Method for constructing lung cancer screening model based on PTR-TOF-MS

Abstract

The present invention is based on the method for constructing a lung cancer screening model based on PTR-TOF-MS, including the steps of: A. exhaled breath sample collection; B. full spectrum analysis of the collected exhaled breath sample by PTR-TOF-MS equipment to form a spectrogram sample; C. data preprocessing: including various routine data preprocessing and related calculations on the obtained spectrogram sample, and selecting suitable features; D. model construction: constructing an integrated learning model, the base classifier sorts the gain importance of each feature, and constitutes the feature set of the integrated learning model; the logistic regression model and the integrated learning model are jointly formed into a comprehensive lung cancer screening prediction model; E. model performance evaluation: predicting the performance of the lung cancer screening prediction model through the confusion matrix, and then screening out the best performing lung cancer screening prediction model. Most of the features selected by the present invention have significant differences, can be used as potential lung cancer markers, and have positive significance for lung cancer screening.

Description

Technical Field

[0001] The present invention belongs to the field of model construction for medicine, and specifically relates to a method for constructing a lung cancer screening model based on PTR-TOF-MS. Background Art

[0002] Lung cancer is one of the most common cancers in the world, second only to breast cancer in women, and the mortality rate caused by lung cancer ranks first. In China, the incidence and mortality of lung cancer rank first among all cancers. The main reason for the high mortality rate of lung cancer is that most patients are already in the late stage of lung cancer when they are diagnosed, and conventional treatment at this time often has little effect. Early screening and diagnosis of lung cancer can improve the treatment effect, which is urgently needed to improve the prognosis of patients with this disease.

[0003] The mortality rate of lung cancer ranks first among all cancers worldwide. This is related to the fact that lung cancer is difficult to detect early and is mostly in the late stage when it is discovered. Therefore, research on early screening of lung cancer is extremely urgent. Screening for lung cancer can be achieved through a variety of methods, among which the commonly used methods in clinical practice include chest X-ray, chest CT, sputum culture test, puncture biopsy, etc. Chest X-ray examination can improve the detection rate and survival rate of early lung cancer, but it cannot reduce the mortality rate, and the overlapping images it displays are prone to missed diagnosis. According to the results of the National Lung Screening Trial in my country, the use of low-dose chest spiral CT (LDCT) to screen lung cancer is the most effective way to reduce the mortality rate of lung cancer. However, the false positive rate of LDCT is high, which may lead to more misdiagnosis and unnecessary examinations and treatments. In addition, the high requirements and high costs of LDCT for equipment and personnel, as well as the possible radiation exposure, limit its practical application. Sputum examination is one of the earliest lung cancer screening methods used in clinical practice, but the sensitivity and specificity of this method are relatively low, and there are no reliable markers, which makes this method unsuitable for large-scale lung cancer screening. The puncture biopsy method is the most accurate and is the gold standard for cancer diagnosis. Once cancer cells are detected, lung cancer can be diagnosed. However, compared with other methods, puncture biopsy is invasive and may cause serious complications, so it cannot be used as the first choice for lung cancer screening. It can be seen that these current lung cancer detection methods have certain drawbacks, so a more convenient, accurate, cheap and fast lung cancer early screening method is urgently needed in clinical practice.

[0004] Volatile organic compounds (VOCs) are volatile organic compounds with a melting point below room temperature and a boiling point between 50-260°C. When pathological changes occur in the human body, VOCs different from those in the normal state may be produced, and some of them are excreted from the lungs through the respiratory tract. Different VOCs form different mixtures, and each disease has a different VOCs pattern, so specific diseases can be screened by detecting the components of the VOCs mixture. Diagnostic technology based on VOCs analysis is considered to be a very promising non-invasive early screening method for lung cancer. Compared with traditional detection technology, it has the advantages of being rapid, non-invasive, and easy to obtain exhaled breath samples. It has been a hot topic in the research of early diagnosis of lung cancer in recent years.

[0005] In the current study of lung cancer-significant VOCs in exhaled breath, gas chromatography-mass spectrometry (GC-MS) is the most common analytical method. However, GC-MS technology has some limitations: first, due to the limitation of its chromatographic separation principle, GC-MS has a certain selectivity when analyzing complex samples, and it is impossible to perform a full spectrum analysis of VOCs of different properties in exhaled breath in one collection, which is not conducive to the selection of characteristic VOCs of lung cancer exhaled breath; second, VOCs in human exhaled breath have the characteristics of low concentration and many types, but the sensitivity of GC-MS alone is limited, and it usually needs to be combined with offline enrichment methods such as solid phase microextraction, which will lead to the loss of some VOCs, and the detection results are greatly affected by the sample pretreatment method; finally, the sample pretreatment, enrichment and chromatographic separation process of GC-MS makes the full process analysis time of a single sample at least 30 minutes, which leads to its shortcomings such as complex and cumbersome analysis process, time-consuming operation, and inability to achieve online analysis. Therefore, traditional GC-MS is difficult to meet the current requirements for rapid detection and analysis of exhaled breath on a large scale, over a wide range, and with a large amount of exhaled breath samples.

