Home
Search
Collections
Journals
About
Contact us
My IOPscience
Application of an artificial neural network model for selection of potential lung cancer biomarkers
This content has been downloaded from IOPscience. Please scroll down to see the full text. 2015 J. Breath Res. 9 027106 (http://iopscience.iop.org/1752-7163/9/2/027106) View the table of contents for this issue, or go to the journal homepage for more
Download details: IP Address: 134.129.120.3 This content was downloaded on 31/05/2015 at 04:02
Please note that terms and conditions apply.
J. Breath Res. 9 (2015) 027106
doi:10.1088/1752-7155/9/2/027106
paper
received
30 November 2014
Application of an artificial neural network model for selection of potential lung cancer biomarkers
re vised
26 March 2015 accep ted for publication
30 March 2015 published
6 May 2015
Tomasz Ligor, Łukasz Pater and Bogusław Buszewski Department of Environmental Chemistry and Bioanalytics, Faculty of Chemistry, Nicolaus Copernicus University, 7 Gagarin St, 87–100 Toruń, Poland E-mail: [email protected] Keywords: volatile organic compounds, breath analysis, artificial neural network, relational sql database, genetic algorithm
Abstract Determination of volatile organic compounds (VOCs) in the exhaled breath samples of lung cancer patients and healthy controls was carried out by SPME-GC/MS (solid phase microextraction- gas chromatography combined with mass spectrometry) analyses. In order to compensate for the volatile exogenous contaminants, ambient air blank samples were also collected and analyzed. We recruited a total of 123 patients with biopsy-confirmed lung cancer and 361 healthy controls to find the potential lung cancer biomarkers. Automatic peak deconvolution and identification were performed using chromatographic data processing software (AMDIS with NIST database). All of the VOCs sample data operation, storage and management were performed using the SQL (structured query language) relational database. The selected eight VOCs could be possible biomarker candidates. In crossvalidation on test data sensitivity was 63.5% and specificity 72.4% AUC 0.65. The low performance of the model has been mainly due to overfitting and the exogenous VOCs that exist in breath. The dedicated software implementing a multilayer neural network using a genetic algorithm for training was built. Further work is needed to confirm the performance of the created experimental model. S Online supplementary data available from stacks.iop.org/JBR/9/027106/mmedia
1. Introduction Lung cancer is one of the main causes of death in developed countries. The important risk factors which are related to lung cancer include exposure to tobacco smoke, radon, asbestos, arsenic or beryllium. The main types of lung cancer are small cell lung carcinoma (SCLC) and non-small cell lung carcinoma (NSCLC), the latter being further divided into three major histologic types, the adenocarcinoma, squamous cell carcinoma and planoepitheliae [1]. The tumors are diagnosed by means of the low dose spiral computed tomography, x-ray, positron emission tomography and then confirmed by biopsy. These techniques are effective, but they cannot be used as a screening method because of health concerns associated with radiation dosage and high cost. Early diagnosis of lung cancer is a vital step towards successful recovery. Nevertheless, there is still no effective and easily available screening method in clinical practice. In the 1970s Pauling detected around 200 different volatile organic compounds (VOCs) in the breath. Thus, he proved that normal human © 2015 IOP Publishing Ltd
breath is a gas of rather complex composition. VOCs are the products of metabolic processes in the human body, which can provide additional information regarding biochemical pathways [2]. Nowadays, gas chromatography combined with mass spectrometry (GC/MS) allows for reliable and sensitive analyses of VOCs in the breath [3]. However, the GC/MS technique needs an additional preconcentration step and separation, which are time consuming [4–7]. On the other hand, numerous instrumental techniques, such as proton transfer reaction mass spectrometry, selected ion flow tube mass spectrometry, tunable diode laser absorption spectroscopy and e-noses are extensively applied in breath analysis [8, 9].The breath profile varies according to the diet, energy expenditure, state of health, and smoking status of the particular person. Recent findings suggest that the composition of the exhaled breath of lung cancer patients is altered in comparison with the healthy controls. There were many attempts to use this difference to recognize the specific pattern of VOCs, which would enable lung cancer persons to be distinguished from the healthy population
J. Breath Res. 9 (2015) 027106
T Ligor et al
[10, 11]. Data analysis and correct classification of the chromatographic data still represent a challenge due to many factors, which includes sample contamination, low concentrations of the VOCs of interest and the sampling procedure. The application of a single compound in breath as a lung cancer biomarker is very limited due to the lack of the data related to their metabolism pathway, biochemical source in the cells or ambient air occurrence. Indeed, investigation of the total number of exhaled compounds is a promising way to obtain reliable data regarding the pathophysiological state of the human body [12]. Principal component (PCA) and discriminant analysis are used for VOCs selection, and to differentiate the people diagnosed with cancer from the healthy controls [13]. Other approaches include fuzzy logic or the artificial neural networks models [14]. Important tasks regarding machine learning techniques in this field have been presented in recent articles [15, 16]. In current work, the SPME (solid phase microextraction) method and GC/MS have been used for the determination of VOCs in the exhaled breath of lung cancer patients, healthy controls and ambient air. AMDIS (automated spectral deconvolution and identification system) software has been applied for automatic deconvolution and identification of the peaks. All of the data obtained from this procedure has been inserted into the structured query language (SQL) database. In order to identify a set of compounds that occur with the greatest frequency among the cancer-patient samples, and therefore that might be considered as lung cancer biomarker candidates, a query to the SQL database has been used. After that, the selected VOCs have been revised and known contaminants, exogenous compounds and tobacco smoke constituents were removed. The resulting VOCs have been used to train the neural network model. The training of a neural network has been done by genetic algorithm.
2. Methods 2.1. Apparatus The analyses were performed on an Agilent 5975 Inert XL MSD coupled with a 6890 N gas chromatograph (Agilent, Waldbronn, Germany) with split-splitless injector. The GC was equipped with a CP-Porabond-Q (Varian) 25 m × 0.25 mm × 3 µm column. The oven temperature programme was as follows: the initial 40 °C was kept for 2 min, and ramped at 10 °C min−1 to 140 °C, and then ramped at 5 °C min−1 to 270 °C, and kept for 5 min.The temperature of the injector was 300 °C. The splitless time was 2 min, while the split ratio was 1:50. Helium was used as a carrier gas at velocity 0.7 ml min−1. The MS analyses were carried out in a full scan mode, with scan range 35–300 amu. A scan rate of 3.4 scan s −1 was applied. Electron impact ionisation at energy 70 eV was used for the MS measurement. The ion source, quadrupole and transfer line temperatures were maintained at 250 °C, 2
Table 1. Histology and demographic data for cancer patients.
