# Fileset

[Manuscript R1(clean).docx](https://mdr.nims.go.jp/filesets/eb889eb9-cb01-4564-a8d9-91212776c9cb/download)

## Creator

Yusuke Saeki, Naoki Maki, Takahiro Nemoto, Katsushige Inada, [Kosuke Minami](https://orcid.org/0000-0003-4145-1118), [Ryo Tamura](https://orcid.org/0000-0002-0349-358X), [Gaku Imamura](https://orcid.org/0000-0002-3130-7190), Yukiko Cho-Isoda, Shinsuke Kitazawa, Hiroshi Kojima, [Genki Yoshikawa](https://orcid.org/0000-0002-9136-8964), Yukio Sato

## Rights

[Creative Commons BY-NC-ND Attribution-NonCommercial-NoDerivs 4.0 International](https://creativecommons.org/licenses/by-nc-nd/4.0/)

## Other metadata

[Lung cancer detection in perioperative patients' exhaled breath with nanomechanical sensor array](https://mdr.nims.go.jp/datasets/539e66a4-4185-412f-9d20-3344295b8113)

## Fulltext

Lung Cancer Detection in Perioperative Patients' Exhaled Breath with Nanomechanical Sensor ArrayAuthorsYusuke Saeki, MD, PhD,a Naoki Maki, PhD,a Takahiro Nemoto, MEd,b,c Katsushige Inada, PhD,d Kosuke Minami, PhD,b,c,e Ryo Tamura, PhD,f,g,h,i Gaku Imamura, PhD,c,f,j Yukiko Cho-Isoda, PhD,d Shinsuke Kitazawa, MD, PhD,a Hiroshi Kojima, MD, PhD,d,k Genki Yoshikawa, PhD,b,c,l Yukio Sato, MD, PhDa,*Institutions and affiliationsa Department of Thoracic Surgery, University of Tsukuba, Ibaraki, Japanb Center for Functional Sensor & Actuator (CFSN), Research Center for Functional Materials, National Institute for Materials Science (NIMS), Ibaraki, Japanc Research Center for Macromolecules and Biomaterials, National Institute for Materials Science (NIMS), Ibaraki, Japand Department of Medical Oncology, Ibaraki Prefectural Central Hospital, Ibaraki, Japane International Center for Young Scientists (ICYS), National Institute for Materials Science (NIMS), Ibaraki, Japanf World Premier International (WPI) Research Center for Materials Nanoarchitectonics (MANA), National Institute for Materials Science (NIMS), Ibaraki, Japang Graduate School of Frontier Sciences, The University of Tokyo, Chiba, Japanh Research and Services Division of Materials Data and Integrated System (MaDIS), National Institute for Materials Science (NIMS), Ibaraki, Japani Center for Basic Research on Materials, National Institute for Materials Science (NIMS), Ibaraki, Japanj Graduate School of Information Science and Technology, Osaka University, Osaka, Japank Ibaraki Clinical Education and Training Center, University of Tsukuba Hospital, Ibaraki, Japanl Materials Science and Engineering, Graduate School of Pure and Applied Science, University of Tsukuba, Ibaraki, Japan*Corresponding authorAddress for correspondence: Yukio Sato, MD, PhD, Department of Thoracic Surgery, University of Tsukuba, 1-1-1 Tennodai, Ibaraki 305-8575, Japan. E-mail: ysato@md.tsukuba.ac.jpDisclosures:All authors have no conflicts of interest to declare.HIGHLIGHTS· Early lung cancer detection is vital due to its high mortality in advanced stage.· We created a novel membrane-type surface stress sensor for breath analysis.· We conducted a pilot study on lung cancer diagnosability from perioperative breath.· We achieved the optimized prediction model with over 80% accuracy.· Breath analysis could be a promising non-invasive screening method for lung cancer.ABSTRACTIntroduction: Breath analysis using a chemical sensor array combined with machine learning algorithms may be applicable for detecting and screening lung cancer. In this study, we examined whether perioperative breath analysis can predict the presence of lung cancer using a Membrane-type Surface stress Sensor (MSS) array and machine learning.Methods: Patients who underwent lung cancer surgery at an academic medical center, Japan, between November, 2018 and November, 2019 were included. Exhaled breaths were collected just before surgery and about one month after surgery, and analyzed using an MSS array. The array had 12 channels with various receptor materials and provided 12 waveforms from a single exhaled breath sample. Boxplots of the perioperative changes in the expiratory waveforms of each channel were generated and Mann-Whitney U test were performed. An optimal lung cancer prediction model was created and validated using machine learning.Results: Sixty-six patients were enrolled of whom 57 were included in the analysis. Through the comprehensive analysis of the entire dataset, a prototype model for predicting lung cancer was created from the combination of array five channels. The optimal accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were 0.809, 0.830, 0.807, 0.806, and 0.812, respectively.Conclusion: Breath analysis with MSS and machine learning with careful control of both samples and measurement conditions provided a lung cancer prediction model, demonstrating its capacity for non-invasive screening of lung cancer.Keywords: Lung cancer, breath analysis, electronic nose (e-nose), Membrane-type Surface stress Sensor (MSS), machine learningIntroductionThe number of lung cancer deaths is the highest among all cancer types worldwide [1]. The 