RU2797334C1 - Method of biohybrid screening of lung cancer, stomach cancer, diabetes mellitus and pulmonary tuberculosis by exhaled air - Google Patents

Abstract

FIELD: medicine.

SUBSTANCE: invention relates to the study and analysis of air exhaled by the subject, and can be used in the screening of socially significant diseases (lung cancer, stomach cancer, diabetes mellitus, pulmonary tuberculosis) in order to provide support for medical decision-making. Either one rat is used for screening for lung cancer, stomach cancer, diabetes mellitus and pulmonary tuberculosis, or four rats for each of these diseases separately. Air intake from the subjects in both cases is carried out in sampling bags. The flow rate of air from the sampling bags to the olfactory organ of the rat is 80–160 ml/s. A microelectrode matrix is placed on the rat's head, which is immersed into the olfactory bulb with eight microelectrodes and provides for the removal of bioelectric potentials from the olfactory analyzer of the rat. The processing of the received signals is carried out by decomposing each of the 8 channels using a wavelet transform using a biorthogonal wavelet spline of the 3rd and 1st orders. For each of the obtained five detailed decompositions of the signal, parameters such as root mean square (RMS) and modulo average (SpM) are calculated, as well as the resulting vector for each of the 8 channels. The vectors from all channels are combined into one resulting vector for a signal of 80 parameters, which is fed to the input of the processing block using the method of principal components. Processing is performed in pairs for all channels. The resulting compressed feature vector with dimension 3 is fed to the input of the trained classification model based on the support vector machine, the result of which is the following outcomes: "There is a risk of disease" and "The risk of disease is not detected". A conclusion on the presence/absence of the risk of the disease is issued.

EFFECT: increased efficiency of indicators "Accuracy", "Completeness", a complex statistical indicator "F-measure", "Sensitivity" and "Specificity" during biohybrid screening of lung cancer, stomach cancer, diabetes mellitus and pulmonary tuberculosis by exhaled air.

1 cl, 5 dwg, 4 tbl

Description

Technical field

The invention relates to medicine, namely to the study and analysis of air exhaled by the subject, and can be used in the screening of socially significant diseases (lung cancer, stomach cancer, diabetes mellitus, pulmonary tuberculosis) in order to provide support for medical decision-making.

State of the art

Known patent RU 2659712 C1 (publication date 2018-07-03) "A method for detecting low concentrations of explosives and narcotic substances in the air based on the analysis of bioelectric potentials of the rat olfactory analyzer"), based on the analysis of bioelectric potentials of the rat olfactory analyzer, which consists in implantation in the upper the surface of the olfactory bulb of a rat with a microelectrode matrix with eight working and one reference electrode, recording the electrocorticographic (ECoG) signal of the olfactory bulb in a given frequency range at the time of inspiration, extracting five groups of features expressed as mathematical values of the mean, variance, asymmetry and kurtosis, calculated for amplitudes of the ECoG signal of each lead, the first derivative of the amplitudes of the ECoG signal of each lead, the frequency-amplitude spectrum of each lead, the cross-correlation coefficients of the amplitudes of the ECoG signal between the leads, the cross-correlation coefficients of the frequency-amplitude spectrum between the leads, the processing of each group of features of a separate multilayer neural network (MNS), while for training each MNS, an array of odor indicators is additionally formed, presented to the rat a given number of times with a given pause duration, which is the output array for learning the MNS, when learning for each MNS, classification weights are calculated according to the backpropagation algorithm and when odor identification, five groups of signs of the ECoG signal process five MNS and calculate the probabilities that the odor presented to the rat belongs to one of the specified odors for each group of signs, calculate the arithmetic mean of the probabilities from five MNS with the choice of odor with the maximum probability as a recognition result.

Known international application WO2017223489 A1 "BIO-ELECTRIC NOSE", (publication date 2017-12-28). The disclosure of the claimed invention generally relates to systems and methods for detecting ligands. According to the description of the application, the bioelectric odor detector has a very high sensitivity compared to electrochemical detectors and exceeds it by 4, 5 or 6 orders of magnitude.

One of the embodiments of the invention relates to a chemical sensor containing: a system of sensors (a matrix of electrodes) and a biological detector - the glomerular complex of the animal's olfactory bulb.

