Prospective Detection of Early Lung Cancer in Patients With COPD in Regular Care by Electronic Nose Analysis of Exhaled Breath

BackgroundPatients with COPD are at high risk of lung cancer developing, but no validated predictive biomarkers have been reported to identify these patients. Molecular profiling of exhaled breath by electronic nose (eNose) technology may qualify for early detection of lung cancer in patients with COPD.Research QuestionCan eNose technology be used for prospective detection of early lung cancer in patients with COPD?Study Design and MethodsBreathCloud is a real-world multicenter, prospective, follow-up study using diagnostic and monitoring visits in day-to-day clinical care of patients with a standardized diagnosis of asthma, COPD, or lung cancer. Breath profiles were collected at inclusion in duplicate by a metal-oxide semiconductor eNose positioned at the rear end of a pneumotachograph (SpiroNose). All patients with COPD were managed according to standard clinical care, and the incidence of clinically diagnosed lung cancer was prospectively monitored for 2 years. Data analysis involved advanced signal processing, ambient air correction, and statistics based on principal component (PC) analysis, linear discriminant analysis, and receiver operating characteristic analysis.ResultsExhaled breath data from 682 patients with COPD and 211 patients with lung cancer were available. Thirty-seven patients with COPD (5.4%) demonstrated clinically manifest lung cancer within 2 years after inclusion. PCs 1, 2, and 3 were significantly different between patients with COPD and those with lung cancer in both training and validation sets with areas under the receiver operating characteristic curve (AUCs) of 0.89 (CI, 0.83-0.95) and 0.86 (CI, 0.81-0.89). The same three PCs showed significant differences (P < .01) at baseline between patients with COPD who did and did not subsequently demonstrate lung cancer within 2 years, with a cross-validation value of 87% and AUC of 0.90 (CI, 0.84-0.95).InterpretationExhaled breath analysis by eNose identified patients with COPD in whom lung cancer became clinically manifest within 2 years after inclusion. These results show that eNose assessment may detect early stages of lung cancer in patients with COPD. Patients with COPD are at high risk of lung cancer developing, but no validated predictive biomarkers have been reported to identify these patients. Molecular profiling of exhaled breath by electronic nose (eNose) technology may qualify for early detection of lung cancer in patients with COPD. Can eNose technology be used for prospective detection of early lung cancer in patients with COPD? BreathCloud is a real-world multicenter, prospective, follow-up study using diagnostic and monitoring visits in day-to-day clinical care of patients with a standardized diagnosis of asthma, COPD, or lung cancer. Breath profiles were collected at inclusion in duplicate by a metal-oxide semiconductor eNose positioned at the rear end of a pneumotachograph (SpiroNose). All patients with COPD were managed according to standard clinical care, and the incidence of clinically diagnosed lung cancer was prospectively monitored for 2 years. Data analysis involved advanced signal processing, ambient air correction, and statistics based on principal component (PC) analysis, linear discriminant analysis, and receiver operating characteristic analysis. Exhaled breath data from 682 patients with COPD and 211 patients with lung cancer were available. Thirty-seven patients with COPD (5.4%) demonstrated clinically manifest lung cancer within 2 years after inclusion. PCs 1, 2, and 3 were significantly different between patients with COPD and those with lung cancer in both training and validation sets with areas under the receiver operating characteristic curve (AUCs) of 0.89 (CI, 0.83-0.95) and 0.86 (CI, 0.81-0.89). The same three PCs showed significant differences (P < .01) at baseline between patients with COPD who did and did not subsequently demonstrate lung cancer within 2 years, with a cross-validation value of 87% and AUC of 0.90 (CI, 0.84-0.95). Exhaled breath analysis by eNose identified patients with COPD in whom lung cancer became clinically manifest within 2 years after inclusion. These