5-year survival rate of lung cancer decreases with the advance of the stage, from stage IA1 (92%) to stage IVB (0%), according to an IASLC report [2]. Therefore, it is essential to detect and treat lung cancer at an early stage. However, lung cancer is difficult to detect in its early stages because it often remains asymptomatic until more advanced stages. Although screening through low-dose computed tomography (CT) has been reported to reduce the lung cancer mortality rate, CT screening is expensive, labor intensive, and the rate of false positives is relatively high, 56–96% [3-5]. In addition, the radiological exposure of CT carries a risk of radiation injury, which limits the frequency of applications for lung cancer screening [5, 6]. Therefore, a non-invasive and cost-efficient method for lung cancer screening is highly sought after.Less invasive attempts with exhaled breath have been reported, as human exhaled breath contains various types of volatile organic compounds (VOCs) [7-10] and some of the VOCs in exhaled breath are produced through physiological or pathological processes. Although some studies using various gas analyzers have reported that some VOCs in exhaled breath may be associated with cancer [11-13], a biomarker for lung cancer has not yet been established with biological/medical evidence. In recent years, chemical sensor arrays are reported to distinguish lung cancer patients from healthy controls through the analyses of breath samples using machine learning algorithms [14-19]. Such chemical sensor array-based platforms can be smaller in size, easier to use, and less expensive compared to gas analyzers, and, thus, are promising candidates for practical screening applications. However, these chemical sensor array-based platforms can be susceptible to various factors including measurement conditions such as the temperature and humidity, as well as patient background factors such as age, sex, metabolic functions, and smoking habits. Therefore, it is challenging to clearly identify the information pertaining to the presence or absence of cancer through the measurements of exhaled breath samples because of interference from these multiform factors.In the present study, to reduce the influence of interfering factors as possible, we combined the following two approaches: 1) well-controlled breath sampling in constant temperature and humidity and measurements according to the previously established protocol using a chemical sensor array; and 2) analyses of breath samples from the same patient before and after surgery to minimize the effects arising from background factors other than the presence of lung cancer. Regarding the former approach, we recently demonstrated its high reproducibility by the analyses of breath samples collected by means of total expiratory breath sampling and measured with room air as a purge gas for more than a year using Membrane-type Surface stress Sensor (MSS) [20]. MSS is a kind of nanomechanical sensor with a piezoresistive electric readout [21, 22]. The mechanical stress induced by sorption of gas molecules on a receptor layer, which is a VOC-sensitive material coated on the MSS sensing membrane, is transduced into an electrical signal providing unique time series data depending on the physical/chemical properties of the interacting gas molecules and receptor materials. Thus, an MSS array with different receptor materials can be used as a sensing platform for an artificial olfactory system. With the advanced performance of MSS, various applications have been demonstrated so far [23] ranging from discrimination of VOCs in body odor of various concentrations [24] to quantification of alcohol content in different types of liquid vapors [25, 26]. Moreover, a pilot study using MSS demonstrated discrimination between head and neck cancer patients and healthy volunteers through breath analysis [27]. Nanomechanical sensors are capable of utilizing almost any material as a receptor layer without the need for electrical conductivity. Among nanomechanical sensors, MSS does not require complicated read-out system [23]. These features render MSS a potential candidate for the development of practical sensors compared to other sensors such as metal oxide semiconductors, chemiresistors, field effect transistors, and quartz crystal microbalances. The comprehensive analyses of these data with machine learning could provide a lung cancer prediction model with significant potential for practical lung cancer screening.Material and MethodsPatientsPatients who underwent lobectomy plus mediastinal lymph node dissection as the curative treatment for primary lung cancer at the University of Tsukuba Hospital from November, 2018 to November, 2019 were enrolled in this study. Patients who underwent sublobar resection with a diagnosis of adenocarcinoma in situ or minimally invasive adenocarcinoma were also included, as sublobar resections are considered as the curative treatment for those early-stage adenocarcinoma. Patients with a history