[0006] Compared with traditional GC-MS technology, Proton Transfer Reaction Time-of-flight Mass Spectrometry (PTR-TOF-MS) technology does not require sample pretreatment, and the proton affinity is greater than that of the reactive ion H3O + According to the basic properties of substances, the proton affinity of most VOCs is greater than that of H2O, so PTR-TOF-MS can achieve non-selective full-spectrum analysis of VOCs in exhaled breath; at the same time, the proton affinity of the main components of air (N2, O2, CO2, etc.) is less than that of H2O, that is, H3O +It can undergo proton transfer reaction with most VOCs without being affected by the main components of air, which makes it very suitable for detection in exhaled breath. In addition, PTR-TOF MS has the advantage of high-throughput and rapid detection. It can detect the full spectrum (m / z 10-500) of exhaled breath samples in just 1 minute. While ensuring rapid detection, it also has the advantage of high sensitivity. The detection capacity of VOCs can reach pptV (volume mixing ratio of 10 -12 = one trillionth = 1pptV) level. Therefore, PTR-TOF-MS has the advantages of full spectrum, high sensitivity, fast response speed, no interference from conventional components in the air, no need for sample pretreatment, simple operation, and real-time online detection. It is very suitable for rapid, non-invasive, low-cost, high-throughput clinical collection and analysis of exhaled breath, providing a solid, accurate and reliable technical guarantee for the scientific exploration of lung cancer-related biomarkers. Summary of the invention

[0007] The present invention provides a method for constructing a lung cancer screening model based on PTR-TOF-MS. By constructing a corresponding lung cancer screening model, volatile organic compounds in exhaled breath are analyzed to find potential exhaled breath lung cancer markers.

[0008] The method for constructing a lung cancer screening model based on PTR-TOF-MS of the present invention is characterized by comprising the steps of:

[0009] A. Exhaled breath sample collection: Collect exhaled breath samples of lung cancer, normal exhaled breath samples and nodule exhaled breath samples that have been confirmed within a set time range, and define normal exhaled breath samples and nodule exhaled breath samples as non-cancer exhaled breath samples;

[0010] B. Performing full spectrum analysis on each exhaled breath sample collected by using a PTR-TOF-MS device to obtain spectrum data of each exhaled breath sample to form a spectrum sample;

[0011] C. Data preprocessing: including data cleaning, missing value mean filling, outlier deletion, calibration of the spectrum sample data to the same level of distribution through calibration gas, environmental background subtraction and peak area calculation, dividing the spectrum sample into training set, validation set and test set, and selecting appropriate features;

[0012] D. Model construction: construct an ensemble learning model, adjust the hyperparameters of the base classifier of the ensemble learning model through Bayesian optimization, and obtain the best model parameter combination; the base classifier of the ensemble learning model selects the top N most important features of information gain importance according to the information gain importance of each feature, and constitutes a feature set of the ensemble learning model, wherein N is a preset natural number; the features in the feature set are applied to the logistic regression model, and the logistic regression model and the ensemble learning model together form a comprehensive lung cancer screening prediction model;

[0013] E. Model performance evaluation: The lung cancer screening prediction model is predicted through the confusion matrix combined with the test set or validation set, and the accuracy, sensitivity and specificity indicators of the lung cancer screening prediction model are calculated to quantify the predictive ability of the lung cancer screening prediction model; then the receiver operating characteristic (ROC) curve of the lung cancer screening prediction model is plotted, and the area under the curve (Area Under the Curve, AUC) is calculated to compare the diagnostic performance of the lung cancer screening prediction model under different parameters, and based on the analysis of the area under the curve, the lung cancer screening prediction model with the best performance is screened out.

[0014] The present invention is based on proton transfer reaction time-of-flight-mass spectrometry (PTR-TOF-MS) to analyze volatile organic compounds (VOCs) in exhaled breath to screen for lung cancer. VOCs of subjects (including non-lung cancer patients and confirmed lung cancer patients) are collected and analyzed by PTR-TOF-MS, and the data set is evaluated by using methods such as confusion matrix and ROC curve. The data is analyzed by using an integrated learning model for several different groups of people such as non-lung cancer and lung cancer, normal people and lung cancer nodules and lung cancer, to find potential exhaled breath lung cancer markers.

[0015] Furthermore, in step B, the steps of full spectrum analysis of each exhaled breath sample collected by the PTR-TOF-MS device are as follows:

[0016] After the PTR-TOF-MS equipment is preheated and the state is stable, the collected exhaled breath samples are connected to the sampling pipeline one by one. In the full spectrum analysis of the exhaled breath samples, two spectral data points are collected for each exhaled breath sample, and the second spectral data of each exhaled breath sample is taken as the spectral sample of the exhaled breath sample.

[0017] Furthermore, in step C, the area below the peak value of all qualitative VOCs in the spectrum sample is normalized to the interval (0, 1), and a data matrix is ​​generated after data processing. The spectrum sample is divided into a training set, a validation set, and a test set according to a set ratio, and then the data is corrected.

[0018] Furthermore, when the data is corrected, the peak area and the standard gas area of ​​the daily calibration gas data within the specified mass-to-charge ratio range are calculated, and then the area data of each mass-to-charge ratio after correction of each exhaled breath sample is obtained, and then all features within the set mass-to-charge ratio area range are selected.

