Method for recognizing VOCs biomarkers in expired air of physiologically abnormal crowd based on machine learning and application

Abstract

The invention discloses a method for identifying VOCs biomarkers in expiration of a physiologically abnormal crowd based on machine learning, and the method comprises the steps: obtaining expiration VOCs data of the physiologically abnormal crowd and a normal crowd through a PTR-TOF-MS online sampling technology, and carrying out the recognition of the VOCs biomarkers in the expiration of the physiologically abnormal crowd and the normal crowd through a propensity matching scoring method in combination with mixed factors of the physiologically abnormal crowd and the normal crowd. Hybrid factor interference of VOCs data in the crowd is eliminated, and the VOCs biomarkers in expiration of the physiologically abnormal crowd are obtained through screening; the crowds with physiological abnormality are crowds with abnormal transglutaminase, crowds with abnormal glucose and / or crowds with abnormal apolipoprotein A; the mixed factor information is age, gender, body mass index and / or living habits. And a model capable of distinguishing the physiological abnormality crowd is constructed based on the screened VOCs biomarkers in the expiration of the physiological abnormality crowd, and the physiological abnormality crowd can be distinguished quickly, noninvasively and accurately by using the model.

Description

Technical Field

[0001] The present invention relates to the technical field of biomarker screening, and in particular to a method and application of identifying VOCs biomarkers in the exhaled breath of people with physiological abnormalities based on machine learning. Background Art

[0002] In recent years, breath analysis has attracted widespread attention as a rapid, non-invasive detection technology. Currently, there are thousands of volatile organic compounds (VOCs) detected in human exhaled breath, and new types of exhaled VOCs are reported every year. There are many types of exhaled VOCs, mainly including alkanes, alkenes, alkynes, alcohols, ketones, aldehydes and esters.

[0003] VOCs in human exhaled air are divided into endogenous and exogenous. Exogenous VOCs mainly come from the environment, enter the human body through inhalation, diet, drugs, skin contact, etc., and are transferred and metabolized in the body. Endogenous VOCs come from various metabolic processes in the human body, including products of the host's physiological metabolic processes, products of the metabolic processes of microbial pathogens in the host, and products of the host's pathological responses to processes such as infection or inflammation. Exhaled VOCs come from the metabolic processes of various organs in the human body. Changes in physiological metabolism will cause changes in exhaled VOCs to a certain extent. Reports have pointed out that pentane, ethane, nonanal and decanal detected in exhaled air are related to lipid peroxidation, and acetone detected is also related to the metabolic state of diabetic patients. In addition, exhaled VOCs have also been used to study asthma, chronic obstructive pulmonary disease, lung cancer and other diseases. Therefore, exhaled VOCs can be used as biomarkers reflecting physical health and physiological metabolism.

[0004] Physical examination indicators can reflect the physiological state of the human body through changes in indicators. Glutamine transferase in the blood is a protein-modifying enzyme involved in physiological and pathological processes. It is essential for the synthesis of glutamine and the production of bile, and can affect the expression of glycolytic enzymes and damage the glycolytic metabolic pathway. Monitoring glutamine transferase levels helps to timely detect disease changes in the liver, pancreas and other organs. Glucose metabolism has been shown to be associated with antiviral response and is an important indicator for the diagnosis of diabetes. Abnormal glucose indicators may reflect problems in the body's metabolic system, such as insufficient insulin secretion or insulin resistance, and may further lead to the occurrence of other metabolic diseases such as hypertension and hyperlipidemia. Apolipoprotein A is the protein part of plasma lipoprotein. It is mainly involved in the transport and metabolism of blood lipids and plays an important role in maintaining blood lipid balance and cardiovascular health. Abnormal apolipoprotein A may reflect abnormal blood lipids in the body. Excessive apolipoprotein A will increase blood viscosity and lead to diseases such as atherosclerosis and hyperlipidemia. At present, there is a lack of research on the relationship between these three indicators and exhaled VOCs, and no method has been found to study their corresponding biomarkers.

[0005] VOCs detection is divided into offline detection and online monitoring. Traditional offline detection methods such as GC-MS not only take time to analyze and detect, but also require complex pre-treatment. Moreover, VOCs are prone to transformation and loss after transportation and storage. Online detection methods can solve these problems well. Proton transfer reaction time-of-flight mass spectrometer (PTR-TOF-MS) has the advantages of short measurement response time, high sensitivity, and no need for complex sample pre-treatment, which can realize continuous real-time analysis of trace VOCs. A large amount of exhaled VOCs online measurement data will also bring another problem, that is, the abnormally large amount of data. The use of classical mathematical statistical analysis methods is not only time-consuming and laborious, but also difficult to select a suitable analysis model. It is impossible to analyze the relationship between exhaled VOCs and abnormal physiological indicators from the perspective of big data. In addition, confounding factors such as gender, age, body mass index and other population information may affect both physiological indicators and exhaled VOCs, thereby masking or exaggerating the relationship between exhaled VOCs and physiological indicators, resulting in false positives or false negatives.

[0006] As an explainable big data analysis method, machine learning can implement machine learning and deep learning based on multiple platforms and automatically test and train models. It can not only greatly reduce analysis time, but also make the conclusions obtained more persuasive.

[0007] Therefore, analyzing the VOCs composition in exhaled breath can provide important information about human health and physiological metabolism. In the context of a large amount of exhaled VOCs data and interference from confounding factors, it is necessary to develop a method that is rapid, non-invasive and can accurately identify exhaled VOCs biomarkers in people with physiological abnormalities. Summary of the invention

[0008] The purpose of the present invention is to overcome the above-mentioned deficiencies of the prior art and to provide a method and application for identifying VOCs biomarkers in the exhaled breath of people with physiological abnormalities based on machine learning.

[0009] The first objective of the present invention is to provide a method and application for identifying VOCs biomarkers in the exhaled breath of people with physiological abnormalities based on machine learning.

[0010] The second object of the present invention is to provide a detection reagent for use in preparing a product for distinguishing people with abnormal transglutaminase.

[0011] The third object of the present invention is to provide a model for distinguishing people with abnormal transglutaminase.

[0012] A fourth object of the present invention is to provide a detection reagent for use in preparing a product for distinguishing people with abnormal glucose levels.

[0013] The fifth object of the present invention is to provide a model for distinguishing people with abnormal glucose levels.

[0014] The sixth object of the present invention is to provide a detection reagent for use in preparing a product for distinguishing people with abnormal apolipoprotein A.

[0015] The seventh object of the present invention is to provide a model for distinguishing people with abnormal apolipoprotein A.

[0016] In order to achieve the above object, the present invention is implemented by the following scheme:

[0017] A method for identifying VOCs biomarkers in the breath of people with physiological abnormalities based on machine learning, comprising the following steps:

[0018] S1. Obtain the exhaled VOCs data of people with physiological abnormalities and normal people, as well as the confounding factor information of people with physiological abnormalities and normal people;

[0019] The confounding factor information is age, gender, body mass index and / or lifestyle habits;

[0020] S2. Use the propensity score-based distance method to eliminate the interference of confounding factors in the VOCs data of the physiologically abnormal population and the normal population, and obtain the VOCs data of the physiologically abnormal population and the normal population after the interference is eliminated;

[0021] S3. Using statistical test methods, based on the VOCs data of the physiologically abnormal population and the normal population after eliminating interference obtained in step S2, difference screening is performed, and VOCs with significant differences are used as exhaled VOCs biomarkers.

[0022] Preferably, the physiologically abnormal population is a population with abnormal glutaminase, a population with abnormal glucose and / or a population with abnormal apolipoprotein A.

[0023] More preferably, when the physiologically abnormal population is a population with abnormal glutaminase, VOCs with VIPpredictive≥2 are used as exhaled VOCs biomarkers;

[0024] When the physiological abnormality population is a glucose abnormality population, VOCs with a W value ≥ 300 and X = 0.9 being TRUE are used as exhaled VOCs biomarkers;

[0025] When the physiological abnormal population is apolipoprotein A abnormal population, -log 10 VOCs with (P-value) ≥ 20 were used as exhaled VOCs biomarkers.

