Document MGoarj2Op6wEorz2xmpZLKyKy

CLIN. CHEM. 31/8,1278-1282 (1985) VolatileOrganicCompoundsin ExhaledAir from Patientswith LungCancer S. M. Gordon,' J. P. Szldon,2 B. K. Krotoszynski,' R. D. Gibbons,3 and H. J. O'Neill' Using a specially developed breath collection technique and computer-assisted gas chromatography/mass spectrometry (GC/MS), we have identified in the exhaled air of lung cancer patients severalvolatileorganic compounds that appear to be associated with the disease. The GCMS profiles of 12 samples from lung cancer patients and 17 control samples were analyzed by using general computerized statistical procedures to distinguish lung cancer patients from controls. The selected volatile compounds had sufficient diagnostic power in the GC/MS profiles to allow almost complete differentiation between the two groups in a limited patient population. AddItional Keyphrases: gas chromatography/mass spectrometty #{1d4i9a}gnostic screening . sample collection dataprocessing The clinical value of early detection of presymptomatic lung cancer has not yet been established. Current controlled trials of screening are aimed, in part, at resolving the issue of whether the observed increased survival of patients with such tumors is real or apparent (1,2). Diagnostic techniques currently used in these experimental studies of early detection (ultraviolet-assisted fiber-optic bronchoscopy, cytology) are costly and invasive, obviating their applicability to large-scale screening efforts. Alternative methods for early detection should therefore be explored. The presence in expired breath of certain volatile organic compounds in above-normal concentrations suggests a means for such medical diagnosis (3). Various analytical procedures have detected and quantified more than 200 compounds in the expired air of human subjects (3-9). The technology has originated from efforts to identify environmental pollutants in ambient air at concentrations of parts per billion (10) or less (10). Although the metabolic pathways by which compounds present in expired air are derived are generally unknown, interest in the clinical diagnostic potential of breath analysis is evidenced by preliminary characterizations of expired air in renal failure (11), liver disease (12-14), and diabetes (9, 15-17). In this study, we explored the possibility that the expired air of subjects with lung cancer might contain unique volatile organic compounds that could be used to develop noninvasive diagnostic screens. We used a computer-assisted methodology to extract characteristic volatile organic compounds from the expired breath of subjects with lung cancer and from controls, combining the use of gas chromatography/mass spectrometry (GC/MS) and multivariate statistics. MaterIals and Methods Selection of Subjects Lung cancer patients. Samples of expired breath were collected from 14 consecutive subjects of both sexes admitted fiT Research Institute, 10 West 35th St., Chicago, IL 60616. 2Department of Medicine, Michael Reese Hospital and Medical Center, 29th Street and Ellis Ave., Chicago, IL 60616. #{176}DepartmoefntPsychiatry, University of illinois at the Medical Center, 912 South Wood St., Chicago, IL 60680. Received March 13, 1985, accepted May 2, 1985. to Michael Reese Hospital and Medical Center for diagnosis and treatment.4 The diagnosis of lung cancer was established in each case by bronchoscopic or fine-needle biopsy of pulmonary lesions. Samples of expired air were collected before chemotherapy or radiation therapy was instituted. Comprehensive records of medications taken during the four- to six-week period preceding sample collection were available. There was no consistent pattern of medication used among these subjects; two subjects took thiazide diuretice, two used propranolol, one used digitalis, and three took aspirin or acetaminophen intermittently. In two cases, the GC!MS data were technically unsatisfactory because of unacceptably high background. None of the subjects had any significant co-morbid factors such as infections or liver! renal failure. The characteristics of the 12 patients studied are summarized in Table 1. 