results show that eNose assessment may detect early stages of lung cancer in patients with COPD. Take-home PointsStudy Question: Can electronic nose (eNose) technology be used for prospective detection of early lung cancer in patients with COPD?Results: The eNose was able to discriminate patients with COPD who subsequently received a lung cancer diagnosis from those who did not receive such a diagnosis with 87% accuracy, an area under the receiver operating characteristic curve of 0.90 (95% CI, 0.84-0.95), 86% sensitivity, and 89% specificity.Interpretation: These results show that eNose assessment may detect early stages of lung cancer in patients with COPD and therefore may be of value in screening this risk group. Study Question: Can electronic nose (eNose) technology be used for prospective detection of early lung cancer in patients with COPD? Results: The eNose was able to discriminate patients with COPD who subsequently received a lung cancer diagnosis from those who did not receive such a diagnosis with 87% accuracy, an area under the receiver operating characteristic curve of 0.90 (95% CI, 0.84-0.95), 86% sensitivity, and 89% specificity. Interpretation: These results show that eNose assessment may detect early stages of lung cancer in patients with COPD and therefore may be of value in screening this risk group. Patients with COPD are at higher risk of lung cancer developing, with studies showing a relative risk of twofold to fourfold compared with the general population.1de Torres J.P. Marín J.M. Casanova C. et al.Lung cancer in patients with chronic obstructive pulmonary disease—incidence and predicting factors.Am J Respir Crit Care Med. 2011; 184: 913-919Crossref PubMed Scopus (225) Google Scholar Although several biomarkers are candidates for lung cancer discovery, such as autoantibodies, complement fragments, microRNA, circulating DNA, DNA methylation, RNA and protein profiling, and metabolomics,2Seijo L.M. Peled N. Ajona D. et al.Biomarkers in lung cancer screening: achievements, promises, and challenges.J Thorac Oncol. 2019; 14: 343-357Abstract Full Text Full Text PDF PubMed Scopus (238) Google Scholar no validated biomarkers have been discovered yet that can identify patients with COPD who are at a higher risk of lung cancer developing. During the past decade, screening studies using low-dose CT (LDCT) imaging have shown significant reductions (up to 39% in women) in lung cancer mortality rates among high-risk individuals, who were defined based on tobacco use history (status, pack-years, quit time) and age (50-75 years).3Aberle D.R. Adams A.M. Berg C.D. et al.Reduced lung-cancer mortality with low-dose computed tomographic screening.N Engl J Med. 2011; 365: 395-409Crossref PubMed Scopus (7282) Google Scholar, 4Ru Zhao Y. Xie X. de Koning H.J. Mali W.P. Vliegenthart R. Oudkerk M. NELSON lung cancer screening study.Cancer Imaging. 2011; 11 (spec no A(1a)): S79-S84Crossref PubMed Scopus (145) Google Scholar, 5de Koning H.J. van der Aalst C.M. de Jong P.A. et al.Reduced lung-cancer mortality with volume CT screening in a randomized trial.N Engl J Med. 2020; 382: 503-513Crossref PubMed Scopus (1347) Google Scholar Interestingly, a recently published study showed that patients with COPD are at higher risk of lung cancer regardless of tobacco use history.6Park H.Y. Kang D. Shin S.H. et al.Chronic obstructive pulmonary disease and lung cancer incidence in never smokers: a cohort study.Thorax. 2020; 75: 506-509Crossref PubMed Scopus (44) Google Scholar Because the potential benefits of screening might be exceeded by the increased risk of death inherent to COPD and its associated comorbidities, concerns exist about the inclusion of these patients in lung cancer screening programs.7Wiener R.S. Schwartz L.M. Woloshin S. Welch H.G. Population-based risk for complications after transthoracic needle lung biopsy of a pulmonary nodule: an analysis of discharge records.Ann Intern Med. 2011; 155: 137-144Crossref PubMed Scopus (368) Google Scholar,8Kozower B.D. Sheng S. O’Brien S.M. et al.STS database risk models: predictors of mortality and major morbidity for lung cancer resection.Ann Thorac Surg. 2010; 90 (discussion 881-883): 875-881Abstract Full Text Full Text PDF PubMed Scopus (261) Google Scholar