The first implementation component relates to a method for determining the concentration of a given chemical, including: binding the chemical to the olfactory receptor located in the olfactory sensory neuron in the epithelium of the nose, and projecting its axon to the olfactory bulb (for example, the olfactory bulb of a normal or genetically modified mammal), detection electrochemical change in the olfactory bulb; and determining the concentration of the chemical based on the detected electrochemical change.

The second implementation component relates to a method for determining the concentration of a mixture of chemicals at various concentrations, including: binding a chemical to an olfactory neuron; detection of an electrochemical change in the olfactory bulb; and determining compounds and chemical concentration based on the detected electrochemical change.

The third implementation component relates to a method for distributed odor mapping of areas, including: providing a bioelectrical activity recording system, wireless communication, GPS, and a chemical sensor in communication with each other; detection of odors by freely moving organisms and recording information about the position of organisms; and mapping the odor of the target area based on the recorded information.

The fourth implementation component relates to a system for driving a vehicle based on odor detection, comprising: linking a vehicle, a control system, an odor and concentration classifier, and said chemical sensor; vehicle approach to odor sources based on the odor processing algorithm and concentration gradients detected by the chemical sensor.

The fifth implementation component relates to a system for non-invasive disease diagnosis, where a disease marker is chemically detected by the system in a sample of inhaled air, tissue, body fluid, feces, urine, or sweat.

The sixth implementation component relates to a transgenic non-human mammal having a gene sequence encoding targeted or modified expression of one or more olfactory receptor genes such that they bind to specific regions of the olfactory bulb to optimize use in the brain-machine interface.

The seventh implementation component relates to a method for creating a chemical sensor, including: modifying the genetic sequence of an organism; directing the expression of olfactory receptors with specificity for the selected chemical; creation of ectopic or multiple dorsal glomeruli having specificity for the selected chemical; and placement of the electrode array adjacent to the dorsal glomeruli.

The eighth implementation component to the method and system containing a receptor mutated to this odorant to increase the affinity for this chemical substance of interest for inclusion in the system is a bioelectric nose.

Known patent CN108760829 "ELECTRONIC NOSE RECOGNITION METHOD BASED ON BIONIC OLFACTORY BULB MODEL AND CONVOLUTIONAL NEURAL NETWORK" (publication date 2018-11-06). The invention relates to a bioelectronic nose recognition method based on a bionic olfactory bulb model and a convolutional neural network. The method includes: selecting an object for recognition using an electronic platform of the nose to obtain a data set S; construction of a bionic olfactory bulb model, in which a bionic olfactory bulb model is formed by connecting a plurality of olfactory glomerulus models, while the number of olfactory glomerulus models in the bionic olfactory bulb model is the same as the number of electronic olfactory sensors, each olfactory bulb glomerulus model is formed by connecting four main neuron models , and the four main neuronal models, respectively, are the olfactory receptor, mitral cell, granule cell, and olfactory pericyte of the glomerulus; inputting a sufficient data set into the bionic olfactory bulb model using the olfactory receptor and processing to obtain a new multivariate pulse time series data set; carrying out data normalization processing; obtaining the corresponding grayscale dataset; definition of a convolutional neural network model and its training. In this method, automatic feature extraction can be achieved, end-to-end learning, and the versatility of the electronic nose recognition algorithm can be improved.

From the analysis of the prior art, it follows that results have been achieved in increasing the level of sensitivity of the bioelectric odor detector in comparison with electrochemical detectors by 4, 5 or 6 orders of magnitude. A method for recognition by a bioelectric nose has been developed, which is based on models of a bionic olfactory bulb and a neural network. A method has been developed based on the analysis of bioelectric potentials of the olfactory analyzer of a rat, with the help of which it is possible to detect low concentrations of explosives and narcotic compounds in the air.