4Hoepitalized subjects were recruited by written informed consent. The protocols were approved by the Michael Reese Human ResearchCommittee. Table 1. DescrIption of Lung Cancer Study Subjects Clgar.tt.smoking Age, history, Occupation pe.rs Diagnosis 48 Steel worker 25 Poorlydifferentiated squamouscell carcinoma, left upper lobe, bilateral mediastinal nodes and cerebral metastasis. 53 Inkfactory 10 Large-cell carcinoma right worker lung, right-sided pleural effusionL. eftmediastinal enlargement. 58 Truck driver 40 Squamous cellcarcinomal,eft upperlobe.Lefthllar enlargementC. erebral metastasis. 68 Officeworker 30 Adenocarcinomaleft upper lobe, upper mediastinum,right hilum. 68 Businessman 82 Aderiocarcinomaleft upper lobe. Left hilarenlargement, hypercalcemia. 54 Salesperson 30 Adenocarclnoma,left hilar mass, predominantly 54 Securityguard 40 extrabronchial. Adenocarcinoma,rightupper lobe. Rib metastasis. 56 78 55 Printer 64 - 60 Squamouscell carcinoma,left upper lobe. Left pulmonary artery invasion. 50 Adenocarcinomaleft lung. Left-sidedpleuraleffusion. 74 Undifferentiatedcarcinoma, cavitary(with Aspergillus hffae in cavity). 50 Poorlydifferentiated adenocarcinomaof tracheal 78 Housewif&' bifurcation. 58 Squamous cellcarcinoma,left upper lobe. Lefthilar metastasis. ay031.5 of cigaretteconsumptionatrateofone packper day. bsjj othersubiects were male. 1278 CUNICAL CHEMISTRY, Vol. 31, No. 8, 1985 Control subjects. We also studied 17 samples of expired air from a control group of nine normal subjects, ages 25 to 70 years, including one woman. Two of the men were heavy cigarette smokers. None of the subjects had a history or physical evidence of pulmonary or systemic disease. Most of these control subjects were hospital personnel. Sampling and Analysis Sample collection. A specially designed system was used to collect samples of expired breath on 60/80 mesh Tenax GC (Alltech Associates, Arlington Heights, IL) sorbent cartridges for later analysis of the organic vapors by GCIMS. The system, described elsewhere (5, 6), was designed to isolate the subject from environmental contaminants temporarily and to provide normal breathing conditions, as shown in Figure 1. In brief, a sterilized Rudolph breathing valve (H. Rudolph, Inc., Kansas City, MO) was connected to the breathing manifold and a spring-loaded clamp attached to the person's nose. For 5 mm the subject inhaled purified air from the reservoir through the mouthpiece attached to the valve and exhaled through the valve into the atmosphere. This served to minimize the presence of any environmental contaminants in the lungs inhaled during ordinary exposure. The Rudolph valve was then connected to the heated expired-air delivery tube, and the subject was asked to fill the 40-L sampling bag. After each of two such fillings, the bag was emptied by applying external pressure to it with air from the spirometer. The third ifiling of the bag to capacity constituted the expired air used as the sample. During this ifiling, the spirometer measured the subject's ventilatory rate (tidal volume and frequency). After collecting the sample, we attached a Tenax GC cartridge to the outlet valve of the sampling bag and used a pump to draw the expired air through the cartridge, measuring the total volume of air transferred (about 20 L) with a wet test meter (Sargent-Welch Scientific Company, Skohe, IL). CC/MS analysis. The Tenax cartridges were analyzed for volatile organic compounds by a thermal desorption GC/MS procedure that has been described in detail elsewhere (10, 18). Desorption was performed by heating the cartridge in an inert gas stream and collecting the volatiles in a trap cooled with liquid nitrogen. Heating the trap rapidly while flushing it with carrier gas transferred the components onto a high-resolution capillary GC column, which was coupled to a mass spectrometer and data system for cyclic scan analysis. Fig. 1. Systemfor collectingexpiredair samples To improve the quality of the mass spectral data and permit comparisons between GC/MS data sets, we applied an efficient spectrum-enhancement algorithm (19) to the raw data. This program automatically locates and extracts components from the data to produce a set of clean spectra, i.e., free of background contributions and contaminating compounds. Once a clean spectrum has been defined by the program, its area is calculated and used to