Lung cancer differential diagnosis system for serum cancer marker negative cases

Abstract

The invention provides a lung cancer differential diagnosis system for serum tumor marker negative cases. Specifically, based on a machine learning method, a differential diagnosis system is constructed by using six characteristic indexes. The result shows that the AUC of the identification and diagnosis system based on the support vector machine in the training set and the verification set is larger than 0.8, and whether the serum tumor marker negative patient suffers from the lung cancer or not can be accurately identified and diagnosed. According to the system, identification and diagnosis can be completed by using a small amount of characteristic indexes, the examination decision time of related patients is greatly shortened, and the system has a wide clinical application prospect.

Description

Technical Field

[0001] This invention relates to the fields of machine learning and disease differential diagnosis, and more specifically, to a method and system for constructing a machine learning-based differential diagnosis model for lung cancer in cases with negative serum tumor markers. Background Technology

[0002] Lung cancer is a highly prevalent malignant tumor worldwide, ranking first in both incidence and mortality among malignant tumors in China, and is a leading cause of death globally. In 2022, its incidence rate in my country continued to rise, accounting for 18.06% of all malignant tumors, with approximately 733,300 deaths, representing 28.5% of all malignant tumor deaths. More than 1,060,600 new cases of lung cancer are diagnosed annually, making it the leading cause of malignant tumor incidence. Due to the lack of specific early symptoms, up to 75% of patients with early-stage lung cancer already have regional lymph node metastasis or distant spread at diagnosis, and approximately 70% of patients have progressed to advanced stages at diagnosis. The prognosis for patients with advanced lung cancer is significantly worse, with a 5-year survival rate of less than 20%, a stark contrast to the 70%-85% cure rate for early-stage patients.

[0003] Currently, early lung cancer screening mainly relies on imaging examinations (low-dose CT (LDCT) and positron emission tomography / computed tomography (PET-CT)) and invasive tests (tumor markers). Among these, LDCT screening can reduce lung cancer mortality by 31%, but its specificity in differentiating early atypical lung cancer lesions is insufficient, often leading to overdiagnosis. PET-CT can further improve diagnostic sensitivity, but promoting this screening method among low-income groups still faces significant challenges. In addition, LDCT and PET-CT imaging also pose considerable challenges in differentiating diseases with mediastinal / hilar lymphadenopathy as the main imaging manifestation (lymph node tuberculosis, lymphoma, sarcoidosis, primary lung cancer, and metastatic lymph node lesions, etc.). The main reason is that early cases of these diseases do not show specific differentiation in clinical and imaging characteristics, resulting in a misdiagnosis rate as high as 30%-40%. Tumor markers (such as CEA) have an auxiliary role in the diagnosis and disease monitoring of lung cancer; however, their application value in early-stage lung cancer is limited, and their sensitivity is significantly reduced (35%-40%).

[0004] Existing studies have attempted to combine multi-omics data for identification and diagnosis, but the models are too complex and difficult to translate into clinical practice. In addition, tumor biomarkers have insufficient sensitivity for the diagnosis of early lung cancer (CEA only 20-50%), and about 30% of patients have all biomarkers that remain normal.

[0005] Therefore, there is an urgent need in this field for a model or system with good specificity and high sensitivity, which can be used to identify and diagnose lung disease patients whose multiple common cancer markers are within the normal range, and a method for constructing such a system. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for constructing a differential diagnostic model for lung cancer in cases with negative serum tumor markers based on machine learning.

[0007] In a first aspect, the present invention provides a method for constructing a differential diagnostic model for lung cancer, the method comprising the steps of:

[0008] (S1) Provide a clinical dataset of disease patients; the disease patients are those with atypical imaging findings and negative tumor markers, including lung cancer patients and non-lung cancer patients; divide the clinical dataset into a training set and a validation set;

[0009] (S2) The clinical dataset is preprocessed to obtain a preprocessed clinical dataset;

[0010] (S3) Extract feature indicators from the preprocessed clinical dataset and filter the feature indicators to obtain preferred feature indicators;

[0011] (S4) The preferred feature indicators in the training set are used to train the machine learning model; when the training result reaches the predetermined termination condition, the training of the model is terminated, thereby obtaining the lung cancer differential diagnosis model.

[0012] In another preferred embodiment, the clinical dataset includes: imaging feature data and peripheral blood test data.

[0013] In another preferred embodiment, the imaging features include: the number of infiltrated lung fields and the extent of hilar / mediastinal lymph node enlargement.

[0014] In another preferred embodiment, "atypical imaging features" refers to imaging features that do not meet the typical diagnostic criteria for the disease, and / or imaging features that make it difficult to make a clear diagnosis based on them.

[0015] In another preferred embodiment, the tumor markers include: alpha-fetoprotein (AFP), carcinoembryonic antigen (CEA), cytokeratin 19 fragment (CYFRA21-1), β2-microglobulin (β2), carbohydrate antigen 153 (CA153), neuron-specific enolase (NSE), carbohydrate antigen 242 (CA242), carbohydrate antigen 50 (CA50), carbohydrate antigen 199 (CA199), carbohydrate antigen 724 (CA724), and squamous cell carcinoma antigen (SCCA).

[0016] In another preferred embodiment, the lung cancer patient includes: adenocarcinoma patient, squamous cell carcinoma patient, small cell lung cancer patient, or a combination thereof.

[0017] In another preferred embodiment, the non-lung cancer patients include: patients with pulmonary tuberculosis, patients with pulmonary sarcoidosis, patients with pneumoconiosis, and patients with pulmonary lymphoma;

[0018] In another preferred embodiment, the clinical dataset also includes the age, sex, and clinical symptoms of the patients with the disease.

[0019] In another preferred embodiment, the preprocessing includes: handling missing values, handling outliers, transforming data, and standardizing the clinical dataset.

[0020] In another preferred embodiment, the preprocessing further includes: using propensity score matching (PSM) to eliminate baseline feature differences between the lung cancer patients and the non-lung cancer patients.