of preoperative treatment, a history of lung cancer treatment within the past five years, discontinued surgery, incomplete resection, and patients who were unable to obtain breath samples in the first month after surgery due to postoperative complications were excluded.This study was approved by the Institutional Review Board of the University of Tsukuba Hospital (H28-193), and informed consent was obtained from all participants.Sample collectionPatients underwent expiratory sampling on the day of surgery and approximately one month after surgery at an outpatient clinic. In this study, a separate room in the ward, that was not used by other patients, was prepared for breath collection. Both preoperative and postoperative breath samples were taken in the same room. The room temperature and humidity were adjusted to 25 °C and 50%, respectively, to minimize the influence of temperature and humidity fluctuations. Preoperative patients were allowed to eat until the night before surgery but were not allowed to eat or drink after midnight. Breath samples were taken early in the morning on the day of surgery. For postoperative measurements, breakfast was not allowed for morning outpatients, and lunch was not allowed for afternoon outpatients. Patients were allowed to drink water.Exhaled breath samples were collected according to the protocol established by Inada et al [20]. New 1,000 mL polyvinyl alcohol-based sampling bags (Smart Bag PA; GL Sciences, Tokyo, Japan) with valve sleeves were used for exhaled breath collection. To minimize the influence of the oral cavity, it was rinsed with 20 ml of saline prior to exhalation. After taking three deep breaths and holding the breath for 6 seconds after inhalation, the patient was asked to blow into the collection bag. To minimize the amount of air from the oral cavity and to collect the exhaled air from the lung, the participants were instructed not to puff up their cheeks during the breath-hold (Fig. 1a). A sufficient amount of room air from the same room was collected for purge gas in the same type of sampling bag using a purpose-made negative pressure generator. Each sample bag was placed in an incubator (SLC-25A; Mitsubishi Electric Engineering, Tokyo, Japan) at 25°C for 30 minutes, after which the sensing measurement was performed using a sensor array (Fig. 1b).Nanomechanical sensor arrayThe construction of the MSS chips and its working principle has been previously reported [21, 22]. Previous studies have shown that the device has advantages over conventional cantilever-based nanomechanical sensors, offering high sensitivity [22], compactness, and stable operation [28, 29]. The MSS chips used in this study were purchased from NanoWorld AG, Switzerland. The nanomechanical sensor array used in this study contained 12 cross-reactive, chemically diverse receptor materials of two types with various time constants in dynamic responses to VOCs such as diffusion and viscoelastic relaxation: (1) silica/titania-based hybrid nanoparticles (STNPs) functionalized with octadecyl (C18-STNPs) and phenyl groups (Ph-STNPs) [25, 30]； and (2) commercially available polymers, poly(ethylene oxide) (PEO), polystyrene (PS), poly(4-methylstyrene) (P4MS), poly(vinylidene fluoride) (PVF), polycaprolactone (PCL), poly(methyl methacrylate) (PMMA), and poly(2,6-diphenyl-1,4-phenylene oxide) [20, 25, 31-33]. Four different layer thicknesses were used as separate channels for poly(2,6-diphenyl-1,4-phenylene oxide). The receptor materials were coated directly onto the MSS membrane using an inkjet spotter (LaboJet-500SP, MICROJET Corporation) with a nozzle (IJHBS-300, MICROJET Corporation). Each receptor material was dissolved in DMF at a concentration of 1 mg/mL, and the resulting solutions were deposited onto each channel of the MSS. A stage of the inkjet spotter was heated at 80°C to promote evaporation of the DMF.The MSS standard measurement module produced by the framework of industry-academia-government called “MSS Alliance” [34]. The MSS chips with receptor materials were placed in a Teflon chamber, which was closely sealed with O-rings. The chamber was connected to a gas flow system consisting of a switching valve connected with sampling and purging gas lines. The sample and purge gas flows were controlled by an aspiration pump with a flow rate adjusted to 30 mL/min. Data were measured at the bridge voltage of –1.0 V and the relative resistance changes of piezoresistors embedded on the four MSS sensing beams [21, 22] were acquired at a sampling frequency of 100 Hz. Temperature and relative humidity (RH) of the sample and purge gases were monitored using a temperature/humidity sensor installed in the MSS measurement module. In this study, the same module was used for all the measurements without changing any components. Measurements of breath samplesAll measurements made on the MSS standard measurement module were performed in the incubator maintained at 25 ℃ in order to avoid the fluctuation of temperature which affects the sensing