[0019] Furthermore, when correcting the data, the peak area of ​​the mass-to-charge ratio of the calibration gas for each day in the range of [78.7, 79.4] is calculated, and the value of the mass-to-charge ratio intensity lower than 90 is set to 0; then the area of ​​the calibration gas with a mass-to-charge ratio of 79 is calculated: a standard value is set, and the quotient of the standard value divided by the area of ​​the calibration gas with a mass-to-charge ratio of 79 for each day is used as the coefficient value, and finally the area of ​​each mass-to-charge ratio of each exhaled breath sample is multiplied by the coefficient value of the corresponding time to obtain the area data of each mass-to-charge ratio after correction of each exhaled breath sample, select all features in the range of mass-to-charge ratio 15 to mass-to-charge ratio 249, delete the feature of mass-to-charge ratio 94, and delete the features with a peak area of ​​0 accounting for more than 90%.

[0020] The beneficial effects of the present invention include:

[0021] 1. Most of the different numbers of features selected by the model have significant differences and can be used as potential lung cancer markers, which has positive significance for lung cancer screening.

[0022] 2. It can well distinguish between lung cancer patients and non-lung cancer patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 The figure is a flow chart of the method for constructing a lung cancer screening model based on PTR-TOF-MS of the present invention.

[0024] Figure 2 This is a graph showing the contribution percentages of the top 30 features in Example 1.

[0025] Figure 3 a is the prediction graph of the confusion matrix for the validation set samples in Example 1.

[0026] Figure 3 b is the ROC curve diagram of the validation set samples in Example 1.

[0027] Figure 4 a is the prediction graph of the confusion matrix for the test set samples in Example 1.

[0028] Figure 4 b is the ROC curve diagram of the test set samples in Example 1.

[0029] Figure 5 These are the 28 characteristic graphs with significant differences in Example 1.

[0030] Figure 6 This is a graph showing the contribution percentages of the top 30 features in Example 2.

[0031] Figure 7 a is the prediction graph of the confusion matrix for the validation set samples in Example 2.

[0032] Figure 7 b is the ROC curve diagram of the validation set samples in Example 2.

[0033] Figure 8 a is the prediction graph of the confusion matrix for the test set samples in Example 2.

[0034] Figure 8 b is the ROC curve diagram of the test set samples in Example 2.

[0035] Fig. 9 These are the 22 characteristic graphs with significant differences in Example 2.

[0036] Fig.10 This is a graph showing the contribution percentages of the top 30 features in Example 3.

[0037] Fig.11 a is the prediction graph of the confusion matrix for the validation set samples in Example 3.

[0038] Fig.11 b is the ROC curve diagram of the validation set samples in Example 3.

[0039] Fig.12 a is the prediction graph of the confusion matrix for the test set samples in Example 3.

[0040] Fig.12 b is the ROC curve diagram of the test set samples in Example 3.

[0041] Fig.13 These are the 28 characteristic graphs with significant differences in Example 3. DETAILED DESCRIPTION

[0042] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application usually described and shown in the drawings here can be arranged and designed in various configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

[0043] like Figure 1 As shown, the method for constructing a lung cancer screening model based on PTR-TOF-MS of the present invention is characterized by comprising the steps of:

[0044] A. Exhaled breath sample collection: exhaled breath samples of lung cancer, normal exhaled breath samples and nodule exhaled breath samples confirmed within a set time range are collected, and normal exhaled breath samples and nodule exhaled breath samples are defined as non-cancer exhaled breath samples.

[0045] The exhaled breath samples in this embodiment were collected from patients who visited the West China Hospital of Sichuan University from October 2022 to April 2023 and whose chest CT showed space-occupying lesions in the lungs as research subjects, and healthy subjects were recruited as a control group at the same time.

[0046] Inclusion criteria: (1) Aged 18-70 years old, regardless of gender; (2) Complete clinical information, including hospitalization / outpatient number, age, gender, clinical diagnosis, etc.; (3) Meet one of the following conditions: a) Subjects suspected of lung cancer; b) Subjects with other benign respiratory diseases; c) Patients with other cancers, such as gastric cancer, esophageal cancer, and intestinal cancer; d) Healthy subjects without respiratory diseases and signs. Subjects who meet all of the above criteria can be included.

[0047] Exclusion criteria: (1) patients who have been diagnosed with lung cancer and received lung cancer treatment; (2) patients who ate after 22:00 the previous night; (3) patients with ventilation or gas exchange dysfunction; (4) patients with severe mental illness or cognitive impairment; (5) women who are pregnant or breastfeeding; (6) patients with a history of drug abuse or current drug abuse; (7) patients who are considered by the researchers to be unsuitable for participation in this study. Subjects who meet any of the above criteria will not be included.

[0048] This study has established the criteria and procedures for subject withdrawal to protect the rights and interests of the subjects. Subjects have the right to withdraw from the trial at any time without any reason. If the subject decides to withdraw from the trial, the researcher should try to find out the reason for the withdrawal and record this reason as the original text data. In order to protect the rights and interests of the subjects, ensure the quality of the trial, and avoid unnecessary economic losses, if the subject encounters the following situations, the researcher determines that the subject is unable to continue this clinical trial, and the researcher has the right to suspend his / her participation in the trial and require him / her to withdraw from the trial early: (1) The subject meets the exclusion criteria (new or previously undiscovered) and cannot continue to participate in the study; (2) The subject encounters other conditions that the researcher believes make it impossible to continue this clinical trial; (3) The subject is unwilling to continue to participate in this clinical trial for any reason.