evaluate the concentration of a component in the sample by comparing the peak area of that component with the peak area of the internal standard. The standard, perfluorotoluene, was add- ed to each Tenax GC cartridge just before analysis. To match peaks and generate composite GC/MS proffles, a second program subjected each unknown spectrum to reten- tion-time scaling, then passed it to an historical library matching program (20). Retention-time scaling is achieved by converting the raw chromatographic retention data (spectrum numbers) into a set of standardized retention indices relative to a set of marker peaks that occur through- out the data base and are easily located. The matching program defines a time window, then compares each un- known spectrum with each reference spectrum in the win- dow of the library by performing a relative retention index (RRI) match and a spectrum match. Matching effectiveness is expressed in terms of a score that is a measure of spectral similarity and RRI proximity. (The historical library is initialized by designating one of the profiles of interest as the library.) When a satisfactory match occurs, the peak is added to the historical library and the relative concentra- tion and RRJ data for the entry are updated. In this way, we prepared composites of individual GC/MS proffles and established statistics on the frequency of occur- rence and reproducibility of the average peak intensities in the two sample populations of interest. We then applied various statistical tests to determine differences between the respective composite GCIMS profiles. Statistical Methods We used three general statistical procedures to reduce the dimensionality of the data and to establish relationships within the data from each group objectively. As a first screening for each of the 297 compounds in the lung cancer and control samples, we used the Fisher exact test to examine the null hypothesis of no association between lung cancer and the frequency of occurrence for each compound separately. A type I error rate of p <.05 indicated sigrnflcance in this initial screen. For the 57 breath compounds that were present in at least 50% of the patients and 35% of the controls, both t-statistics and Mann-Whitney U statistics were computed to test the null hypothesis of no difference in the concentration of each compound between patients and controls. In light of these multiple comparisons, we allowed an overall false-positive rate of 5% (i.e., p <.05) for individual compounds that have a probability of p <.001 (based on the Bonferoni inequality). Thus, probability values <.001 are significant in and of themselves, despite the large number of comparisons performed. These statistical procedures allowed us to select 22 peaks as a starting point for developing a discriminant between CC/MS proffles from lung cancer patients and normal persons. We constructed a discriminant function to yield a linear combination of compounds that maximally differenti- ated patients from controls, then tested the accuracy of the estimated function both statistically and by examining the concordance between the observed and estimated classification. We considered significant a classification accuracy of greater than 80%. CLINICALCHEMISTRY, Vol. 31, No. 8, 1985 1279 Ideally, validation of the result should be based on the classification of an independent data set. However, because no unknown samples were available for prediction, we used the `jackknife" validation procedure (24) in an effort to munimize this bias. In this approach, a classifier is constructed on the basis of 28 (of the 29 known) samples and the remaining sample is classified. This sample is then returned to the sample set, another sample is removed, a new classifier is developed for this set, and the deleted sample is classified. This procedure continues until all samples in the set have been deleted and classified once. Results The CC/MS profiles of expired air from lung cancer patients and controls were complex, generally containing about 150 peaks each. Figure 2 shows typical CC/MS profiles (plots of total ion current vs spectrum number or retention time) obtained for a lung cancer patient and a control subject; obviously, visual inspection alone is unlikely to