In addition, concerns regarding increased overdiagnosis by LDCT imaging and invasive follow-up investigations limit its applicability to patients with COPD.9Patz Jr., E.F. Pinsky P. Gatsonis C. et al.Overdiagnosis in low-dose computed tomography screening for lung cancer.JAMA Intern Med. 2014; 174: 269-274Crossref PubMed Scopus (570) Google Scholar Therefore, an urgent need exists for an accurate and noninvasive test that can be implemented at the point of care. Specifically, a test that can refine the selection procedure of high-risk individuals for further follow-up screening tests is likely to prevail in clinical practice. Molecular profiling of exhaled breath by electronic nose (eNose) technology may qualify for the early detection of lung cancer.2Seijo L.M. Peled N. Ajona D. et al.Biomarkers in lung cancer screening: achievements, promises, and challenges.J Thorac Oncol. 2019; 14: 343-357Abstract Full Text Full Text PDF PubMed Scopus (238) Google Scholar eNose technology is an appealing noninvasive approach that applies advanced pattern recognition algorithms for analysis of the mixture of volatile organic compounds (VOCs) in exhaled breath.10Rocco G. Pennazza G. Santonico M. et al.Breathprinting and early diagnosis of lung cancer.J Thorac Oncol. 2018; 13: 883-894Abstract Full Text Full Text PDF PubMed Scopus (23) Google Scholar,11Farraia M.V. Cavaleiro Rufo J. Paciência I. Mendes F. Delgado L. Moreira A. The electronic nose technology in clinical diagnosis: a systematic review.Porto Biomed J. 2019; 4: e42Crossref PubMed Google Scholar The thousands of VOCs present in exhaled breath reflect the metabolic processes occurring in the host both locally in the airways and systemically.12Bos L. Sterk P. Fowler S. Breathomics in asthma and COPD.J Allergy Clin Immunol. 2016; 138PubMed Google Scholar Comprehensive analysis of these VOC patterns (breathomics) provides opportunities for noninvasive biomarker discovery in lung cancer.10Rocco G. Pennazza G. Santonico M. et al.Breathprinting and early diagnosis of lung cancer.J Thorac Oncol. 2018; 13: 883-894Abstract Full Text Full Text PDF PubMed Scopus (23) Google Scholar,13Lamote K. Van den Heuvel M.M. Where the nose is going to help the eye: sniffing lung cancer.Lung Cancer. 2021; 154: 195-196Abstract Full Text Full Text PDF PubMed Scopus (0) Google Scholar,14Haick H. Hashoul D. Lung cancer breath tests.Expert Rev Respir Med. 2019; 13: 597-599Crossref PubMed Scopus (6) Google Scholar By using eNoses, differences in exhaled VOC patterns of patients with COPD, patients with lung cancer, and healthy individuals already have been demonstrated by multiple research groups.11Farraia M.V. Cavaleiro Rufo J. Paciência I. Mendes F. Delgado L. Moreira A. The electronic nose technology in clinical diagnosis: a systematic review.Porto Biomed J. 2019; 4: e42Crossref PubMed Google Scholar,15Dragonieri S. Annema J.T. Schot R. et al.An electronic nose in the discrimination of patients with non-small cell lung cancer and COPD.Lung Cancer. 2009; 64: 166-170Abstract Full Text Full Text PDF PubMed Scopus (315) Google Scholar, 16Krauss E. Haberer J. Barreto G. Degen M. Seeger W. Guenther A. Recognition of breathprints of lung cancer and chronic obstructive pulmonary disease using the Aeonose® electronic nose.J Breath Res. 2020; Crossref PubMed Scopus (13) Google Scholar, 17Poli D. Carbognani P. Corradi M. et al.Exhaled volatile organic compounds in patients with non-small cell lung cancer: cross sectional and nested short-term follow-up study.Respir Res. 2005; 6: 71Crossref PubMed Scopus (326) Google Scholar, 18Krilaviciute A. Heiss J.A. Leja M. Kupcinskas J. Haick H. Brenner H. Detection of cancer through exhaled breath: a systematic review.Oncotarget. 2015; 6: 38643-38657Crossref PubMed Scopus (127) Google Scholar, 19Tsou P.-H. Lin Z.-L. Pan Y.-C. et al.Exploring volatile organic compounds in breath for high-accuracy prediction of lung cancer.Cancers. 