Diagnosing Non-Small Cell Lung Cancer by Exhaled Breath Profiling Using an Electronic Nose

BackgroundDespite the potential of exhaled breath analysis of volatile organic compounds to diagnose lung cancer, clinical implementation has not been realized, partly due to the lack of validation studies.Research QuestionThis study addressed two questions. First, can we simultaneously train and validate a prediction model to distinguish patients with non-small cell lung cancer from non-lung cancer subjects based on exhaled breath patterns? Second, does addition of clinical variables to exhaled breath data improve the diagnosis of lung cancer?Study Design and MethodsIn this multicenter study, subjects with non-small cell lung cancer and control subjects performed 5 min of tidal breathing through the aeoNose, a handheld electronic nose device. A training cohort was used for developing a prediction model based on breath data, and a blinded cohort was used for validation. Multivariable logistic regression analysis was performed, including breath data and clinical variables, in which the formula and cutoff value for the probability of lung cancer were applied to the validation data.ResultsA total of 376 subjects formed the training set, and 199 subjects formed the validation set. The full training model (including exhaled breath data and clinical parameters from the training set) were combined in a multivariable logistic regression analysis, maintaining a cut off of 16% probability of lung cancer, resulting in a sensitivity of 95%, a specificity of 51%, and a negative predictive value of 94%; the area under the receiver-operating characteristic curve was 0.87. Performance of the prediction model on the validation cohort showed corresponding results with a sensitivity of 95%, a specificity of 49%, a negative predictive value of 94%, and an area under the receiver-operating characteristic curve of 0.86.InterpretationCombining exhaled breath data and clinical variables in a multicenter, multi-device validation study can adequately distinguish patients with lung cancer from subjects without lung cancer in a noninvasive manner. This study paves the way to implement exhaled breath analysis in the daily practice of diagnosing lung cancer.Clinical Trial RegistrationThe Netherlands Trial Register; No.: NL7025; URL: https://trialregister.nl/trial/7025 Despite the potential of exhaled breath analysis of volatile organic compounds to diagnose lung cancer, clinical implementation has not been realized, partly due to the lack of validation studies. This study addressed two questions. First, can we simultaneously train and validate a prediction model to distinguish patients with non-small cell lung cancer from non-lung cancer subjects based on exhaled breath patterns? Second, does addition of clinical variables to exhaled breath data improve the diagnosis of lung cancer? In this multicenter study, subjects with non-small cell lung cancer and control subjects performed 5 min of tidal breathing through the aeoNose, a handheld electronic nose device. A training cohort was used for developing a prediction model based on breath data, and a blinded cohort was used for validation. Multivariable logistic regression analysis was performed, including breath data and clinical variables, in which the formula and cutoff value for the probability of lung cancer were applied to the validation data. A total of 376 subjects formed the training set, and 199 subjects formed the validation set. The full training model (including exhaled breath data and clinical parameters from the training set) were combined in a multivariable logistic regression analysis, maintaining a cut off of 16% probability of lung cancer, resulting in a sensitivity of 95%, a specificity of 51%, and a negative predictive value of 94%; the area under the receiver-operating characteristic curve was 0.87. Performance of the prediction model on the validation cohort showed corresponding results with a sensitivity of 95%, a specificity of 49%, a negative predictive value of 94%, and an area under the receiver-operating characteristic curve of 0.86. Combining exhaled breath data and clinical variables in a