signals [20]. Sampling bags filled with exhaled breath and room air were stored at 25℃ for 30 minutes after sampling. The exhaled breath sample and room air sample bags were connected to the sample and purge injection ports of the MSS module, respectively (Fig. 1b). When measuring the breath samples, an injection sequence that cycled 10 times between 5-second injections of breath sample followed by 5-second purges with room air was used (Fig. 1c).Feature extraction and data analysis To extract stable signals without initial fluctuations observed at the beginning of sample flow, the signal values at 95 seconds (the last point measured for the breath sampling in the 10th cycle, where the sensors show stable repetitive responses) subtracted by the values at 90 seconds (the first point in the 10th cycle) were extracted from all 12 channels: Si, where subscript i denotes the channel number (i.e., i=1,2,⋯,12). Because each channel is coated with a different receptor layer and responds uniquely to VOCs, the differences in the response of each channel should reflect the information regarding VOC composition in each exhaled breath sample. Accordingly, the features of a breath sample were defined by the differences between two signal values of the signals of ith and jth channels as . Twelve different types of receptor materials were prepared and, thus, one exhaled breath sample is represented by the n-dimensional feature obtained from all the combinations of channels used for analysis, where k is the number of channels used for analysis and n is the number of combinations of two channels:  (e.g.,  in the case of the 5 channels used for this analysis). The preoperative and postoperative features were boxplotted, and Mann-Whitney U test was performed to evaluate the statistical significance. A two-tailed test finding of p < 0.01 was considered statistically significant. Statistical tests were performed using SPSS statistical software (IBM, Armonk, NY, US).Machine learning classification methods can be used to determine whether a cancer patient is preoperative or postoperative from signal features of MSS, that is, a binary classification. To measure the potential of the constructed classification model for the case of  patients, we evaluated the prediction accuracy using the following procedures. 1) Data from about 10% of randomly-selected patients (i.e., five patients in this study) was used as testing data. 2) The remaining patient data was used to create a machine learning classification model trained to predict preoperative and postoperative where the number of training data is . 3) Using the trained model, the preoperative or postoperative for the five patients in the preselected testing data set were predicted, and the accuracy, precision, recall, and F1-score, which are defined by: accuracy = (TP + TN) / (TP + TN + FP + FN),  (1) precision = TP / (TP + FP),   recall = TP / (TP + FN),   F1-score = 2 (precision × recall) / (precision + recall), where TP, TN, FP, and FN denote true positive, true negative, false positive, and false negative, respectively, are evaluated. 4) To allow for statistical analyses, procedures 1)–3) were repeated 100 times, and the average accuracy was calculated. Because there are 12 different channels, procedures 1)–4) were evaluated for all combinations, i.e.,  4083.In this study, a classification model was constructed using random forest classifier implemented on scikit-learn (https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html). Hereby, n_estimators and max_depth were considered as hyper parameters, and these parameters were optimized by 5-fold cross validation. The random forest classifier used in this study is a method that can construct a machine learning model with high prediction accuracy even with a small number of training data. There are other methods for binary classification, such as support vector machine (SVM) and logistic regression. While the selection of the machine learning method is not necessarily optimal, we have confirmed that the random forest classifier shows better prediction accuracy than SVM and logistic regression in the present sample set. If the number of samples increases, it would be possible to provide more accurate machine-learning models by using neural networks or deep learning methods.ResultsA total of 66 patients were enrolled in this study. Among them, one patient with concurrent multiple lung cancers, three patients with incomplete resection, three patients with postoperative complication (ARDS [N =1], ischemia at bronchial anastomosis [N =1], postoperative pulmonary fistula [N =1]), and two patients who could not obtain exhaled breath in the first month after surgery for reasons transferring to a different hospital were excluded. Samples from the remaining 57 patients were used for the analysis.The backgrounds of the 57 patients are shown in Table 1. Twenty patients were nonsmokers, 35 were former smokers, and 2 were current smokers. The histological types were adenocarcinoma (45 cases), squamous cell carcinoma (11 cases), and adenosquamous cell carcinoma (1 case). In adenocarcinoma, the numbers of invasive adenocarcinoma, minimally invasive adenocarcinoma in situ, and adenocarcinoma in situ were 47, 8, and 2, respectively. Lymph node metastasis was observed in 9 cases. The numbers of Stage 0, I, II, and III were 2, 38, 9, and 8, respectively.The results of the univariate analysis of the features obtained from the difference between all channels are summarized in Figure S1 and Table S1. Some channel differences showed statistically significant different perioperative features.The accuracy of the machine learning-based lung cancer prediction model for each combination of all channels is shown in Figure 2. The highest accuracy (0.809 ± 0.132) was obtained using the five channels: Ch. 1, 4, 5, 6, and 7. The optimized machine learning-based lung cancer prediction model is shown in Table 2 (see also Fig. S2 and Tables S1 and S2 for results of all prediction models). The optimal sensitivity, specificity, positive predictive value, and negative predictive value were 0.830, 0.807, 0.806, and 0.812, respectively.DiscussionTo date, no lung cancer VOC markers have been identified, although a number of studies using advanced gas analyzers such as gas chromatography-mass spectrometry (GC-MS), selected ion flow tube (SIFT)-MS, and proton transfer reaction (PTR)-MS have been published [11]. Two major reasons have been put forward to account for this difficulty in detecting lung cancer VOC markers. Firstly, exhaled breath components are affected not only by individual factors, but also by foreign gases in the inhaled air (i.e., room air) [35]. Secondly, analysis results are affected not only by the presence of lung cancer, but also by other individual factors such as age, sex, smoking habits, and metabolic functions of the liver and kidney. Therefore, in this study, a strict, reproducible, protocol was adopted for breath sample collection and measurement. Here, samples were collected and measured in a separate room with a consistent temperature and humidity, and measurements were made using the same person’s breath before and after lung cancer surgery to overcome limitations described above. As previously reported [20], a “total expiratory breath sampling” method was employed for the breath collection, and the room air, where the breath was collected, was used as the purge gas during the measurements. In this way, most of the foreign gases present in the room air would be included in both the sampling gas (expiratory breath) and the purge gas (room air), and thus have little effect on the sensor signals in the repeated sampling and purging measurements. Therefore, in the current study, the sensor signals mainly reflect the effect of endogenous components contained only in exhaled breath. Additionally, chemical sensor arrays, including the nanomechanical sensor MSS used in this study, are in principle or essentially, affected by temperature and humidity. Thus, we sought to minimize these effects by preparing a special temperature- and humidity-controlled room in which breath samples were collected. Moreover, because the two perioperative (before and after) breath samples were taken from the same person, the effects of demographic factors that might influence the data, such as age, gender, smoking history, liver, kidney functions, and so on, was minimized. Recently, Wang et al. studied measurements of breath before and after surgery for lung cancer using GC-MS, and reported the prediction accuracy to be 86.9% [35]. However, in the current study, we used a simpler sensor system, and succeeded in constructing a prediction model with an accuracy of 80.9%, which is comparable to those reported using advanced gas analyzers; furthermore, we also achieved high sensitivity (0.830), high specificity (0.807), high positive predictive value (0.806), and high negative predictive value (0.812). These results were comparable to those reported in past studies on electronic noses. In this study, we employed a random forest classifier to construct a machine learning model with high predictive accuracy from a small amount of training data [39]. However, as the number of training data increases, there is potential to provide even more accurate machine learning models by utilizing neural networks or deep learning methods. Therefore, we consider that the relatively simple system used in this study could be practically applied to lung cancer screening in the future.Currently, low-dose CT is used for lung cancer screening, and has been reported to reduce the mortality rate by more than 20% [3, 4]. While low-dose CT is considered a useful screening test for lung cancer, further improvement in specificity has been demanded, although recent deep learning-based analysis demonstrated high accuracy [36-38]. The disadvantages of low-dose CT screening include radiation exposure and the high cost. In contrast, breath analysis using MSS is a simple, low-cost, radiation-free, and