2021; 13: 1431Crossref PubMed Scopus (23) Google Scholar Therefore, we hypothesized that metabolic and molecular changes that occur in early stage asymptomatic lung cancer can be detected from exhaled breath using an eNose. This study aimed to determine the diagnostic accuracy of exhaled breath analysis by eNose for (1) the discrimination between patients with COPD and those with lung cancer in a training and validation set and (2) the prospective prediction of early lung cancer in COPD. By using this stepwise approach and transparent reporting, the study follows the recommendations of the Standards for Reporting of Diagnostic Accuracy Studies guidelines20Cohen J.F. Korevaar D.A. Altman D.G. et al.STARD 2015 guidelines for reporting diagnostic accuracy studies: explanation and elaboration.BMJ Open. 2016; 6e012799Crossref Scopus (1078) Google Scholar and the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis statement.21Collins G.S. Reitsma J.B. Altman D.G. Moons K.G.M. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement.BMC Med. 2015; 13: 1Crossref PubMed Scopus (613) Google Scholar BreathCloud is a real-world multicenter observational study in healthy control participants and patients with a suspected or established diagnosis of asthma, COPD, or lung cancer.22de Vries R. Dagelet Y.W.F. Spoor P. et al.Clinical and inflammatory phenotyping by breathomics in chronic airway diseases irrespective of the diagnostic label.Eur Respir J. 2018; 51Crossref Scopus (87) Google Scholar All patients who visited the lung function departments for diagnostic and monitoring assessments in day-to-day care were recruited sequentially. The presently reported data include results from 682 patients with COPD and 211 patients with lung cancer who were included between May 2017 and November 2018. Patients with COPD were characterized according to the Global Initiative for Chronic Obstructive Lung Disease criteria23Vogelmeier C.F. Criner G.J. Martinez F.J. et al.Global Strategy for the Diagnosis, Management, and Prevention of Chronic Obstructive Lung Disease 2017 Report. GOLD Executive Summary.Am J Respir Crit Care Med. 2017; 195: 557-582Crossref PubMed Scopus (2078) Google Scholar and an established medical diagnosis of lung cancer was based on current guidelines.24Postmus P.E. Kerr K.M. Oudkerk M. et al.Early and locally advanced non-small-cell lung cancer (NSCLC): ESMO clinical practice guidelines for diagnosis, treatment and follow-up.Ann Oncol. 2017; 28: iv1-iv21Abstract Full Text Full Text PDF PubMed Scopus (1111) Google Scholar,25Früh M. De Ruysscher D. Popat S. Crinò L. Peters S. Felip E. Small-cell lung cancer (SCLC): ESMO clinical practice guidelines for diagnosis, treatment and follow-up.Ann Oncol. 2013; 24: vi99-vi105Abstract Full Text Full Text PDF PubMed Scopus (474) Google Scholar Exclusion criteria for participating in this study were the recent (< 12 h) intake of alcohol or if patients were not willing or able to participate. No further restrictions (eg, eating, drinking, tobacco use, or medication use) on participation were made to increase the applicability of breath analysis by eNose in clinical practice. The ethics board of all participating centers concluded in writing that Dutch legislation on human participation in research was not considered to be applicable, given the noninvasive and minimally bothering nature of this study that merely added exhaled breath analysis by eNose to standard diagnostic procedures.22de Vries R. Dagelet Y.W.F. Spoor P. et al.Clinical and inflammatory phenotyping by breathomics in chronic airway diseases irrespective of the diagnostic label.Eur Respir J. 2018; 51Crossref Scopus (87) Google Scholar,26de Vries R. Muller M. van der Noort V. et al.Prediction of response to anti-PD-1 therapy in patients with non-small-cell lung cancer by electronic nose analysis of exhaled breath.Ann Oncol. 2019; 30: 1660-1666Abstract Full Text Full Text PDF PubMed Scopus (50) Google Scholar,27de Vries R. Brinkman P. van der Schee M.P. et al.Integration of electronic nose technology with spirometry: validation of a new approach for exhaled breath analysis.J Breath Res. 2015; 9046001Crossref PubMed Scopus (95) Google Scholar Despite the waiver that was provided by the ethics review board, the purpose of adding the eNose to routine diagnostics was explained to the patients, all of whom gave their oral consent. The study had an observational, prospective, follow-up design. At baseline all patients with COPD underwent eNose assessment. The patients with COPD were treated according to standard care, and the subsequent incidence of clinically diagnosed lung cancer was assessed prospectively based on documented clinical records by following up the patients prospectively for 2 years. An established medical diagnosis of lung cancer was confirmed with CT scan imaging and was based on current guidelines.24Postmus P.E. Kerr K.M. Oudkerk M. et al.Early and locally advanced non-small-cell lung cancer (NSCLC): ESMO clinical practice guidelines for diagnosis, treatment and follow-up.Ann Oncol. 