Type of cancer
Age (year)
Median (age)
NSCLC
56–74
59
Male 8
Female
Total
12
20
SCLC
51–87
63.5
11
5
16
Adenocarcinoma
56–78
68
17
7
24
Planoepitheliae
41–86
64
33
11
44
Others
36–77
61
10
9
19
200 °C and 250 °C, respectively. The acquisition of chromatographic data was performed by means of Chemstation (Agilent), and a mass spectrum library NIST 2005 (Gaithersburg, USA) was applied to identification. Deconvolution and automatic peak finding were done by AMDIS version 2.64 (NIST). Carboxen/Polydimethylsiloxane coated fibre (Supelco) were used for the SPME method. All the chromatographic standards were purchased from Sigma–Aldrich, Chemsampco and Penta MFG. Extractions of VOC from gas samples were done in Tedlar bags at 22 °C for 10 min, afterwards volatiles were desorbed in the hot GC injector for 2 min. 2.2. Standard preparation Gaseous standards were prepared by injecting 1 or 3 µL of liquid compounds into a glass bulb (Supelco, Bellefonte, USA). The liquid was then evaporated. Next, the mixture was subsequently diluted in tedlar bags filled with argon to obtain concentration in the range 0.5–200 ppb. This procedure has been previously described [17]. 2.3. Human subjects A cohort of 123 lung cancer patients (109 smokers and 14 nonsmokers) at stages III (117 patients) and IV (6 patients) was recruited. The medical diagnosis of each patient was confirmed by biopsy. All individuals gave informed consent for participation in the study. The patients completed a questionnaire describing their current smoking status (active smokers, nonsmokers). Volunteers inhaled ambient air. The classification as smoker/non-smoker/ex-smoker is based on the self-declaration of the volunteers. Samples were collected in the hospital. The cancer patients’ breath was compared with 361 healthy volunteers. They also gave informed consent for participation in the study and declared their smoking habits. All patients and volunteers consumed food not later than 1 h before breath sampling. No special dietary regimes were applied. The samples were collected in two hospital rooms at different times independent of the time of meals and were processed within 2–8 h. Every breath sample was collected with parallel collection of ambient air. Finally, we measured 484 ambient air samples. The study was approved by the Nicolaus Copernicus University Ethic Commission. The demographic data of all study subjects and histology are shown in tables 1 and 2.
J. Breath Res. 9 (2015) 027106
T Ligor et al
Table 2. Demographic data for healthy volunteers. Healthy controls
Male
Female
Age
Total
Nonsmokers
patient VOC occurrence the number of the single VOC in the patients' group = the total number of patients samples
91
126
20–52
217
Passive smokers
9
2
23–58
11
Active smokers
53
41
20–55
94
healthy VOC occurrence the number of the single VOC in the health’s group = the total number of health’s control samples
2.4. Breath collection The samples of mixed breath gas were collected in Tedlar bags with a parallel collection of ambient air. Breath gas samples were obtained after a 10 min rest of the volunteer. Each subject provided two breath samples. Before collection of breath, all bags were cleaned to remove volatiles by flushing with argon, and then stored at 60 °C in the laboratory oven.
3. Results 3.1. Data analysis In order to extract more significant information and to improve reliability and precision, AMDIS 2.64 computer software was used. Our methodology included two steps both performed by AMDIS with standard configuration. The first step involved a deconvolution procedure, while the chromatograms were automatically processed for peak searching. The second one included identification of the peaks by a comparison with the NIST mass spectra library. Ultimately, in every single breath, 60–120 VOCs were tentatively detected and identified. All of the data generated from the AMDIS as well as information regarding every breath and indoor air sample (sampling room, smoking status, diagnosis, age, sex, etc) were transferred into the relational SQL database. The list of VOCs present in each sample was checked for any possible errors, which could be caused by the mismatch of the retention times and mass spectrum (MS) extracted from the sample with the computer MS library and retention time database. From the peak areas of the identified compounds present in every breath sample, the corresponding ambient air samples were subtracted. We assumed that the ideal biomarker should be detected only in the breaths of the patients and not in the ambient air or healthy controls. The list of the most frequently occurring VOCs in the breaths of the patients’ group was selected from the SQL database. For every compound in this selection, occurrences in ambient air and exhaled air samples were calculated. The occurrences of VOCs were calculated as follows: 3
× 100
(1)
(2)
ambient air VOC occurrence the number of the single VOC in the ambient air = the total number of the ambient air samples
× 100
(3)
DPH = Patients occurrence —Healthy occurrence (4) Acetone and isoprene have been identified in every breath sample. These compounds are always present in the largest amount in the exhaled air. Concentrations of both were in the range 60–689.4 ppb, 37–234 ppb, for acetone and isoprene, respectively (table S2 (stacks. iop.org/JBR/9/027106/mmedia)). These compounds are typical volatile metabolites generated inside the human body. Other substances are observed in a higher concentration in breath than in the inhaled air, however, these are not only of the endogenous origin. Patients suffering from lung cancer often have a smoking habit. Therefore, tobacco smoke constituents are frequently identified in breath [18–20]. In our investigations, acetonitrile, 2,5-dimethylfuran, furan, 2-methylfuran, 1,3-cyclohexadiene, 1,3-cyclopentadiene, 1,4- pentadiene, 1,3-pentadiene, 2-methyl-1-butene are observed in the breath of smokers. The appearance of 2,5-dimethylfuran correlated with acetonitrile and these substances were detected in the breath of those patients with a smoking habit. The concentrations of acetonitrile varied from 6.8 ppb to 86.4 ppb, and the concentration of 2,5-dimethylfuran ranged from 5.8 to 12.8 ppb. In the case of monocyclic aromatic hydrocarbons, the concentration of benzene correlates better with smoking behavior than other alkyl substituted compounds (toluene and xylenes). Benzene in the range 4.2–28.2 ppb is observed in the breath of smokers. These ones should be considered as indicators of smoking. Moreover, substances found in cigarette smoke and the majority of VOCs tentatively identified have been omitted in the further part of the discussion and have not been used in the search for biomarkers. Propyl alcohols (1-propanol and 2-propanols) were often detected in the breath of lung cancer patients. These two compounds, and usually ethyl alcohol, are used as typical disinfectants in the clinical environment. However, 2-propanol in 70% of patients’ breath and in 69% of healthy ones was detected. In the case of 1-propanol, this compound is observed at 64% of patients and only at 20% of healthy. The observed amount of 2-propanol
2.5. Method validation All calibration data including concentration range of the calibration curves, limits of detection (LOD) and quantification (LOQ), precision (RSD) as a relative standard deviation estimated for peak areas are presented in table S1 (stacks.iop.org/JBR/9/027106/ mmedia).