[0049] In this example, a total of 4325 lung cancer samples and normal samples were collected from West China Hospital of Sichuan University, including 360 lung cancer (Lung cancer, referred to as Lc) samples, 995 normal samples, and 2970 nodule (Node) samples. Normal samples and nodule samples were defined as non-lung cancer (Non-Lc) samples, totaling 3965 samples.

[0050] The clinical data of patients in this study were mainly collected from the HIS medical record system (Hospital Information System, HIS), and the clinical data of healthy physical examination populations were collected from the physical examination system. It mainly includes basic information of patients and examination results: gender, age, exposure history of risk factors such as smoking history, imaging examination, pathology report (histological classification, pathological stage), tumor location and size, laboratory examination (including: blood routine, biochemistry, tumor markers, etc.).

[0051] To ensure that the quality of exhaled breath collection meets the requirements, the exhaled breath is collected through an exhaled breath sample bag, and the following conditions must be met: 1) Exhaled breath is collected before surgery, chemotherapy, radiotherapy, etc.; 2) Fasting for more than 8 hours in the morning; 3) No smoking or drinking within 2 hours; 4) Use an offline sampling device to collect the subject's alveolar gas sample. Strict standard procedures for the collection of exhaled breath samples and testing and analysis of subjects have been formulated for the collection of exhaled breath samples, the collection of ambient air, and the transportation of exhaled breath sample bags to ensure the accuracy and scientificity of the test.

[0052] B. Perform full spectrum analysis on each exhaled breath sample collected by PTR-TOF-MS equipment to obtain the spectrum data of each exhaled breath sample and form a spectrum sample:

[0053] After the exhaled breath sample collection bag containing the exhaled breath sample is sent to the laboratory according to standard procedures, the exhaled breath sample can be tested and analyzed. Mass spectrometry conditions and parameter settings of the PTR-TOF-MS equipment: Equipment model: PTR-TOF-MS2500, injection tube and ion source temperature 70°C, ion source pressure 265Pa, drift tube voltage 800V, analysis time 50s; Exhaled breath sample detection and analysis test process: (1) After the PTR-TOF-MS equipment is preheated and the state is stable, connect the exhaled breath sample sampling bag to the injection pipeline, open the bag valve, start the full spectrum analysis of the exhaled breath sample, collect two spectrum data points for each exhaled breath sample, and take the second spectrum data of each exhaled breath sample as the result; (2) After the spectrum collection is completed, close the bag valve, start the nitrogen cleaning pipeline controlled by the mass flow controller (Mass Flow Controller, MFC), execute the pipeline cleaning process of the above two data points, and discharge the gas residue in the pipeline with clean high-purity nitrogen; (3) After the pipeline cleaning process is completed, the above steps can be repeated to start the injection analysis process of the next exhaled breath sample.

[0054] C. Data preprocessing: including data cleaning of the obtained spectrum samples, missing value mean filling processing, outlier deletion, calibration of the spectrum sample data to the same level distribution through calibration gas, environmental background subtraction and peak area calculation:

[0055] The area below the peak value (VOCs content) of all qualitative VOCs (volatile organic compounds) in the spectrum sample is normalized to the interval (0, 1). After data processing, a data matrix is ​​generated. The spectrum sample is divided into a training set, a validation set, and a test set in a ratio of 3:1:1, and then the data is corrected.

[0056] When correcting the data, calculate the peak area and standard gas area of ​​the M / Z (mass-to-charge ratio) of the daily calibration gas data in the range of [78.7, 79.4], set the M / Z value with an intensity value lower than 90 to 0, and calculate the standard gas area of ​​M / Z_79: take 3,000,000 as the standard value, and retain the quotient of the standard value divided by the area of ​​the daily standard gas M / Z_79 as the coefficient value. Finally, multiply the area of ​​each M / Z of each exhaled breath sample by the coefficient value of the corresponding time to obtain the area data of each M / Z after the exhaled breath sample is corrected. Select all features in the range of M / Z_15 to M / Z_249, and a total of 235 features are selected. Finally, delete the contaminated M / Z_94 in the exhaled breath sample bag itself and the features with a ratio of 0 greater than 90%.

[0057] D. Build a model: Build an integrated learning model, which can include a Random Forest (RF) model, a CatBoost model, or an XGBoost model. Accurate modeling is performed through training sets and validation sets to ensure the reliability and accuracy of the model. During the model training process, the base classifiers of the model are trained using the training set, and their performance is evaluated using the validation set. Then, Bayesian optimization techniques are applied to fine-tune the hyperparameters of these base classifiers (such as the learning rate and the maximum depth of the tree) to ensure that the best combination of model parameters is selected.

[0058] To further improve the performance and interpretability of the model, the information gain importance of each feature is scored based on the base classifier. By training the base classifier and collecting these scores, a feature importance curve is plotted, which arranges the features in descending order of their importance and shows their cumulative contribution. This curve is then analyzed to identify those features that have a decreasing effect on the performance of the model, which is usually reflected in a significant inflection point on the curve. Based on this inflection point, the top N most important features are selected to form the feature set of the model, where N is a preset natural number.