2017; 28: iv1-iv21Abstract Full Text Full Text PDF PubMed Scopus (1111) Google Scholar,25Früh M. De Ruysscher D. Popat S. Crinò L. Peters S. Felip E. Small-cell lung cancer (SCLC): ESMO clinical practice guidelines for diagnosis, treatment and follow-up.Ann Oncol. 2013; 24: vi99-vi105Abstract Full Text Full Text PDF PubMed Scopus (474) Google Scholar Clinical assessment of patients with COPD was performed using the Clinical COPD Questionnaire.28Reda A.A. Kotz D. Kocks J.W. Wesseling G. van Schayck C.P. Reliability and validity of the clinical COPD questionnaire and chronic respiratory questionnaire.Respir Med. 2010; 104: 1675-1682Abstract Full Text Full Text PDF PubMed Scopus (62) Google Scholar In addition, the personal best FEV1 % predicted after bronchodilation was used from data collected up to 12 months before inclusion. Other clinical data were collected for routine clinical care and subsequently were handled by complying with the Wet Bescherming Persoonsgegevens (Dutch Personal Data Protection Act). Exhaled breath measurements were performed in duplicate using a technically and clinically validated eNose, the SpiroNose.22de Vries R. Dagelet Y.W.F. Spoor P. et al.Clinical and inflammatory phenotyping by breathomics in chronic airway diseases irrespective of the diagnostic label.Eur Respir J. 2018; 51Crossref Scopus (87) Google Scholar,26de Vries R. Muller M. van der Noort V. et al.Prediction of response to anti-PD-1 therapy in patients with non-small-cell lung cancer by electronic nose analysis of exhaled breath.Ann Oncol. 2019; 30: 1660-1666Abstract Full Text Full Text PDF PubMed Scopus (50) Google Scholar,27de Vries R. Brinkman P. van der Schee M.P. et al.Integration of electronic nose technology with spirometry: validation of a new approach for exhaled breath analysis.J Breath Res. 2015; 9046001Crossref PubMed Scopus (95) Google Scholar,29Moor C.C. Oppenheimer J.C. Nakshbandi G. et al.Exhaled breath analysis by use of eNose technology: a novel diagnostic tool for interstitial lung disease.Eur Respir J. 2020; 2002042Crossref PubMed Scopus (25) Google Scholar The SpiroNose consists of seven different cross-reactive metal-oxide semiconductor sensors (sensors 1-7) for sampling of exhaled air. Another set of the same sensors sampled ambient air for background correction (Fig 1). SpiroNose measurements comprise five tidal breaths, followed by an inspiratory capacity maneuver to total lung capacity, a 5-s breath hold, followed by a slow expiration (< 0.4 L/s) to residual volume. The raw SpiroNose sensor signals were sent in real time to an online analysis platform for automated data analysis. An in-depth description of the measurement setup, and the verification of the sensor stability is published elsewhere.22de Vries R. Dagelet Y.W.F. Spoor P. et al.Clinical and inflammatory phenotyping by breathomics in chronic airway diseases irrespective of the diagnostic label.Eur Respir J. 2018; 51Crossref Scopus (87) Google Scholar,26de Vries R. Muller M. van der Noort V. et al.Prediction of response to anti-PD-1 therapy in patients with non-small-cell lung cancer by electronic nose analysis of exhaled breath.Ann Oncol. 2019; 30: 1660-1666Abstract Full Text Full Text PDF PubMed Scopus (50) Google Scholar The processing of the eNose sensor deflections was carried out automatically using the standard eNose software as was published previously.22de Vries R. Dagelet Y.W.F. Spoor P. et al.Clinical and inflammatory phenotyping by breathomics in chronic airway diseases irrespective of the diagnostic label.Eur Respir J. 2018; 51Crossref Scopus (87) Google Scholar,27de Vries R. Brinkman P. van der Schee M.P. et al.Integration of electronic nose technology with spirometry: validation of a new approach for exhaled breath analysis.J Breath Res. 2015; 9046001Crossref PubMed