× 100
J. Breath Res. 9 (2015) 027106
T Ligor et al
Table 3. Compounds selected for the highest performance of model. Rank
VOC occurrence in models
1
Butane
2
Butane, 2-methyl-
3
Octane, 4-methyl-
4
Propane
5
2-Pentanone
6
Propanal
7
Heptane, 2,4-dimethyl-
8
Propene
can result from using common disinfectants which also contain 1-propanol or ethyl alcohol. On the other hand, isopropanol is postulated to be a product of enzyme mediated reduction of acetone [21]. The biochemical pathways of 1-propanol is unknown. However, we cannot also exclude the endogenous origin of 1- and 2-propanols. 3.2. Proposed model The initially selected list of the VOCs presented in table S3 (stacks.iop.org/JBR/9/027106/mmedia), suggests that there is no single volatile compound which would clearly determine the presence of lung cancer with high performance. Instead, the occurrence of several VOCs in combination can be used to increase both overall sensitivity and specificity of the diagnostic test. In our approach, a feedforward multilayer perceptron neural network with one hidden layer was used. Each input value represented the peak area of the compound. The hidden layer consisted of ten neurons. For every neuron in the network, a sigmoid function was utilized as an activation function. A genetic algorithm was used for the efficient training of the neural network. The list of the hypothetical biomarkers is presented in table S4 along with the references to other papers. There are mostly branched alkanes. In order to narrow down the input parameter space PCA analysis has been used. The scree plots have been used to evaluate the variables, which explain most of the data variability (figure S1 (stacks.iop.org/ JBR/9/027106/mmedia)). Most of the overall variability is determined by ten compounds. In order to compensate for feature correlation, data agglomerative cluster analysis by Ward method with euclidean distance measurement has been conducted. The top ten features for each cross-validation model are presented in table S5 (stacks.iop.org/JBR/9/027106/ mmedia). According to the feature ranks, eight compounds were selected, which occurred in every cross-validation model (table 3). The same compounds are mostly observed in each cross validation model. Summarizing, the agglomerative tree chart of overall compounds, showing their 4
relevance, is presented in figure S2 (stacks.iop.org/ JBR/9/027106/mmedia). These compounds were selected to make new cross validation models. However, the VOCs selected in this way were not enough to train and validate accurate models. The low performance of the eight feature model has been mainly due to overfitting. The influence of the exogenous VOCs existing in breath samples is also an important factor. We have decided to remove particular compounds from the study (tobacco smoke markers, halogenated compounds, etc). Nevertheless, the background correction, which has been done by the subtraction of each peak of ambient air from every breath sample seems to be insufficient to eliminate exogenous compounds.
4. Discussion Malondialdehyde, saturated aldehydes and simple alkanes have been identified as products of lipid peroxidation and their concentration increases during inflammation and oxidative stress [21]. Therefore, one can assume that the elevated level of branched hydrocarbons in breath indicates cancer. As such, a hypothesis was introduced by Philips [22]. During the study, the increased concentration of pentane in patients with lung cancer has not been notified. Pentane is a well known marker of oxidative stress and may appear during inflammation, which is observed in advanced cancer states. It is difficult, however, to confirm or exclude the hypothesis that inflammation may be the cause of the formation and the presence in the breath of methylated alkanes. On the other hand, the source of methylated alkanes may be extensive, e.g. plastics and transportation (fuel and exhaust). Taking into account the lipophilicity of these substances, there is the possibility of their accumulation in lipid-rich tissues and slow excretion from the body. This statement could be explained by the presence of methylated alkanes in the breath and their absence in the inhaled air during sample taking. Taking into consideration butane and propane, their presence can be caused by bacterial flora present in the gut. Morever, there has been more frequent appearance of propanal observed in ill persons and at higher concentrations, rather than in healthy ones. It can be caused by oxidative stress, which accompanies cancer. It should be noted that Poli work confirms that propanal may be a biomarker of lung cancer [23]. Although the 11 compounds presented in table S4 of the supplementary data (stacks. iop.org/JBR/9/027106/mmedia), have already been claimed as possible lung cancer biomarkers in other papers, we do not claim that all of the VOCs selected by the genetic algorithm are biomarkers. Additionally, we did not find statistically significant differences in the concentration and frequency of the occurrence of VOCs between different cancer stages and their histological type. However, research in large numbers of patients is indispensable to confirm these results.
J. Breath Res. 9 (2015) 027106
T Ligor et al
Figure 1. Model—(a) training—80% data set. Sensitivity of 0.69, specificity of 0.62 and AUC of 0.66; (b) test—20% data set. Sensitivity of 0.79, specificity 0.67 and AUC of 0.71.
Figure 2. Model 2—(a) training—80% data set. Sensitivity of 0.76, specificity of 0.61 and AUC of 0.68; (b) test—20% data set. Sensitivity of 0.79, specificity of 0.52 and AUC of 0.66.
Figure 3. Model 3—(a) training—80% data. Sensitivity of 0.30, specificity of 0.76 and AUC of 0.61; (B) test—20%. Sensitivity of 0.54, specificity of 0.79 and AUC of 0.66.
4.1. Validation In order to assess the robustness of the model, which uses selected possible biomarkers, a 5-fold cross-validation 5
and permutation test was carried out. The standard evaluation of any diagnostic test is a presentation of its performance, as a receiver-operating characteristic
J. Breath Res. 9 (2015) 027106
T Ligor et al
Figure 4. Model 4—(a) training—80% data set. Sensitivity of 0.70, specificity of 0.59 and AUC of 0.64; (b) test—20% data set. Sensitivity of 0.45, specificity of 0.81 and AUC of 0.64.
Figure 5. Model 5—(a) training—80% data set. Sensitivity of 0.67, specificity of 0.65 and AUC of 0.67; (b) test—20% data set. Sensitivity of 0.59, specificity of 0.83 and AUC of 0.62.
(ROC) curve. During the permutation test, when the labels of the patients were chosen randomly, the ROC curve fell to the main diagonal. We divided our training and test data into five sets. Each of them gave a ROC curve of similar shape. The developed neural network model in 5-fold crossvalidation reached an average sensitivity of 62.2%, a specificity of 66.6% and AUC of 0.65 on training data and a sensitivity of 63.5%, specificity of 72.4% and AUC of 0.65 on test data. Figures 1–5 demonstrates the ROC curves for 5-fold cross validation. We also have to notice that model performance is mostly determined by the selection of the VOCs. The exclusion of popular disinfectant compounds such as 1-propanol and isopropanol, which had a great significance in determining patients from the hospital, causes the noticeable loss of specificity. Unfortunately, our healthy controls are younger than the patients. Therefore, the findings could have also been affected by the age of the healthy 6
controls. Certainly age and gender are confounding variables, which have to be evaluated carefully [31, 32].
5. Conclusions The SPME-GC/MS method is relatively simple and effective for the analysis of the exhaled air oriented on the search for lung cancer biomarkers. The application of AMDIS software in combination with the SQL relational database improves peak recognition, identification of the compounds and enables new possibilities in the handling of large amounts of chromatographic data. The artificial neural network trained by genetic algorithms proved useful for easy creation of a model that can distinguish cancer patients. However, the hypothetical biomarkers selected by the model exhibit limited quality for distinguishing cancer patients from healthy controls. Nevertheless, further work is needed to confirm the predictive capability of
J. Breath Res. 9 (2015) 027106
T Ligor et al
the model using a larger number of breath samples. Usage of the multilayer artificial neural networks also enhances the predictive capability of the model. Currently, we are working on further optimization of the model, taking into account application of new data. As a result, we expect to increase the quality of the predictive power.
Acknowledgments This work was supported by The National Centre for Research and Development—grant Sensormed NCBiR No PBS/A3/7/2012 (2012–2015).
