CC BY
14DEEP LEARNING FOR HUMAN BREATH RESEARCH USING INFRARED QUANTUM CASCADE LASER
SPECTROSCOPY
IGOR FUFURIN1, PAVEL BEREZHANSKIY2, IGOR GOLYAK1, DMITRIY ANFIMOV1, ANASTASIYA SCHERBAKOVA1, PAVEL DEMKIN1, OLGA NEBRITOVA,1 AND ANDREY MOROZOV1
1Department of Physics, BMSTU, Russian Federation
2Morozov Children's Clinical Hospital, State Budgetary Healthcare Institution, Moscow Healthcare Pulmonology
Department, Moscow, Russia igfil@mail.ru
ABSTRACT
Nowadays, approximately 10.5% of the world's population aged 20-79 years live with diabetes. Express analysis of a number of socially significant diseases, including diabetes mellitus, is currently an urgent task. In the present paper we describe method based on infrared laser spectroscopy using a quantum cascade laser emitting in pulsed mode with a peak power of up to 150 mW in the spectral range of 5.3-12.8 microns, and a multi-pass Herriot gas cell with an optical path length of 76 m. According to the results of the analysis of 1200 infrared spectra of exhaled air in 60 healthy volunteers (control group) and 60 volunteers with confirmed type 1 diabetes mellitus (T1DM) using a one-layer convolutional neural network, the results of classification accuracy 99.7%, recall 99.6% and the area under the AUC curve 99.9% are shown. Additionally, 30 volunteers with bronchial asthma were examined. The use of the support vector method gives an accuracy value of about 95% when classifying volunteers into 3 groups (healthy, type 1 diabetes, asthma).
Noninvasive diagnostics is one of the most important directions of development of modern medicine. Worldwide, there are 537 million adults aged 20-79 years suffer from diabetes in 2021. According to estimates of the International Diabetes Federation (IDF), in 2021 the number of children (0-19 years old) and T1DM teenagers will be about 1.2 million. According to forecasts, by 2030 this number will increase to 643 million, and by 2045 to 783 million [1]. The possibility of analyzing glucose levels by skin secretions and in human respiration is shown [2,3]. Recently, an area related to the analysis of volatile organic compounds released from the human body has been actively developing
[4]. It has been shown that such compounds are isolated from exhaled breath (872 compounds), saliva (359 compounds), blood (154 compounds), milk (256 compounds), skin secretions (532 compounds), urine (279 compounds) and feces (381 compounds). A number of volatile organic compounds represent biomarkers of a specific human disease. Thus, the average concentration of acetone in healthy breathing ranges from 293 to 870 ppb, and ethanol — from 27 to 153 ppb
[5]. For diabetes mellitus patients the average concentration of acetone may exceed 1800 ppb [6]. It was shown [7] that the measurement of the level of exhaled carbon monoxide (eCO) is applicable for assessing the condition of infants and toddlers with stable asthma and during an acute asthma attack. In [8-10] the application of machine learning methods for analyzing the spectra of multicomponent gas mixtures, including human exhaled air is shown. In [11] the possibilities of using deep learning for the diagnosis of type 1 diabetes mellitus by infrared spectra of exhaled air is shown.
Our method is based on the analysis of volunteers' exhaled breath. Figure 2 shows the basic principle of the developed diagnostic method. An infrared laser spectroscopy to analyze human breath was proposed. The breath sample is collected in a Urine Bag ST 1300102 (Meridian, Moscow, Russia), it is then passed through a Nafion dryer and placed into a Herriot multipass gas cell. IR radiation is emitted by an external cavity quantum cascade laser then it enters to the gas cell and after the required number of reflections is collected at the photodetector. The measured IR spectrum undergoes preprocessing procedures and then comes to the convolutional neural network. A neural network trained on the control and the target groups can classify healthy and T1DM volunteers by their IR breath spectra. The two mass flow controllers (MFC) type FC-201CV and GE50A (Bronkhorst High-Tech B.V., Bronkhorst, The Netherlands), the pressure controller P-602CV (Bronkhorst High-Tech B.V., Bronkhorst, The Netherlands), and the vacuum pump MVP 015-2 DC (Vacuumbrand GMBH and CO KG, Wertheim, Germany) with pressures up to 3.5 mbar are used. The normal operating pressure is approximately 500 mbar. The exhaled breath must be dehydrated after collection and for this purpose a Nafion gas dryer MD series (Perma Pure LLC, 197 Lakewood, NJ, USA) is used. The pure nitrogen with a flow rate about 40 standard cubic centimetres per minute (sccm) to dry the breath sample and a flow rate of about 20 sccm to place the breath sample into a pre-vacuumed gas cell is used.
Figure 1. Basic scheme for breath sample analysis method
In the present paper, the shallow Convolutional Neural Network (CNN) that is a well known deep learning architecture inspired by the natural visual perception mechanism of living organisms is used. Figure 2 shows the shallow convolutional neural network that was created in this work. The proposed CNN model contains an input layer, a single convolutional layer, a max-pooling layer, a fully connected MLP layer (FCL), and the output layer. Dataset consists of 600 spectra of healthy volunteers and 600 T1DM volunteers. Results of CNN performance is shown in table 1.