The prototype of the present invention was patent RU 2666873 C1 (publication date: 2018-09-12). "Method for diagnosing lung cancer by analyzing the air exhaled by the patient based on the analysis of bioelectric potentials of the olfactory analyzer of the rat"), based on the analysis of the bioelectric potentials of the olfactory analyzer of the rat. The invention relates to the study and analysis of gaseous biological materials, in particular, respiratory products, and can be used to diagnose lung cancer in humans using the registration of a bioelectric signal from the olfactory bulb of a rat. The technical result of the invention is to increase the accuracy of disease recognition by the air exhaled by the patient by separating the system response into positive, associated with the presence of a lung cancer biomarker, and negative, associated with the absence of a lung cancer biomarker, and simplifying the method by eliminating the training of a rat to recognize substances in exhaled air . The first level of processing the bioelectric signal of the olfactory bulb of the rat contains blocks of wavelet transforms of each channel of the eight-channel bioelectric signal with the subsequent calculation of root-mean-square and modulo-average values for each feature introduced in the described patent. The second level contains a processing block using the principal component method, where the channel vectors are processed in pairs, the result of the work is a feature space of a lower dimension than the original signals. The third level is the actual classifier based on a machine learning model using the support vector machine.

Registration of the bioelectric signal of the olfactory bulb of the rat is carried out in the frequency range of 1-250 Hz.

The disadvantages (prototype) of the presented diagnostic method were insufficient throughput, the risk of infection with various diseases, including a new coronavirus infection (COVID-19), when the patient exhales air into a tube through a funnel installed in front of the rat's nose. Due to the fact that the patient exhales air unevenly through the tube to the rat, there is a possibility of a decrease in the accuracy of recognition of the air exhaled by the patient, for a positive or negative reaction of the rat, for the presence or absence of a lung cancer biomarker. In addition, the disadvantage of the processing method is the use of the Fourier transform for the analysis of a non-stationary signal. This leads to the problem of spectrum spreading, in addition, this type of signal requires a clear localization of events in the time domain. The disadvantages also include the use of a multilayer convolutional neural network, since this method requires significantly larger training samples and data diversity than simpler machine learning models such as the support vector machine.

The purpose and technical result of the invention is to increase the effectiveness of the indicators "Accuracy", "Completeness", a complex statistical indicator "F-measure", "Sensitivity" and "Specificity" in biohybrid screening of lung cancer, stomach cancer, diabetes mellitus and pulmonary tuberculosis by exhaled air being examined.

Disclosure of invention

The goals and technical result are achieved by the fact that biosensors are used in a paradigm: either one biosensor for the nosologies lung cancer, stomach cancer, diabetes mellitus and pulmonary tuberculosis, or biosensors for each of the nosologies lung cancer, stomach cancer, diabetes mellitus and pulmonary tuberculosis separately, sampling air from the subjects is carried into sampling bags, the air flow rate from the sampling bags to the olfactory organ of the biosensors is 80-160 ml/s, signal processing and conclusions are determined by decomposing the signals of each of the 8 channels to remove bioelectric potentials using wavelet transform, a biorthogonal spline wavelet of the 3rd and 1st orders is used, for each of the 5 detailed decompositions obtained, 2 parameters are calculated: root mean square (RMS) and modulo average (SpM), the resulting vector for each channel will be 10 values of RMS and SpM , the resulting vector for the signal consists of 80 parameters fed to the input of the processing block using the principal component method (processing is performed in pairs for all 8 channels with bioelectric signals), the resulting compressed feature vector with dimension 3 is fed to the input of the trained classification model based on the support vector machine.

The claimed method is carried out as follows.

The air exhaled by the subject is collected in a sampling bag 1 (Fig. 1) by means of an automatic deflation system 2, uniformly supplied to the olfactory organ of the biosensor 3 with an air flow rate of 80-160 ml/s from the sampling bag. On the head of the biosensor 3 there is a microelectrode matrix 4, which is immersed in the olfactory bulb (OL) with eight microelectrodes and is used to remove bioelectric potentials; one channel for monitoring the physiological parameters of the biosensor, one channel - a mark for the start of air flow from the sampling package (the total number of channels is 10). The bioelectric potentials of the olfactory analyzer of the biosensor taken from the microelectrodes are converted by an amplifier 5, an analog-to-digital converter (ADC) 6 and enter the signal processing and conclusion output unit (BOS and VZ) 7 with the help of which the received data is processed, and as a result, on the computer screen information about the presence/absence of the "Risk of the disease" is displayed.