[0021] In another preferred embodiment, step (S3) specifically includes the following steps:

[0022] (S3.1) Perform correlation analysis on each indicator in the preprocessed clinical dataset and remove redundant indicators with r > 0.8;

[0023] (S3.2) Perform univariate analysis on the indicators obtained in step (S3.1) and retain the indicators with P < 0.05;

[0024] (S3.3) Further filter the indicators obtained in step (S3.2), including:

[0025] Perform LASSO regression analysis on the indicators obtained in step (S3.2) to obtain the candidate indicator set A;

[0026] Boruta regression analysis was performed on the indicators obtained in step (S3.2) to obtain the candidate indicator set B;

[0027] (S3.4) Take the intersection of the candidate index set A and the candidate index set B to obtain the preferred feature index.

[0028] In another preferred embodiment, the optimal penalty coefficient is determined using 10-fold cross-validation in the LASSO regression analysis.

[0029] In another preferred embodiment, in step (S4), the training set is tuned using K-fold cross-validation.

[0030] In another preferred embodiment, in step (S4), the training set is tuned using 10-fold cross-validation.

[0031] In another preferred embodiment, the preferred characteristic indicators are carcinoembryonic antigen (CEA), monocyte percentage (MON%), peripheral blood lymphocyte count (Lym#), mean corpuscular hemoglobin (MCH), red blood cell distribution width standard deviation (RDW-SD), and albumin / globulin ratio (ALB / GLB).

[0032] In another preferred embodiment, the machine learning model includes: Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Naive Bayes (NBM), Logistic Regression (Logistics), and K-Nearest Neighbors (KNN).

[0033] In another preferred embodiment, the machine learning model is selected from the group consisting of: Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Naive Bayes (NBM), Logistic Regression, or K-Nearest Neighbors (KNN).

[0034] In another preferred embodiment, the machine learning model is selected from the group consisting of: Support Vector Machine (SVM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost).

[0035] In another preferred embodiment, the machine learning model is a support vector machine (SVM).

[0036] In another preferred embodiment, the method further includes the step of: (S5) validating the lung cancer differential diagnosis model in the validation set.

[0037] A second aspect of the present invention provides a lung cancer differential diagnosis system, the lung cancer differential diagnosis system comprising:

[0038] An input unit configured to input data, the data including clinical data of the subject being tested;

[0039] A differential diagnosis unit is configured as a differential diagnosis model to obtain a diagnostic result for the object to be tested; wherein the differential diagnosis model is constructed using the method described in the first aspect of the present invention;

[0040] An output unit is configured to output the diagnostic results of the differential diagnosis unit.

[0041] A third aspect of the present invention provides an electronic device, including a processor and a memory, the memory having a plurality of executable instructions, the processor being configured to read the instructions and perform the following steps:

[0042] (1) Provide clinical data of a subject to be tested; the clinical data includes characteristic index values;

[0043] (2) Input the clinical data into the diagnostic model, and the diagnostic model predicts whether the subject to be tested is a lung cancer patient based on the clinical data, thereby obtaining a risk probability value;

[0044] The characteristic indicators include: carcinoembryonic antigen (CEA), monocyte percentage (MON%), peripheral blood lymphocyte count (Lym#), mean corpuscular hemoglobin (MCH), red blood cell distribution width standard deviation (RDW-SD), and albumin / globulin ratio (ALB / GLB).

[0045] A fourth aspect of the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when read and executed by a processor, perform the following steps:

[0046] (1) Provide clinical data of a subject to be tested; the clinical data includes characteristic index values;

[0047] (2) Input the clinical data into the diagnostic model, and the diagnostic model predicts whether the subject to be tested is a lung cancer patient based on the clinical data, thereby obtaining a risk probability value;

[0048] The characteristic indicators include: carcinoembryonic antigen (CEA), monocyte percentage (MON%), lymphocyte count (Lym#), mean corpuscular hemoglobin (MCH), red blood cell distribution width standard deviation (RDW-SD), and albumin / globulin ratio (ALB / GLB).

[0049] A fifth aspect of the present invention provides a computer program product comprising computer-executable instructions, which, when executed by a processor, perform the following steps:

[0050] (1) Provide clinical data of a subject to be tested; the clinical data includes characteristic index values;

[0051] (2) Input the clinical data into the diagnostic model, and the diagnostic model predicts whether the subject to be tested is a lung cancer patient based on the clinical data, thereby obtaining a risk probability value;

[0052] The characteristic indicators include: carcinoembryonic antigen (CEA), monocyte percentage (MON%), lymphocyte count (Lym#), mean corpuscular hemoglobin (MCH), red blood cell distribution width standard deviation (RDW-SD), and albumin / globulin ratio (ALB / GLB).

[0053] It should be understood that, within the scope of this invention, the above-described technical features of this invention and the technical features specifically described below (such as in the embodiments) can be combined with each other to form new or preferred technical solutions. Due to space limitations, they will not be described in detail here. Attached Figure Description

[0054] Figure 1 The feature selection process and results are shown. a: Correlation heatmap between variables; b: LASSO regression variable selection plot; c: Boruta analysis plot; d: Venn diagram of feature variables for LASSO and Boruta analyses.

[0055] Figure 2 The visualization analysis of the model is shown (SHAP value visualization). a: Waterfall plot; b: Peak group plot; c: Scatter plot; d: Bar chart.

[0056] Figure 3 This shows a comparison of the ROC curves of Logistic Regression (Logistics), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Naive Bayes (NBM) models on the training set. Sensitivity is the sensitivity level.

[0057] Figure 4 The decision curve analysis of each machine learning model on the training set is shown. ProbabilityThreshold: probability threshold; Logistics: logistic regression; SVM: support vector machine; KNN: K-nearest neighbor; random: random forest; NBM: Naive Bayes.

[0058] Figure 5 The comparison of ROC curves for Logistic Regression (Logistics), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Naive Bayes (NBM) models on the validation set is shown. Sensitivity is the sensitivity level.

