the parameters of the random forest algorithm include the number of decision trees, the minimum number of leaf nodes, replacement sampling, and the number of samples randomly selected during node splitting; Wherein, the step S2 specifically includes the following sub-steps: S21: Construct a two-dimensional space based on the lung cancer classification and TNM stage of the subject sample, randomly select 2-5 anchor points in the two-dimensional space, and use the nearest neighbor distance method to determine their error , error If the value is less than the threshold, then go to step S3, otherwise go to the next step; in, is the number of anchor points, is the actual number of nearest neighbors, is the theoretical nearest neighbor number; S22: According to the number of subject samples in different TNM stages, oversampling and / or undersampling are performed on the corresponding subject samples so that the uniformity of the subject samples meets the requirements, and then proceed to the next step; the oversampling method is the SMOTE method.

2. The construction method according to claim 1, wherein The characteristic variables are related to TNM staging, including categorical variables and numerical variables; the categorical variables include age, gender, smoking, drinking, nodule density, nodule distribution and antibody test results; the antibody test results include GAGE7, MAGE A1 and GBU4-5.

3. The construction method according to claim 2, wherein: The numerical variables are constructed from original numerical variables through principal component analysis; the original numerical variables include nodule size, tumor marker levels, liver and kidney function test results, blood routine test results and electrolyte test results; the tumor marker levels include CEA and CA19-9.

4. The construction method according to claim 1, wherein After step S3, the method further includes step S4: evaluating the lung cancer staging diagnosis model using the ROC curve and the AUC index. If the model does not meet the evaluation criteria, the method returns to step S3.

5. A lung cancer staging diagnostic model obtained by the construction method according to any one of claims 1 to 4.

6. A system for constructing a lung cancer staging diagnosis model, which is applied to the construction method according to any one of claims 1 to 4, characterized in that: include: a data acquisition module for obtaining subject samples; A data processing module for performing integrated correction on the subject sample; A model building module used to construct a lung cancer staging diagnosis model using subject samples to form a sample set and use the random forest algorithm for machine learning.

Images