The block diagram of preparing the biosensor(s) (BS) for operation and biohybrid screening for nosologies: lung cancer (LC), stomach cancer (GC), diabetes mellitus (DM) and pulmonary tuberculosis (TL) (Fig. 2) is implemented in the paradigm : either one biosensor for the nosologies lung cancer, stomach cancer, diabetes mellitus and pulmonary tuberculosis, or biosensors for each of each nosology lung cancer, stomach cancer, diabetes mellitus and pulmonary tuberculosis separately.

The work of the cycle - "Preparation of the biosensor(s) for work" is carried out as follows. One biosensor is being prepared for the nosologies of lung cancer, stomach cancer, diabetes mellitus and pulmonary tuberculosis, or four biosensors for each of the nosologies of RL, GC, DM, TL. The cycle is divided into two steps.

The first step is to control the functional state of the biosensor. It is carried out on the basis of the analysis of the registered breath channel. After the introduction of the biosensor into anesthesia, the measured breathing signal is used to calculate the respiratory rate (RR) in real time, on the basis of which a conclusion is made about the possibility of using the biosensor. Excursions of the animal's chest are recorded by the pressure sensor 24PC01SMPT. The signal from the pressure sensor is amplified by an instrumental amplifier and fed to the input of an analog-to-digital converter. The received data is filtered to highlight the operating frequency range of 0.5 - 5 Hz.

Next, the moving average of the signal is calculated using the formula:

Where

T - averaging interval,

X _i - current pressure value,

U - the current value of the moving average,

U _i - the current value of the signal.

To determine the rising edge of the respiration signal phase, the moment when the current pressure value X _i exceeds the value of the moving average pressure is determined

by threshold with value

To determine the falling edge of the phase of the respiration signal, the moment when the current value of pressure X _i becomes less than the value of the moving average is determined

threshold pressure

Here

_РН - the upper threshold for the selection of the rising front of the respiratory phase,

P _L is the lower threshold for detecting the falling edge of the respiration phase.

The respiratory rate F _i is determined through the time interval T _i between the rising phases of the respiration signal.

When the respiratory rate reaches 60-70 per minute and there are no sharp deviations from these values, both upwards, which corresponds to the presence of frequent sniffing and the biosensor comes out of anesthesia, and downwards, which can signal a respiratory arrest, the biosensor is considered to be operational.

The second step is to implement 3 series of database set and train the classifier in a 50/50 ratio (50% of patients diagnosed with LC, GC, DM, TL and 50% of apparently healthy volunteers, which makes it possible to achieve the target value of the F-measure ( the sufficiency of the volume and completeness of the generated base), settings of the Signal Processing Unit and conclusions.

The signal processing and conclusion output block contains several modules:

1. The data acquisition module interacts through the file system with the ADC program for data from the amplifier. The data is written to the BOS and VZ file system.

2. Data preprocessing and storage module. The application monitors file system directory changes. When a new file appears, data blocks are read from it. When record label 1 appears in the file in channel 10, data blocks are constantly read from the current position in one second, processed and the results are accumulated in the BFB and VZ. After the appearance of record label 0 in channel 10, the distributions of classifier responses for each one-second block are compared for the current record and for the calibration record. Such an implementation is necessary to ensure real-time operation.

3. The classification module implements data processing methods using a classifier trained in advance within the framework of the model.

4. Module for displaying the results of the classifier on the laptop screen, generating final reports containing the classification results, and saving to the laptop memory.

The main component of the signal processing and conclusion output block is a classifier based on machine learning technologies. The architecture and operation of the classifier are described below.

The successive processing steps of the classifier signal are shown in Fig. 3:

1. The wavelet transform makes it possible to obtain for each channel a given number of signal decompositions consisting of signal representation coefficients (projections onto a new orthogonal basis of functions). The number of decompositions can be chosen empirically (by estimating the energy in each decomposition and comparing with a given threshold), or based on signal properties. In this case, the second approach was chosen - an estimate of the number of decompositions based on the properties of the signal. Each decomposition makes it possible to obtain signal components on a certain frequency scale (determined by the sampling rate of the original signal), and at each step the frequency of the original signal is divided by 2 (due to the properties of the discrete wavelet transform itself, as shown in [1]). The processed signal, according to [2], has a maximum energy in the range of 10 - 125 Hz, therefore, the 5th decomposition with a sampling frequency of 15.625 Hz (since the original signal has a sampling frequency of 500 Hz, this value is obtained after a 5-fold division 2, in the process of obtaining signal decompositions) allows analyzing the signal in the band of 7.8125 Hz (according to the Kotelnikov theorem), subsequent decompositions reflect low-frequency oscillations of the signal isoline level. In this case, a biorthogonal spline wavelet of the 3rd and 1st orders (bior3.1) was used. In this case, the approximating decomposition is excluded from the obtained decompositions to eliminate the low-frequency trend (Fig. 3, blocks 8 and 9).

where ψ(t) is the selected wavelet basis bior3.1, s ₁ (t)…s ₈ (t) are signal samples from channels 1 to 8,

С1…С8 - resulting arrays of approximating and detailing coefficients for 5 decompositions. Only detailed coefficients are processed further.

2. For each of the 5 detailed decompositions obtained, 2 parameters are calculated for each channel: root mean square (RMS) and modulo average (Am) (Fig. 3, block 10).

The arrays of coefficients for each processed channel are:

where Cd ₁ ...Cd ₅ - detail coefficients for decompositions 1-5.

For each array, RMS and SpM values are calculated for each decomposition using the formulas:

where N is the total number of detail coefficients in the decomposition.

The result of calculations for each channel will be vectors of the form:

Next, the value of the variable i (channel number) is incremented and a check is made for the excess of the number of available channels (Fig. 3, blocks 11 and 12). If the channel is not the last one, a return is made to the step of calculating the parameters (Fig. 4, block 10). If the last channel is processed, a transition is made to the joint processing of the received parameters (Fig. 4, block 13).

3. The resulting vector for each channel will be 10 RMS and SpM values.

4. Vectors from all channels are combined into one vector of 80 parameters, which is fed to the BOS and VZ using the principal component method. This method allows you to reduce the dimension of data with the least loss of information.

Processing is performed in pairs for all channels, with the reduction of the space dimension to 3 (Fig. 3, block 13).

To reduce the dimension, we select the vectors of the following pairs of channels (indicated by numbers):

C13, C24, C35, C46, C57, C68, C15, C26, C37, C48.

Then, for each vector of a pair of channels, the dimension is reduced through expansion by the method of principal components [3]. The given resulting dimension is 3. The dimension of the feature vector was chosen based on obtaining the best results of the classification model, taking into account acceptable computational complexity. When training a classification model with 4 or more inputs, no increase in accuracy was found, however, the computational complexity increases according to [9]:

At the same time, it is impossible to reduce the size of the training sample for the model - this leads to an increase in the probability of retraining the model.

5. The resulting compressed feature vector is fed to the input of the classification model based on the support vector machines (support vector machines [4]) (Fig. 3, block 14).

With the help of this classification model (this classifier), in the process of training the model, a hyperplane is built that divides the input data into two classes.

The result of the classification model is two outputs: "There is a risk of disease" and "Risk of the disease is not detected" (Fig. 3, block 15).

The work of the second cycle "Biohybrid screening" (Fig. 2) is implemented in the paradigm: either one biosensor for nosology (RL), (RJ), (SD), (TL); or, four biosensors for each of the nosologies of RL, RG, DM, TL. The cycle consists of three steps, the first of which is the intake of air from the subject into the sampling bag (air intake from the subjects is not shown in the figures). The second step is the control of the functional state (implemented according to the example of the cycle “Preparing the biosensor (s) for operation”, Step 1 (described above)). The third step is to conduct a biohybrid screening, issuing conclusions on the presence / absence of the “Risk of the disease”.

Brief description of the figures

Fig. 1. Block diagram of a device for implementing the proposed method.

Fig. Fig. 2. Block diagram of biosensor(s) (BS) preparation for operation and biohybrid screening by nosologies: lung cancer (LC), gastric cancer (GC), diabetes mellitus (DM) and pulmonary tuberculosis (TL).

Fig. 3. Successive stages of classifier processing.

Fig. 4. Dependence of throughput on the option of sampling the air exhaled by the examined.