References [1] Edelman M J and Gandara D R 2009 Lung cancer Manual of Clinical Oncology ed D E Casciato and M C Territo (Philadelphia: LWW) p 169 [2] Pauling L, Robinson A B, Teranishi R and Cary P 1971 Quantitative analysis of urine vapor and breath by gas–liquid partition chromatography Proc. Natl Acad. Sci. USA 68 2374–6 [3] Miekisch W and Schubert J K 2006 From highly sophisticated analytical techniques to life-saving diagnostics: technical developments in breath analysis Trends Anal. Chem. 25 665–73 [4] Mieth M, Schubert J K, Groger T, Sabel B, Kischkel S, Fuchs P, Hein D, Zimmermann R and Miekisch W 2009 Automated needle trap heart-cut GC/MS and needle trap comprehensive 2D GC/TOF-MS for breath gas analysis in the clinical environment Anal. Chem. 81 5851–7 [5] Sanchez J M and Sacks R D 2006 Development of a multibed sorption trap, comprehensive 2D gas chromatography, and time-of-flight mass spectrometry system for the analysis of volatile organic compounds in human breath Anal. Chem. 78 3046–54 [6] Trefz P, Kischkel S, Hein D, James E S, Schubert J K and Miekisch W 2012 Needle trap micro-extraction for VOC analysis: effects of packing materials and desorption parameters J. Chromatogr. A 1219 29–38 [7] Ueta I, Samsudin E L, Mizuguchi A, Takeuchi H, Shinki T, Kawakubo S and Saito Y 2014 Double-bed-type extraction needle packed with activated-carbon-based sorbents for very volatile organic compounds J. Pharm. Biomed. Anal. 88 423–28 [8] Buszewski B, Kęsy M, Ligor T and Amann A 2007 Human exhaled air analytics: biomarkers of diseases Biomed. Chromatogr. 21 553–66 [9] Buszewski B, Grzywiński D, Ligor T, Stacewicz T, Bielecki Z and Wojtas J 2013 Detection of volatile organic compounds as biomarkers in breath analysis by different analytical techniques Bioanalysis 5 2287–306 [10] Kim K H, Jahan S A and Kabir E 2012 A review of breath analysis for diagnosis of human health Trends Anal. Chem. 33 1–8 [11] Ulanowska A, Kowalkowski T, Trawińska E and Buszewski B 2011 The application of statistical methods using VOCs to identify patients with lung cancer J. Breath Res. 5 046008 [12] Boots A W, van Berkel J J B N, Dallinga J W, Smolinska A, Wouters E F and van Schooten F J 2012 The versatile use of exhaled volatile organic compounds in human health and disease J. Breath Res. 6 027108 [13] Philips M, Gleeson K, Hughes J M B, Greenberg J, Cataneo R N and Baker L 1999 Volatile organic compounds in breath as markers of lung cancer: a cross-sectional study Lancet 353 1930–33 [14] Phillips M, Herrera J, Krishnan S, Zain M, Greenberg J and Cataneo R N 1999 Variation in volatile organic compounds in the breath of normal humans J. Chromatogr. B 729 75–88 [15] Smolinska A, Hauschild A-Ch, Fijten R R R, Dallinga J W, Baumbach J and van Schooten F J 2014 Current breathomics:
7
a review on data pre-processing techniques and machine learning in metabolomics breath analysis J. Breath Res. 8 027105 [16] Hauschild A-Ch, Kopczynski D, D’Addario M, Baumbach J I, Rahmann S and Baumbach J 2013 Peak detection method evaluation for ion mobility spectrometry by using machine learning approaches Metabolites 3 277–93 [17] Ligor T, Ligor M, Amann A, Ager C, Bachler M, Dzien A and Buszewski B 2008 The analysis of healthy volunteers’ exhaled breath by use of solid phase microextraction and GC-MS J. Breath Res. 2 046006 [18] Buszewski B, Ulanowska A, Ligor T, Denderz N and Amann A 2009 Analysis of exhaled breath from smokers, passive smokers and non-smokers by solidphase microextraction gas chromatography/mass spectrometry Biomed. Chromatogr. 23 551–6 [19] Ligor M et al 2009 Determination of volatile organic compounds in exhaled breath of patients with lung cancer using solid phase microextraction and gas chromatography mass spectrometry Clin. Chem. Lab. Med. 47 550–60 [20] Filipiak W et al 2012 Dependence of exhaled breath composition on exogenous factors, smoking habits and exposure to air pollutants J. Breath Res. 6 036008 [21] Miekisch W, Schubert J K, Noeldge-Schomburg G F E 2004 Diagnostic potential of breath analysisfocus on volatile organic compounds Clin. Chim. Acta 347 25–39 [22] Phillips M, Cataneo R N, Greenberg J, Gunawardena R, Naidu A and Rahbari-Oskoui F 2000 Effect of age on the breath methylated alkane contour, a display of apparent new markers of oxidative stress J. Lab. Clin. Med. 136 243–9 [23] Poli D, Goldoni M, Corradi M, Acampa O, Carbognani P, Internullo E, Casalini A and Mutti A 2010 Determination of aldehydes in exhaled breath of patients with lung cancer by means of on-fiber-derivatisation SPME–GC/MS J. Chromatogr. B 878 2643–51 [24] Song G, Qin T, Liu H, Xu G B, Pan Y Y, Xiong F X, Gu K S, Sun G P and Chen Z D 2010 Quantitative breath analysis of volatile organic compounds of lung cancer patients Lung Cancer 67 227–31 [25] Philips M et al 2008 Detection of lung cancer using weighted digital analysis of breath biomarkers Clin. Chim. Acta 393 76–84 [26] Bajtarevic A et al 2009 Noninvasive detection of lung cancer by analysis of exhaled breath BMC Cancer 9 348 [27] Filipiak W, Sponring A, Filipiak A, Ager C, Schubert J, Miekisch W, Amann A and Troppmair J 2010 TD-GC-MS Analysis of volatile metabolites of human lung cancer and normal cells in vitro Cancer Epidemiol. Biomarkers 19 182–95 [28] Westhoff M, Litterst P, Maddula S, Bödeker B, Rahmann S, Davies A N and Baumbach J I 2010 Differentiation of chronic obstructive pulmonary disease (COPD) including lung cancer from healthy control group by breath analysis using ion mobility spectrometry Int. J. Ion Mobil. Spectrom. 13 131–39 [29] Philips M, Cataneo R N, Ditkoff B A, Fisher P, Greenberg J, Gunawardena R, Kwon C S, Rahbari-Oskoui F and Wong C 2003 Volatile markers of breath cancer in the breath Breast J. 9 184–91 [30] Hakim M, Billan S, Tisch U, Peng G, Dvrokind I, Marom O, Abdah-Bortnyak R, Kuten A and Haick H 2011 Diagnosis of head-and-neck cancer from exhaled breath Br. J. Cancer 104 1649–55 [31] Kushch I et al 2008 Breath isoprene: aspects of normal physiology related to age, gender and cholesterol profile as determined in a proton transfer reaction mass spectrometry study Clin. Chem. Lab. Med. 46 1011–8 [32] Schwarz K, Pizzini A, Arendacká B, Zerlauth K, Filipiak W, Schmid A, Dzien A, Neuner S, Lechleitner M and SchollBürgi S 2009 Breath acetone: aspects of normal physiology related to age and gender as determined in a PTR-MS study J. Breath Res. 3 027003
Search
Collections
Journals
About
Contact us
My IOPscience
Application of an artificial neural network model for selection of potential lung cancer biomarkers
This content has been downloaded from IOPscience. Please scroll down to see the full text. 2015 J. Breath Res. 9 027106 (http://iopscience.iop.org/1752-7163/9/2/027106) View the table of contents for this issue, or go to the journal homepage for more
Download details: IP Address: 134.129.120.3 This content was downloaded on 31/05/2015 at 04:02
Please note that terms and conditions apply.