[0026] Further preferably, the population with abnormal transglutaminase is a population with a transglutaminase level of <5U / L or >54U / L;

[0027] The abnormal glucose population refers to those with glucose levels <3.9mmol / L or >6.1mmol / L;

[0028] The population with abnormal apolipoprotein A includes males with apolipoprotein A levels <1.1 g / L or >1.72 g / L and females with apolipoprotein A levels <1.2 g / L or >1.9 g / L.

[0029] Preferably, the living habits in step S1 include whether one smokes, drinks alcohol, drinks coffee, drinks tea, coughs, expectorates, and wheezes.

[0030] Preferably, the step S1 of obtaining VOCs data of people with physiological abnormalities and normal people specifically includes the following steps:

[0031] After sitting quietly for ≥5 minutes, the subjects used the exhalation buffer sampling device to blow air. The exhalation time period was not less than 20 seconds. The samples were collected at least 5 times, and the interval between each breath was not less than 60 seconds until a stable exhalation peak platform appeared. The exhaled VOCs were collected and monitored online by PTR-TOF-MS.

[0032] More preferably, the exhaled breath buffer sampling device is an exhaled breath buffer sampling device matched with a PTR-TOF-MS, and the parameters of the PTR-TOF-MS are set as follows: the ion source is H 3 O + , the charge-to-mass ratio (m / z) range is 30-600, the drift tube voltage is 580-620V, the drift tube pressure is 2.1-2.4mbar, the drift tube temperature is 80-150℃, the injection tube temperature is 80-110℃, and the E / N (E is the battery strength, N is the number density of the neutral gas) is 120-150TD.

[0033] More preferably, the standard for online collection and monitoring of exhaled VOCs is: in the VOCs data, each breath sample generates an exhalation profile, based on the exhalation curve indicated by acetone, isoprene, acetonitrile, furan and / or dimethyl sulfide in each breathing cycle, and the data at least 5s after exhalation of the exhalation peak platform is selected.

[0034] Preferably, the method of eliminating the interference of confounding factors using the propensity score-based distance method in step S2 comprises the following steps:

[0035] Based on the RStudio platform, the R package MatchIt was used, the matchit function was adopted, the method was set to nearest, optimal, full and genetic, the sampling ratio was set to a natural number, and the distance metric distance was set to mahalanobis, euclideam and scaled_euclidean, to construct the propensity score matching method;

[0036] Based on the exhaled VOCs data of people with physiological abnormalities and normal people and the confounding factor information of people with physiological abnormalities and normal people, the propensity score matching method is used to eliminate the interference of confounding factors in the VOCs data of people with physiological abnormalities and normal people based on the size of distance measurement.

[0037] The present invention also claims the use of a detection reagent in the preparation of a product for distinguishing people with abnormal glutaminase, wherein the detection reagent is a reagent for detecting the following VOCs; the VOCs are VOCs in exhaled breath, including C 20 H 23 N 3 O 2 , C 9 H 14 O 6 , C 9 H 6 O 4 , C 3 H 2 N 2 S.C. 16 H22 N 2 O 6 , C 10 H 10 N 6 O 11 , C 6 H 6 IN、C 3 H 4 , C 18 H 22 , C 13 H 8 N 2 O 12 , C 2 H 3 N, C 11 H 9 N 7 and C 2 H 4 O.

[0038] Among them, the above-mentioned VOCs are the VOCs shown in the test results of PTR-TOF-MS (proton transfer reaction time-of-flight mass spectrometer) and its matching exhaled breath buffer sampling device; the instrument manufacturer is: IoniconAnalutik Gmbh / Austria; the instrument model is: 1000.

[0039] The present invention also claims a model for distinguishing people with abnormal glutaminase, including a data acquisition module, a judgment module and a result output module;

[0040] The data acquisition module is used to obtain VOCs concentration data of the population to be distinguished; the VOCs are VOCs in the breath of the population to be distinguished, including C 20 H 23 N 3 O 2 , C 9 H 14 O 6 , C 9 H 6 O 4 , C 3 H 2 N 2 S.C. 16 H 22 N 2 O 6 , C 10 H 10 N 6 O 11 , C 6 H 6 IN、C 3 H 4 , C 18H 22 , C 13 H 8 N 2 O 12 , C 2 H 3 N, C 11 H 9 N 7 and C 2 H 4 O;

[0041] The judgment module uses the VOCs concentration data of the to-be-differentiated population acquired by the data acquisition module as input, and inputs it into the trained machine learning model to judge whether the to-be-differentiated population is a population with abnormal glutaminase or a population with normal glutaminase;

[0042] The trained machine learning model is a machine learning model trained based on VOCs concentration data in the training set and the test set samples; the VOCs are the VOCs shown in the data acquisition module; the training set and the test set are sample sets containing both people with abnormal glutamine transferase and people with normal glutamine transferase;

[0043] The result output module is used to output the result obtained by the judgment module.

[0044] Preferably, the machine learning model is a random forest, a gradient boosting machine or a support vector machine.

[0045] More preferably, when the machine learning model is a random forest, the parameters in the random forest training process are set as:

[0046] The R package h2o is used for machine learning. The random forest algorithm uses the h2o.randomForest function. The parameters nfolds are set to 0, ntrees are set to 50, max_depth is set to 20, seed is a random natural number greater than 0, sample_rate is set to 0.632, col_sample_rate_per_tree is set to 1, and col_sample_rate_change_per_level is set to 1.

[0047] More preferably, when the machine learning model is a gradient boosting machine, the parameters in the gradient boosting machine training process are set as:

[0048] The R package h2o is used for machine learning. The gradient boosting algorithm uses the h2o.gbm function. The parameters nfolds are set to 0, ntrees are set to 50, max_depth is set to 20, seed is a random natural number greater than 0, sample_rate is set to 0.632, col_sample_rate_per_tree is set to 1, and col_sample_rate_change_per_level is set to 1.

[0049] The present invention also claims protection for the use of a detection reagent in the preparation of a product for distinguishing people with abnormal glucose levels, wherein the detection reagent is a reagent for detecting the following VOCs; the VOCs are VOCs in exhaled breath, including CH 2 O.C 9 H 14 O 6 , C 3 H 2 N 2 S.C. 18 H 22 , CH 2 O 2 , C 4 H 4 SO, C 9 H 6 O 4 , C 3 H 4 , C 4 H 2 O 4 , C 6 H 6 IN、C 2 H 4 O.C 19 H 38 , C 2 H 2 O、CHNO、C 3 H 3 NO 4 , C 6 H 4 O 5 , C 3 H 14 NO 10 , C 2 H 2 O 3 , C 11 H 9 N 7 , CHF 3 , C 4 H 10 S 3 , C 4 H4 , C 3 H 4 O.C 2 H 4 S.C. 7 H 10 O 3 , C 3 H 4 O 3 , CHO 2 , C 3 H 4 ClN、C 8 H 4 O 3 , C 6 H 2 NO 12 , C 2 H 3 NO, C 3 H 5 Cl, C 13 H 18 S.C. 6 H 4 N 2 O 4 , C 10 H 14 N 2 O 4 , C 6 H 4 Cl 2 and C 7 H 5 ClO.

[0050] Among them, the above-mentioned VOCs are the VOCs shown in the test results of PTR-TOF-MS (proton transfer reaction time-of-flight mass spectrometer) and its matching exhaled breath buffer sampling device; the instrument manufacturer is: IoniconAnalutik Gmbh / Austria; the instrument model is: 1000.

[0051] The present invention also claims a model for distinguishing people with abnormal glucose levels, comprising a data acquisition module, a judgment module and a result output module;

[0052] The data acquisition module is used to obtain VOCs concentration data of the population to be distinguished; the VOCs are VOCs in the breath of the population to be distinguished, including CH 2 O.C 9 H 14 O 6 , C 3 H 2 N 2 S.C. 18 H 22 , CH 2O 2 、C 4 H 4 SO、C 9 H 6 O 4 、C 3 H 4 、C 4 H 2 O 4 、C 6 H 6 IN、C 2 H 4 O、C 19 H 38 、C 2 H 2 O、CHNO、C 3 H 3 NO 4 、C 6 H 4 O 5 、C 3 H 14 NO 10 、C 2 H 2 O 3 、C 11 H 9 N 7 、CHF 3 、C 4 H 10 S 3 、C 4 H 4 、C 3 H 4 O、C 2 H 4 S、C 7 H 10 O 3 、C 3 H 4 O 3 、CHO 2 、C 3 H 4 ClN、C 8 H 4 O 3 、C 6 H 2 NO 12 、C 2 H 3 NO、C 3 H 5 Cl、C 13 H 18 S、C 6 H 4 N2 O 4 , C 10 H 14 N 2 O 4 , C 6 H 4 Cl 2 and C 7 H 5 ClO;

[0053] The judgment module uses the VOCs concentration data of the to-be-differentiated population acquired by the data acquisition module as input, inputs it into the trained machine learning model, and judges whether the to-be-differentiated population is a glucose abnormality population or a glucose normal population;

[0054] The trained machine learning model is a machine learning model trained based on VOCs concentration data in the training set and the test set samples; the VOCs are the VOCs shown in the data acquisition module; the training set and the test set are sample sets containing both people with abnormal glucose and people with normal glucose;

[0055] The result output module is used to output the result obtained by the judgment module.