reveal subtle differences in the relative intensities of peaks in the data sets. A list of the 49 peaks identified by RRI for which there were statistically significant differences in percent peak occurrence and (or) concentration, based on the statistical screening procedures described above, is available on request from the authors or from the Editorial Office of this journal. In this group of lung cancer patients, four peaks occtirred in more than half of the samples but were completely absent from the controls. If this observation is confirmed in a large 600000 500000 (a) 400000 z z ; 300000 0 200000 100000 a ii 200 400 600 600 1000 1200 1400 1600 SPECTRUM NUMBER (b) 600 600 1000 SPECTRUM NUMBER 1200 1400 1600 Ag. 2. Plots of total ion current vs spectrum number for GC/MS analysis of expired air samples from (a)lungcancer patient and (b) normal (smoker) subject population, the presence of these four peaks might alone provide sufficient diagnostic power in the CC/MS profiles to allow unambiguous differentiation between the lung cancer and control groups. However, the fact that the peaks did not occur in all the lung cancer patients indicates that a diagnostic approach based exclusively on their presence would probably have a false-negative rate that would be unacceptable for screening purposes. To obviate this possibility, we selected the 22 peaks showing the largest differences in peak percent occurrence and (or) concentration (Table 2) and used them to develop a linear discriminant function for classifying CC/MS proffles from either lung cancer or control subjects. Mass spectra,l analysis showed that the candidate substances included 16 oxygen-containing compounds and four sulfur-containing compounds. With the linear discriminant ftmction we developed, all 29 samples in the data set (17 from controls and 12 from lung cancer patients) were correctly classified; moreover, results were identical for both raw and jackknifed values. Although we initially entered 22 variables into the analysis, only seven of the peaks were needed for full discrimination; their REIs were 611, 653,690,739, 794,841, and 1248. However, three of these peaks occurred in fewer than 35% in one of the two sample groups. Because these low frequencies could have biased the result, we removed from the original list of 22 peaks all variables with occurrences below 50%. Consequently, 10 peaks (RRIs 524, 611, 680, 739, 789, 841, 1280, 1299, 1374, and 1499) were used to define a new discrirninant function. This function accurately classified 93% of the samples (27/29) with use of only three peaks (RRIs 611,680, and 841). Thus, despite the restriction on the peaks and their frequencies of occurrence, the linear discriminant analysis procedure was able to distinguish between the lung cancer and control samples highly accurately, based on significant differences in the concentrations of Table 2. PotentIally Significant DiagnostIc Peaks Based on FIsher Exact and Group Comparison Tests Relative retention lndex Peak presence/absence % Occurrence Lung cancer Control p VIIU.b Group comparison, p V.lUee Mann-. Whitney I-test 524 91.7 82.4 0.0010 0.056 611 100.0 100.0 0.0013 0.004 653 75.0 35.3 0.0407 0.01 76 0.038 659 50.0 0.0 0.0019 680 100.0 100.0 0.0213 0.001 690 8.3 64.7 0.0030 723 0.0 58.8 0.0010 725 91.7 35.3 0.0030 0.0067 0.002 739 100.0 88.2 0.0002 0.000 764 0.0 58.8 0.0010 779 100.0 35.3 0.0004 0.0242 0.000 789 100.0 94.1 0.0300 0.002 794 33.3 82.4 0.0106 829 66.7 17.6 0.0106 841 100.0 82.4 0.0003 0.000 1095 50.0 0.0 0.0019 1248 75.0 0.0 0.0001 1280 91.7 76.5 0.0039 0.022 1299 100.0 100.0 0.0040 0.003 1374 91.7 58.8 0.0019 0.010 1499 100.0 94.1 0.0026 0.001 1512 66.7 41.2 0.0001 0.000 asee text for definition.bFisherExact Probability Test.Group compari- sons calculated onlyfor those cases in whichoccurrence exceeded 5O% for lungcancergroupand 35% forcontrolgroup. 