Scopus (95) Google Scholar Signal processing included signal detrending, filtering, ambient air correction, automatic peak detection, and parameter selection. From each sensor signal, two variables were determined: (1) the highest sensor peak normalized to the most stable sensor, sensor 2, to minimize inter array differences and (2) the ratio between the sensor peak and the breath hold point.22de Vries R. Dagelet Y.W.F. Spoor P. et al.Clinical and inflammatory phenotyping by breathomics in chronic airway diseases irrespective of the diagnostic label.Eur Respir J. 2018; 51Crossref Scopus (87) Google Scholar,26de Vries R. Muller M. van der Noort V. et al.Prediction of response to anti-PD-1 therapy in patients with non-small-cell lung cancer by electronic nose analysis of exhaled breath.Ann Oncol. 2019; 30: 1660-1666Abstract Full Text Full Text PDF PubMed Scopus (50) Google Scholar The sensor peak and peak to breath hold ratios were used for statistical analysis. A principal component analysis (PCA) was performed to merge the variables of interest into a multivariate component. According to the Kaiser criterion, all principal components (PCs) with an eigenvalue of > 1 were retained.30Yeomans K.A. Golder P.A. The Guttman-Kaiser criterion as a predictor of the number of common factors.Journal of the Royal Statistical Society Series D (The Statistician). 1982; 31: 221-229Crossref Google Scholar The processed sensor variables, the original sensor peaks, and the peak to breath hold ratios were restructured to four PCs that captured 78.4% of the variance within the dataset (PC 1, 39.8%; PC 2, 19.5%; PC 3, 11.1%; and PC 4, 8.0%). PCs were constructed for all participants (training and validation set) based on eNose data from participants within the training set (e-Table 1). Reducing the dimensionality of the data before machine learning is preferred to reduce the risk of overfitting.31Bikov A. Lázár Z. Horvath I. Established methodological issues in electronic nose research: how far are we from using these instruments in clinical settings of breath analysis?.J Breath Res. 2015; 9034001Crossref Scopus (61) Google Scholar,32Gromski P.S. Correa E. Vaughan A.A. Wedge D.C. Turner M.L. Goodacre R. A comparison of different chemometrics approaches for the robust classification of electronic nose data.Anal Bioanal Chem. 2014; 406: 7581-7590Crossref PubMed Scopus (61) Google Scholar Analyses were performed in both a training and validation set defined by random split analysis (2:1), as recommended for metabolomics experiments.33Broadhurst D.I. Kell D.B. Statistical strategies for avoiding false discoveries in metabolomics and related experiments.Metabolomics. 2006; 2: 171-196Crossref Scopus (626) Google Scholar The obtained PCs were compared between groups using independent t tests. The PCs that discriminated (P < .05) between groups were selected for further analysis. The t tests were validated internally by 1,000 bootstrap iterations . Subsequently, linear discriminant analysis was performed using the selected variables. The aim of linear discriminant analysis is to maximize the separability of the defined categories and represents a relatively simple and portable algorithm. The latter is recommended for eNose research31Bikov A. Lázár Z. Horvath I. Established methodological issues in electronic nose research: how far are we from using these instruments in clinical settings of breath analysis?.J Breath Res. 2015; 9034001Crossref Scopus (61) Google Scholar and often is used in clinical studies performing exhaled breath analysis with the SpiroNose.26de Vries R. Muller M. van der Noort V. et al.Prediction of response to anti-PD-1 therapy in patients with non-small-cell lung cancer by electronic nose analysis of exhaled breath.Ann Oncol. 2019; 30: 1660-1666Abstract Full Text Full Text PDF PubMed Scopus (50) Google Scholar,27de Vries R. Brinkman P. van der Schee M.P. et al.Integration of electronic nose technology with spirometry: validation of a new approach for exhaled breath analysis.J Breath Res. 2015; 9046001Crossref PubMed Scopus (95) Google Scholar,34de Vries R. Vigeveno R.M. Mulder S. et al.Ruling out SARS-CoV-2 infection using exhaled breath analysis by electronic nose in a public health setting.medRxiv. 