[0056] Preferably, the machine learning model is a random forest, a gradient boosting machine or a support vector machine.

[0057] More preferably, when the machine learning model is a random forest, the parameters in the random forest training process are set as:

[0058] The R package h2o is used for machine learning. The random forest algorithm uses the h2o.randomForest function. The parameters nfolds are set to 0, ntrees are set to 50, max_depth is set to 20, seed is a random natural number greater than 0, sample_rate is set to 0.632, col_sample_rate_per_tree is set to 1, and col_sample_rate_change_per_level is set to 1.

[0059] More preferably, when the machine learning model is a gradient boosting machine, the parameters in the gradient boosting machine training process are set as:

[0060] Set the random seed to a random natural number greater than 0, use the R package h2o for machine learning, use the h2o.gbm function for the gradient boosting algorithm, set training_frame to the training set, nfolds to 0, ntrees to 50, max_depth to 20, min_rows to 1, sample_rate to 0.632, and col_sample_rate_per_tree to 1.

[0061] The present invention also claims protection for the use of a detection reagent in the preparation of a product for distinguishing people with abnormal apolipoprotein A, wherein the detection reagent is a reagent for detecting the following VOCs; the VOCs are VOCs in exhaled breath, including CH 2 O.C 3 H 4 、CHNO、C 2 H 4 O.C 2 H 5 OS, C 4 H 4 OS, C 8 H 6 NOF 3 , C 12 H 17 NO, C 13 H 18 S.C. 12 H 17 NO 2 , C 3 H 2 N 2 S.C. 9 H 6 O 4 , C 9 H 14 O 6 , C 20 H 23 N 3 O 2 and C 16 H 22 N 2 O 6 .

[0062] Among them, the above-mentioned VOCs are the VOCs shown in the test results of PTR-TOF-MS (proton transfer reaction time-of-flight mass spectrometer) and its matching exhaled breath buffer sampling device; the instrument manufacturer is: IoniconAnalutik Gmbh / Austria; the instrument model is: 1000.

[0063] The present invention also claims a model for distinguishing people with abnormal apolipoprotein A, comprising a data acquisition module, a judgment module and a result output module;

[0064] The data acquisition module is used to obtain VOCs concentration data of the population to be distinguished; the VOCs are VOCs in the breath of the population to be distinguished, including CH 2 O.C 3 H 4 、CHNO、C 2 H 4 O.C 2 H 5 OS, C 4 H 4 OS, C 8 H 6 NOF 3 , C 12 H 17 NO, C 13 H 18 S.C. 12 H 17 NO 2 , C 3 H 2 N 2 S.C. 9 H 6 O 4 , C 9 H 14 O 6 , C 20 H 23 N 3 O 2 and C 16 H 22 N 2 O 6 ;

[0065] The judgment module uses the VOCs concentration data of the to-be-differentiated population acquired by the data acquisition module as input, and inputs it into the trained machine learning model to judge whether the to-be-differentiated population is a population with abnormal apolipoprotein A or a population with normal apolipoprotein A;

[0066] The trained machine learning model is a machine learning model trained based on VOCs concentration data in the training set and the test set samples; the VOCs are the VOCs shown in the data acquisition module; the training set and the test set are sample sets containing both people with abnormal apolipoprotein A and people with normal apolipoprotein A;

[0067] The result output module is used to output the result obtained by the judgment module.

[0068] Preferably, the machine learning model is a random forest, a gradient boosting machine or a support vector machine.

[0069] More preferably, when the machine learning model is a random forest, the parameters in the random forest training process are set as:

[0070] The R package h2o is used for machine learning. The random forest algorithm uses the h2o.randomForest function. The parameters nfolds are set to 0, ntrees are set to 50, max_depth is set to 20, seed is a random natural number greater than 0, sample_rate is set to 0.632, col_sample_rate_per_tree is set to 1, and col_sample_rate_change_per_level is set to 1.

[0071] More preferably, when the machine learning model is a gradient boosting machine, the parameters in the gradient boosting machine training process are set as:

[0072] Set the random seed to a random natural number greater than 0, use the R package h2o for machine learning, use the h2o.gbm function for the gradient boosting algorithm, set training_frame to the training set, nfolds to 0, ntrees to 50, max_depth to 20, min_rows to 1, sample_rate to 0.632, and col_sample_rate_per_tree to 1.

[0073] Compared with the prior art, the present invention has the following beneficial effects:

[0074] The present invention provides a method for identifying VOCs biomarkers in the exhaled breath of people with physiological abnormalities based on machine learning, and the method comprises the following steps: S1. Acquiring exhaled VOCs data of people with physiological abnormalities and normal people, as well as confounding factor information of the people with physiological abnormalities and normal people; the confounding factor information is age, gender, body mass index and / or living habits; S2. Using a propensity score-based distance method to eliminate the interference of confounding factors in the VOCs data of the people with physiological abnormalities and normal people, and obtaining the VOCs data of the people with physiological abnormalities and normal people after the interference is eliminated; S3. Using a statistical test method, based on the VOCs data of the people with physiological abnormalities and normal people after the interference is eliminated obtained in step S2, performing difference screening, and using VOCs with significant differences as exhaled VOCs biomarkers.

[0075] The method of the present invention utilizes PTR-TOF-MS online sampling technology to obtain exhaled VOCs data of people with physiological abnormalities and normal people, and combines the confounding factors of the people with physiological abnormalities and normal people with a propensity matching score method, thereby eliminating the interference of confounding factors of VOCs data in the population and screening out VOCs biomarkers in the exhaled breath of people with physiological abnormalities; and based on the screened VOCs biomarkers in the exhaled breath of people with physiological abnormalities, a model that can be used to distinguish people with physiological abnormalities is constructed, and the model can be used to quickly, non-invasively and accurately distinguish people with physiological abnormalities. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 This is a flowchart of the construction of the random forest model for distinguishing people with abnormal transglutaminase in Example 1;

[0077] Figure 2 This is the OPLS-DA scatter plot after 200 permutation tests shown in step (5) of Example 1;

[0078] Figure 3 The AUROC curve diagram of the random forest model training process used to distinguish people with abnormal transglutaminase in Example 1;

[0079] Figure 4 The AUROC curve diagram of the gradient boosting model training process for distinguishing people with abnormal transglutaminase in Example 2;

[0080] Figure 5 The AUROC curve diagram of the machine learning process of the random forest model used to distinguish people with abnormal glucose in Example 3;

[0081] Figure 6 The AUROC curve diagram of the gradient boosting model training process for distinguishing people with abnormal glucose in Example 4;

[0082] Figure 7 It is a volcano map drawn based on the RStudio platform in Example 5;

[0083] Figure 8 The AUROC curve diagram of the machine learning process of the random forest model used to distinguish people with abnormal apolipoprotein A in Example 5;

[0084] Fig. 9 The AUROC curve diagram of the gradient boosting model training process for distinguishing people with abnormal apolipoprotein A in Example 6;

[0085] Fig.10 is an AUROC curve diagram of the machine learning process of the random forest model 1 used to distinguish people with abnormal transglutaminase in Example 7;

[0086] Fig.11 This is an AUROC curve diagram of the gradient boosting machine 1 model training process used to distinguish people with abnormal glutaminase in Example 8. DETAILED DESCRIPTION

[0087] The present invention is further described in detail below in conjunction with the accompanying drawings and specific examples of the specification. The examples are only used to explain the present invention and are not used to limit the scope of the present invention. The test methods used in the following examples are conventional methods unless otherwise specified; the materials and reagents used are reagents and materials that can be obtained from commercial channels unless otherwise specified.