1280 CLINICAL CHEMISTRY, Vol. 31, No. 8, 1985 the compounds between the two groups. The statistical analyses described here show that the development of a discriminant that separates the lung cancer proffles from the normal profilesdoes not require a knowledge of the chemical identity of the selected peaks. Nevertheless, the availability of the mass spectral data for each sample allows us readily to identify these as well as other peaks that may have clinical diagnostic value. In the above analysis, for example, the three peaks that gave a classification accuracy of 93% correspond to acetone, methyl ethyl ketone, and n-propanol. Discussion The results of this study suggest that unique volatile compounds of potential diagnostic usefulness are present in expired air of patients with lung cancer. Data from a small sample of subjects diagnosed late in their disease were contrasted with those from a small control population. We are aware that experimental variables such as donor gender, donor age, and GC column batch can confound the chemical information and affect the classification process (21), and that misleading results are possible if the ratio of the number of samples to the number of variables (peaks) is less than about three (21,22). In addition, other confounding factors may have played an uncertain role. We doubt that the hospital environment was a significant factor, because most of the controls were hospital personnel. Similarly, there was no consistent pattern of medication usage that could have accounted for differences in expired air composition. However, it is possible that systematic differences in diet existed between controls and patients with lung cancer, and these may have contributed to some of the discriminant power. Nonetheless, we have obtained sufficient relevant information to justify the need for a large-scale, systematic study in which these and other confounding factors are taken into account by use of statistical methods. The analytical and statistical approaches described here clearly provide a powerful means of selecting diagnostically significant compounds in expired breath and establishing their diagnostic power and limitations for the classification of GCIMS profiles. Applying these procedures to a larger sample population should help confirm, refute, or expand the significance of the compounds selected in this preliminary study and establish diagnostic criteria for general use. The classification function finally developed could then be used in a dedicated analytical system to classify unknown samples rapidly and reliably. Most earlier work in diagnostic applications of expired gas analysis has contributed information of minor clinical value (3). However, the approach has generally been to study one or a few constituents, selected on the basis of their presumed relevance. Our approach contrasts with earlier studies in that we search concurrently for multiple volatile compounds, using state-of-the-art computer-assisted analytical technology and taking advantage of the potential diagnostic value of simultaneous changes in the concentrations of several constituents. A disadvantage of the present approach is its cost and the reliance on sophisticated technology. Our results suggest, however, that the presence of a discrete number of diagnostic compounds may suffice for identifying subjects with lung cancer. If this is confirmed by a larger study, perhaps a device based on nothing more than a high-resolution gas-chromatograph equipped with compound-specific detectors could be developed as an effective diagnostic tool. Such a device could be simple enough for potential use in a physician's consulting room as part of a routine pulmonary test. Even if the discriminant function requires the assessment of more than a few compounds, there is still potential for simplification. Technology for such measurements is already available with the advent of the tandem mass spectrometry (MS/MS) technique. This technique, coupled with direct-sampling atmospheric pressure ionization, is capable of real-time (i.e., instantaneous) detection and identification of extremely low concentrations (parts per trillion, 10 12) of compounds in air (23). Although these systems are still relatively expensive, the technology lends itself to miniaturization and a significant reduction incest, once the performance requirements and mode of operation have been clearly defined. The incidence of lung cancer continues to increase. Prob- lems with early diagnosis and treatment still represent major clinical and epidemiological challenges. The data obtained here indicate that the analysis of expired air