multicenter, multi-device validation study can adequately distinguish patients with lung cancer from subjects without lung cancer in a noninvasive manner. This study paves the way to implement exhaled breath analysis in the daily practice of diagnosing lung cancer. The Netherlands Trial Register; No.: NL7025; URL: https://trialregister.nl/trial/7025 FOR EDITORIAL COMMENT, SEE PAGE 479Take-home PointsStudy Question: Can exhaled breath patterns of patients with NSCLC and without NSCLC adequately be discriminated with an electronic nose in a multicenter, multi-device validation study?Results: Exhaled breath data can adequately distinguish patients with lung cancer from subjects without lung cancer in a noninvasive manner in this multicenter, multi-device study that included 575 subjects. Adding clinical variables relevantly improved the diagnostic performance to diagnose lung cancer.Interpretation: Validation of a prediction model, as performed in this study, is a pivotal step for clinical integration of exhaled breath analysis in the diagnostic path of lung cancer. FOR EDITORIAL COMMENT, SEE PAGE 479 Study Question: Can exhaled breath patterns of patients with NSCLC and without NSCLC adequately be discriminated with an electronic nose in a multicenter, multi-device validation study? Results: Exhaled breath data can adequately distinguish patients with lung cancer from subjects without lung cancer in a noninvasive manner in this multicenter, multi-device study that included 575 subjects. Adding clinical variables relevantly improved the diagnostic performance to diagnose lung cancer. Interpretation: Validation of a prediction model, as performed in this study, is a pivotal step for clinical integration of exhaled breath analysis in the diagnostic path of lung cancer. Lung cancer is the leading cause of cancer mortality worldwide.1Siegel R.L. Miller K.D. Jemal A. Cancer statistics, 2020.CA Cancer J Clin. 2020; 70: 7-30Crossref PubMed Scopus (14740) Google Scholar,2Dagenais G.R. Leong D.P. Rangarajan S. et al.Variations in common diseases, hospital admissions, and deaths in middle-aged adults in 21 countries from five continents (PURE): a prospective cohort study.Lancet. 2020; 395: 785-794Abstract Full Text Full Text PDF PubMed Scopus (363) Google Scholar Its high mortality rate is generally a consequence of advanced-stage disease at the time of initial diagnosis. Despite striking progress in treatment options in advanced-stage lung cancer, such as molecular-targeted therapies and immunotherapy, an essential step to reducing lung cancer mortality is early detection through noninvasive, point-of-care strategies.3Broodman I. Lindemans J. van Sten J. Bischoff R. Luider T. Serum protein markers for the early detection of lung cancer: a focus on autoantibodies.J Proteome Res. 2017; 16: 3-13Crossref PubMed Scopus (38) Google Scholar, 4Jiang R. Dong X. Zhu W. et al.Combining PET/CT with serum tumor markers to improve the evaluation of histological type of suspicious lung cancers.PLoS One. 2017; 12e0184338Crossref Scopus (10) Google Scholar, 5Rolfo C. Russo A. Liquid biopsy for early stage lung cancer moves ever closer.Nat Rev Clin Oncol. 2020; 17: 523-524Crossref PubMed Scopus (33) Google Scholar, 6Seijo 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 (274) Google Scholar Exhaled breath contains a gas mixture of thousands of volatile organic compounds (VOCs) in low concentrations that reflect metabolic processes at the tissue level.7Mansurova M. Ebert B.E. Blank L.M. Ibanez A.J. A breath of information: the volatilome.Curr Genet. 2018; 64: 959-964Crossref PubMed Scopus (31) Google Scholar,8Boots A.W. van Berkel J.J. Dallinga J.W. Smolinska A. Wouters E.F. van Schooten F.J. The versatile use of exhaled volatile organic compounds in human health and disease.J Breath Res. 2012; 6027108Crossref PubMed Scopus (243) Google Scholar Exhaled breath analysis is based on shifts of this VOC composition due to biochemical changes in different (patho) physiological processes. This method has been investigated extensively in clinical research as a noninvasive tool to diagnose a variety of conditions.9Nakhleh M.K. Amal H. Jeries R. et al.Diagnosis and classification of 17 diseases from 1404 subjects via pattern analysis of exhaled molecules.ACS Nano. 2017; 11: 112-125Crossref PubMed Scopus (353) Google Scholar, 10Wilson A.D. Advances in electronic-nose technologies for the detection of volatile biomarker metabolites in the human breath.Metabolites. 