2021; (2021.02.14.21251712)https://doi.org/10.1101/2021.02.14.21251712Crossref Scopus (0) Google Scholar Based on the differentiating PCs, a discriminant function was calculated that best distinguished between groups. The accuracy of this model was defined as the percentage of correctly classified patients in the training set. Leave-one-out cross-validation was used to calculate the cross-validated accuracy value (percentage). The discriminant scores were used to construct a receiver operating characteristic curve, including the area under the receiver operating characteristic curve (AUC) and corresponding 95% CI. Finally, the discriminant function obtained from the training set was examined in the independent validation set and compared mutually based on AUC curves. As a first step in the assessment of early lung cancer, the training and validation sets were combined and the accuracy, sensitivity, and specificity for the discrimination between COPD and lung cancer were assessed across the entire study population. A PCA plot was created to visualize the data. Finally, the influence of comorbid COPD in the lung cancer group on the accuracy of distinguishing lung cancer from COPD was assessed by removing all patients with a double diagnosis from the analysis (e-Appendix 1, e-Fig 1, e-Table 2). To avoid drawing conclusions based on a single classification model, we used three additional and powerful machine learning techniques to classify the eNose data: gradient boosting machine, adaptive least absolute shrinkage and selection operator, and sparse partial least squares discriminant analysis. These machine learning techniques all have been used before in metabolomics research35Bouwmeester R. Martens L. Degroeve S. Comprehensive and empirical evaluation of machine learning algorithms for small molecule LC retention time prediction.Anal Chem. 2019; 91: 3694-3703Crossref PubMed Scopus (47) Google Scholar, 36Cuperlovic-Culf M. Machine learning methods for analysis of metabolic data and metabolic pathway modeling.Metabolites. 2018; 8: 4Crossref PubMed Scopus (104) Google Scholar, 37Determan Jr., C.E. Optimal algorithm for metabolomics classification and feature selection varies by dataset.Int J Biol. 2015; 7: 100Google Scholar and provide an estimate of the robustness of the statistical performance across different models (e-Appendix 1, e-Tables 3, 4). To contribute to the standardization of the SpiroNose’s methods, we chose to present the linear discriminant analysis results throughout herein. Similar analyses including PCA, independent t tests, linear discriminant analysis, and receiver operating characteristic curve analysis were used to determine the diagnostic accuracy of eNose analysis for the discrimination between patients with COPD who did and did not receive a clinical diagnosis of lung cancer within 2 years after inclusion. In addition, sensitivity, specificity, a positive likelihood ratio (LR), and a negative LR were calculated. A PCA plot was created to visualize the data. A duplicate of this plot was created, in which the COPD group that did receive a clinical diagnosis of lung cancer within 2 years was divided into the different lung cancer stages (e-Fig 2). A multiple linear regression was calculated to predict eNose data based on the clinical metadata. More information on the analysis and results is presented in e-Tables 5-7. In total, 893 patients (682 patients with COPD and 211 patients with lung cancer) were included in this study, of whom 596 patients (455 patients with COPD and 141 patients with lung cancer) were included for training of the results and 297 patients (227 patients with COPD and 70 patients with lung cancer) for validation of the results (Table 1). In the training set, only pack-years of tobacco use and FEV1 showed significant differences (P < .05) between patients with COPD and those with lung cancer. No significant differences were found between these groups in baseline characteristics in the validation set. In addition, the baseline characteristics of patients in the training and validation set were similar, except for significantly different COPD staging (P < .05) and lung cancer pathologic features (P < .05) (Table 1).Table 1Baseline Characteristics of Patients With COPD and Those With Lung