is feasible and could provide a noninvasive diagnostic method for use in mass screening at a relatively low cost. Further studies to develop and refine this approach are justified. We gratefully acknowledge the support, assistance, and suggestions of Dr. Demetrios Moachandreas. Drs. Charles Shapiro, Jacob Bitran, and Richard Evans contributed by referring their patients. We are also indebted to Dr. Terrance Kane for collecting the expired air samples and to Ms. Louise Brousek and Ms. Sally Amagai for performing the CC/MS analyses. References 1. NIH Publication No. 79, NC! Cooperative Early Lung Cancer Group, U.S. Dept. of Health, Bethesda, MD, 1979. 2. Bailar JC. Screening for lung cancer-where are we now? Am Rev Resp Dir 130, 541-542 (1984). 3. Manolis A. The diagnostic potential of breath analysis. Clin Chem 29, 5-15 (1983). Review. 4. Conkle JP, Camp BJ, Welch BE. Trace composition of human respiratory breath. Arch Environ Health 30, 290-295 (1975). 5. Krotoazynski BK, Gabriel G, O'Neill HJ, Claudio MPA. Characterization of human expired air A promising investigative and diagnostic technique. J Chromatogr Sci 15, 239-244 (1977). 6. Kroto#{128}zyBnsKk,i Bruneau GM, O'Neill HJ. Measurement of chemical inhalation exposure in an urban population in the presence of endogenous effluents. JAnal Toxicol 3,225-234(1979). 7. Krotoezynski BK, O'Neill HJ. Involuntary bioaccumulation of environmental pollutants in nonsmoking heterogeneous human population. J Environ Sci Health A17, 855-883 (1982). 8. Barkley J, Bunch J, Bursey JT, et al. Gas chromatography-mass spectrometry-computer analysis of volatile halogenated hydrocarbans in man and his environment-a multimedia environmental study. Biomed Mass Spectrom 7, 139-147 (1980). 9. Lovett AM, Reid NM, Buckley JA, et al. Real-time analysis of breath using an atmospheric pressure ionization mass spectrometer. Biomed Mass Spectrom 6, 91-97 (1979). 10. Pellizzari ED. Analysis of organic air pollutants by gas chromatography and mass spectroscopy. U.S. Environmental Protection Agency, EPA-600/2-77-100, June 1977. 11. Simenhoff ML, Burke JF, Saukkonen JJ, et al. Biochemical proffle of uremic breath. N EngI J Med 297, 132-135 (1977). 12. Kaji H, Hisamura M, Saito N, Murao M. Evaluation of volatile sulfur compounds in the expired alveolar gas in patients with liver cirrhosis. Clin Chim Ada 85, 279-284 (1978). 13. Chen S, Mahadevan V, Zieve L. Volatile fatty acids in the breath of patients with cirrhosis of the liver. J Lab Clin Med 75, 622-627 (1970). 14. Routh JI. Liver function. In Fundamentals of Clinical Chemistry, NW Tietz, Ed., WB Saunders, Philadelphia, PA, 1976, pp 10261062. 15. Levey 5, Balchuin OJ, Medrano V, Jung R. Studies of metabolic products in expired air-il. Acetone. J Lab Clin Med 63, 574-584 (1964). CLINICAL CHEMISTRY, Vol.31, No. 8, 1985 1281 16. Rooth G, Ostenson S. Acetone in alveolar air and the control of diabetes. Lancet ii, 1102-1105 (1966). 17. Crofford OB, Mallard RE, Winton RE, et al. Acetone in breath and blood. Trans Am Clin Climatol Assoc 88, 128-139 (1977). 18. Gordon SM. GC/MS characterization of volatile organic pollutants in ambient air. First Annual Nati. Symp. on Monitoring Hazardous Organic Pollutants in Air, Raleigh, NC, April 1981. 19. Dromey RG, Stefik MJ, Rindfleisch TC, Duffleld AM. Extraction of mass spectra free of background and neighboring component contributions from gas chromatography/mass spectrometry data. Anal Chem 48, 1368-1375 (1976). 20. Smith DH, Achenbach M, Yeager WJ, et al. Quantitative comparison of combined gas chromatographic/mass spectrometric proffles of complex mixtures. Anal Chem 49, 1623-1632 (1977). 21. Pino JA, McMurry JE, Jurs PC, et al. Application of pyrolysis/ gas chromatography/pattern recognition to the detection of cystic fibrosis heterozygotes. Anal Chem 57, 295-302 (1985). 22. Tou JT, Gonzalez RC. Pattern-Recognition Principles, AddisonWesley, Reading, MA, 1974. 23. For example: Maugh TH. Separations by MS speed up, simpli1r analysis. Science 209, 675-677 (1980); McLafferty FW. Tandem mass spectrometry. Science 214, 280-287 (1981). 24. Dixon WJ, Brown MB, Eds. BMDP Biomedical Computer Programs P-Series 1979, University of California Press, Berkeley, CA, 1979. 1282 CLINICAL CHEMISTRY, Vol.31,No. 8, 1985