2015; 5: 140-163Crossref PubMed Scopus (175) Google Scholar, 11Dragonieri 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 (320) Google Scholar Studies on pattern recognition for classification of VOC mixtures through nonspecific cross-reactive sensors mimicking human and animal olfaction (eg, electronic noses) as well as identifying individual VOCs by using separation methods (eg, gas chromatography mass spectrometry) have shown promising results in pilot studies for the diagnosis of lung cancer.12Gasparri R. Santonico M. Valentini C. et al.Volatile signature for the early diagnosis of lung cancer.J Breath Res. 2016; 10016007Crossref PubMed Scopus (106) Google Scholar, 13Kort S. Tiggeloven M.M. Brusse-Keizer M. et al.Multi-centre prospective study on diagnosing subtypes of lung cancer by exhaled-breath analysis.Lung Cancer. 2018; 125: 223-229Abstract Full Text Full Text PDF PubMed Scopus (38) Google Scholar, 14Phillips M. Altorki N. Austin J.H. et al.Prediction of lung cancer using volatile biomarkers in breath.Cancer Biomark. 2007; 3: 95-109Crossref PubMed Scopus (271) Google Scholar, 15van de Goor R. van Hooren M. Dingemans A.M. Kremer B. Kross K. Training and validating a portable electronic nose for lung cancer screening.J Thorac Oncol. 2018; 13: 676-681Abstract Full Text Full Text PDF PubMed Scopus (95) Google Scholar, 16Mazzone P.J. Wang X.F. Lim S. et al.Progress in the development of volatile exhaled breath signatures of lung cancer.Ann Am Thorac Soc. 2015; 12: 752-757Crossref PubMed Scopus (24) Google Scholar, 17Rocco 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 (28) Google Scholar, 18Tirzite M. Bukovskis M. Strazda G. Jurka N. Taivans I. Detection of lung cancer with electronic nose and logistic regression analysis.J Breath Res. 2018; 13016006Crossref PubMed Scopus (53) Google Scholar In addition, studies based on imaging techniques have been shown to be effective for screening purposes in the diagnosis of lung cancer in high-risk asymptomatic subjects. Significant mortality reduction in high-risk subjects was observed in the National Lung Screening Trial (NLST) and the Dutch-Belgian Lung Cancer Screening Trial (NELSON).19Aberle D.R. Adams A.M. et al.National Lung Screening Trial Research TeamReduced lung-cancer mortality with low-dose computed tomographic screening.N Engl J Med. 2011; 365: 395-409Crossref PubMed Scopus (7646) Google Scholar,20de 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 (1573) Google Scholar However screening of high-risk subjects has not yet been implemented in Europe. Furthermore, determination of accurate screening criteria remains debatable because only subjects at the highest risk for lung cancer are targeted in current screening programs. The aeoNose (the eNose Company) is a handheld electronic nose device featuring an array of three metal-oxide sensors that enables real-time breath analysis. The technology and breath sampling method have been described previously in detail.21Bruins M.G.J. Gerritsen J.W. Van de Sande W. Van Belkum A. Bos A. Enabling a transferable calibration model for metal-oxide type electronic noses.Sensors Actuators B Chemical. 2013; 188: 1187-1195Crossref Scopus (39) Google Scholar,22Kort S. Brusse-Keizer M. Gerritsen J.W. van der Palen J. Data analysis of electronic nose technology in lung cancer: generating prediction models by means of Aethena.J Breath Res. 2017; 11026006Crossref PubMed Scopus (38) Google Scholar Following exposure to VOCs, consecutive conductivity changes at the sensors are recorded, resulting in a digital exhaled breath profile consisting of conductivity values. Exhaled breath profiles of patients with lung cancer can then be distinguished from profiles of non-lung cancer subjects by using artificial intelligence techniques. Once a model has been developed for separating the groups, a new breath profile can be classified using this model. In previous studies, several malignant and nonmalignant conditions have been investigated using the aeoNose.13Kort S. Tiggeloven M.M. Brusse-Keizer M. et al.Multi-centre prospective study on diagnosing subtypes of lung cancer by exhaled-breath analysis.Lung Cancer. 