[0059] The features in the feature set are applied to a logistic regression model, which is a statistical learning method for solving classification problems. The training goal of the logistic regression model is to maximize the likelihood function, and optimization algorithms such as gradient descent are usually used to solve the parameters. Once the model training is completed, it can be used to classify new samples and output the corresponding category probabilities. The logistic regression model and the ensemble learning model together form a comprehensive lung cancer screening prediction model. By analyzing the weights of the independent variables in the logistic regression model, key factors closely related to the occurrence of lung cancer can be identified. These factors and their weights will become the core indicators for predicting the risk of individual lung cancer, thereby being able to more accurately predict the possibility of lung cancer.

[0060] E. Model performance evaluation: The lung cancer screening prediction model is predicted by combining the confusion matrix with the test set or validation set, and the accuracy, sensitivity and specificity indicators of the lung cancer screening prediction model are calculated to quantify the prediction ability of the lung cancer screening prediction model; then the receiver operating characteristic (ROC) curve of the lung cancer screening prediction model is plotted, and the area under the curve (Area Under the Curve, AUC) is calculated to compare the diagnostic performance of the lung cancer screening prediction model under different parameters, and the best performing lung cancer screening prediction model is screened out based on the analysis of the area under the curve. The closer the AUC value is to 1, the higher the diagnostic accuracy of the model. Based on the analysis of the AUC value, the best performing diagnostic model can be screened out.

[0061] The confusion matrix, also known as the error matrix or likelihood matrix, is a standard format for expressing accuracy evaluation in the field of machine learning. The confusion matrix is ​​a visualization tool, especially used in supervised learning. In structured accuracy evaluation, it is mainly used to compare classification results with actual measured values, and the accuracy of classification results can be displayed in a confusion matrix.

[0062] The confusion matrix means: each column of the confusion matrix represents the predicted category, and the total number of each column represents the number of data predicted to be in that category; each row represents the true category of the data, and the total number of data in each row represents the number of data instances in that category; the value in each column represents the number of true data predicted to be in that category. The structure of the confusion matrix is ​​generally shown in Table 1.

[0063] Table 1:

[0064]

[0065] In Table 1, TP represents true positive, that is, the number of samples that are actually positive and correctly predicted as positive; FN represents false negative, that is, the number of samples that are actually positive but incorrectly predicted as negative; FP represents false positive, that is, the number of samples that are actually negative but incorrectly predicted as positive; TN represents true negative, that is, the number of samples that are actually negative and correctly predicted as negative.

[0066] Sensitivity = true positive / (true positive + false negative) × 100% = TP / (TP + FN) × 100%;

[0067] Specificity = true negative / (false positive + true negative) × 100% = TN / (FP + TN) × 100%;

[0068] Correct rate = (true positive + true negative) / (true positive + false negative + false positive + true negative) × 100% = (TP + TN) / (TP + FP + FN + TN) × 100%.

[0069] Together, these metrics form an overall assessment of the performance of a classification model. Accuracy gives the overall proportion of the model’s predictions that are correct, while sensitivity and specificity measure the model’s ability to identify true positives and true negatives, respectively.

[0070] AUC (Area Under the Curve) is the area under the ROC (Receiver Operating Characteristic) curve, usually ranging between 0.5 and 1. This metric is used to intuitively evaluate the performance of a classifier, where the value closer to 1 indicates better performance of the classifier. AUC is an important performance evaluation criterion because, compared to the ROC curve itself, it provides a more explicit comparison of classifier performance in the form of a single value.

[0071] The calculation of AUC usually adopts the trapezoidal rule, that is, to calculate the total area under the ROC curve. The value of AUC is used to judge the quality of the classifier, and the specific explanations are as follows:

[0072] AUC = 1: It means the classifier is perfect and can make perfect predictions by setting thresholds.

[0073] 0.5 < AUC < 1: It means the performance of the classifier is better than random guessing and has practical predictive value.

[0074] AUC = 0.5: It means the prediction effect of the classifier is the same as random guessing, that is, it has no practical predictive value.

[0075] AUC < 0.5: It means the performance of the classifier is worse than random guessing, but if its prediction results can be interpreted reversely, it may be better than random guessing.

[0076] This method makes AUC a concise and effective tool for measuring and comparing the performance of different classifiers when dealing with imbalanced datasets.

[0077] The present invention is based on proton transfer reaction time-of-flight mass spectrometry (PTR-TOF-MS) to analyze volatile organic compounds (VOCs) in exhaled breath for lung cancer screening. By collecting and analyzing the VOCs of subjects (including non-lung cancer patients and diagnosed lung cancer patients) through PTR-TOF-MS, methods such as confusion matrices and ROC curves are used to evaluate the dataset, and data analysis is performed on several different paired populations such as non-lung cancer and lung cancer, normal people and lung cancer nodules and lung cancer using an ensemble learning model to find potential exhaled breath lung cancer markers.

[0078] The following further illustrates the present invention through multiple clinical data:

[0079] Example 1:

[0080] The prediction model based on PTR-TOF-MS can effectively distinguish lung cancer patients from non-lung cancer patients (healthy people and nodule patients), as follows:

[0081] In order to match the two types of data samples of lung cancer patients and non-lung cancer patients and improve the generalization ability of the model, 3965 non-lung cancer samples were stratified and sampled according to age and gender, 400 non-lung cancer samples were selected, and then 760 samples (360 lung cancer samples, 400 non-lung cancer samples) were randomly divided into training set, validation set and test set according to 6:2:2: 456 samples in the training set; 152 samples in the validation set, including 74 lung cancer samples and 78 non-lung cancer samples; 152 samples in the test set, including 63 lung cancer samples and 89 non-lung cancer samples. The XGBoost (eXtreme Gradient Boosting) algorithm was used to establish a binary classification model.