Fig. 5. Dependence F - measures on the air flow rate from the sampling bag.

An example of the use (implementation) of the proposed method for each nosology: lung cancer, stomach cancer, diabetes mellitus and pulmonary tuberculosis.

Test design of the method for biohybrid screening of lung cancer, stomach cancer, diabetes mellitus and pulmonary tuberculosis by the exhaled air of the examined air: one-stage (on-line).

In total, 106 volunteers took part in testing the method of biohybrid screening of lung cancer, stomach cancer, diabetes mellitus and pulmonary tuberculosis by exhaled air. Of these, 58 volunteers participated in the test of the biohybrid screening method for lung cancer and stomach cancer, 24 volunteers participated in the tests of the biohybrid screening method for type II diabetes mellitus, and 24 volunteers participated in the tests of the biohybrid screening method for pulmonary tuberculosis. The sampling of exhaled air from all volunteer subjects was carried out using disposable sampling bags with a volume of 5 l.

Of the 58 volunteer subjects who took part in trials of the biohybrid screening method for lung cancer and stomach cancer:

- conditionally healthy - 50 people (control group);

- with a diagnosis of lung cancer - 5 patients in the early stages of the disease (T1N0M0, T1N0M0, T1N0M0, T2N0M0, T2N1M0);

- with a diagnosis of stomach cancer - 3 patients (T3N1M0, T3N2M0, T3N0M1).

The average age of apparently healthy volunteers was 59.4±8.1 years. The average age of patients with diagnosed oncological diseases was 60.8±4.0 years.

Upon completion of testing the method of biohybrid screening of lung cancer and stomach cancer and the formation of a conclusion about the presence or absence of risk of these diseases, 30 out of 50 apparently healthy volunteers underwent a medical examination (compliance was 60%). Of these, 11 people were examined using both spiral computed tomography of the chest (SCT) and esophagogastroduodenoscopy (EGDS); 14 - only with the use of SRCT and 4 - only with the use of endoscopy.

Confirmation of the risk of lung cancer in apparently healthy volunteer subjects was assessed in accordance with the Lung Imaging Reporting and Data System (LungRADS™) [10] (the risk of lung cancer was considered confirmed at levels LR=2 and above), as well as using the recommendations of the Fleischner Society, 2017 [11].

Confirmation of the risk of gastric cancer in apparently healthy volunteer subjects was assessed in accordance with international guidelines for the treatment of precancerous conditions and changes in the stomach (MAPS II) [12] and Clinical guidelines of the Russian Ministry of Health: Gastric cancer [13].

Used:

- Recommendation 1 (MAPS II): Patients with chronic atrophic gastritis or intestinal metaplasia are at risk of developing gastric adenocarcinoma (high level of evidence).

- Recommendation 2 (MAPS II): histologically confirmed intestinal metaplasia is the most reliable marker of atrophy of the gastric mucosa (high level of evidence).

- Recommendation (MAPS II): Patients with extensive intestinal metaplasia, as well as persistent H. pylori infection or incomplete intestinal metaplasia, or especially those with a family history of gastric cancer in the next of kin, should have more frequent endoscopic surveillance to rule out risk gastric cancer (moderate level of evidence).

Of the 24 volunteer subjects who took part in the trials of the biohybrid screening method for diabetes mellitus:

- conditionally healthy - 20 people (control group), in which the concentration of glucose in the venous blood on an empty stomach after the test was within the normative values of this indicator;

- with a diagnosis of "diabetes mellitus type II" - 4 patients who had an increased concentration of glucose in the venous blood on an empty stomach after the test: 6.8; 8.2; 10.6 and 21.7 mmol/l.

Of the 24 volunteers who took part in the trials of the biohybrid screening method for pulmonary tuberculosis:

- conditionally healthy - 20 people (control group), who underwent fluorography/X-ray examination of the chest organs 1-3 months before the test and received the conclusion: "Chest organs without pathological changes";

- 4 patients diagnosed with "Pulmonary tuberculosis (BC "-")".