non-invasive method suitable for screening tests. By combining breath analysis with CT-screened cases, the false-positive rate is expected to be further reduced. Additionally, the evidence regarding lung cancer screening with low-dose CT is currently limited to heavy smokers over 55 years old because of the above-mentioned disadvantages of this procedure [40]. Therefore, screening by breath analysis is expected to be utilized for those who are not applicable for CT screening.In constructing the machine learning model, the highest prediction accuracy was achieved when an appropriate number of features (in this study, five channels with different receptor materials having cross-selectivity) was used. This may be due to the presence of multiple types of specific exhaled breath components derived from lung cancer, as suggested in the report by Wang et al [35]. If the number of features is too small, the information regarding the behavior of lung cancer marker molecules may be limited, as the univariate analysis showed no significant differences between the before and after surgery measurements. Conversely, too many features would increase the influence of components related to the factors other than lung cancer. In this study, the MSS channels coated with PEO, PMMA, PCL, PS, and PVF gave the highest prediction accuracy, suggesting that lung cancer markers may be exhaled components that are partially responsive to these receptor materials. Since each of these receptor materials includes a variety of different chemical properties from hydrophilic (PEO and PMMA) to hydrophobic (PCL, PS, and PVF), polar (PMMA, PCL, and PVF) to non-polar (PEO and PS), and aliphatic (PEO, PMMA, PCL, and PVF) to aromatic (PS), these receptor materials respond to a large variety of volatile compounds, such as alkanes, alcohols, aldehydes, ketones, esters, carboxylic acids, and aromatic compounds [23, 27, 32-33, 41-44]. Therefore, it is likely that the volatile components that correlate with the presence of lung cancer will also include a variety of different chemical properties [45]. On the other hand, since a certain level of prediction accuracy was obtained even when other channels with different materials were used, there is a possibility that more information can be extracted for higher prediction accuracy by optimizing the feature values. At present, specific lung cancer marker molecules have not been identified. To this end, we are trying to clarify such marker molecules through GC-MS and PTR-MS measurements, taking account of changes in metabolic pathways caused by cancer. There are several limitations to this study. Firstly, as this is still a pilot study, all the measurements were conducted at a single center using the same measurement module. The capability for screening at early stages was not adequately examined because of the limited number of samples. Secondly, the distribution of lung cancer patients in the measured samples could be different from the actual cases for practical screening applications. Finally, it is possible that the lung cancer cells may not have been completely removed during surgery, leading to some different predictions. This is often the case in advanced lung cancer, and even if there is no residual tumor observable on imaging, there may be residual cancer cells on microscopic examination.In conclusion, using an MSS array, with the closely control of both samples and measurements conditions, we demonstrated a significant prediction accuracy for detecting the presence of lung cancer in patients who underwent lung cancer surgery. These results suggest the potential of MSS analysis for application in lung cancer screening. To elucidate the precise signaling pattern of MSS response to the VOC from patients with lung cancer, and detect the specific VOC as a lung cancer marker, we are planning to conduct a combined analysis of MSS, GC-MS, and PTR-MS, the results of which will hopefully be applied to the development of a safe, non-invasive, and cost-effective screening for lung cancer even at an early stage.FundingThis study was financially supported by a Grant-in-Aid for Scientific Research (A), MEXT, Japan (No. 18H04168); Grant-in-Aid for Scientific Research (B), MEXT, Japan (No. 21H01971); Grant-in-Aid for Scientific Research (C), MEXT, Japan (No. 20K05345); Grant-in-Aid for Scientific Research (C), MEXT, Japan (No. 22K05324); Grant-in-Aid for Exploratory Research, MEXT, Japan (No. 21K18859); Grant-in-Aid for Scientific Research (C), MEXT Japan (no. 21K08879); Grant-in-Aid for Challenging Research (Pioneering) (No. 20K20554); Fostering Joint International Research (B), MEXT, Japan (No. JP19KK0141); the Public/Private R&D Investment Strategic Expansion Program (PRISM), Cabinet Office, Japan; and Center for Functional Sensor & Actuator (CFSN), NIMS.AcknowledgementThe authors thank Thomas Mayers of the Medical English Communications Center, University of Tsukuba, for revision of this manuscript.Appendix A. Supplementary dataThe following are