[0082] The top 30 ranking results of the model's feature importance are as follows Figure 2 As shown in the Feature contribution percentage chart, Figure 2 It can be seen that feature M / Z_101_area has the highest contribution percentage to the model, feature M / Z_89_area has the second highest contribution to the model, feature M / Z_32_area has the third highest contribution to the model, and so on for other features. M / Z_41_area has the lowest contribution to the model.

[0083] The feature contribution percentage plot is a visualization method to show the impact of each feature in the model on the prediction results. Features are labeled M / Z plus a series of numbers. The importance of each feature is expressed as a percentage and is presented in the graph as a length. The longer the length of the bar, the greater the impact of the feature on the model's prediction results. The longest bar represents the most important feature, while the shortest represents the less influential feature.

[0084] Use the trained model to predict the validation set samples, and use the confusion matrix ( Figure 3 a) and ROC curve for evaluating diagnostic performance ( Figure 3 b) The prediction of the validation set samples is as follows Figure 3 a and Figure 3 As shown in b.

[0085] exist Figure 3In a, 2 of the 78 lung cancer samples (True value vertical Lc) were predicted as non-lung cancer samples (Predicted value horizontal Non_lc), and 2 of the 74 non-lung cancer samples (True value vertical Non_lc) were predicted as lung cancer samples (Predicted value horizontal Lc). The overall accuracy was 97.37%, the sensitivity was 97.44%, and the specificity was 97.30%; Figure 3 The ROC area AUC value in b is 0.9969, which is very close to 1, and the modeling effect is excellent.

[0086] Use the trained model to predict the test set samples to evaluate the model performance, and use the confusion matrix ( Figure 4 a) and ROC curve for evaluating diagnostic performance ( Figure 4 b) Prediction of test set samples, such as Figure 4 a and Figure 4 As shown in b.

[0087] exist Figure 4 Among the 63 lung cancer samples in a (Lc in the vertical direction of True value), 1 was predicted as a non-lung cancer sample (Non_lc in the horizontal direction of Predicted value), and 7 of the 89 non-lung cancer samples (Non_lc in the vertical direction of True value) were predicted as lung cancer samples (Lc in the horizontal direction of Predicted value). The overall accuracy was 94.74%, the sensitivity was 98.41%, and the specificity was 92.13%; Figure 4 The ROC area AUC value in b is 0.9954, which is very close to 1, and the modeling effect is excellent.

[0088] The first 30 features of the XGBoost model were selected for non-lung cancer patients and cancer patients for non-parametric Mann-Whitney U test. There were 28 features with significant P values ​​less than 0.05. These 28 features were plotted as box plots for non-lung cancer patients and lung cancer patients, and the P values ​​of the non-parametric Mann-Whitney U test were displayed in the title of the figure, as shown in the figure below: Figure 5 shown.

[0089] exist Figure 5The data contains distribution comparison diagrams of multiple different features, each of which shows the distribution of a specific mass-to-charge ratio (M / Z) feature in the two categories of "non-lung cancer" and "lung cancer". Most features have very low P values ​​(close to 0) between the two categories, far below the commonly used statistical significance threshold of 0.05. This suggests that these features may be related to lung cancer status, and there are multiple features with significant differences between non-lung cancer and lung cancer patients, which can well distinguish these two groups of people.

[0090] Overall results: The accuracy of the training set was 100%, the accuracy of the validation set was 97.37%, and the accuracy of the test set was 94.74%. The prediction effect of the three data sets was good, with the accuracy rate greater than 90%, and the model generalization ability was good. It was proved that the prediction model based on PTR-TOF-MS can effectively distinguish lung cancer patients from non-lung cancer patients (healthy people and nodule patients).

[0091] Embodiment 2:

[0092] The prediction model based on PTR-TOF-MS can effectively distinguish lung cancer patients from healthy people, as follows:

[0093] In order to match the two types of data samples of lung cancer patients and healthy people and improve the generalization ability of the model, 995 normal samples were stratified and sampled according to age and gender, 398 normal samples were selected, and then 758 samples (360 lung cancer samples, 398 normal samples) were randomly divided into training set, validation set and test set according to 6:2:2: 454 samples in the training set; 152 samples in the validation set, including 64 lung cancer samples and 88 normal samples; 152 samples in the test set, including 77 lung cancer samples and 75 normal samples. The XGBoost algorithm was used to establish a binary classification model.

[0094] The top 30 ranking results of the model's feature importance are as follows Figure 6 As shown in the Feature contribution percentage chart, feature M / Z_101_area has the highest contribution percentage to the model, feature M / Z_129_area has the second highest contribution percentage to the model, feature M / Z_42_area has the third highest contribution percentage to the model, and so on for other features.

[0095] Use the trained model to predict the validation set samples, and use the confusion matrix ( Figure 7 a) and ROC curve for evaluating diagnostic performance ( Figure 7 b) The prediction of the validation set samples is as follows Figure 7 a and Figure 7 As shown in b.