J. Breath Res. 9 (2015) 027106
doi:10.1088/1752-7155/9/2/027106
paper
received
30 November 2014
Application of an artificial neural network model for selection of potential lung cancer biomarkers
re vised
26 March 2015 accep ted for publication
30 March 2015 published
6 May 2015
Tomasz Ligor, Łukasz Pater and Bogusław Buszewski Department of Environmental Chemistry and Bioanalytics, Faculty of Chemistry, Nicolaus Copernicus University, 7 Gagarin St, 87–100 Toruń, Poland E-mail: [email protected] Keywords: volatile organic compounds, breath analysis, artificial neural network, relational sql database, genetic algorithm
Abstract Determination of volatile organic compounds (VOCs) in the exhaled breath samples of lung cancer patients and healthy controls was carried out by SPME-GC/MS (solid phase microextraction- gas chromatography combined with mass spectrometry) analyses. In order to compensate for the volatile exogenous contaminants, ambient air blank samples were also collected and analyzed. We recruited a total of 123 patients with biopsy-confirmed lung cancer and 361 healthy controls to find the potential lung cancer biomarkers. Automatic peak deconvolution and identification were performed using chromatographic data processing software (AMDIS with NIST database). All of the VOCs sample data operation, storage and management were performed using the SQL (structured query language) relational database. The selected eight VOCs could be possible biomarker candidates. In crossvalidation on test data sensitivity was 63.5% and specificity 72.4% AUC 0.65. The low performance of the model has been mainly due to overfitting and the exogenous VOCs that exist in breath. The dedicated software implementing a multilayer neural network using a genetic algorithm for training was built. Further work is needed to confirm the performance of the created experimental model. S Online supplementary data available from stacks.iop.org/JBR/9/027106/mmedia
1. Introduction Lung cancer is one of the main causes of death in developed countries. The important risk factors which are related to lung cancer include exposure to tobacco smoke, radon, asbestos, arsenic or beryllium. The main types of lung cancer are small cell lung carcinoma (SCLC) and non-small cell lung carcinoma (NSCLC), the latter being further divided into three major histologic types, the adenocarcinoma, squamous cell carcinoma and planoepitheliae [1]. The tumors are diagnosed by means of the low dose spiral computed tomography, x-ray, positron emission tomography and then confirmed by biopsy. These techniques are effective, but they cannot be used as a screening method because of health concerns associated with radiation dosage and high cost. Early diagnosis of lung cancer is a vital step towards successful recovery. Nevertheless, there is still no effective and easily available screening method in clinical practice. In the 1970s Pauling detected around 200 different volatile organic compounds (VOCs) in the breath. Thus, he proved that normal human © 2015 IOP Publishing Ltd
breath is a gas of rather complex composition. VOCs are the products of metabolic processes in the human body, which can provide additional information regarding biochemical pathways [2]. Nowadays, gas chromatography combined with mass spectrometry (GC/MS) allows for reliable and sensitive analyses of VOCs in the breath [3]. However, the GC/MS technique needs an additional preconcentration step and separation, which are time consuming [4–7]. On the other hand, numerous instrumental techniques, such as proton transfer reaction mass spectrometry, selected ion flow tube mass spectrometry, tunable diode laser absorption spectroscopy and e-noses are extensively applied in breath analysis [8, 9].The breath profile varies according to the diet, energy expenditure, state of health, and smoking status of the particular person. Recent findings suggest that the composition of the exhaled breath of lung cancer patients is altered in comparison with the healthy controls. There were many attempts to use this difference to recognize the specific pattern of VOCs, which would enable lung cancer persons to be distinguished from the healthy population
J. Breath Res. 9 (2015) 027106
T Ligor et al
[10, 11]. Data analysis and correct classification of the chromatographic data still represent a challenge due to many factors, which includes sample contamination, low concentrations of the VOCs of interest and the sampling procedure. The application of a single compound in breath as a lung cancer biomarker is very limited due to the lack of the data related to their metabolism pathway, biochemical source in the cells or ambient air occurrence. Indeed, investigation of the total number of exhaled compounds is a promising way to obtain reliable data regarding the pathophysiological state of the human body [12]. Principal component (PCA) and discriminant analysis are used for VOCs selection, and to differentiate the people diagnosed with cancer from the healthy controls [13]. Other approaches include fuzzy logic or the artificial neural networks models [14]. Important tasks regarding machine learning techniques in this field have been presented in recent articles [15, 16]. In current work, the SPME (solid phase microextraction) method and GC/MS have been used for the determination of VOCs in the exhaled breath of lung cancer patients, healthy controls and ambient air. AMDIS (automated spectral deconvolution and identification system) software has been applied for automatic deconvolution and identification of the peaks. All of the data obtained from this procedure has been inserted into the structured query language (SQL) database. In order to identify a set of compounds that occur with the greatest frequency among the cancer-patient samples, and therefore that might be considered as lung cancer biomarker candidates, a query to the SQL database has been used. After that, the selected VOCs have been revised and known contaminants, exogenous compounds and tobacco smoke constituents were removed. The resulting VOCs have been used to train the neural network model. The training of a neural network has been done by genetic algorithm.
2. Methods 2.1. Apparatus The analyses were performed on an Agilent 5975 Inert XL MSD coupled with a 6890 N gas chromatograph (Agilent, Waldbronn, Germany) with split-splitless injector. The GC was equipped with a CP-Porabond-Q (Varian) 25 m × 0.25 mm × 3 µm column. The oven temperature programme was as follows: the initial 40 °C was kept for 2 min, and ramped at 10 °C min−1 to 140 °C, and then ramped at 5 °C min−1 to 270 °C, and kept for 5 min.The temperature of the injector was 300 °C. The splitless time was 2 min, while the split ratio was 1:50. Helium was used as a carrier gas at velocity 0.7 ml min−1. The MS analyses were carried out in a full scan mode, with scan range 35–300 amu. A scan rate of 3.4 scan s −1 was applied. Electron impact ionisation at energy 70 eV was used for the MS measurement. The ion source, quadrupole and transfer line temperatures were maintained at 250 °C, 2
Table 1. Histology and demographic data for cancer patients.