[0088] The exhaled breath buffer sampling device used in the embodiment of the present invention is a PTR-TOF-MS (proton transfer reaction time-of-flight mass spectrometer) and its matching exhaled breath buffer sampling device, the manufacturer is: IoniconAnalytik Gmbh / Austria; the instrument model is 1000.

[0089] The sample population used in the embodiment of the present invention comes from a physical examination population in a petrochemical industry in a certain place in my country, and the population data has been informed and consent has been obtained before use.

[0090] Example 1 A random forest model based on machine learning for distinguishing people with abnormal transglutaminase

[0091] The flowchart of the construction of the random forest model for distinguishing people with abnormal glutaminase is shown in the figure. Figure 1 As shown, the details are as follows:

[0092] 1. Screening of exhaled VOCs biomarkers in people with abnormal transglutaminase

[0093] 1. Experimental methods

[0094] (1) Work before sampling

[0095] The proton transfer reaction time-of-flight mass spectrometer (PTR-TOF-MS) was assembled with a buffer tube sampler. The injection tube temperature and buffer tube temperature were set to 80 °C, the drift tube temperature was set to 110 °C, the drift tube voltage was set to 600 V, the drift tube pressure was set to 2.3 mbar, and H 3 O + As the ion source for PTR-TOF-MS.

[0096] Based on the detection of the ambient atmosphere, the signal quality of VOCs components and concentrations is checked to determine whether the exhaled VOCs data can be collected normally, whether normal operation can be continued, and whether stable and continuous sampling can be performed; the instrument needs to be kept turned on, the instrument parameters must be stable and can operate normally 4 hours before the start of sampling.

[0097] (2) Exhaled VOCs sampling

[0098] A total of 915 subjects were selected as the subject population. Each subject filled out an informed consent form and a questionnaire before the breath VOCs test to obtain information on confounding factors such as age, gender, body mass index (BMI), and lifestyle (whether they smoke, drink alcohol, drink coffee, drink tea, cough, expectorate, and wheeze). Subsequently, a physical examination was performed to obtain physiological indicator information (including glutamine transferase level, glucose level, and apolipoprotein A level).

[0099] Based on the normal range of glutaminase level of 5-54 U / L, the subjects were divided into 129 people with abnormal glutaminase and 786 people with normal glutaminase.

[0100] A total of 915 subjects were tested for exhaled VOCs. The test time was from 8 to 10 in the morning, and the test was conducted on an empty stomach. All subjects did not eat, drink any liquid other than purified water, or smoke for 4 hours before the test. After the subjects sat quietly for 5 minutes, the PTR-TOF-MS coupled exhaled buffer sampling device was used to obtain the exhaled VOCs composition and concentration level of the crowd in real time. The exhalation time period of each subject was at least not less than 20 seconds. According to the stability of the data, the data was collected more than 5 times. To ensure the stability of the lung exhalation state, the interval between each breath was not less than 60 seconds.

[0101] After sampling for each subject, the exhaled VOCs components and concentrations were immediately checked and recorded. For subjects with incomplete exhaled profiles or unclear exhaled platforms, re-sampling was performed.

[0102] (3) Sampling data processing

[0103] For the VOCs data obtained in step (2), an exhalation profile was generated for each subject. Based on the exhalation curves indicated by acetone and isoprene in each respiratory cycle, the data 5 to 10 seconds after the exhalation peak platform was selected as the VOCs component and concentration corresponding to the subject, and the obtained exhaled VOCs concentration was logarithmized, that is, the exhaled VOCs concentration was logarithmized. 10 N represents (N is the VOCs concentration, the unit is ppb), and the exhaled VOCs data and concentration (the concentration is expressed in logarithm) of each subject are obtained.

[0104] (4) Propensity score matching

[0105] The confounding factor information obtained in step (2) and the exhaled VOCs data and concentration of each subject obtained in step (3) are processed by propensity score matching, and the normal glutaminase group and the abnormal glutaminase group are matched according to the propensity score, so that the difference in confounding factors between individuals of the two groups is minimized and the distance is shortest, and the VOCs data of the abnormal glutaminase group and the glutaminase group after eliminating interference are obtained;

[0106] The propensity score matching code is as follows: based on the RStudio platform, using the R package—MatchIt, using the matchit function, the method is set to nearest, the sampling ratio is set to 1, and the distance metric distance is set to mahalanobis;

[0107] Load the data and run the propensity score matching code to eliminate the interference of confounding factors. After the code is run, the matching results are output to obtain samples of normal and abnormal glutamine transferase populations with the interference of confounding factors eliminated. The corresponding sample size is 1:1 (that is, after eliminating the interference, it includes 129 people with abnormal glutamine transferase and 129 people with normal glutamine transferase, recorded as sample set 1).

[0108] (5) Screening of exhaled VOCs biomarkers for people with abnormal glutaminase

[0109] Based on the VOCs data of the glutaminase abnormality population and the glutaminase normal population after eliminating interference obtained in step (4), supervised orthogonal partial least squares discriminant analysis was performed on SMICA. After 200 permutation tests, the OPLS-DA scatter plot was used to evaluate the differences between the glutaminase abnormality population and the glutaminase normal population. The output parameters R2(x), R2(y) and Q2 of the OPLS-DA model fitting were all required to be greater than 0.5, and VOCs with parameters VIP predictive ≥ 2 were selected as exhaled VOCs biomarkers for the glutaminase abnormality population.

[0110] 2. Experimental results

[0111] The confounding factor information of the glutaminase abnormality population and the glutaminase population after eliminating the interference shown in step (4) is shown in Table 1.

[0112] Table 1 Confounding factor information

[0113] Enzyme indicators gender age BMI Smoking … cough Expectoration respite abnormal male 29 24.44 Never smoke … no no no normal male 30 26.58 Never smoke … no no no abnormal male 58 25.86 Never smoke … no no no normal male 59 25.69 Never smoke … no yes no abnormal male 52 28.43 Never smoke … no no no normal male 39 23.32 Never smoke … no no no … … … … … … … … … abnormal male 45 30.41 Smoking now … yes yes no normal male 52 25.96 Smoking now … yes yes yes

[0114] The OPLS-DA scatter plot after 200 permutation tests shown in step (5) is as follows Figure 2As shown, the results show that the model has excellent classification effect and can well distinguish between people with abnormal glutamine transferase and people with normal glutamine transferase.

[0115] The exhaled VOCs biomarkers of people with abnormal glutaminase based on VIPpredictive≥2 are shown in Table 2.

[0116] Table 2 Exhaled VOCs biomarkers in people with abnormal transglutaminase

[0117] Chemical formula VIPpredictive <![CDATA[C 20 H 23 N 3 O 2 ]]> 3.5 <![CDATA[C 9 H 14 O 6 ]]> 2.8 <![CDATA[C 9 H 6 O 4 ]]> 2.7 <![CDATA[C 3 H 2 N 2 S]]> 2.5 <![CDATA[C 16 H 22 N 2 O 6 ]]> 2.4 <![CDATA[C 10 H 10 N 6 O 11 ]]> 2.2 <![CDATA[C 6 H 6 IN]]> 2.2 <![CDATA[C 3 H 4 ]]> 2.2 <![CDATA[C 18 H 22 ]]> 2.1 <![CDATA[C 13 H 8 N 2 O 12 ]]> 2.1 <![CDATA[C 2 H 3 N]]> 2.1 <![CDATA[C 11 H 9 N 7 ]]> 2.0 <![CDATA[C 2 H 4 O]]> 2.0

[0118] 2. A random forest model for distinguishing people with abnormal transglutaminase

[0119] 1. Experimental methods

[0120] (1) Model training

[0121] 150 random seeds (random natural numbers greater than 0) are set, each of which corresponds to a sampling scheme. The sample set 1 is divided into 150 training sets 1 and test sets 1 in a ratio of 7:3. The logarithm of the VOCs concentration of each sample in the training set 1 is obtained as shown in Table 2, and used as the input of the random forest model.