[0096] Depend on Figure 7 aIt can be seen that among the 64 lung cancer samples in the validation set (True value vertical Lc), 1 was predicted as a normal sample (Predicted value horizontal Normal), and among the 88 normal samples (True value vertical Normal), 3 were predicted as lung cancer samples (Predicted value horizontal Lc). The overall accuracy is 97.37%, the sensitivity is 98.44%, and the specificity is 96.60%. Figure 7 The area under the ROC curve in (b) is 0.9980, and the AUC is greater than 0.90, indicating that the classifier works very well.

[0097] Use the trained model to predict the test set samples to evaluate the model performance, and use the confusion matrix ( Figure 8 a) and ROC curve for evaluating diagnostic performance ( Figure 8 b) Prediction of test set samples, such as Figure 8 a and Figure 8 As shown in b.

[0098] Depend on Figure 8 aIt can be seen that 4 out of 77 lung cancer samples (Lc in the vertical direction of True value) in the test set were predicted as normal samples (Normal in the horizontal direction of Predicted value), and 2 out of 75 normal samples (Normal in the vertical direction of True value) were predicted as lung cancer samples (Lc in the horizontal direction of Predicted value). The overall accuracy is 96.05%, the sensitivity is 94.81%, and the specificity is 97.33%. Figure 8 The area under the ROC curve in (b) is 0.9950, and the AUC is greater than 0.90, indicating that the classifier works very well.

[0099] The first 30 features of the XGBoost model were selected for normal people and cancer patients for non-parametric Mann-Whitney U test. There were 22 features with significant P values ​​less than 0.05. The box plots of normal people and lung cancer patients were drawn for these 22 features, and the P values ​​of the non-parametric Mann-Whitney U test were displayed in the title of the figure, as shown in the figure below: Fig. 9 shown.

[0100] exist Fig. 9It contains distribution comparison graphs of multiple different features, each of which represents the distribution of specific mass-to-charge ratio (M / Z) features under the two categories of "normal" and "lung cancer". Most features have very low P values ​​(close to 0) between the two categories, far below the commonly used statistical significance threshold of 0.05. This suggests that these features may be associated with lung cancer status and can be used as biomarkers to distinguish healthy and disease samples.

[0101] Overall results: The accuracy of the training set was 100%, the accuracy of the validation set was 97.37%, and the accuracy of the test set was 96.05%. The prediction effects of the three data sets were all good, with the accuracy rates all greater than 95%, and the model generalization ability was good. It was proved that the prediction model based on PTR-TOF-MS can effectively distinguish lung cancer patients from healthy people.

[0102] Embodiment 3:

[0103] The prediction model based on PTR-TOF-MS can effectively distinguish patients with lung cancer from patients with benign lung nodules, as follows:

[0104] In order to match the two types of data samples and improve the generalization ability of the model, 2970 nodule samples were stratified and sampled according to age and gender, 401 nodule samples were selected, and then 761 samples (360 lung cancer samples, 401 nodule samples) were randomly divided into training set, validation set and test set according to 6:2:2: 456 samples in the training set; 153 samples in the validation set, including 75 lung cancer samples and 78 nodule samples; 152 samples in the test set, including 74 lung cancer samples and 78 nodule samples. The XGBoost algorithm was used to establish a binary classification model.

[0105] The top 30 ranking results of the model's feature importance are as follows Fig.10 As shown in the Feature contribution percentage chart, Fig.10 It can be seen that feature M / Z_101_area contributes the most to the model, feature M / Z_89_area contributes the second most, feature M / Z_37_area contributes the third most, and so on for other features. M / Z_42_area contributes the least to the model.

[0106] Use the trained model to predict the validation set samples, and use the confusion matrix ( Fig.11 a) and ROC curve for evaluating diagnostic performance ( Fig.11 b) The prediction of the validation set samples is as follows Fig.11 a and Fig.11 As shown in b.

[0107] Fig.11In a, 6 of 75 lung cancer samples (True value vertical Lc) were predicted as nodule samples (Predicted value horizontal Node), and 3 of 78 nodule samples (True value vertical Node) were predicted as lung cancer samples (Predicted value horizontal Lc). The overall accuracy was 94.12%, the sensitivity was 92.00%, and the specificity was 96.15%; Fig.11 The AUC value of the ROC curve in b is 0.9897, which is very close to 1, and the modeling effect is excellent.

[0108] Use the trained model to predict the test set samples, and use the confusion matrix ( Fig.12 a) and ROC curve for evaluating diagnostic performance ( Fig.12 b) The prediction of the test set samples is as follows Fig.12 a and Fig.12 As shown in b.

[0109] exist Fig.12 In a, 3 of the 66 lung cancer samples (True value vertical Lc) were predicted as nodule samples (Predicted value horizontal Node), and 3 of the 86 nodule samples (True value vertical Node) were predicted as lung cancer samples (Predicted value horizontal Lc). The overall accuracy was 96.05%, the sensitivity was 95.45%, and the specificity was 96.51%; Fig.12 The AUC value of the ROC curve in b is 0.9924, which is very close to 1, and the modeling effect is excellent.