2018; 125: 223-229Abstract Full Text Full Text PDF PubMed Scopus (38) Google Scholar,23Peters Y. Schrauwen R.W.M. Tan A.C. Bogers S.K. de Jong B. Siersema P.D. Detection of Barrett's oesophagus through exhaled breath using an electronic nose device.Gut. 2020; 69: 1169-1172Crossref PubMed Scopus (42) Google Scholar, 24van de Goor R. Hardy J.C.A. van Hooren M.R.A. Kremer B. Kross K.W. Detecting recurrent head and neck cancer using electronic nose technology: a feasibility study.Head Neck. 2019; 41: 2983-2990Crossref PubMed Scopus (26) Google Scholar, 25Uslu H.I. Dolle A.R. Dullemen H.M. Aktas H. Kolkman J.J. Venneman N.G. Pancreatic ductal adenocarcinoma and chronic pancreatitis may be diagnosed by exhaled-breath profiles: a multicenter pilot study.Clin Exp Gastroenterol. 2019; 12: 385-390Crossref PubMed Scopus (4) Google Scholar We have previously reported the results of a proof-of-concept multicenter study performed with the aeoNose in which a prediction model, based on exhaled breath profiles, was developed using supervised machine learning techniques to discriminate subjects with and without non-small cell lung cancer (NSCLC) in a hospital setting.13Kort S. Tiggeloven M.M. Brusse-Keizer M. et al.Multi-centre prospective study on diagnosing subtypes of lung cancer by exhaled-breath analysis.Lung Cancer. 2018; 125: 223-229Abstract Full Text Full Text PDF PubMed Scopus (38) Google Scholar An artificial neural network (ANN) trained with 290 subjects was able to classify breath samples with a sensitivity of 94%, a specificity of 33%, and an area under the receiver-operating characteristic curve (AUC-ROC) of 0.76. Resampling techniques, including leave-10%-out cross-validation and bootstrapping, were incorporated to reduce the risk of overfitting of the diagnostic model. Adding readily available clinical information (ie, sex, age, number of pack-years, smoking status, COPD status) to the exhaled breath data resulted in a relevant improvement in diagnosing patients with lung cancer.26Kort S. Brusse-Keizer M. Gerritsen J.W. et al.Improving lung cancer diagnosis by combining exhaled-breath data and clinical parameters.ERJ Open Res. 2020; 6: 00221-2019Crossref PubMed Scopus (21) Google Scholar To date, no single breath test has yet been approved for clinical practice to diagnose lung cancer. For this, validation studies are required, preferably involving multiple devices in multiple centers, where part of the data are used for developing a diagnostic model, and the remainder remain blinded to validate the model. Several studies on external validation of breath biomarkers in lung cancer have been performed; however, these studies were aimed at identification of specific VOCs rather than exhaled breath patterns.14Phillips M. Altorki N. Austin J.H. et al.Prediction of lung cancer using volatile biomarkers in breath.Cancer Biomark. 2007; 3: 95-109Crossref PubMed Scopus (271) Google Scholar,27Long Y. Wang C. Wang T. et al.High performance exhaled breath biomarkers for diagnosis of lung cancer and potential biomarkers for classification of lung cancer.J Breath Res. 2021; 15016017Crossref PubMed Scopus (9) Google Scholar,28Phillips M. Bauer T.L. Cataneo R.N. et al.Blinded validation of breath biomarkers of lung cancer, a potential ancillary to chest CT screening.PLoS One. 2015; 10e0142484Crossref Scopus (39) Google Scholar Regarding pattern recognition techniques, Fens et al29Fens N. Roldaan A.C. van der Schee M.P. et al.External validation of exhaled breath profiling using an electronic nose in the discrimination of asthma with fixed airways obstruction and chronic obstructive pulmonary disease.Clin Exp Allergy. 