[0122] Based on 150 training sets 1, machine learning was performed and the average AUROC value was calculated to obtain a random forest model that can be used to distinguish people with abnormal glutaminase, as follows:

[0123] The R package h2o is used for machine learning. The random forest algorithm uses the h2o.randomForest function. The parameters nfolds are set to 0, ntrees are set to 50, max_depth is set to 20, seed is a random natural number greater than 0, sample_rate is set to 0.632, col_sample_rate_per_tree is set to 1, and col_sample_rate_change_per_level is set to 1.

[0124] (2) Model Verification

[0125] The 150 test sets 1 divided from the sample set 1 are input into the random forest model obtained in step (1), and the random forest model that can be used to distinguish people with abnormal glutaminase is run, and the AUROC is output in a readable form to obtain the model verification effect.

[0126] 2. Experimental results

[0127] The AUROC curve during the random forest model training process is as follows Figure 3As shown, the results showed that the average AUROC value of 150 machine learning was 0.93±0.04 (95%CI: 0.93, 0.94), indicating that after machine learning, the random forest model combined with the breath VOCs in the samples as shown in Table 2 can well distinguish between people with abnormal glutamine transferase and people with normal glutamine transferase.

[0128] The confusion matrix of the random forest model used to distinguish people with abnormal glutaminase for distinguishing one of the test set 1 samples is shown in Table 3. The results show that all 38 samples with normal glutaminase were accurately predicted as normal glutaminase, and 39 samples with abnormal glutaminase were predicted as abnormal glutaminase, and only 4 were incorrectly predicted as normal glutaminase.

[0129] Table 3 Confusion matrix of the random forest model used to distinguish people with abnormal transglutaminase

[0130] Predicting abnormalities in transglutaminase Predicted normal transglutaminase True transglutaminase abnormality 39 4 True glutaminase normal 0 38

[0131] Example 2 A gradient boosting model based on machine learning for distinguishing people with abnormal transglutaminase

[0132] 1. Experimental Methods

[0133] 1. Gradient boosting machine model construction

[0134] Based on the exhaled VOCs biomarkers of people with abnormal glutaminase as shown in Table 2 obtained in step 1 of Example 1, 150 random seeds (random natural numbers greater than 0) are set, each random seed corresponds to a sampling scheme, and the sample set 1 is divided into 150 training sets 2 and test sets 2 in a ratio of 7:3. The logarithm of the VOCs concentration of each sample in the above training set 2 as shown in Table 2 is obtained and used as the input of the gradient boosting machine model.

[0135] The VOCs data of each sample in the 150 training set 2 as shown in Table 2 were used as the input of the gradient boosting machine model, and whether the glutamyl transferase was normal or not was used as the prediction. Machine learning was performed and the average AUROC value was calculated to obtain the gradient boosting machine model for distinguishing people with abnormal glutamyl transferase, as follows:

[0136] Set the random seed to a random natural number greater than 0, use the R package h2o for machine learning, use the h2o.gbm function for the gradient boosting algorithm, set the parameters nfolds to 0, ntrees to 50, max_depth to 20, seed to a random natural number greater than 0, sample_rate to 0.632, col_sample_rate_per_tree to 1, and col_sample_rate_change_per_level to 1.

[0137] 2. Model Validation

[0138] The 150 test sets 2 divided from the sample set 1 are input into the input of the gradient boosting machine model obtained in step (1), and the gradient boosting machine for distinguishing people with abnormal glutaminase is run, and the AUROC is output in a readable form to obtain the corresponding model verification effect.

[0139] 2. Experimental Results

[0140] The AUROC curve of the gradient boosting model training process for distinguishing people with abnormal glutaminase is as follows Figure 4 As shown in Table 2, the average AUROC of 150 machine learning was 0.92±0.04 (95%CI: 0.92, 0.93), indicating that after machine learning, the gradient boosting model combined with the samples can well distinguish the people with abnormal glutamine transferase from the people with normal glutamine transferase.

[0141] The confusion matrix of the gradient boosting model used to distinguish people with abnormal glutaminase when distinguishing one of the test set 2 samples is shown in Table 4. The results show that all 38 samples with normal glutaminase were accurately predicted as normal glutaminase, and 39 samples with abnormal glutaminase were predicted as abnormal glutaminase, and only 4 were incorrectly predicted as normal glutaminase.

[0142] Table 4 Confusion matrix of the gradient boosting model used to distinguish people with abnormal glutaminase

[0143] Predicting abnormalities in transglutaminase Predicted normal transglutaminase True transglutaminase abnormality 39 4 True glutaminase normal 0 38

[0144] Example 3 A random forest model based on machine learning for distinguishing people with abnormal glucose levels

[0145] 1. Screening of exhaled VOCs biomarkers in people with abnormal glucose levels

[0146] 1. Experimental methods

[0147] Assuming that the normal range of glucose level is 3.9-6.1 mmol / L, the 915 subjects in Example 1 were divided into 155 subjects with abnormal glucose and 760 subjects with normal glucose.

[0148] According to the method shown in (1) to (4) in step 1 of Example 1, VOCs sampling was performed on the subject population and propensity score matching was performed based on the confounding factor information of each subject and the exhaled VOCs data and concentration of each subject to obtain samples of normal and abnormal glucose populations with the interference of confounding factors eliminated, and the corresponding sample size was 1:1 (i.e., after eliminating the interference, 155 abnormal glucose populations and 155 normal glucose populations were included, recorded as sample set 2).

[0149] Construct the ANCOM analysis code: obtain the feature matrix (with samples as row information and normal / abnormal glucose of samples as column information) and change matrix (with samples as row information and logarithm of VOCs concentration of samples as column information) of each sample in sample set 2. Based on the RStudio platform, use the built-in function feature_table_pre_process of the R package ancom_v2.1 to process the feature matrix and change matrix of each sample in sample set 2. Then use the ancom function to calculate the matrix to obtain the W value of each VOCs and whether X=0.9 is TRUE. Select VOCs with W value ≥300 and X=0.9 is TRUE as the exhaled VOCs biomarkers for people with abnormal glucose.

[0150] 2. Experimental results

[0151] The exhaled VOCs biomarkers of the glucose abnormality population obtained by screening based on W value ≥ 300 and X = 0.9 being TRUE are shown in Table 5.

[0152] Table 5 Exhaled VOCs biomarkers in people with abnormal glucose levels

[0153]

[0154]

[0155] 2. A random forest model for distinguishing people with abnormal glucose levels

[0156] 1. Experimental methods

[0157] (1) Model training

[0158] 150 random seeds (random natural numbers greater than 0) are set, each random seed corresponds to a sampling scheme, and sample set 2 is divided into training set 3 and test set 3 in a ratio of 7:3. The logarithm of the VOCs concentration of each sample in training set 3 is obtained as shown in Table 5, and it is used as the input of the random forest model.

[0159] Machine learning was performed based on 150 training sets 3, and the average AUROC value was calculated to obtain a random forest model that can be used to distinguish people with abnormal glucose levels, as follows:

[0160] The R package h2o is used for machine learning. The random forest algorithm uses the h2o.randomForest function. The parameters nfolds are set to 0, ntrees are set to 50, max_depth is set to 20, seed is a random natural number greater than 0, sample_rate is set to 0.632, col_sample_rate_per_tree is set to 1, and col_sample_rate_change_per_level is set to 1.

[0161] (2) Model Verification

[0162] The 150 test sets 1 divided from the sample set 2 are input into the random forest model obtained in step (1) that can be used to distinguish people with abnormal glucose levels, and the AUROC is output in a readable form to obtain the model verification effect.

[0163] 2. Experimental results

[0164] The AUROC curve of the random forest model used to distinguish people with abnormal glucose levels during machine learning is shown in Figure 2. Figure 5 As shown, the average AUROC value of 150 machine learnings was 0.89±0.04 (95%CI: 0.89, 0.90), indicating that after machine learning, the random forest model combined with the breath VOCs in the samples as shown in Table 5 can well distinguish between people with abnormal glucose and people with normal glucose.

[0165] The confusion matrix of the random forest model used to distinguish people with abnormal glucose levels when distinguishing one of the test set 3 samples is shown in Table 6. The results show that among the 47 samples with normal glucose levels, all were accurately predicted to have normal glucose levels; among the 49 samples with abnormal glucose levels, 42 samples were predicted to have abnormal glucose levels, and only 7 samples were predicted to have normal glucose levels.