[0059] Figure 6 The decision curve analysis of each machine learning model on the validation set is shown. ProbabilityThreshold: probability threshold; Logistics: logistic regression; SVM: support vector machine; KNN: K-nearest neighbor; random: random forest; NBM: Naive Bayes. Detailed Implementation

[0060] Through extensive and in-depth research, the inventors extracted feature indicators from clinical characteristics, laboratory tests, and imaging data, and for the first time constructed a machine learning-based differential diagnostic model and system for lung cancer. Specifically, this invention uses six previously screened feature indicators—MON%, Lym#, MCH, RDW-SD, ALB / GLB, and CEA—and utilizes six machine learning models to construct diagnostic models. Among them, the support vector machine-based diagnostic model has an AUC greater than 0.8 on both the training and validation sets, demonstrating good discriminative power and accurately identifying and diagnosing whether patients with negative serum tumor markers have lung cancer. The model and system of this invention can complete identification and diagnosis using only a small number of feature indicators, significantly shortening the time required for relevant patient examination decisions. Based on this, this invention was completed.

[0061] It should be understood that the specific methods and experimental conditions of the invention described below in varying degrees of detail are intended to provide a substantive understanding of the invention. Definitions of certain terms used in this specification are provided below. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0062] the term

[0063] As used herein, the terms “containing” or “including (comprise)” can be open-ended, semi-closed, or closed-ended. In other words, the terms also include “consistently made of” or “made of”.

[0064] As used herein, the term “and / or” refers to and covers any and all possible combinations of one or more of the related listed items.

[0065] As used in this article, the term "significant" means that, in a hypothesis test, the observed effect (such as the difference between the experimental and control groups) is unlikely to be caused solely by random error. A hypothesis test includes: the null hypothesis (H0), which assumes that the observed effect does not exist (such as no difference between the experimental and control groups); the p-value, which is the probability of observing the current or more extreme effect when H0 is true; and the significance threshold (α). The significance threshold is typically used to determine whether a hypothesis test is significant. Generally, the significance threshold is 0.05. If the p-value ≤ α, then H0 is rejected, meaning the observed effect exists, and the result is called "significant."

[0066] As used herein, the term "positive" means that the level / proportion / activity of the target substance exceeds the normal range recommended by the kit used for the test, thus indicating a risk or possibility of disease. Conversely, the term "negative" means that the level / proportion / activity of the target substance has not changed abnormally (within the normal range), thus indicating a low risk or possibility of disease. Generally, the target substance being tested is a tumor marker. A negative tumor marker result means that the detected tumor marker value does not exceed the normal range recommended by the kit for that tumor marker. When a tumor marker is detected, it indicates that the subject has an increased risk or possibility of having a tumor / cancer. In actual testing, "false positives" or "false negatives" may occur. A "false positive" means that there is no abnormality but a false alarm occurs; conversely, a "false negative" means that there is an abnormality but it is not detected. The diagnostic system provided by this invention has a significant effect on the identification and diagnosis of "false negative" patients, that is, it can accurately diagnose lung cancer patients with negative results for multiple tumor markers. Preferably, the tumor marker is selected from the group consisting of: alpha-fetoprotein (AFP), carcinoembryonic antigen (CEA), cytokeratin 19 fragment (CYFRA21-1), β2-microglobulin (β2, or β2-MG), carbohydrate antigen 153 (CA153), neuron-specific enolase (NSE), carbohydrate antigen 242 (CA242), carbohydrate antigen 50 (CA50), carbohydrate antigen 199 (CA199), carbohydrate antigen 724 (CA724), squamous cell carcinoma antigen (SCCA), or combinations thereof.

[0067] Early lung cancer screening primarily relies on imaging examinations (low-dose CT, positron emission tomography / computed tomography (PET / CT)) and invasive tests (tumor markers). This paper demonstrates that by detecting only six characteristic indicators, early-stage lung cancer cases among several common lung diseases with similar imaging features and negative tumor markers can be differentiated. This approach reduces diagnostic costs, decreases the complexity of diagnostic models, and avoids invasive testing.

[0068] The computer system is equipped with at least one processor and a memory. The processor invokes a sequence of computer-executable instructions stored in the memory to implement the evaluation process defined in the claims. Although the process describes the operation steps in a specific logical order, in actual execution, the steps may be processed in parallel, their order adjusted, or some may be omitted. As long as such adjustments do not deviate from the core features of the technical solution described in the claims and achieve the same technical effect, they all fall within the scope of protection of this invention. This flexibility in execution order is determined by the programmable nature of computer instructions.

[0069] Construction method of the present invention

[0070] This article provides a method for constructing a differential diagnostic model for lung cancer, the method comprising the following steps:

[0071] (S1) Provide a clinical dataset of disease patients; the disease patients are those with atypical imaging findings and negative cancer markers, including lung cancer patients and non-lung cancer patients; divide the clinical dataset into a training set and a validation set;

[0072] (S2) The clinical dataset is preprocessed to obtain a preprocessed clinical dataset;

[0073] (S3) Extract feature indicators from the preprocessed clinical dataset and filter the feature indicators to obtain preferred feature indicators;

[0074] (S4) The preferred feature indicators in the training set are used to train the machine learning model; when the training result reaches the predetermined termination condition, the training of the model is terminated, thereby obtaining the lung cancer differential diagnosis model.

[0075] Preferably, the clinical dataset includes: imaging feature data and peripheral blood test data.

[0076] The imaging features include: the number of infiltrated lung fields and the degree of hilar / mediastinal lymph node enlargement. The term "lung field" refers to the uniformly transparent area of ​​both lungs filled with air on a chest X-ray. When the normally air-containing areas of a lung field are partially filled with fluid, cells, tissue, or other substances, causing an increase in density and appearing as a blurred, cloud-like white shadow, this is called "lung infiltration." The number of infiltrated lung fields can be determined by a professional physician based on the patient's pathological images. In this invention, the number of infiltrated lung fields is 0-6. Similarly, the degree of hilar / mediastinal lymph node enlargement can also be determined by a professional physician based on the patient's pathological images. In this invention, the degree of hilar / mediastinal lymph node enlargement is divided into three categories: "no enlargement," "symmetrical enlargement," and "unilateral enlargement."