2011; 41: 1371-1378Crossref PubMed Scopus (122) Google Scholar and Bos et al30Bos L.D. Schultz M.J. Sterk P.J. Exhaled breath profiling for diagnosing acute respiratory distress syndrome.BMC Pulm Med. 2014; 14: 72Crossref PubMed Scopus (34) Google Scholar assessed validation of exhaled breath molecular patterns in pulmonary diseases other than lung cancer, based on previous created training sets, showing moderate to high accuracy. The objective of this prospective multicenter study using multiple devices was to train and subsequently validate a prediction model to distinguish NSCLC patients from subjects initially suspected of lung cancer but considered negative and healthy control subjects, based on their exhaled breath patterns. Participants suspected of having lung cancer were recruited from seven outpatient pulmonary departments between May 2018 and April 2020. The participating hospitals comprised Medisch Spectrum Twente Enschede, Radboud UMC Nijmegen, Medisch Centrum Leeuwarden, Martini Ziekenhuis Groningen, Catharina Ziekenhuis Eindhoven, Sint Antonius Ziekenhuis Utrecht (all in The Netherlands), and Universitätsspital Basel (Switzerland). Each center used one aeoNose device, except for the Basel site, which used two devices. A single aeoNose device needs, as a rule of thumb, a minimum number of 30 observations in the smallest group (in this case, positive measurements) to calibrate the device and hence form reliable conclusions considering the training data; thus, data from devices with an insufficient number of measurements were not used for further analyses. Subjects suspected of having lung cancer, based on symptom reports or abnormal imaging, were divided into a group with confirmed NSCLC based on pathology and a group with a rejected diagnosis of lung cancer (control subjects) based on imaging and/or pathology. Types of lung cancer other than NSCLC were excluded. Additional healthy control subjects with a minimum age of 55 years were recruited through an alert at the hospitals’ websites. In case of pathologically confirmed lung cancer, staging was established according to the eighth edition of the American Joint Committee on Cancer TNM staging system.31Goldstraw P. Chansky K. Crowley J. et al.The IASLC Lung Cancer Staging Project: proposals for revision of the TNM stage groupings in the forthcoming (Eighth) Edition of the TNM Classification for Lung Cancer.J Thorac Oncol. 2016; 11: 39-51Abstract Full Text Full Text PDF PubMed Scopus (2801) Google Scholar Patients suspected of lung cancer in whom pathology (ie, the gold standard) was not performed due to insufficient clinical performance were excluded from the analyses. Demographic data and data on some highly prevalent comorbidities (ie, COPD, diabetes mellitus, hypertension) were collected for all subjects. All participants were asked to complete a short questionnaire on recent smoking, eating, and alcohol intake, and were instructed to perform tidal breathing through the non-rebreathing aeoNose device for 5 min with their nose clipped. The study protocol was approved by the institutional review board of Medisch Spectrum Twente and the board of directors of all participating institutions (e-Appendix 1). All eligible patients provided written informed consent. The second-generation, CE-certified aeoNose device was used in this study. Because the training study was performed with the first-generation, CE-uncertified device, these previously collected data were deemed not compatible and therefore not used.13Kort S. Tiggeloven M.M. Brusse-Keizer M. et al.Multi-centre prospective study on diagnosing subtypes of lung cancer by exhaled-breath analysis.Lung Cancer. 2018; 125: 223-229Abstract Full Text Full Text PDF PubMed Scopus (38) Google Scholar Instead, we decided to create a split-sample study design in which we enabled development and subsequent validation of new prediction models, which conform to the European Respiratory Society criteria for exhaled biomarkers.32Horvath I. Barnes P.J. Loukides S. et al.A European Respiratory Society technical standard: exhaled biomarkers in lung disease.Eur Respir J. 2017; 49Crossref Scopus (396) Google Scholar Collected breath data were split into a training cohort for supervised learning and internal cross-validation, and a validation cohort, which was kept blinded, for model validation. A random subset of subjects was assigned to the validation cohort, based on the sample size calculation of this validation cohort. Also, an equal