[0166] Table 6 Confusion matrix of random forest model used to distinguish people with abnormal glucose

[0167] Predicting glucose abnormalities Predicting normal glucose True glucose abnormality 40 9 True glucose normal 0 47

[0168] Example 4 A gradient boosting model based on machine learning for distinguishing people with abnormal glucose levels

[0169] 1. Experimental Methods

[0170] 1. Gradient boosting machine model construction

[0171] Based on the exhaled VOCs biomarkers of the glucose abnormality population as shown in Table 5 obtained in step 1 of Example 3, 150 random seeds (random natural numbers greater than 0) are set, each random seed corresponds to a sampling scheme, and the sample set 2 is divided into 150 training sets 4 and test sets 4 in a ratio of 7:3. The logarithm of the VOCs concentration of each sample in the above training set 4 as shown in Table 5 is obtained and used as the input of the gradient boosting machine model.

[0172] The VOCs data of each sample in the 150 training set 4 as shown in Table 5 were used as the input of the gradient boosting machine model. Whether the glucose level was normal or not was used as the prediction, machine learning was performed, and the average AUROC value was calculated to obtain the gradient boosting machine model for distinguishing people with abnormal glucose levels, as follows:

[0173] Set the random seed to a random natural number greater than 0, use the R package h2o for machine learning, use the h2o.gbm function for the gradient boosting algorithm, set training_frame to the training set, nfolds to 0, ntrees to 50, max_depth to 20, min_rows to 1, sample_rate to 0.632, and col_sample_rate_per_tree to 1.

[0174] 2. Model Validation

[0175] The 150 test sets 4 divided from the sample set 2 are input into the gradient boosting machine model for distinguishing people with abnormal glucose levels, and the AUROC of the model is output to obtain the verification result.

[0176] 2. Experimental Results

[0177] The AUROC curve during the training of the gradient boosting model for distinguishing people with abnormal glucose levels is as follows: Figure 6 As shown in Table 5, the average AUROC value of 150 machine learning results was 0.89±0.04 (95%CI: 0.88, 0.89), indicating that after machine learning, the gradient boosting model combined with the samples can well distinguish the breath VOCs of people with abnormal glucose from those with normal glucose.

[0178] The confusion matrix of the gradient boosting machine model used to distinguish people with abnormal glucose levels when distinguishing one of the test set 4 samples is shown in Table 7. The results show that among the 47 samples with normal glucose levels, 45 samples were accurately predicted to be normal glucose levels, and only 2 samples were predicted to be abnormal glucose levels; among the 49 samples with abnormal glucose levels, 42 samples were predicted to be abnormal glucose levels, and only 7 samples were predicted to be normal glucose levels.

[0179] Table 7 Confusion matrix of the gradient boosting model used to distinguish people with abnormal glucose levels

[0180] Predicting glucose abnormalities Predicting normal glucose True glucose abnormality 42 7 True glucose normal 2 45

[0181] Example 5 A random forest model based on machine learning for distinguishing people with abnormal apolipoprotein

[0182] 1. Screening of exhaled VOCs biomarkers in people with abnormal apolipoprotein A

[0183] 1. Experimental methods

[0184] Assuming that the normal range of apolipoprotein A level is 1.1-1.72 g / L (male) and 1.2-1.9 g / L (female), the 915 subjects in Example 1 were divided into 164 subjects with abnormal apolipoprotein A and 751 subjects with normal apolipoprotein A.

[0185] According to the method shown in (1) to (4) in step 1 of Example 1, VOCs sampling was performed on the subject population and propensity score matching was performed based on the confounding factor information of each subject and the exhaled VOCs data and concentration of each subject to obtain samples of normal and abnormal apolipoprotein A populations with confounding factors eliminated, and the corresponding sample size was 1:1 (i.e., after eliminating the interference, 164 subjects with abnormal apolipoprotein A and 164 subjects with normal apolipoprotein A were included, recorded as sample set 3).

[0186] Based on the VOCs data and logarithms of the concentration of each sample in sample set 3, the R package DESeq2 was used to calculate the indicator information corresponding to the volcano plot based on the RStudio platform, and the log2 (Fold Change), -log10 (P-value), p-adj parameters and group parameters were obtained. VOCs with log2 (Fold Change) ≥ 0.05 or log2 (Fold Change) ≤ -0.05, p-adj ≤ 0.05 and -log10 (P-value ≥ 20 were selected as exhaled VOCs biomarkers for people with abnormal apolipoprotein A.

[0187] 2. Experimental results

[0188] The volcano map drawn based on the RStudio platform is as follows Figure 7 As shown in Table 8, the exhaled VOCs biomarkers of the population with abnormal apolipoprotein A screened based on log2 (Fold Change) ≥ 0.05 or log2 (Fold Change) ≤ -0.05, p-adj ≤ 0.05 and -log10 (P-value) ≥ 20 are shown.

[0189] Table 8 Exhaled VOCs biomarkers in people with abnormal apolipoprotein A

[0190] Chemical formula Log2(FoldChange) -Log10(P-value) Group <![CDATA[CH 2 The]]> -1.04 26.41 SignificantlyDown <![CDATA[C 3 H 4 ]]> -0.58 23.25 SignificantlyDown CHNO -0.55 24.83 SignificantlyDown <![CDATA[C 2 H 4 O]]> -0.58 36.94 SignificantlyDown <![CDATA[C 2 H 5 OS]]> -0.23 22.45 SignificantlyDown <![CDATA[C 4 H 4 OS]]> 0.38 27.49 SignificantlyDown <![CDATA[C 8 H 6 SLEEP 3 ]]> -0.34 23.60 SignificantlyDown <![CDATA[C 12 H 17 NO]]> -0.62 27.07 SignificantlyDown <![CDATA[C 13 H 18 S]]> -0.46 25.25 SignificantlyDown <![CDATA[C 12 H 17 NO 2 ]]> -0.33 20.99 SignificantlyDown <![CDATA[C 3 H 2 N 2 S]]> 0.89 21.66 SignificantlyUp <![CDATA[C 9 H 6 O 4 ]]> 1.23 26.31 SignificantlyUp <![CDATA[C 9 H 14 O 6 ]]> 0.63 21.97 SignificantlyUp <![CDATA[C 20 H 23 N 3 O 2 ]]> 1.02 29.18 SignificantlyUp <![CDATA[C 16 H 22 N 2 O 6 ]]> 0.52 21.08 SignificantlyUp

[0191] 2. A random forest model for distinguishing people with abnormal apolipoprotein A

[0192] 1. Experimental methods

[0193] (1) Model training,

[0194] 150 random seeds (random natural numbers greater than 0) are set, each random seed corresponds to a sampling scheme, and sample set 3 is divided into training set 5 and test set 5 in a ratio of 7:3. The logarithm of the VOCs concentration of each sample in training set 5 is obtained as shown in Table 8, and it is used as the input of the random forest model.

[0195] Based on 150 training sets 3, machine learning was performed and the average AUROC value was calculated to obtain a random forest model that can be used to distinguish people with abnormal apolipoprotein A, as follows:

[0196] The R package h2o is used for machine learning. The random forest algorithm uses the h2o.randomForest function. The parameters nfolds are set to 0, ntrees are set to 50, max_depth is set to 20, seed is a random natural number greater than 0, sample_rate is set to 0.632, col_sample_rate_per_tree is set to 1, and col_sample_rate_change_per_level is set to 1.

[0197] (2) Model Verification

[0198] The 150 test sets 5 divided from the sample set 3 are input into the random forest model obtained in step (1) that can be used to distinguish people with abnormal apolipoprotein A, and the AUROC is output in a readable form to obtain the model verification effect.

[0199] 2. Experimental results

[0200] The AUROC curve of the random forest model used in the machine learning process for distinguishing people with abnormal apolipoprotein A is shown in the figure Figure 8 As shown in Figure 8, the average AUROC value of 150 machine learning runs was 0.89±0.04 (95% CI: 0.88, 0.89), indicating that after machine learning, the random forest model combined with the exhaled VOCs in the samples as shown in Table 8 can well distinguish between people with abnormal apolipoprotein A and people with normal apolipoprotein A.

[0201] The confusion matrix of the random forest model used to distinguish people with abnormal apolipoprotein A when distinguishing one of the test set 5 samples is shown in Table 9. The results show that all 47 samples with normal apolipoprotein A were accurately predicted as normal apolipoprotein A; among the 52 samples with abnormal apolipoprotein A, 43 samples were predicted as abnormal apolipoprotein A, and only 9 samples were predicted as normal apolipoprotein A.