[0077] Those skilled in the art are familiar with various common peripheral blood test data. In this document, the peripheral blood test data include: hemoglobin (HGB), neutrophil percentage (NEU%), mean corpuscular hemoglobin concentration (MCHC), total serum protein (TP), red blood cells (RBC), white blood cells (WBC), lymphocyte percentage (Lym%), monocyte percentage (MON%), eosinophil percentage (EO%), basophil percentage (BASO%), neutrophil count (NEU#), lymphocyte count (Lym#), neutrophils (NeuLym), monocyte count (Mono#), monocyte / lymphocyte count (MonoLym), eosinophil count (EO#), basophil count (BASO#), hematocrit (HCT), mean corpuscular volume (MCV), mean corpuscular hemoglobin concentration (MCH), platelet count (PLT#), and platelet / lymphocyte ratio (PLT / Lym). Platelet-to-neutrophil ratio (PLT / Neu), red blood cell distribution width standard deviation (RDW-SD), red blood cell distribution width coefficient of variation (RDW-CV), platelet-to-hematocrit (PCT), platelet distribution width (PDW), large platelet ratio (PLCR), mean platelet volume (MPV), systemic immune inflammatory index (SII), systemic inflammatory response index (SIRI), serum albumin (ALB), prognostic nutritional index (PNI), albumin / globulin ratio (ALB / GLB), random blood glucose (GLU), serum calcium (CA), serum potassium (K), prothrombin time (PTs), international normalized ratio (INR-S), fibrinogen (FIBs), activated partial thromboplastin time (APTTS), thrombin time (TTs), antithrombin III (AT3S), D-dimer-S (Ddimer-S), fibrin degradation products (FDPS), etc.

[0078] In this invention, the disease patient is one with atypical imaging findings and negative tumor markers. The term "atypical imaging findings" refers to imaging characteristics that do not meet the typical diagnostic criteria for the disease, and / or imaging findings that make it difficult to make a definitive diagnosis. Generally, the diagnosis is made by a professional physician based on the imaging images. Preferably, the cancer markers include: alpha-fetoprotein (AFP), carcinoembryonic antigen (CEA), cytokeratin 19 fragment (CYFRA21-1), β2-microglobulin (β2), carbohydrate antigen 153 (CA153), neuron-specific enolase (NSE), carbohydrate antigen 242 (CA242), carbohydrate antigen 50 (CA50), carbohydrate antigen 199 (CA199), carbohydrate antigen 724 (CA724), and squamous cell carcinoma antigen (SCCA).

[0079] Preferably, the preprocessing further includes: using propensity score matching (PSM) to eliminate baseline feature differences between the lung cancer patients and the non-lung cancer patients.

[0080] Preferably, step (S3) specifically includes the following steps: (S3.1) performing correlation analysis on each indicator in the preprocessed clinical dataset and removing redundant indicators with r > 0.8; (S3.2) performing univariate analysis on the indicators obtained in step (S3.1) and retaining indicators with P < 0.05; (S3.3) further screening the indicators obtained in step (S3.2), including: performing LASSO regression analysis on the indicators obtained in step (S3.2) to obtain candidate indicator set A; performing Boruta regression analysis on the indicators obtained in step (S3.2) to obtain candidate indicator set B; (S3.4) taking the intersection of the candidate indicator set A and the candidate indicator set B to obtain the preferred feature indicators.

[0081] Preferably, this paper uses six machine learning models for model training. The machine learning models include: logistic regression, support vector machine (SVM), Naive Bayes (NB), K-nearest neighbors (KNN), extreme gradient boosting (XGBoost), and random forest (RF).

[0082] The term "logistic regression" is a classic statistical model that uses the sigmoid function to map the linear regression results to the (0,1) interval, establishing a non-linear relationship between features and target probabilities. Its core lies in constructing a linear decision boundary: Z = w1x1 + w2x2 + ... + w n x n +b; through the sigmoid function σ(z) = 1 / (1+e -z To achieve probability transformation, the maximum likelihood estimation is used to optimize the weight parameter w.

[0083] The term "Support Vector Machine" is a binary classification model. Its basic idea is to find a hyperplane that separates data points of different classes while maximizing the distance (margin) between the data points of each class to the hyperplane. SVMs can handle not only linearly separable problems but also non-linearly separable problems through kernel functions.

[0084] The term "Naive Bayes," also known as the Naive Bayes classifier, is based on Bayes' theorem and the assumption of conditional independence of features. Bayes' theorem is an important theorem in probability theory that describes the probability of an event occurring given certain conditions. The calculation formula is:

[0085] The term "K-Nearest Neighbors" is a basic instance-based supervised learning algorithm that can be used for both classification and regression tasks. Its core idea is "like attracts like"—similar feature vectors tend to have similar output values, and the prediction result is determined by "voting" or "averaging" the neighboring samples.

[0086] The term "extreme gradient boosting" can be used interchangeably with "XGBoost," and it's a highly efficient implementation of the Gradient Boosting Decision Tree (GBDT) algorithm. It incorporates several optimizations and improvements over traditional GBDT, using an additive model to sequentially construct multiple weak learners (usually decision trees) and combine them into a strong predictor. Each new tree aims to correct the residual error of the previous tree. The calculation formula is: Among them, f k Let be the predicted output of the Kth tree, the loss function, L: measure the prediction error (such as mean squared error, cross entropy), Ω: control the complexity of a single tree.

[0087] The term "random forest" is an ensemble learning method based on Bagging (Bootstrap Aggregating) and Random Subspace, which improves the robustness of the model by constructing multiple decision trees and combining their predictions.

[0088] Preferably, the machine learning model is a support vector machine.

[0089] The main advantages of this invention include:

[0090] (1) This invention extracts feature indicators from data such as clinical characteristics, laboratory tests, and imaging tests, thereby constructing a lung cancer differential diagnosis system based on machine learning. The system is based on a support vector machine model, and the AUC of both the training set and the validation set is greater than 0.8, which has good discriminative power and can better identify and diagnose the risk probability value of lung cancer in patients with negative serum cancer markers.