[0110] The first 30 features of the XGBoost model were selected for nodule patients and cancer patients for non-parametric Mann-Whitney U test. The significant P value was less than 0.05. There were 25 features in total. These 25 features were used to draw box plots of nodule patients and lung cancer patients, and the P value of the non-parametric Mann-Whitney U test was displayed in the title of the figure, as shown in the figure below: Fig.13 shown.

[0111] exist Fig.13The data contains distribution comparison graphs of multiple different features. Each comparison graph shows the distribution of specific mass-to-charge ratio (M / Z) features in the two categories of "nodules" and "lung cancer". Most features have very low P values ​​(close to 0) between the two categories, far below the commonly used statistical significance threshold of 0.05. This suggests that these features may be related to lung cancer status. There are multiple features with significant differences between nodules and lung cancer patients, which can well distinguish these two groups of people.

[0112] Overall results: The accuracy of the training set was 100%, the accuracy of the validation set was 94.12%, and the accuracy of the test set was 96.05%; the prediction effects of the three data sets were all good, with the accuracy rates all greater than 90%, and the model generalization ability was good. It was proved that the prediction model based on PTR-TOF-MS can effectively distinguish lung cancer patients from nodule patients.

[0113] In the above embodiments, the obtained exhaled breath samples were divided into lung cancer samples, nodule samples and normal samples, and the VOCs in the exhaled breath samples were analyzed using PTR-TOF-MS. A model was built and the model was evaluated using a confusion matrix and a ROC curve.

[0114] Table 2: Effectiveness evaluation indicators of validation sets and test sets for different sample combinations

[0115]

[0116] From the evaluation results shown in Table 2, the modeling effects of the three research objects are excellent, and most of the different numbers of features selected have significant differences and can be used as potential lung cancer markers.

[0117] The above embodiments only express the specific implementation methods of the present application, and the description is relatively specific and detailed, but it cannot be understood as limiting the scope of the present application. For ordinary technicians in this field, several modifications and improvements made without departing from the concept of the present application all belong to the protection scope of the present application.

Claims

1. Construction method of lung cancer screening model based on PTR-TOF-MS, Its characteristics are: including the steps of: A. Exhaled breath sample collection: Collect exhaled breath samples of lung cancer, normal exhaled breath samples and nodule exhaled breath samples that have been confirmed within a set time range, and define normal exhaled breath samples and nodule exhaled breath samples as non-cancer exhaled breath samples; B. Performing full spectrum analysis on each exhaled breath sample collected by using a PTR-TOF-MS device to obtain spectrum data of each exhaled breath sample to form a spectrum sample; C. Data preprocessing: including data cleaning, missing value mean filling, outlier deletion, calibration of the spectrum sample data to the same level distribution through calibration gas, environmental background subtraction and peak area calculation, dividing the spectrum sample into training set, validation set and test set, and selecting appropriate features; including: The area below the peak value of all qualitative VOCs in the spectrum sample is normalized to the interval (0, 1), and a data matrix is ​​generated after data processing. The spectrum sample is divided into a training set, a validation set, and a test set according to a set ratio, and then the data is corrected; When correcting the data, the peak area and the standard gas area of ​​the daily calibration gas data within the specified mass-to-charge ratio range are calculated, and then the area data of each mass-to-charge ratio after correction of each exhaled breath sample is obtained, and then all features within the set mass-to-charge ratio area range are selected; the steps include: Calculate the peak area of ​​the mass-to-charge ratio of the calibration gas for each day in the range of [78.7, 79.4], and set the value of the mass-to-charge ratio intensity lower than 90 to 0; then calculate the area of ​​the calibration gas with a mass-to-charge ratio of 79: set the standard value, divide the standard value by the area of ​​the calibration gas with a mass-to-charge ratio of 79 every day as the coefficient value, and finally multiply the area of ​​each mass-to-charge ratio of each exhaled breath sample by the coefficient value of the corresponding time to obtain the area data of each mass-to-charge ratio after correction for each exhaled breath sample, select all features in the range of mass-to-charge ratio 15 to mass-to-charge ratio 249, delete the feature of mass-to-charge ratio 94, and delete the features with a peak area of ​​0 accounting for more than 90%; D. Model construction: construct an ensemble learning model, adjust the hyperparameters of the base classifier of the ensemble learning model through Bayesian optimization, and obtain the best model parameter combination; the base classifier of the ensemble learning model selects the top N most important features of information gain importance according to the information gain importance of each feature, and constitutes a feature set of the ensemble learning model, wherein N is a preset natural number; the features in the feature set are applied to the logistic regression model, and the logistic regression model and the ensemble learning model together form a comprehensive lung cancer screening prediction model; E. Model performance evaluation: The best performing lung cancer screening prediction model was screened out through the confusion matrix and the receiver operating characteristic curve of the lung cancer screening prediction model.

2. The method for constructing a lung cancer screening model based on PTR-TOF-MS according to claim 1, characterized in that: In step B, the steps of full spectrum analysis of each exhaled breath sample collected by the PTR-TOF-MS device are as follows: After the PTR-TOF-MS equipment is preheated and the state is stable, the collected exhaled breath samples are connected to the sampling pipeline one by one. In the full spectrum analysis of the exhaled breath samples, two spectral data points are collected for each exhaled breath sample, and the second spectral data of each exhaled breath sample is taken as the spectral sample of the exhaled breath sample.

Images