Type of cancer
Age (year)
Median (age)
NSCLC
56–74
59
Male 8
Female
Total
12
20
SCLC
51–87
63.5
11
5
16
Adenocarcinoma
56–78
68
17
7
24
Planoepitheliae
41–86
64
33
11
44
Others
36–77
61
10
9
19
200 °C and 250 °C, respectively. The acquisition of chromatographic data was performed by means of Chemstation (Agilent), and a mass spectrum library NIST 2005 (Gaithersburg, USA) was applied to identification. Deconvolution and automatic peak finding were done by AMDIS version 2.64 (NIST). Carboxen/Polydimethylsiloxane coated fibre (Supelco) were used for the SPME method. All the chromatographic standards were purchased from Sigma–Aldrich, Chemsampco and Penta MFG. Extractions of VOC from gas samples were done in Tedlar bags at 22 °C for 10 min, afterwards volatiles were desorbed in the hot GC injector for 2 min. 2.2. Standard preparation Gaseous standards were prepared by injecting 1 or 3 µL of liquid compounds into a glass bulb (Supelco, Bellefonte, USA). The liquid was then evaporated. Next, the mixture was subsequently diluted in tedlar bags filled with argon to obtain concentration in the range 0.5–200 ppb. This procedure has been previously described [17]. 2.3. Human subjects A cohort of 123 lung cancer patients (109 smokers and 14 nonsmokers) at stages III (117 patients) and IV (6 patients) was recruited. The medical diagnosis of each patient was confirmed by biopsy. All individuals gave informed consent for participation in the study. The patients completed a questionnaire describing their current smoking status (active smokers, nonsmokers). Volunteers inhaled ambient air. The classification as smoker/non-smoker/ex-smoker is based on the self-declaration of the volunteers. Samples were collected in the hospital. The cancer patients’ breath was compared with 361 healthy volunteers. They also gave informed consent for participation in the study and declared their smoking habits. All patients and volunteers consumed food not later than 1 h before breath sampling. No special dietary regimes were applied. The samples were collected in two hospital rooms at different times independent of the time of meals and were processed within 2–8 h. Every breath sample was collected with parallel collection of ambient air. Finally, we measured 484 ambient air samples. The study was approved by the Nicolaus Copernicus University Ethic Commission. The demographic data of all study subjects and histology are shown in tables 1 and 2.
J. Breath Res. 9 (2015) 027106
T Ligor et al
Table 2. Demographic data for healthy volunteers. Healthy controls
Male
Female
Age
Total
Nonsmokers
patient VOC occurrence the number of the single VOC in the patients' group = the total number of patients samples
91
126
20–52
217
Passive smokers
9
2
23–58
11
Active smokers
53
41
20–55
94
healthy VOC occurrence the number of the single VOC in the health’s group = the total number of health’s control samples
2.4. Breath collection The samples of mixed breath gas were collected in Tedlar bags with a parallel collection of ambient air. Breath gas samples were obtained after a 10 min rest of the volunteer. Each subject provided two breath samples. Before collection of breath, all bags were cleaned to remove volatiles by flushing with argon, and then stored at 60 °C in the laboratory oven.
3. Results 3.1. Data analysis In order to extract more significant information and to improve reliability and precision, AMDIS 2.64 computer software was used. Our methodology included two steps both performed by AMDIS with standard configuration. The first step involved a deconvolution procedure, while the chromatograms were automatically processed for peak searching. The second one included identification of the peaks by a comparison with the NIST mass spectra library. Ultimately, in every single breath, 60–120 VOCs were tentatively detected and identified. All of the data generated from the AMDIS as well as information regarding every breath and indoor air sample (sampling room, smoking status, diagnosis, age, sex, etc) were transferred into the relational SQL database. The list of VOCs present in each sample was checked for any possible errors, which could be caused by the mismatch of the retention times and mass spectrum (MS) extracted from the sample with the computer MS library and retention time database. From the peak areas of the identified compounds present in every breath sample, the corresponding ambient air samples were subtracted. We assumed that the ideal biomarker should be detected only in the breaths of the patients and not in the ambient air or healthy controls. The list of the most frequently occurring VOCs in the breaths of the patients’ group was selected from the SQL database. For every compound in this selection, occurrences in ambient air and exhaled air samples were calculated. The occurrences of VOCs were calculated as follows: 3
× 100
(1)
(2)
ambient air VOC occurrence the number of the single VOC in the ambient air = the total number of the ambient air samples
× 100
(3)
DPH = Patients occurrence —Healthy occurrence (4) Acetone and isoprene have been identified in every breath sample. These compounds are always present in the largest amount in the exhaled air. Concentrations of both were in the range 60–689.4 ppb, 37–234 ppb, for acetone and isoprene, respectively (table S2 (stacks. iop.org/JBR/9/027106/mmedia)). These compounds are typical volatile metabolites generated inside the human body. Other substances are observed in a higher concentration in breath than in the inhaled air, however, these are not only of the endogenous origin. Patients suffering from lung cancer often have a smoking habit. Therefore, tobacco smoke constituents are frequently identified in breath [18–20]. In our investigations, acetonitrile, 2,5-dimethylfuran, furan, 2-methylfuran, 1,3-cyclohexadiene, 1,3-cyclopentadiene, 1,4- pentadiene, 1,3-pentadiene, 2-methyl-1-butene are observed in the breath of smokers. The appearance of 2,5-dimethylfuran correlated with acetonitrile and these substances were detected in the breath of those patients with a smoking habit. The concentrations of acetonitrile varied from 6.8 ppb to 86.4 ppb, and the concentration of 2,5-dimethylfuran ranged from 5.8 to 12.8 ppb. In the case of monocyclic aromatic hydrocarbons, the concentration of benzene correlates better with smoking behavior than other alkyl substituted compounds (toluene and xylenes). Benzene in the range 4.2–28.2 ppb is observed in the breath of smokers. These ones should be considered as indicators of smoking. Moreover, substances found in cigarette smoke and the majority of VOCs tentatively identified have been omitted in the further part of the discussion and have not been used in the search for biomarkers. Propyl alcohols (1-propanol and 2-propanols) were often detected in the breath of lung cancer patients. These two compounds, and usually ethyl alcohol, are used as typical disinfectants in the clinical environment. However, 2-propanol in 70% of patients’ breath and in 69% of healthy ones was detected. In the case of 1-propanol, this compound is observed at 64% of patients and only at 20% of healthy. The observed amount of 2-propanol
2.5. Method validation All calibration data including concentration range of the calibration curves, limits of detection (LOD) and quantification (LOQ), precision (RSD) as a relative standard deviation estimated for peak areas are presented in table S1 (stacks.iop.org/JBR/9/027106/ mmedia).