[0091] (2) Through feature extraction and screening, this invention discovered six feature indicators that significantly contribute to model performance: MON%, Lym#, MCH, RDW-SD, ALB / GLB, and CEA. The model constructed using these feature indicators exhibits superior performance, overcoming the limitations of single-indicator diagnostic efficacy.

[0092] (3) The system of the present invention can complete identification and diagnosis using a small number of feature indicators, which greatly shortens the time for relevant patients to make examination decisions.

[0093] The present invention will be further illustrated below with reference to specific embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. It is also understood that the purpose of describing the present invention in conjunction with the embodiments is to cover other options or modifications that may be derived based on the claims of the present invention. To provide a thorough understanding of the invention, many specific details will be included in the following description. The invention may also be practiced without using these details. Furthermore, some specific details will be omitted in the description to avoid confusion or obscuring the focus of the invention. Experimental methods in the following embodiments that do not specify specific conditions are generally performed under conventional conditions, such as those described in Sambrook et al., Molecular Cloning: A Laboratory Manual (New York: Cold Spring Harbor Laboratory Press, 1989), or according to the manufacturer's recommendations. Unless otherwise stated, percentages and parts are weight percentages and parts by weight.

[0094] Data sources and methods

[0095] 1. Data Sources and Arrangement:

[0096] This is a retrospective observational study. Data were obtained from the electronic medical record system of Shanghai Pulmonary Hospital affiliated to Tongji University. The study included patients with early-stage lung cancer and non-lung cancer (tuberculosis, sarcoidosis, pneumoconiosis, lymphoma) who were admitted for the first time between January 2012 and January 2024 and whose imaging manifestations were atypical and difficult to diagnose. The following indicators were independently assessed by two physicians at the deputy director level or above from the tuberculosis department and the respiratory department: difficulty of disease diagnosis (based on clinical manifestations, imaging features and pathological results), hilar / mediastinal lymph node enlargement (assessed by CT or PET-CT), and number of infiltrated lung fields.

[0097] The patient's demographic characteristics (age and gender), clinical diagnosis, pathological diagnosis, and first peripheral blood test data upon initial admission were entered into the hospital's electronic medical record system. This included complete blood count (white blood cells, neutrophils, lymphocytes, etc.), biochemical indicators (TP, LDL-C, etc.), inflammatory indicators (erythrocyte sedimentation rate, C-reactive protein, etc.), coagulation (PT, APTT, D-dimer, etc.), and tumor markers (CEA, CYFRA21-1, β2-MG, CA153, NSE, CA242, CA50, CA199, CA724, SCC).

[0098] The criteria for determining the difficulty of disease diagnosis and the inclusion criteria for cases in this study are as follows:

[0099] According to pathology, the staging of mediastinal lymph node tuberculosis can be divided into four stages. In pathological stages 1 and 2, the CT manifestations of mediastinal lymph node tuberculosis are relatively atypical, which may make it difficult to distinguish from the mediastinal lymph node enlargement in sarcoidosis. At the same time, extensively calcified lymph nodes are considered as easy to distinguish, while partially calcified lymph nodes are considered as difficult to distinguish.

[0100] Sarcoidosis can be divided into 5 stages according to the Scadding staging system. In stages 0, I, and II, the CT and imaging manifestations are relatively atypical. At the same time, lymph node calcification presents as amorphous or punctate calcification (partial calcification), and hilar / mediastinal lymph node enlargement occurs.

[0101] Pneumoconiosis is difficult to diagnose and differentiate from other diseases based solely on imaging findings.

[0102] Lung cancer cases are characterized by multiple invasive lesions in the lungs and involvement of multiple lymph nodes, and are morphologically similar to sarcoidosis cases.

[0103] Lymphoma (especially Hodgkin's lymphoma (HL) and non-Hodgkin's lymphoma (NHL)) presents with enlargement and fusion of hilar / mediastinal lymph nodes; intrapulmonary lymphatic infiltration. On enhanced CT, lymphoma appears as mild to moderate homogeneous enhancement, similar to proliferative tuberculous lymphadenopathy and sarcoidosis.

[0104] Inclusion criteria: newly diagnosed cases with no history of related diseases (such as tuberculosis or tumors) or treatment (hormones, antibiotics not used within the past two weeks, etc.) before admission. The initial imaging findings were atypical (without a clear diagnosis), and the peripheral blood test results were complete at the first admission. All cancer markers were negative, and the following criteria were met: (1) Lung cancer group: Pathological diagnosis was clear during the follow-up period; (2) Tuberculosis group: Mycobacterium tuberculosis smear was negative at admission, and molecular diagnosis (Mycobacterium tuberculosis DNA or RNA detection) was negative, but Mycobacterium tuberculosis culture and species identification were positive during the follow-up period, or smear was positive after at least 1 month of follow-up; (3) Lymphoma group: Cases with uniform enhancement and less fusion on imaging were selected, and the diagnosis was confirmed by histopathology during the follow-up period; (4) Sarcoidosis group: The diagnosis criteria for stage 0-II pulmonary sarcoidosis were met according to international consensus; (5) Pneumoconiosis group: There was a clear history of occupational exposure, but the imaging findings were atypical (other lung diseases need to be excluded), and the diagnosis of pneumoconiosis was confirmed during the follow-up period.

[0105] Exclusion criteria (those meeting any of the following conditions will be excluded): (1) Treatment interference: those who have received anti-tuberculosis, anti-tumor or hormone therapy or have a confirmed diagnosis before admission, or those who have used antibiotics other than cephalosporins within 2 weeks; (2) Typical imaging findings (such as lobulation sign, spiculation sign, etc.); (3) Any of the diseases in the differential diagnosis or other malignant tumors or active infections (such as HIV, HBV); (4) Missing data, incomplete key clinical indicators (such as blood routine, tumor marker test, imaging report).