[0202] Table 9 Confusion matrix of the apolipoprotein A model used to distinguish people with abnormal glucose levels

[0203] Predicting apolipoprotein A abnormalities Predicted Apolipoprotein A normal True apolipoprotein A abnormality 43 9 True Apolipoprotein A normal 0 47

[0204] Example 6 A gradient boosting machine model based on machine learning for distinguishing people with abnormal apolipoprotein A

[0205] 1. Experimental Methods

[0206] 1. Gradient boosting machine model construction

[0207] Based on the exhaled VOCs biomarkers of people with abnormal apolipoprotein A as shown in Table 8 obtained in step 1 of Example 5, 150 random seeds (random natural numbers greater than 0) are set, each random seed corresponds to a sampling scheme, and the sample set 3 is divided into 150 training sets 6 and test sets 6 in a ratio of 7:3. The logarithm of the VOCs concentration of each sample in the above training set 6 as shown in Table 8 is obtained and used as the input of the gradient boosting machine model.

[0208] The VOCs data of each sample in the 150 training sets 6 as shown in Table 8 were used as the input of the gradient boosting machine model. Whether apolipoprotein A is normal or not was used as the prediction, machine learning was performed, and the average AUROC value was calculated to obtain the gradient boosting machine model for distinguishing people with abnormal apolipoprotein A, as follows:

[0209] Set the random seed to a random natural number greater than 0, use the R package h2o for machine learning, use the h2o.gbm function for the gradient boosting algorithm, set training_frame to the training set, nfolds to 0, ntrees to 50, max_depth to 20, min_rows to 1, sample_rate to 0.632, and col_sample_rate_per_tree to 1.

[0210] 2. Model Validation

[0211] The 150 test set 6 divided from the sample set 3 is input into the gradient boosting machine model for distinguishing people with abnormal apolipoprotein A, and the AUROC of the model is output to obtain the verification result.

[0212] 2. Experimental Results

[0213] The AUROC curve of the gradient boosting model training process for distinguishing people with abnormal apolipoprotein A is as follows Fig. 9 As shown in Table 8, the average AUROC value of 150 machine learning was 0.88±0.04 (95%CI: 0.88, 0.89), indicating that after machine learning, the gradient boosting model combined with the samples can well distinguish the apolipoprotein A abnormal group from the apolipoprotein A normal group.

[0214] The confusion matrix of the gradient boosting machine used to distinguish people with abnormal apolipoprotein A when distinguishing one of the test set 6 samples is shown in Table 10. The results show that among the 55 samples with normal apolipoprotein A, 52 samples were accurately predicted to be normal apolipoprotein A, and only 3 samples were predicted to be abnormal apolipoprotein A; among the 40 samples with abnormal apolipoprotein A, 32 samples were predicted to be abnormal apolipoprotein A, and only 8 samples were predicted to be normal apolipoprotein A.

[0215] Table 10 Confusion matrix of the gradient boosting model used to distinguish people with abnormal apolipoprotein A

[0216] Predicting apolipoprotein A abnormalities Predicted Apolipoprotein A normal True apolipoprotein A abnormality 32 8 True Apolipoprotein A normal 3 52

[0217] Example 7 A random forest model based on machine learning for distinguishing people with abnormal transglutaminase

[0218] 1. Experimental Methods

[0219] 1. Model training

[0220] In the propensity score matching of step 1 (4) of Example 1, the confounding factor information is replaced by the gender information, and the rest remains unchanged, to obtain the population with abnormal glutaminase and the population with normal glutaminase after eliminating the interference of gender (that is, after eliminating the interference, 129 people with abnormal glutaminase and 129 people with normal glutaminase are included, recorded as sample set 4).

[0221] 150 random seeds (random natural numbers greater than 0) are set, each of which corresponds to a sampling scheme. The sample set 4 is divided into 150 training sets 7 and test sets 7 in a ratio of 7:3. The logarithm of the VOCs concentration of each sample in the training set 7 is obtained as shown in Table 2, and is used as the input of the random forest model.

[0222] Machine learning was performed based on 150 training sets7, and the average AUROC value was calculated to obtain a random forest model that can be used to distinguish people with abnormal glutaminase, as follows:

[0223] The R package h2o is used for machine learning. The random forest algorithm uses the h2o.randomForest function. The parameters nfolds are set to 0, ntrees are set to 50, max_depth is set to 20, seed is a random natural number greater than 0, sample_rate is set to 0.632, col_sample_rate_per_tree is set to 1, and col_sample_rate_change_per_level is set to 1.

[0224] 2. Model Validation

[0225] The 150 test sets 7 divided from the sample set 4 are input into the random forest model 1 obtained in step 1 for distinguishing people with abnormal glutaminase, and the AUROC is output in a readable form to obtain the model verification effect.

[0226] 2. Experimental Results

[0227] The AUROC curve of the machine learning process of the random forest model 1 used to distinguish people with abnormal transglutaminase is shown in the figure Fig.10 As shown, the average AUROC value of 150 machine learnings was 0.85±0.05 (95% CI: 0.84, 0.85), indicating that after machine learning, the random forest model 1 combined with the breath VOCs in the samples as shown in Table 2 can distinguish between people with abnormal glutamine transferase and people with normal glutamine transferase.

[0228] The confusion matrix of the random forest model 1 used to distinguish people with abnormal glutaminase when distinguishing one of the test set 7 samples is shown in Table 11. The results show that among the 39 samples with normal glutaminase, 31 samples were accurately predicted to have normal glutaminase, and only 8 samples were predicted to have abnormal glutaminase; among the 36 samples with abnormal glutaminase, 31 samples were predicted to have abnormal glutaminase, and only 4 samples were incorrectly predicted to have normal glutaminase.

[0229] Table 11 Confusion matrix of random forest model 1 used to distinguish people with abnormal transglutaminase

[0230] Predicting abnormalities in transglutaminase Predicted normal transglutaminase True transglutaminase abnormality 36 4 True glutaminase normal 8 31

[0231] Example 8 A gradient boosting model based on machine learning for distinguishing people with abnormal transglutaminase

[0232] 1. Experimental Methods

[0233] 1. Model training

[0234] Based on the exhaled VOCs biomarkers of the population with abnormal glutamyl transferase as shown in Table 2 obtained in step 1 of Example 1, 150 random seeds (random natural numbers greater than 0) are set, each random seed corresponds to a sampling scheme, and the sample set 4 is divided into 150 training sets 8 and test sets 8 in a ratio of 7:3, and the logarithm of the VOCs concentration of each sample in the above training set 8 as shown in Table 2 is obtained as the input of the gradient boosting machine model, and whether the glutamyl transferase is normal or not is used as a prediction, and machine learning is performed, and the average AUROC value is calculated to obtain a gradient boosting machine model 1 for distinguishing people with abnormal glutamyl transferase, which is specifically as follows:

[0235] Set the random seed to a random natural number greater than 0, use the R package h2o for machine learning, use the h2o.gbm function for the gradient boosting algorithm, set the parameters nfolds to 0, ntrees to 50, max_depth to 20, seed to a random natural number greater than 0, sample_rate to 0.632, col_sample_rate_per_tree to 1, and col_sample_rate_change_per_level to 1.

[0236] 2. Model Validation

[0237] The 150 test sets 8 divided from the sample set 4 are input into the gradient boosting machine model 1 obtained in step 1 for distinguishing people with abnormal glutaminase, and the AUROC of the model is output to obtain the verification result.

[0238] 2. Experimental Results

[0239] The AUROC curve of the gradient boosting machine 1 model training process for distinguishing people with abnormal glutaminase is as follows Fig.11 As shown in Table 2, the average AUROC of 150 machine learning results was 0.92±0.04 (95%CI: 0.92, 0.93), indicating that after machine learning, the gradient boosting model combined with the samples can distinguish between people with abnormal glutamine transferase and people with normal glutamine transferase in breath VOCs.

[0240] The confusion matrix of the gradient boosting machine model 1 used to distinguish people with abnormal glutaminase when distinguishing one of the test set 8 samples is shown in Table 12. The results show that among the 38 samples with normal glutaminase, 29 samples were accurately predicted to have normal glutaminase, and only 9 samples were predicted to have abnormal glutaminase; among the 373 samples with abnormal glutaminase, 35 samples were predicted to have abnormal glutaminase, and only 2 were incorrectly predicted to have normal glutaminase.