Indicators:

- "accuracy" (formulated according to the results of testing the method of biohybrid screening of the conclusion),

- "completeness" (extractions from the total sample of volunteer subjects with established diagnoses and individuals at risk of diseases),

- a complex statistical indicator "F-measure", determined depending on the accuracy of the method and the completeness of extracting target indicators from the total sample,

- the sensitivity of the method,

- the specificity of the method,

- predictive value of a negative result (conclusions about the presence or absence of a risk of the disease),

- the predictive value of a positive result is calculated according to the formulas given in [14,15,16,17].

The risk of systematic errors in testing the method of biohybrid screening for lung cancer, stomach cancer, diabetes mellitus and pulmonary tuberculosis by exhaled air was assessed as "low" (10 points on the QUADAS scale) [18].

The claimed method of biohybrid screening of lung cancer, stomach cancer, diabetes mellitus and pulmonary tuberculosis by exhaled air examined eliminates the disadvantages of the prototype, namely, through the use of sampling packages, the possibility of infection with various diseases (including new coronavirus infection (COVID-19)) is eliminated , the throughput increases (Fig. 4). The use of an automatic deflation system makes it possible to achieve a uniform supply of air to the olfactory organ of the biosensor with an air flow rate of 80-160 ml/s from the sampling bag (see Fig. 5, where the boundaries of the air flow interval are denoted by the letters A and B, and the F-measure has the maximum value ), which in turn increases the accuracy of identifying the risk of diseases (RL, RG, DM, TL). In addition, the shortcomings of the processing method were solved, namely, the use of the Fourier transform for analyzing a non-stationary signal, due to signal processing by decomposing each of the 8 channels using a wavelet transform, using a biorthogonal wavelet spline of the 3rd and 1st orders , for each of the obtained 5 detailed decompositions of calculating 2 parameters: root mean square (RMS) and mean modulo (SpM), the resulting vector for each channel is 10 RMS and SpM values, the resulting vector for the signal consists of 80 parameters fed to the input of the block processing using the method of principal components (processing is performed in pairs for all channels). In addition, the disadvantages of the classification model, a multilayer convolutional neural network, are solved by feeding the resulting compressed feature vector with dimension 3 to the input of the trained classification model based on the support vector machines.

Information sources.

1. Chibli Mallat, Stephane Mallat. A Wavelet Tour of Signal Processing. - Elsevier Science, 1999, 637 pp.

2. Patent RU 2666873 C1, registration date: 09/12/2018 "Method for diagnosing lung cancer by analyzing the air exhaled by the patient based on the analysis of bioelectric potentials of the olfactory analyzer of the rat", authors Medvedev Dmitry Sergeevich, Kiroy Valery Nikolaevich, Ilinykh Andrey Sergeevich, Shepelev Igor Evgenievich, Alexey Evgenievich Matukhno, Alexei Borisovich Smolikov, Vladimir Vasilievich Zolotukhin, Nadezhda Ruslanovna Minyaeva.

3. Jolliffe I.T. Principal Component Analysis, Series: Springer Series in Statistics, 2nd ed., Springer, NY, 2002, XXIX, 487 p.28 illus.

4. Alex J. Smola, Bernhard Scholkopf. A tutorial on support vector regression // Statistics and Computing 14 // 2004 // pp 199-222.

5. Patent RU 2659712 C1, registration date: 07/03/2018 "A method for detecting low concentrations of explosives and narcotic substances in the air based on the analysis of bioelectric potentials of the olfactory analyzer of a rat", authors Medvedev Dmitry Sergeevich, Kiroy Valery Nikolaevich, Ilinykh Andrey Sergeevich, Shaposhnikov Dmitry Grigorievich, Alimov Ruslan Rustamovich, Vdovyuk Artem Vladlenovich, Zolotukhin Vladimir Vasilievich, Matukhno Alexey Evgenievich, Smolikov Alexei Borisovich, Stadnikov Evgeniy Nikolaevich, Minyaeva Nadezhda Ruslanovna

6. Patent No. US 20050065446 A1 - Methods of collecting and analyzing human breath. Application date: 29.01.2002. Patent publication date: 03/24/2005. Authors: Talton; James D. (Gainesville, FL).

7. Patent No. 10,455,817 B2 - Animal olfactory detection of disease as control for health metrics collected by medical toilet. Application date: 04.10.2016. Effective date: 04/05/2018. Patent publication date: 10/29/2019. Authors: Hall D.R., Fox J., Pearman T. United States Patent.