[0106] 2. Model selection and evaluation:

[0107] This study selected six machine learning algorithms—Naive Bayes, Logistic Regression, Support Vector Machine, XGBoost, K-Nearest Neighbors, and Random Forest—to develop a differential diagnostic model for lung cancer among pulmonary diseases. The diagnostic model was built using data from various indicators obtained through infiltrative lung fields, hilar / mediastinal lymph node enlargement, clinical symptoms, age, sex, and peripheral blood tests. The dataset was randomly divided into a training set (model building) and a validation set (independent testing) in a 7:3 ratio. The model's discriminative ability was evaluated using the following metrics: Area Under the Receiver Operating Characteristic (AUC), Sensitivity, Specificity, Positive Predictive Value (PPV), Negative Predictive Value (NPV), Recall, F1 score, and Accuracy. Clinical applicability: Decision curve analysis (DCA) was used to assess the model's net benefit. Interpretability: Shapley Additive Explanations (SHAP) analysis was used to visualize the model and enhance its interpretability.

[0108] 3. Statistical analysis and feature selection:

[0109] For quantitative data (such as BMI and laboratory indicators) that conform to a normal distribution, the values ​​are expressed as mean ± standard deviation (Mean ± SD) and analyzed using an independent samples t-test. For non-normally distributed data, the values ​​are expressed as median (M) and interquartile range (IQR) and analyzed using the Mann-Whitney U test. Count data (such as gender and past medical history) are described as frequencies (percentages) and analyzed using a chi-square test (χ²) or Fisher's exact test (if the expected frequency < 5). Correlation analysis was used to remove highly correlated (r > 0.8) redundant variables to avoid multicollinearity. This study employed a two-stage feature selection strategy: the first stage used univariate analysis (P < 0.05) to initially screen potential predictive indicators; the second stage combined LASSO regression and the Boruta algorithm for feature selection, where LASSO regression used 10-fold cross-validation to determine the optimal penalty coefficient (λ) and retained variables with non-zero coefficients. Finally, the common variables from both methods were integrated as the feature set of the predictive model. Model evaluation and visualization employed ROC curves to assess the diagnostic efficacy of individual indicators and combined models (AUC comparisons used the DeLong test). Basic statistical analysis was performed using SPSS 26.0, and machine learning modeling, PSM analysis, and visualization were conducted using R 4.4.3 (RStudio). Data analysis, image and chart creation were performed using SPSS 26.0 and RStudio.

[0110] 4. Propensity Score Matching (PSM):

[0111] This is a retrospective study. There may be significant differences in baseline characteristics such as age, sex, number of infiltrated lung fields, and hilar / mediastinal lymph node enlargement between the case group (lung cancer) and the control group (sarcoidosis, tuberculosis, pneumoconiosis, lymphoma).

[0112] To minimize bias caused by these confounding factors, propensity score matching (PSM) was used to adjust for confounding factors. Multivariate logistic regression (covariates: age, sex) was used to estimate individual propensity scores, with the caliper width set to 0.05 times the standard deviation of the propensity score logit to avoid poor matching. Finally, a balance test was performed. After matching, the standardized mean difference (SMD) of each variable was calculated. If SMD < 0.2, the inter-group balance was considered acceptable. If the key confounding factors such as age and sex were balanced between the case and control groups (SMD < 0.1), selection bias was significantly reduced, ensuring model reliability.

[0113] Example 1: Characteristics of Research Participants

[0114] A total of 1143 cases of various lung diseases were collected for the study, and 440 cases were ultimately included based on the inclusion and exclusion criteria and PSM. Among them, there were 174 cases of lung cancer and 266 cases of control (48 cases of tuberculosis, 165 cases of sarcoidosis, 31 cases of lymphoma, and 22 cases of pneumoconiosis). The median age of the patients was 57.50 years (interquartile range [IQR] 14), and the male proportion was 43.6%. 72 patients (16.3%) smoked, and 22 patients (5.0%) drank alcohol. Among the lung cancer patients in the final analysis, there were 107 cases of non-small cell lung cancer (83 cases of adenocarcinoma, 20 cases of squamous cell carcinoma, and 4 cases without subtype), 7 cases of small cell lung cancer, and the remaining 60 cases were unclassified; of these, 71 cases were early stage (carcinoma in situ, invasive carcinoma, stage I, stage II), 51 cases were advanced stage (stage III, stage IV), and the remaining 52 cases were unstaged. The most common site of tumor infiltration was the upper lobe of the right lung, in 53 (30.5%) cases.

[0115] Due to the significant age difference between the two groups, age matching was performed. Baseline imbalance before PSM was performed based on age was mainly due to confounding factors such as sex, age, extent of lung enlargement, and number of infiltrated lung fields. Baseline balance was achieved after PSM. Univariate analysis was performed using independent samples t-test and Pearson chi-square test, and the results showed no statistically significant difference between sex and age after PSM (Table 1).

[0116] Table 1. Univariate analysis after PSM

[0117]

[0118]

[0119]

[0120]

[0121]

[0122] Example 2: Feature Extraction

[0123] For indicators statistically significant in univariate analysis, LASSO analysis was performed to reduce the dimensionality of the indicators, thereby enabling indicator selection and complexity adjustment. Simultaneously, Boruta regression analysis was conducted on the data to identify common variables from two features. The final variables included in the model are CEA, RDW-SD, Lym%, ALB / GLB, MON%, and MCH.

[0124] Figure 1 The diagram shows a heatmap of correlations among variables (a), a graph of LASSO regression variable selection (b), a Boruta analysis (c), and a Venn diagram of the characteristic variables from the LASSO and Boruta analyses (d). The LASSO and Boruta analyses had eight overlapping characteristic variables. These eight variables were ranked according to the absolute value of their coefficients (β values) in the LASSO regression results, and six characteristic variables were ultimately selected for model construction.

[0125] Example 3: SHAP Analysis and Interpretation

[0126] SHAP was used to analyze each feature. The SHAP visualization results are as follows: Figure 2 As shown.