× 100
J. Breath Res. 9 (2015) 027106
T Ligor et al
Table 3. Compounds selected for the highest performance of model. Rank
VOC occurrence in models
1
Butane
2
Butane, 2-methyl-
3
Octane, 4-methyl-
4
Propane
5
2-Pentanone
6
Propanal
7
Heptane, 2,4-dimethyl-
8
Propene
can result from using common disinfectants which also contain 1-propanol or ethyl alcohol. On the other hand, isopropanol is postulated to be a product of enzyme mediated reduction of acetone [21]. The biochemical pathways of 1-propanol is unknown. However, we cannot also exclude the endogenous origin of 1- and 2-propanols. 3.2. Proposed model The initially selected list of the VOCs presented in table S3 (stacks.iop.org/JBR/9/027106/mmedia), suggests that there is no single volatile compound which would clearly determine the presence of lung cancer with high performance. Instead, the occurrence of several VOCs in combination can be used to increase both overall sensitivity and specificity of the diagnostic test. In our approach, a feedforward multilayer perceptron neural network with one hidden layer was used. Each input value represented the peak area of the compound. The hidden layer consisted of ten neurons. For every neuron in the network, a sigmoid function was utilized as an activation function. A genetic algorithm was used for the efficient training of the neural network. The list of the hypothetical biomarkers is presented in table S4 along with the references to other papers. There are mostly branched alkanes. In order to narrow down the input parameter space PCA analysis has been used. The scree plots have been used to evaluate the variables, which explain most of the data variability (figure S1 (stacks.iop.org/ JBR/9/027106/mmedia)). Most of the overall variability is determined by ten compounds. In order to compensate for feature correlation, data agglomerative cluster analysis by Ward method with euclidean distance measurement has been conducted. The top ten features for each cross-validation model are presented in table S5 (stacks.iop.org/JBR/9/027106/ mmedia). According to the feature ranks, eight compounds were selected, which occurred in every cross-validation model (table 3). The same compounds are mostly observed in each cross validation model. Summarizing, the agglomerative tree chart of overall compounds, showing their 4
relevance, is presented in figure S2 (stacks.iop.org/ JBR/9/027106/mmedia). These compounds were selected to make new cross validation models. However, the VOCs selected in this way were not enough to train and validate accurate models. The low performance of the eight feature model has been mainly due to overfitting. The influence of the exogenous VOCs existing in breath samples is also an important factor. We have decided to remove particular compounds from the study (tobacco smoke markers, halogenated compounds, etc). Nevertheless, the background correction, which has been done by the subtraction of each peak of ambient air from every breath sample seems to be insufficient to eliminate exogenous compounds.
4. Discussion Malondialdehyde, saturated aldehydes and simple alkanes have been identified as products of lipid peroxidation and their concentration increases during inflammation and oxidative stress [21]. Therefore, one can assume that the elevated level of branched hydrocarbons in breath indicates cancer. As such, a hypothesis was introduced by Philips [22]. During the study, the increased concentration of pentane in patients with lung cancer has not been notified. Pentane is a well known marker of oxidative stress and may appear during inflammation, which is observed in advanced cancer states. It is difficult, however, to confirm or exclude the hypothesis that inflammation may be the cause of the formation and the presence in the breath of methylated alkanes. On the other hand, the source of methylated alkanes may be extensive, e.g. plastics and transportation (fuel and exhaust). Taking into account the lipophilicity of these substances, there is the possibility of their accumulation in lipid-rich tissues and slow excretion from the body. This statement could be explained by the presence of methylated alkanes in the breath and their absence in the inhaled air during sample taking. Taking into consideration butane and propane, their presence can be caused by bacterial flora present in the gut. Morever, there has been more frequent appearance of propanal observed in ill persons and at higher concentrations, rather than in healthy ones. It can be caused by oxidative stress, which accompanies cancer. It should be noted that Poli work confirms that propanal may be a biomarker of lung cancer [23]. Although the 11 compounds presented in table S4 of the supplementary data (stacks. iop.org/JBR/9/027106/mmedia), have already been claimed as possible lung cancer biomarkers in other papers, we do not claim that all of the VOCs selected by the genetic algorithm are biomarkers. Additionally, we did not find statistically significant differences in the concentration and frequency of the occurrence of VOCs between different cancer stages and their histological type. However, research in large numbers of patients is indispensable to confirm these results.
J. Breath Res. 9 (2015) 027106
T Ligor et al
Figure 1. Model—(a) training—80% data set. Sensitivity of 0.69, specificity of 0.62 and AUC of 0.66; (b) test—20% data set. Sensitivity of 0.79, specificity 0.67 and AUC of 0.71.
Figure 2. Model 2—(a) training—80% data set. Sensitivity of 0.76, specificity of 0.61 and AUC of 0.68; (b) test—20% data set. Sensitivity of 0.79, specificity of 0.52 and AUC of 0.66.
Figure 3. Model 3—(a) training—80% data. Sensitivity of 0.30, specificity of 0.76 and AUC of 0.61; (B) test—20%. Sensitivity of 0.54, specificity of 0.79 and AUC of 0.66.
4.1. Validation In order to assess the robustness of the model, which uses selected possible biomarkers, a 5-fold cross-validation 5
and permutation test was carried out. The standard evaluation of any diagnostic test is a presentation of its performance, as a receiver-operating characteristic
J. Breath Res. 9 (2015) 027106
T Ligor et al
Figure 4. Model 4—(a) training—80% data set. Sensitivity of 0.70, specificity of 0.59 and AUC of 0.64; (b) test—20% data set. Sensitivity of 0.45, specificity of 0.81 and AUC of 0.64.
Figure 5. Model 5—(a) training—80% data set. Sensitivity of 0.67, specificity of 0.65 and AUC of 0.67; (b) test—20% data set. Sensitivity of 0.59, specificity of 0.83 and AUC of 0.62.
(ROC) curve. During the permutation test, when the labels of the patients were chosen randomly, the ROC curve fell to the main diagonal. We divided our training and test data into five sets. Each of them gave a ROC curve of similar shape. The developed neural network model in 5-fold crossvalidation reached an average sensitivity of 62.2%, a specificity of 66.6% and AUC of 0.65 on training data and a sensitivity of 63.5%, specificity of 72.4% and AUC of 0.65 on test data. Figures 1–5 demonstrates the ROC curves for 5-fold cross validation. We also have to notice that model performance is mostly determined by the selection of the VOCs. The exclusion of popular disinfectant compounds such as 1-propanol and isopropanol, which had a great significance in determining patients from the hospital, causes the noticeable loss of specificity. Unfortunately, our healthy controls are younger than the patients. Therefore, the findings could have also been affected by the age of the healthy 6
controls. Certainly age and gender are confounding variables, which have to be evaluated carefully [31, 32].
5. Conclusions The SPME-GC/MS method is relatively simple and effective for the analysis of the exhaled air oriented on the search for lung cancer biomarkers. The application of AMDIS software in combination with the SQL relational database improves peak recognition, identification of the compounds and enables new possibilities in the handling of large amounts of chromatographic data. The artificial neural network trained by genetic algorithms proved useful for easy creation of a model that can distinguish cancer patients. However, the hypothetical biomarkers selected by the model exhibit limited quality for distinguishing cancer patients from healthy controls. Nevertheless, further work is needed to confirm the predictive capability of
J. Breath Res. 9 (2015) 027106
T Ligor et al
the model using a larger number of breath samples. Usage of the multilayer artificial neural networks also enhances the predictive capability of the model. Currently, we are working on further optimization of the model, taking into account application of new data. As a result, we expect to increase the quality of the predictive power.
Acknowledgments This work was supported by The National Centre for Research and Development—grant Sensormed NCBiR No PBS/A3/7/2012 (2012–2015).