prevalence of patients with lung cancer in both sets was taken into consideration. Clinical characteristics are reported as means with SDs in case of a normal distribution, or as medians with interquartile ranges. Nominal variables are reported as numbers with corresponding percentages. To assess differences between the groups, t tests, U tests, or χ2 tests were applied, as appropriate. Analysis of exhaled breath data was executed by Aethena, a proprietary software package, incorporating data pre-processing, data compression, machine learning algorithms for classification (eg, ANN, Support Vector Machine, Random Forest [RF], XGBoost, logistic regression), internal validation techniques (leave-10%-out cross-validation and bootstrapping), and model selection. Analyses yielded values between –1 and 1 per subject, indicating the degree to which the subject was classified as having lung cancer (maximum value, 1) or not having lung cancer (minimum value, –1). Details on the software package Aethena have been published previously.22Kort S. Brusse-Keizer M. Gerritsen J.W. van der Palen J. Data analysis of electronic nose technology in lung cancer: generating prediction models by means of Aethena.J Breath Res. 2017; 11026006Crossref PubMed Scopus (38) Google Scholar We selected and trained five different models (each using a different classifier: ANN, Logistic Regression, RF, RF Extreme, and XGBoost), with each showing proper discriminative performance. Because the various classifying techniques could interpret the data differently, we envisioned that averaging results over these five models would increase classification robustness. A cutoff value for the probability of lung cancer was determined for the training set to obtain a high sensitivity and negative predictive value (NPV), together with an acceptable number of false-positive cases, as deemed relevant for clinical practice. ROC curves were composed and AUCs were calculated with 95% CIs. Subsequently, clinical variables (ie, sex, age, number of pack-years, COPD, diabetes, hypertension, BMI, the absolute value obtained from the aeoNose [between –1 and 1]) were entered in a multivariable logistic regression analysis. Nonsignificant variables were eliminated according to the backward method until the fit of the model decreased significantly, based on the –2 log likelihood. Age and sex were included regardless of their significance. A cutoff value for the probability of lung cancer based on this multivariable model was again chosen to obtain a high sensitivity and NPV together with an acceptable number of false-positive cases. The diagnostic performance of this final logistic regression model, based on the training data, was validated on the blinded data set, where the β-coefficients were fixed. The same cutoff value, chosen for the training data to determine the presence of lung cancer, was applied to the logistic regression analysis in the validation set. Results are expressed as sensitivity, specificity, predictive values, and AUC-ROC. A calibration plot was constructed to show how well the predicted probability of lung cancer matches the observed probability of lung cancer. Stratification for variables to evaluate possible influences on exhaled breath outcomes was performed in explorative analyses for sex, age, presence of COPD, lung cancer stage, and type of histology. Early stage lung cancer was classified as either stage I or II, and late-stage lung cancer was classified as stage III or IV. To compare the final prediction model including breath data and clinical variables vs a nodule calculator, we calculated Spearman rho and the diagnostic performance of both models expressed in terms of sensitivity, specificity, and predictive values. Taking into consideration a sensitivity of 95%, as acceptable in clinical practice, estimated with a precision of 5%, with a prevalence of lung cancer in our population of 40%, and an expected specificity of 50%, we would need 183 subjects in our validation cohort. This would lead to a negative predictive value of 93% (95% CI, 0.85-0.98). SPSS version 24.0 (IBM SPSS Statistics, IBM Corporation) was used for analysis. All statistical tests were two-sided with