[0127] Figure 2 The chart (a) shows a waterfall plot, providing representative prediction cases for early-stage lung cancer to illustrate the interpretability of the model. Yellow arrows indicate a decrease in predicted values, while purple arrows indicate an increase in predicted values. Figure 2 Figure b shows a peak cluster plot, illustrating the impact of multiple features and revealing the relationship between features and predicted values. Darker yellow indicates higher risk, while darker purple indicates lower risk. Furthermore, Figure 2 c presents a scatter plot showing partial dependencies among the six most influential variables to illustrate the relationships between them. Figure 2 d shows the average absolute SHAP value of each feature across all predictions, which is a measure of feature importance. The discriminative factors of the six extracted features are ranked, with carcinoembryonic antigen (CEA) being the most important factor.

[0128] Example 4: Model Construction

[0129] This embodiment involves training a model on a training set and optimizing the model through cross-validation.

[0130] The collected data undergoes data cleaning, including handling missing values, outlier handling, data transformation, and standardization; the standardization formula is as follows: Where min(X) is the minimum value of the data, and max(X) is the maximum value of the data.

[0131] All participants were divided into training and validation sets in a 7:3 ratio. Using six variables identified through LASSO and Boruta regression analyses, six machine learning algorithms—logistics, support vector machine (SVM), K-nearest neighbor (KNN), Naive Bayes (NB), random forest (RF), and gradient boosting (XGBoost)—were employed to construct and validate predictive models for the training set based on routine peripheral blood biometrics analysis of lung cancer and non-lung cancer (pneumoconiosis, sarcoidosis, tuberculosis, lymphoma) data. During training, 10-fold cross-validation was used to fine-tune the models.

[0132] In plotting the AUC curves of the features used in the final training of each model, it was found that SVM had the best predictive performance. The Support Vector Machine model showed the best predictive performance on the training set, with an AUC of 0.813. Table 2 lists a summary of the overall performance of each model; Figure 3 The ROC curves of each model on the training set are shown. The final DCA results indicate that when the threshold probability is between 0.10 and 0.08, all six models predict a net return greater than 0 for STME on the training set. Among the training set models, Random Forest and XGBoost provide the highest net returns for most threshold ranges. Figure 4 ).

[0133] Table 2. Overall performance of each model in the training set

[0134]

[0135] Example 5: Validation of Model Performance

[0136] This embodiment involves validating the model obtained in Example 4 on a validation set. The results show that the support vector machine model also performs well on the validation set, with an AUC of 0.827, and the specificity for discrimination and diagnosis on both the training and validation sets is higher than 80%. Table 3 lists a summary of the overall performance of each model on the validation set; Figure 5 The ROC curves of each model on the training set are shown.

[0137] DCA results show that when the threshold probability is between 0.10 and 0.08, the net return of STME predicted by all six models is also greater than 0 in the validation set.

[0138] The support vector machine (SVM) model provided the highest net benefit on the validation set, indicating that it has good clinical utility. Figure 6 ).

[0139] Table 3. Overall performance of each model in the validation set

[0140]

[0141] discuss

[0142] Peripheral blood biomarker detection plays an important role in tumor screening, differential diagnosis, prognostic assessment, and progression monitoring. Current research involves a variety of biomarkers, such as carcinoembryonic antigen (CEA), carbohydrate antigen 125 (CA125), cytokeratin 19 fragment (CYFRA21-1), neuron-specific enolase (NSE), and squamous cell carcinoma antigen (SCCA), which are routinely used in clinical practice.

[0143] However, existing biomarkers also have significant limitations. When used alone, their diagnostic sensitivity is low. For example, in patients with locally advanced non-small cell lung cancer, the positive rates for CYFRA21-1 and SCCA are 45.6% and 54.9%, respectively. Elevated CEA is more common in patients with breast and lung cancer. Its diagnostic sensitivity is 54%-68% (more pronounced in adenocarcinoma), and elevated serum SCCA can predict the occurrence of squamous cell carcinoma of the lung. SCCA has strong specificity for squamous cell carcinoma of the lung (positive rate approximately 54.9%). Combined detection can significantly improve diagnostic efficacy. The sensitivity of combined detection of ten tumor markers (CYFRA21-1, CEA, NSE, SCCA, CA125, CA153, CA199, CA50, CA242, and CA724) can reach 82%, with a negative predictive value exceeding 90%. In comprehensive predictive models, approximately 50% of patients are positive for at least one biomarker. Furthermore, tumor biomarkers have insufficient sensitivity for the early diagnosis of lung cancer (CEA only 20-50%), approximately 30% of patients have consistently normal levels of all markers, and inflammatory states and benign lung diseases can easily lead to false positives (e.g., CEA can be elevated in tuberculosis infection). Diagnosing lung cancer patients with normal tumor markers remains a clinical challenge.

[0144] This study retrospectively analyzed the clinical characteristics and peripheral blood parameters of lung diseases. LASSO regression was used for feature selection, and six routine blood test factors—CEA, WBC, TT-s, and MON%—were chosen to develop a machine learning model for lung cancer prediction. Model validation showed that the support vector machine-based prediction model exhibited strong discriminative power and clinical utility potential.

[0145] While conventional wisdom holds that tumor markers must exceed clinical cutoff values ​​to be diagnostically significant, this study found that even when all 10 detected tumor markers, including carcinoembryonic antigen (CEA), were within the normal reference range, CEA still demonstrated significant discriminative ability in the machine learning model. Among the predictive models developed in this study, the support vector machine (SVM) model performed best on both the test and validation sets; furthermore, SHAP visualization analysis indicated that CEA contributed the most to the model's outcome.

[0146] The model described in this paper has shortcomings such as a small sample size, bias in feature matching, false positive interference, and lack of long-term time series analysis. However, this paper is the first to conduct differential diagnosis on lung disease patients whose cancer indicators are within the normal range and are difficult to identify, and constructs the model and system described in this paper, which has strong clinical reference value for lung disease specialist hospitals.