References [1] Edelman M J and Gandara D R 2009 Lung cancer Manual of Clinical Oncology ed D E Casciato and M C Territo (Philadelphia: LWW) p 169 [2] Pauling L, Robinson A B, Teranishi R and Cary P 1971 Quantitative analysis of urine vapor and breath by gas–liquid partition chromatography Proc. Natl Acad. Sci. USA 68 2374–6 [3] Miekisch W and Schubert J K 2006 From highly sophisticated analytical techniques to life-saving diagnostics: technical developments in breath analysis Trends Anal. Chem. 25 665–73 [4] Mieth M, Schubert J K, Groger T, Sabel B, Kischkel S, Fuchs P, Hein D, Zimmermann R and Miekisch W 2009 Automated needle trap heart-cut GC/MS and needle trap comprehensive 2D GC/TOF-MS for breath gas analysis in the clinical environment Anal. Chem. 81 5851–7 [5] Sanchez J M and Sacks R D 2006 Development of a multibed sorption trap, comprehensive 2D gas chromatography, and time-of-flight mass spectrometry system for the analysis of volatile organic compounds in human breath Anal. Chem. 78 3046–54 [6] Trefz P, Kischkel S, Hein D, James E S, Schubert J K and Miekisch W 2012 Needle trap micro-extraction for VOC analysis: effects of packing materials and desorption parameters J. Chromatogr. A 1219 29–38 [7] Ueta I, Samsudin E L, Mizuguchi A, Takeuchi H, Shinki T, Kawakubo S and Saito Y 2014 Double-bed-type extraction needle packed with activated-carbon-based sorbents for very volatile organic compounds J. Pharm. Biomed. Anal. 88 423–28 [8] Buszewski B, Kęsy M, Ligor T and Amann A 2007 Human exhaled air analytics: biomarkers of diseases Biomed. Chromatogr. 21 553–66 [9] Buszewski B, Grzywiński D, Ligor T, Stacewicz T, Bielecki Z and Wojtas J 2013 Detection of volatile organic compounds as biomarkers in breath analysis by different analytical techniques Bioanalysis 5 2287–306 [10] Kim K H, Jahan S A and Kabir E 2012 A review of breath analysis for diagnosis of human health Trends Anal. Chem. 33 1–8 [11] Ulanowska A, Kowalkowski T, Trawińska E and Buszewski B 2011 The application of statistical methods using VOCs to identify patients with lung cancer J. Breath Res. 5 046008 [12] Boots A W, van Berkel J J B N, Dallinga J W, Smolinska A, Wouters E F and van Schooten F J 2012 The versatile use of exhaled volatile organic compounds in human health and disease J. Breath Res. 6 027108 [13] Philips M, Gleeson K, Hughes J M B, Greenberg J, Cataneo R N and Baker L 1999 Volatile organic compounds in breath as markers of lung cancer: a cross-sectional study Lancet 353 1930–33 [14] Phillips M, Herrera J, Krishnan S, Zain M, Greenberg J and Cataneo R N 1999 Variation in volatile organic compounds in the breath of normal humans J. Chromatogr. B 729 75–88 [15] Smolinska A, Hauschild A-Ch, Fijten R R R, Dallinga J W, Baumbach J and van Schooten F J 2014 Current breathomics:
7
a review on data pre-processing techniques and machine learning in metabolomics breath analysis J. Breath Res. 8 027105 [16] Hauschild A-Ch, Kopczynski D, D’Addario M, Baumbach J I, Rahmann S and Baumbach J 2013 Peak detection method evaluation for ion mobility spectrometry by using machine learning approaches Metabolites 3 277–93 [17] Ligor T, Ligor M, Amann A, Ager C, Bachler M, Dzien A and Buszewski B 2008 The analysis of healthy volunteers’ exhaled breath by use of solid phase microextraction and GC-MS J. Breath Res. 2 046006 [18] Buszewski B, Ulanowska A, Ligor T, Denderz N and Amann A 2009 Analysis of exhaled breath from smokers, passive smokers and non-smokers by solidphase microextraction gas chromatography/mass spectrometry Biomed. Chromatogr. 23 551–6 [19] Ligor M et al 2009 Determination of volatile organic compounds in exhaled breath of patients with lung cancer using solid phase microextraction and gas chromatography mass spectrometry Clin. Chem. Lab. Med. 47 550–60 [20] Filipiak W et al 2012 Dependence of exhaled breath composition on exogenous factors, smoking habits and exposure to air pollutants J. Breath Res. 6 036008 [21] Miekisch W, Schubert J K, Noeldge-Schomburg G F E 2004 Diagnostic potential of breath analysisfocus on volatile organic compounds Clin. Chim. Acta 347 25–39 [22] Phillips M, Cataneo R N, Greenberg J, Gunawardena R, Naidu A and Rahbari-Oskoui F 2000 Effect of age on the breath methylated alkane contour, a display of apparent new markers of oxidative stress J. Lab. Clin. Med. 136 243–9 [23] Poli D, Goldoni M, Corradi M, Acampa O, Carbognani P, Internullo E, Casalini A and Mutti A 2010 Determination of aldehydes in exhaled breath of patients with lung cancer by means of on-fiber-derivatisation SPME–GC/MS J. Chromatogr. B 878 2643–51 [24] Song G, Qin T, Liu H, Xu G B, Pan Y Y, Xiong F X, Gu K S, Sun G P and Chen Z D 2010 Quantitative breath analysis of volatile organic compounds of lung cancer patients Lung Cancer 67 227–31 [25] Philips M et al 2008 Detection of lung cancer using weighted digital analysis of breath biomarkers Clin. Chim. Acta 393 76–84 [26] Bajtarevic A et al 2009 Noninvasive detection of lung cancer by analysis of exhaled breath BMC Cancer 9 348 [27] Filipiak W, Sponring A, Filipiak A, Ager C, Schubert J, Miekisch W, Amann A and Troppmair J 2010 TD-GC-MS Analysis of volatile metabolites of human lung cancer and normal cells in vitro Cancer Epidemiol. Biomarkers 19 182–95 [28] Westhoff M, Litterst P, Maddula S, Bödeker B, Rahmann S, Davies A N and Baumbach J I 2010 Differentiation of chronic obstructive pulmonary disease (COPD) including lung cancer from healthy control group by breath analysis using ion mobility spectrometry Int. J. Ion Mobil. Spectrom. 13 131–39 [29] Philips M, Cataneo R N, Ditkoff B A, Fisher P, Greenberg J, Gunawardena R, Kwon C S, Rahbari-Oskoui F and Wong C 2003 Volatile markers of breath cancer in the breath Breast J. 9 184–91 [30] Hakim M, Billan S, Tisch U, Peng G, Dvrokind I, Marom O, Abdah-Bortnyak R, Kuten A and Haick H 2011 Diagnosis of head-and-neck cancer from exhaled breath Br. J. Cancer 104 1649–55 [31] Kushch I et al 2008 Breath isoprene: aspects of normal physiology related to age, gender and cholesterol profile as determined in a proton transfer reaction mass spectrometry study Clin. Chem. Lab. Med. 46 1011–8 [32] Schwarz K, Pizzini A, Arendacká B, Zerlauth K, Filipiak W, Schmid A, Dzien A, Neuner S, Lechleitner M and SchollBürgi S 2009 Breath acetone: aspects of normal physiology related to age and gender as determined in a PTR-MS study J. Breath Res. 3 027003