[0241] Table 4 Confusion matrix of gradient boosting model 1 used to distinguish people with abnormal glutaminase

[0242] Predicting abnormalities in transglutaminase Predicted normal transglutaminase True transglutaminase abnormality 35 2 True glutaminase normal 9 29

[0243] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the protection scope of the present invention. For ordinary technicians in this field, other different forms of changes or modifications can be made based on the above descriptions and ideas. It is not necessary and impossible to list all the implementation methods here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A method for identifying VOCs biomarkers in the breath of people with physiological abnormalities based on machine learning, characterized in that: The following steps are involved: S1. Obtain the exhaled VOCs data of people with physiological abnormalities and normal people, as well as the confounding factor information of people with physiological abnormalities and normal people; The confounding factor information is age, gender, body mass index and / or lifestyle habits; S2. Use the propensity score-based distance method to eliminate the interference of confounding factors in the VOCs data of the physiologically abnormal population and the normal population, and obtain the VOCs data of the physiologically abnormal population and the normal population after the interference is eliminated; S3. Using statistical test methods, based on the VOCs data of the physiologically abnormal population and the normal population after eliminating interference obtained in step S2, difference screening is performed, and VOCs with significant differences are used as exhaled VOCs biomarkers.

2. The method according to claim 1, characterized in that The physiologically abnormal population includes a population with abnormal glutaminase, a population with abnormal glucose and / or a population with abnormal apolipoprotein A.

3. The method according to claim 2, characterized in that When the physiological abnormal population is a population with abnormal glutaminase, VOCs with VIPpredictive≥2 are used as exhaled VOCs biomarkers; When the physiological abnormality population is a glucose abnormality population, VOCs with a W value ≥ 300 and X = 0.9 being TRUE are used as exhaled VOCs biomarkers; When the physiological abnormal population is apolipoprotein A abnormal population, -log 10 VOCs with (P-value) ≥ 20 were used as exhaled VOCs biomarkers.

4. The method according to claim 3, characterized in that The population with abnormal glutamyl transferase refers to the population with glutamyl transferase level <5U / L or >54U / L; The abnormal glucose population refers to those with glucose levels <3.9mmol / L or >6.1mmol / L; The population with abnormal apolipoprotein A includes males with apolipoprotein A levels <1.1 g / L or >1.72 g / L and females with apolipoprotein A levels <1.2 g / L or >1.9 g / L.

5. Use of a detection reagent in the preparation of a product for distinguishing people with abnormal transglutaminase, characterized in that: The detection reagent is a reagent for detecting the following VOCs; the VOCs are VOCs in exhaled air, including C 20 H 23 N3O2、C9H 14 O6, C9H6O4, C3H2N2S, C 16 H 22 N2O6、C 10 H 10 N6O 11 、C6H6IN、C3H4、C 18 H 22 , C 13 H8N2O 12 、C2H3N、C 11 H9N7 and C2H4O.

6. A model based on machine learning for distinguishing people with abnormal transglutaminase, characterized in that: It includes a data acquisition module, a judgment module and a result output module; The data acquisition module is used to obtain VOCs concentration data of the population to be distinguished; the VOCs are VOCs in the breath of the population to be distinguished, including C 20 H 23 N3O2、C9H 14 O6, C9H6O4, C3H2N2S, C 16 H 22 N2O6、C 10 H 10 N6O 11 、C6H6IN、C3H4、C 18 H 22 , C 13 H8N2O 12 、C2H3N、C 11 H9N7 and C2H4O; The judgment module uses the VOCs concentration data of the to-be-differentiated population acquired by the data acquisition module as input, and inputs it into the trained machine learning model to judge whether the to-be-differentiated population is a population with abnormal glutaminase or a population with normal glutaminase; The trained machine learning model is a machine learning model trained based on VOCs concentration data in the training set and the test set samples; the VOCs are the VOCs shown in the data acquisition module; the training set and the test set are sample sets containing both people with abnormal glutamine transferase and people with normal glutamine transferase; The result output module is used to output the result obtained by the judgment module.

7. Use of a detection reagent in the preparation of a product for distinguishing people with abnormal glucose levels, characterized in that: The detection reagent is a reagent for detecting the following VOCs; the VOCs are VOCs in exhaled air, including CH2O, C9H 14 O6, C3H2N2S, C 18 H 22 , CH2O2, C4H4SO, C9H6O4, C3H4, C4H2O4, C6H6IN, C2H4O, C 19 H 38 , C2H2O, CHNO, C3H3NO4, C6H4O5, C3H 14 NO 10 、C2H2O3、C 11 H9N7, CHF3, C4H 10 S3, C4H4, C3H4O, C2H4S, C7H 10 O3, C3H4O3, CHO2, C3H4ClN, C8H4O3, C6H2NO 12 、C2H3NO、C3H5Cl、C 13 H 18 S, C6H4N2O4, C 10 H 14 N2O4, C6H4Cl2 and C7H5ClO.

8. A model based on machine learning for distinguishing people with abnormal glucose levels, characterized in that: It includes a data acquisition module, a judgment module and a result output module; The data acquisition module is used to obtain VOCs concentration data of the population to be distinguished; the VOCs are VOCs in the breath of the population to be distinguished, including CH2O, C9H 14 O6, C3H2N2S, C 18 H 22 , CH2O2, C4H4SO, C9H6O4, C3H4, C4H2O4, C6H6IN, C2H4O, C 19 H 38 , C2H2O, CHNO, C3H3NO4, C6H4O5, C3H 14 NO 10 、C2H2O3、C 11 H9N7, CHF3, C4H 10 S3, C4H4, C3H4O, C2H4S, C7H 10 O3, C3H4O3, CHO2, C3H4ClN, C8H4O3, C6H2NO 12 、C2H3NO、C3H5Cl、C 13 H 18 S, C6H4N2O4, C 10 H 14 N2O4, C6H4Cl2 and C7H5ClO; The judgment module uses the VOCs concentration data of the to-be-differentiated population acquired by the data acquisition module as input, inputs it into the trained machine learning model, and judges whether the to-be-differentiated population is a glucose abnormality population or a glucose normal population; The trained machine learning model is a machine learning model trained based on VOCs concentration data in the training set and the test set samples; the VOCs are the VOCs shown in the data acquisition module; the training set and the test set are sample sets containing both people with abnormal glucose and people with normal glucose; The result output module is used to output the result obtained by the judgment module.

9. Use of a detection reagent in the preparation of a product for distinguishing people with abnormal apolipoprotein A, characterized in that: The detection reagent is a reagent for detecting the following VOCs; the VOCs are VOCs in exhaled air, including CH2O, C3H4, CHNO, C2H4O, C2H5OS, C4H4OS, C8H6NOF3, C 12 H 17 NO, C 13 H 18 S.C. 12 H 17 NO2, C3H2N2S, C9H6O4, C9H 14 O6, C 20 H 23 N3O2 and C 16 H 22 N2O6.

10. A model based on machine learning for distinguishing people with abnormal apolipoprotein A, characterized in that: It includes a data acquisition module, a judgment module and a result output module; The data acquisition module is used to obtain VOCs concentration data of the population to be distinguished; the VOCs are VOCs in the breath of the population to be distinguished, including CH2O, C3H4, CHNO, C2H4O, C2H5OS, C4H4OS, C8H6NOF3, C 12 H 17 NO, C 13 H 18 S.C. 12 H 17 NO2, C3H2N2S, C9H6O4, C9H 14 O6, C 20 H 23 N3O2 and C 16 H 22 N2O6; The judgment module uses the VOCs concentration data of the to-be-differentiated population acquired by the data acquisition module as input, and inputs it into the trained machine learning model to judge whether the to-be-differentiated population is a population with abnormal apolipoprotein A or a population with normal apolipoprotein A; The trained machine learning model is a machine learning model trained based on VOCs concentration data in the training set and the test set samples; the VOCs are the VOCs shown in the data acquisition module; the training set and the test set are sample sets containing both people with abnormal apolipoprotein A and people with normal apolipoprotein A; The result output module is used to output the result obtained by the judgment module.

Images