the Supplementary data to this article:Figure S1, Figure S2, Table S1, Table S2REFERENCES1. Siegel RL, Miller KD, Jemal A. Cancer statistics, 2015. CA Cancer J Clin 2015: 65(1): 5-29;https://doi.org/10.3322/caac.21254.2. Goldstraw 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(1): 39-51;https://doi.org/10.1016/j.jtho.2015.09.009.3. Aberle DR, Adams AM, Berg CD, et al. Reduced lung-cancer mortality with low-dose computed tomographic screening. The New England journal of medicine 2011: 365(5): 395-409;https://doi.org/10.1056/NEJMoa1102873.4. de Koning HJ, van der Aalst CM, de Jong PA, et al. Reduced Lung-Cancer Mortality with Volume CT Screening in a Randomized Trial. The New England journal of medicine 2020: 382(6): 503-513;https://doi.org/10.1056/NEJMoa1911793.5. Nawa T, Fukui K, Nakayama T, et al. A population-based cohort study to evaluate the effectiveness of lung cancer screening using low-dose CT in Hitachi city, Japan. Jpn J Clin Oncol 2019: 49(2): 130-136;https://doi.org/10.1093/jjco/hyy185.6. Nardi-Agmon I, Peled N. Exhaled breath analysis for the early detection of lung cancer: recent developments and future prospects. Lung Cancer (Auckl) 2017: 8: 31-38;https://doi.org/10.2147/LCTT.S104205.7. Miekisch W, Schubert JK, Noeldge-Schomburg GF. Diagnostic potential of breath analysis--focus on volatile organic compounds. Clin Chim Acta 2004: 347(1-2): 25-39;https://doi.org/10.1016/j.cccn.2004.04.023.8. Hakim M, Broza YY, Barash O, et al. Volatile organic compounds of lung cancer and possible biochemical pathways. Chem Rev 2012: 112(11): 5949-5966;https://doi.org/10.1021/cr300174a.9. Konvalina G, Haick H. Sensors for breath testing: from nanomaterials to comprehensive disease detection. Accounts of chemical research 2014: 47(1): 66-76;https://doi.org/10.1021/ar400070m.10. Pereira J, Porto-Figueira P, Cavaco C, et al. Breath analysis as a potential and non-invasive frontier in disease diagnosis: an overview. Metabolites 2015: 5(1): 3-55;https://doi.org/10.3390/metabo5010003.11. Amann A, Mochalski P, Ruzsanyi V, et al. Assessment of the exhalation kinetics of volatile cancer biomarkers based on their physicochemical properties. J Breath Res 2014: 8(1): 016003;https://doi.org/10.1088/1752-7155/8/1/016003.12. Hanna GB, Boshier PR, Markar SR, et al. Accuracy and Methodologic Challenges of Volatile Organic Compound-Based Exhaled Breath Tests for Cancer Diagnosis: A Systematic Review and Meta-analysis. JAMA Oncol 2019: 5(1): e182815;https://doi.org/10.1001/jamaoncol.2018.2815.13. Nakhleh MK, 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(1): 112-125;https://doi.org/10.1021/acsnano.6b04930.14. D'Amico A, Pennazza G, Santonico M, et al. An investigation on electronic nose diagnosis of lung cancer. Lung Cancer 2010: 68(2): 170-176;https://doi.org/10.1016/j.lungcan.2009.11.003.15. Peled N, Hakim M, Bunn PA, Jr., et al. Non-invasive breath analysis of pulmonary nodules. J Thorac Oncol 2012: 7(10): 1528-1533;https://doi.org/10.1097/JTO.0b013e3182637d5f.16. Broza YY, Haick H. Nanomaterial-based sensors for detection of disease by volatile organic compounds. Nanomedicine (London, England) 2013: 8(5): 785-806;https://doi.org/10.2217/nnm.13.64.17. Broza YY, Kremer R, Tisch U, et al. A nanomaterial-based breath test for short-term follow-up after lung tumor resection. Nanomedicine 2013: 9(1): 15-21;https://doi.org/10.1016/j.nano.2012.07.009.18. Nardi-Agmon I, Abud-Hawa M, Liran O, et al. Exhaled Breath Analysis for Monitoring Response to Treatment in Advanced Lung Cancer. J Thorac Oncol 2016: 11(6): 827-837;https://doi.org/10.1016/j.jtho.2016.02.017.19. Shlomi D, Abud M, Liran O, et al. Detection of Lung Cancer and EGFR Mutation by Electronic Nose System. J Thorac Oncol 2017: 12(10): 1544-1551;https://doi.org/10.1016/j.jtho.2017.06.073.20. Inada K, Kojima H, Cho-Isoda Y, et al. Statistical Evaluation of Total Expiratory Breath Samples Collected throughout a Year: Reproducibility and Applicability toward Olfactory Sensor-Based Breath Diagnostics. Sensors (Basel) 2021: 21(14);https://doi.org/10.3390/s21144742.21. Yoshikawa G, Akiyama T, Gautsch S, et al. Nanomechanical membrane-type surface stress sensor. Nano Lett 2011: 11(3): 1044-1048;https://doi.org/10.1021/nl103901a.22. Yoshikawa G, Akiyama T, Loizeau F, et al. Two dimensional array of piezoresistive nanomechanical Membrane-type Surface Stress Sensor (MSS) with improved sensitivity. Sensors (Basel) 2012: 12(11): 15873-15887;https://doi.org/10.3390/s121115873.23. Minami K, Imamura G, Tamura R, et al. Recent Advances in Nanomechanical Membrane-Type Surface Stress Sensors towards Artificial Olfaction. Biosensors (Basel) 2022: 12(9);https://doi.org/10.3390/bios12090762.24. Minami K, Shiba K, Yoshikawa G. Discrimination of structurally similar odorous molecules with various concentrations by using a nanomechanical sensor. Analytical Methods 2018: 10(30): 3720-3726;https://doi.org/10.1039/C8AY01224E.25. Shiba K, Tamura R, Imamura G, et al. Data-driven nanomechanical sensing: specific information extraction