Convolution layer
Input layer
Convolution
(Rei.U) : Pooling layer ^"V
connected layer (RcLL")
Output layer
(So Umax)
128x1
%lxl
Figure 2. Scheme of the shallow Convolutional Neural Network Table 1. Results (%) of CNN performance for described dataset
Accuracy Precision Recall
99.7 99.5 99.6
Of additional interest is the possibility of dividing patients into several classes at the same time. This approach will significantly expand the capabilities of the method and the interest of potential consumers. For this purpose, 30 patients suffering from asthma were additionally studied, so the data set consisted of 60 healthy volunteers (600 spectra),
60 volunteers suffering from diabetes mellitus (600 spectra) and 30 volunteers suffering from bronchial asthma (300 spectra). Machine learning (ML) methods were used to separate volunteers by IR spectra of exhaled air. Results of PCA and t-SNE methods performance for our dataset (3 classes — healthy, diabetes, asthma) is shown in Figure 3.
Principal Component Analysis for diseases
. * *« a *
„.J 0 o
• - . * V _ o
-
° e =t>
o •
o o °
-80 -60 -40 -20 0 20 40 60 SO PCI (Q.6B)
• Healthy
Q Diabetes
• Astma
t-SNE method for diseases
yJ ■ * * o • ■ • •
• o • 0 • o
o • • » !i • I ° * • o • * m • . • o * o • • o *
rt o v • • • o • • o ív;. • O O O O 0 O O o
O O o o *¡ o o »o*.
• o • Healthy Diabetes Astma • o • . Ö o 0 6 ° ■ é
-75 -50 >5 ~~ 25 50 75
Dimension 1
Figure 3. ML methods for dividing volunteers into 3 classes
Support vector machine method for our dataset gives accuracy no less then 0,95 for dividing dataset into 3 classes. Results for SVM method is presented in table 2.
Table 2. SVM method results for described dataset
Class Precision Recall F1-score
Healthy 0.90 1.0 0.95
Diabetes 1.0 0.88 0.93
Asthma 1.0 1.0 1.0
The work was carried out within the framework of the strategic academic leadership program "Priority 2030", approved by the Decree of the Government of the Russian Federation No. 729 dated May 13, 2021.
REFERENCES
[1] International Diabetes Federation. IDF Diabetes Atlas, 10th ed.; International Diabetes Federation: Brussels, Belgium, 2021.
[2] Bayrakli I. Breath analysis using external cavity diode lasers: a review // Journal of Biomedical Optics. SPIE-Intl Soc Optical Eng, 2017. Vol. 22, № 4. P. 040901.
[3] Turner C. Potential of breath and skin analysis for monitoring blood glucose concentration in diabetes // Expert Review of Molecular Diagnostics. Informa UK Limited, 2011. Vol. 11, № 5. P. 497-503.
[4] de Lacy Costello B. et al. A review of the volatiles from the healthy human body // Journal of Breath Research. IOP Publishing, 2014. Vol. 8, № 1. P. 014001.
[5] Diskin A.M., pan l P., Smith D. Time variation of ammonia, acetone, isoprene and ethanol in breath: a quantitative SIFT-MS study over 30 days // Physiological Measurement. IOP Publishing, 2003. Vol. 24, № 1. P. 107-119.
[6][ DENG C. et al. Determination of acetone in human breath by gas chromatography-mass spectrometry and solid-phase microextraction with on-fiber derivatization // Journal of Chromatography B. Elsevier BV, 2004. Vol. 810, № 2. P. 269-275.
[7] Ohara Y. et al. Exhaled carbon monoxide levels in infants and toddlers with episodic asthma // FUKUSHIMA JOURNAL OF MEDICAL SCIENCE. The Fukushima Society of Medical Science, 2020. Vol. 66, № 2. P. 78-87.
[8] Fufurin I.L. et al. Numerical techniques for infrared spectra analysis of organic and inorganic volatile compounds for biomedical applications // Optical Engineering. SPIE-Intl Soc Optical Eng, 2021. Vol. 60, № 08.
[9] Golyak I.S. et al. Numerical methods of spectral analysis of multicomponent gas mixtures and human exhaled breath // Computer Optics. Samara National Research University, 2022. Vol. 46, № 4. P. 650-658.
[10] Fufurin I.L. et al. Machine learning applications for spectral analysis of human exhaled breath for early diagnosis of diseases // Optics in Health Care and Biomedical Optics X / ed. Luo Q. et al. SPIE, 2020.
[11] Fufurin I. et al. Deep Learning for Type 1 Diabetes Mellitus Diagnosis Using Infrared Quantum Cascade Laser Spectroscopy // Materials. MDPI AG, 2022. Vol. 15, № 9. P. 2984.