8. BIO - ELECTRIC NOSE Pub. No.: US 2019/0227053 Al Jul. 25, 2019 Inventors: Dmitry RINBERG, New York, NY (US); Erez SHOR, New York, NY (US); Thomas BOZZA, New York, NY (US).

9. Ulyanov M.V. Resource-efficient computer algorithms. Development and analysis. Tutorial. M.: SCIENCE, FIZMATLIT, 2007. -376 p.

10. Application of the LUNG-RADS system in lung cancer screening. Methodological recommendations No. 3. Department of Health of Moscow / Nikolaev A.E., Blokhin N.A., Gonchar A.P. and others // Series "Best" practices of radiological and instrumental diagnostics". Issue 34. - M., 2020. - 22 p.

11 MacMahon et al. Guidelines for Management of Incidental Pulmonary Nodules Detected on CT Images: From the Fleischner Society 2017 // Radiology (2017) DOI10.1148/radiol.2017161659. Quoted from URL. http://24radiology.ru/grudnaya-kletka/rekomendatsii-obshhestva-fleischner/ (date of access: 20.08.2021).

12. Pimentel-Nunes P., Libanio D., Marcos-Pinto R. et al. Management of epithelial precancerous conditions and lesions in the stomach (MAPS II): European Society of Gastrointestinal Endoscopy (ESGE), European Helicobacter and Microbiota Study Group (EHMSG), European Society of Pathology (ESP), and Sociedade Portuguesa de Endoscopia Digestiva (SPED) ) guideline update 2019 // Endoscopy. 2019; 51.

13 URL. https://cr.minzdrav.gov.ru/recomend/574_1 (Accessed 20.08.2021).

14. David M.W. powers. Evaluation: From Precision, Recall and F-factor to ROC, Informedness, Markedness and Correlation // Technical Report SIE-07-001 December 2007// School of Informatics and Engineering Flinders University, Adelaide, Australia.

15.D.G. Altman, J.M. Bland. Statistics Notes: Diagnostic tests 1: sensitivity and specificity // BMJ // 1994 // 308:1552 doi: 10.1136/bmj.308.6943.1552.

16.D.G. Altman, J.M. Bland. Diagnostic tests 2: Predictive values // BMJ // 1994 Jul 9//309(6947): 102.

17. Nasledov A.D. Mathematical methods of psychological research. - St. Petersburg: Speech, 2004.

18. Rebrova O.Yu., Fedyaeva V.K. Assessing the risk of bias in cross-sectional studies of diagnostic tests: Russian version of the QUADAS questionnaire. Medical technologies. Evaluation and choice. - 2017. - no. 1. - S. 11-14.

Claims (7)

A method for biohybrid screening of lung cancer, stomach cancer, diabetes mellitus and pulmonary tuberculosis by exhaled air examined, which consists in using either one rat for screening lung cancer, stomach cancer, diabetes mellitus and pulmonary tuberculosis, or four rats for each of the listed diseases separately; while the air intake from the subjects in both cases is carried out in the sampling bags and the air flow from the sampling bags to the olfactory organ of the rat is 80-160 ml/s; a microelectrode matrix is placed on the head of the rat, which is immersed in the olfactory bulb with eight microelectrodes and provides for the removal of bioelectric potentials of the olfactory analyzer of the rat; the processing of the received signals is carried out by decomposing each of the 8 channels using a wavelet transform, while using a biorthogonal spline wavelet of the 3rd and 1st orders, for each of the five detailed decompositions of the signal obtained, parameters such as the mean square value (RMS ) and mean modulo (SpM), the resulting vector for each of the 8 channels has the following form:

Where

- detail coefficients for decompositions 1-5, vectors from all channels are combined into one resulting vector for a signal of 80 parameters, which is fed to the input of the processing block using the principal component method, and processing is performed in pairs for all channels; the resulting compressed feature vector with dimension 3 is fed to the input of the trained classification model based on the support vector machine, the result of which is the outputs "There is a risk of disease" and "Risk of the disease is not detected", and a conclusion is issued about the presence/absence of the risk of the disease.

Images