a significance level at .05. A total of 575 subjects were enrolled in the analyses (Fig 1). Approximately two-thirds formed the training set (376 subjects [160 patients with lung cancer, 51 suspected but negative, and 165 healthy control subjects]), and the remaining one-third comprised the validation set (199 subjects [79 patients with lung cancer, 32 suspected but negative, and 88 healthy control subjects]). Subject characteristics are described in Table 1. Data were obtained using five aeoNose devices.Table 1Clinical Characteristics of All Enrolled SubjectsCharacteristicTraining Set (n = 376)Validation Set (n = 199)Lung Cancer (n = 160)Control Subjects (n = 216)P ValueLung Cancer (n = 79)Control Subjects (n = 120)P ValueAge, mean ± SD, y68.4 ± 8.664.6 ± 8.2< .00169.0 ± 7.963.4 ± 9.4< .001Male97 (60.6)131 (60.6).99649 (62.0)59 (49.2).075Smoking status< .001.001 Current smoker48 (30.0)41 (19.0)30 (38.0)27 (22.5) Ex-smoker103 (64.4)130 (60.2)45 (57.0)65 (54.2) Never smoker9 (5.6)45 (20.8)4 (5.1)28 (23.3)Pack-yearsaFive missing subjects.< .001< .001 08 (5.1)45 (20.8)2 (2.6)28 (23.3) 1-2037 (23.4)56 (25.9)14 (18.4)38 (31.7) 21-4052 (32.9)55 (25.5)28 (36.8)19 (15.8) > 4061 (38.6)60 (27.8)32 (42.1)35 (29.2)COPD71 (44.4)94 (43.5).86937 (46.8)52 (43.3).627HypertensionbOne missing subject.66 (41.3)74 (34.3).16627 (34.6)38 (31.9).695DiabetesbOne missing subject.15 (9.4)22 (10.2).79411 (13.9)10 (8.4).217BMI, mean ± SD, kg/m226.4 ± 4.425.8 ± 4.7.21026.2 ± 5.025.7 ± 4.4.402Type of NSCLC Adenocarcinoma101 (63.1)39 (50.0)…… Squamous cell carcinoma43 (26.9)32 (41.0)…… Large cell carcinoma6 (3.8)4 (5.1)…… NOS10 (6.3)3 (3.8)……StagecAccording to the eighth edition of the American Joint Committee on Cancer TNM staging system. I54 (33.8)21 (26.6)…… II23 (14.4)15 (19.0)…… III38 (23.8)19 (24.1)…… IV45 (28.2)24 (30.4)……Hospital MST66 (41.3)69 (31.9)30 (38.0)30 (25.0)… Radboud UMC31 (19.4)29 (13.4)20 (25.3)12 (10.0)… MCL Leeuwarden29 (18.1)34 (15.7)17 (21.5)33 (27.5)… US Basel34 (21.3)84 (38.9)12 (15.2)45 (37.5)…Data are presented as No. (%) unless otherwise indicated. MCL = Medisch Centrum Leeuwarden; MST = Medisch Spectrum Twente; NOS = not otherwise specified; NSCLC = non-small cell lung cancer.a Five missing subjects.b One missing subject.c According to the eighth edition of the American Joint Committee on Cancer TNM staging system. Open table in a new tab Data are presented as No. (%) unless otherwise indicated. MCL = Medisch Centrum Leeuwarden; MST = Medisch Spectrum Twente; NOS = not otherwise specified; NSCLC = non-small cell lung cancer. The training model, exclusively based on breath data from the aeoNose, showed, at a cutoff value of –0.36, an AUC-ROC of 0.83 (95% CI, 0.79-0.87), a sensitivity of 91%, a specificity of 54%, and an NPV of 89%. The diagnostic performance of the aeoNose, maintaining the same cutoff value in the validation set, reached an AUC-ROC of 0.79 (95% CI, 0.72-0.85), with a sensitivity of 88%, a specificity of 52%, and an NPV of 87%, which conforms to the training model. Due to the multicollinearity of smoking status and number of pack-years, we chose to include number of pack-years in these analyses because this parameter contained the most detailed information. The multivariable analysis based on solely clinical data from the training set, including sex, age, and number of pack-years, yielded an AUC-ROC of 0.67 (95% CI, 0.61-0.72); the validation set yielded an AUC-ROC of 0.75 (95% CI, 0.68-0.82). Exhaled breath data and clinical parameters from the training set were combined in a multivariable logistic regression analysis, maintaining a cutoff of 16% probability of lung cancer, resulting in a sensitivity of 95%, a specificity of 51%, and an NPV of 94%, which was based on clinical relevance (Tables 2-3). This corresponded to an AUC-ROC of 0.87 (95% CI, 0.83-0.90). When applying the identical multivariable logistic regression model on the validation set, maintaining the selected cutoff probability of 16%, we observed a sensitivity of 95%, a specificity of 49%, a positive predictive value of 54%, and an NPV of 94%, with a corresponding AUC-ROC of 0.86 (95% CI, 0.81-0.91) (Fig 2, Table 3). In case of this cutoff probability of 16%