[0147] All documents mentioned in this invention are incorporated herein by reference as if each document were individually incorporated by reference. Furthermore, it should be understood that after reading the foregoing teachings of this invention, those skilled in the art can make various alterations or modifications to this invention, and these equivalent forms also fall within the scope defined by the appended claims.

Claims

1. A method for constructing a differential diagnosis model of lung cancer, characterized in that, The method comprises the steps of: (S1) providing a clinical data set of a disease patient; the disease patient is a disease patient with atypical imaging manifestations and negative tumor markers, and the disease patient includes a lung cancer patient and a non-lung cancer patient; dividing the clinical data set into a training set and a validation set; (S2) preprocessing the clinical data set to obtain a preprocessed clinical data set; (S3) extracting feature indicators from the preprocessed clinical data set, and screening the feature indicators to obtain preferred feature indicators; (S4) using the preferred feature indicators in the training set for training of a machine learning model; when a training result reaches a predetermined termination condition, the training of the model is terminated to obtain the lung cancer differential diagnosis model.

2. The method of claim 1, wherein, The clinical data set includes imaging feature data and peripheral blood detection data; the imaging feature data includes the number of infiltrated lung fields and the enlargement of hilar / mediastinal lymph nodes; the "atypical imaging manifestations" refer to imaging features that do not conform to the typical diagnostic criteria of the disease and / or imaging manifestations that are difficult to make a clear diagnosis; the tumor markers include alpha-fetal protein (AFP), carcinoembryonic antigen (CEA), cytokeratin 19 fragment (CYFRA21-1), beta 2-microglobulin (β2), carbohydrate antigen 153 (CA153), neuron-specific enolase (NSE), carbohydrate antigen 242 (CA242), carbohydrate antigen 50 (CA50), carbohydrate antigen 199 (CA199), carbohydrate antigen 724 (CA724), and squamous cell carcinoma antigen (SCCA); the lung cancer patient includes a lung adenocarcinoma patient, a lung squamous cell carcinoma patient, a small cell lung cancer patient, or a combination thereof; and the non-lung cancer patient includes a pulmonary tuberculosis patient, a pulmonary sarcoidosis patient, a pneumoconiosis patient, and a pulmonary lymphoma patient.

3. The method of claim 1, wherein, The preprocessing further comprises eliminating the baseline feature differences between the lung cancer patient and the non-lung cancer patient by using propensity score matching (PSM).

4. The method of claim 1, wherein, In step (S3), specifically comprising the steps of: (S3.1) performing correlation analysis on each indicator in the preprocessed clinical data set, and eliminating redundant indicators with r>0.8; (S3.2) performing single-factor analysis on the indicators obtained in step (S3.1), and retaining indicators with P<0.05; (S3.3) further screening the indicators obtained in step (S3.2), including: performing LASSO regression analysis on the indicators obtained in step (S3.2) to obtain a candidate indicator set A; performing Boruta regression analysis on the indicators obtained in step (S3.2) to obtain a candidate indicator set B; (S3.4) taking the intersection of the candidate indicator set A and the candidate indicator set B to obtain the preferred feature indicators.

5. The method of claim 1, wherein, The preferred feature indicators are carcinoembryonic antigen (CEA), monocyte percentage (MON%), peripheral blood lymphocyte count (Lym#), mean corpuscular hemoglobin (MCH), red blood cell distribution width standard deviation (RDW-SD), and albumin / globulin ratio (ALB / GLB).

6. The method of claim 1, wherein, The machine learning model is selected from the group consisting of a support vector machine (SVM), a random forest (RF), an extreme gradient boosting (XGBoost), a naive Bayes (NBM), a logistics, or a K-nearest neighbor (KNN).

7. A lung cancer differential diagnosis system characterized by comprising: The lung cancer differential diagnosis system comprises: An input unit configured to input data, the data comprising clinical data of a subject to be tested; A differential diagnosis unit configured as a differential diagnosis model to obtain a diagnosis result of the subject to be tested; wherein the differential diagnosis model is constructed by the method of claim 1; An output unit configured to output the diagnosis result of the differential diagnosis unit.

8. An electronic device comprising a processor and a memory, characterized in that The memory has a plurality of executable instructions, and the processor is configured to read the instructions and perform the following steps: (1) providing clinical data of a subject to be tested; the clinical data comprising characteristic index values; (2) inputting the clinical data into a diagnosis model, the diagnosis model predicting whether the subject to be tested is a lung cancer patient according to the clinical data, thereby obtaining a risk probability value; The characteristic indexes include carcinoembryonic antigen (CEA), monocyte percentage (MON%), peripheral blood lymphocyte count (Lym#), mean corpuscular hemoglobin (MCH), red blood cell distribution width standard deviation (RDW-SD), and albumin / globulin ratio (ALB / GLB).

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer executable instructions, and the computer executable instructions are read and executed by the processor to implement the following steps: (1) providing clinical data of a subject to be tested; the clinical data comprising characteristic index values; (2) inputting the clinical data into a diagnosis model, the diagnosis model predicting whether the subject to be tested is a lung cancer patient according to the clinical data, thereby obtaining a risk probability value; The characteristic indexes include carcinoembryonic antigen (CEA), monocyte percentage (MON%), peripheral blood lymphocyte count (Lym#), mean corpuscular hemoglobin (MCH), red blood cell distribution width standard deviation (RDW-SD), and albumin / globulin ratio (ALB / GLB).

10. A computer program product comprising computer executable instructions, characterised in that, The computer executable instructions are executed by the processor to implement the following steps: (1) providing clinical data of a subject to be tested; the clinical data comprising characteristic index values; (2) inputting the clinical data into a diagnosis model, the diagnosis model predicting whether the subject to be tested is a lung cancer patient according to the clinical data, thereby obtaining a risk probability value; The characteristic indexes include carcinoembryonic antigen (CEA), monocyte percentage (MON%), peripheral blood lymphocyte count (Lym#), mean corpuscular hemoglobin (MCH), red blood cell distribution width standard deviation (RDW-SD), and albumin / globulin ratio (ALB / GLB).

Images