from a complex system. Sci Rep 2017: 7(1): 3661;https://doi.org/10.1038/s41598-017-03875-7.26. Shiba K, Tamura R, Sugiyama T, et al. Functional Nanoparticles-Coated Nanomechanical Sensor Arrays for Machine Learning-Based Quantitative Odor Analysis. ACS Sens 2018: 3(8): 1592-1600;https://doi.org/10.1021/acssensors.8b00450.27. Lang HP, Loizeau F, Hiou-Feige A, et al. Piezoresistive Membrane Surface Stress Sensors for Characterization of Breath Samples of Head and Neck Cancer Patients. Sensors (Basel) 2016: 16(7);https://doi.org/10.3390/s16071149.28. Yoshikawa G, Loizeau F, Lee CJ, et al. Double-side-coated nanomechanical membrane-type surface stress sensor (MSS) for one-chip-one-channel setup. Langmuir 2013: 29(24): 7551-7556;https://doi.org/10.1021/la3046719.29. Loizeau F, Akiyama T, Gautsch S, et al. Comparing membrane- and cantilever-based surface stress sensors for reproducibility. Sensors and Actuators A: Physical 2015: 228: 9-15;https://doi.org/10.1016/j.sna.2015.02.039.30. Shiba K, Sugiyama T, Takei T, et al. Controlled growth of silica-titania hybrid functional nanoparticles through a multistep microfluidic approach. Chem Commun (Camb) 2015: 51(87): 15854-15857;https://doi.org/10.1039/c5cc07230a.31. Xu H, Kitai K, Minami K, et al. Determination of quasi-primary odors by endpoint detection. Sci Rep 2021: 11(1): 12070;https://doi.org/10.1038/s41598-021-91210-6.32. Minami K, Shiba K, Yoshikawa G. Sorption-induced static mode nanomechanical sensing with viscoelastic receptor layers for multistep injection-purge cycles. Journal of Applied Physics 2021: 129(12);https://doi.org/10.1063/5.0039045.33. Imamura G, Shiba K, Yoshikawa G, et al. Free-hand gas identification based on transfer function ratios without gas flow control. Sci Rep 2019: 9(1): 9768;https://doi.org/10.1038/s41598-019-46164-1.34. MSS Forum.  2022  [cited 2022 11 July]; Available from: https://mss-forum.com/35. Wang P, Huang Q, Meng S, et al. Identification of lung cancer breath biomarkers based on perioperative breathomics testing: A prospective observational study. EClinicalMedicine 2022: 47: 101384;https://doi.org/10.1016/j.eclinm.2022.101384.36. Hosny A, Parmar C, Coroller TP, et al. Deep learning for lung cancer prognostication: A retrospective multi-cohort radiomics study. PLoS Med 2018: 15(11): e1002711;https://doi.org/10.1371/journal.pmed.1002711.37. Yu KH, Lee TM, Yen MH, et al. Reproducible Machine Learning Methods for Lung Cancer Detection Using Computed Tomography Images: Algorithm Development and Validation. J Med Internet Res 2020: 22(8): e16709;https://doi.org/10.2196/16709.38. Ardila D, Kiraly AP, Bharadwaj S, et al. End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography. Nat Med 2019: 25(6): 954-961;https://doi.org/10.1038/s41591-019-0447-x.39.  S. Vanstraelen, D.R. Jones, G. Rocco, Breathprinting analysis and biomimetic sensor technology to detect lung cancer, J Thorac Cardiovasc Surg 166(2) (2023) 357-361 e1;https://doi.org/10.1016/j.jtcvs.2023.02.029.40. Field JK, Hansell DM, Duffy SW, et al. CT screening for lung cancer: countdown to implementation. The Lancet Oncology 2013: 14(13): e591-e600;https://doi.org/10.1016/s1470-2045(13)70293-6.41. Imamura G, Shiba K, Yoshikawa G, et al. Analysis of nanomechanical sensing signals; physical parameter estimation for gas identification. AIP Advances 2018:8(7):075007;https://doi.org/10.1063/1.5036686.42. Minami K, Kobayashi H, Matoba M, et al. Measurement of Volatile Fatty Acids in Silage through Odors with Nanomechanical Sensors. Biosensors 2023:13(2):152;https://doi.org/10.3390/bios13020152.43. Shiba K, Imamura G, Yoshikawa G. Odor-Based Nanomechanical Discrimination of Fuel Oils Using a Single Type of Designed Nanoparticles with Nonlinear Viscoelasticity. ACS Omega 2021:6(36):23389-98;https://doi.org/10.1021/acsomega.1c03270.44. K. Minami, Y. Zhou, G. Imamura, et al. Sorption Kinetic Parameters from Nanomechanical Sensing for Discrimination of 2-Nonenal from Saturated Aldehydes. ACS Sensors 2024 in press;https://doi.org/10.1021/acssensors.3c01888.Table 1. Patient characteristics (N=57) Indicator Value Sex, n  Male 36 Female 21 Age   Range, y 40–83 Mean age ± SD, y 68.9 ± 9.2 Mean BMI ± SD, kg/m2 23.3 ± 3.2 Smoking status  Average pack-years ± SD 28.7 ± 31.1 Never, n 20 Former, n 35 Current, n 2 Histological diagnosis, n  Adenocarcinoma 45 Minimally invasive adenocarcinoma 8 Adenocarcinoma in situ 2 Squamous cell carcinoma 11 Adenosquamous carcinoma 1 Tumor size  Range, mm 10–105 Average size ± SD, mm 29.1 ± 15.9 Invasive size  Range, mm 0–105 Average size ± SD, mm 22.8 ± 20.2 T factor  Tis 2 T1 33 T1mi 8 T2 13 T3 7 T4 2 N factor  N0 48 N1 4 N2 4 N3 1 Stage, n  0 2 IA1 17 IA2 6 IA3 7 IB 8 IIA 1 IIB 8 IIIA 6 IIIB 2Table 2. Results of machine learning analysis No. of used channels(Used channels) Accuracy a Precision b Recall b F1-score b TP b FN b FP b TN b 2 (9,10) 0.692 ± 0.143 0.676 ± 0.147 0.806 ± 0.178 0.723 ± 0.128 2.89 ± 1.22 2.11 ± 1.22 0.97 ± 0.89 4.03 ± 0.89 3 (5,11,12) 0.759 ± 0.134 0.764 ± 0.158 0.804 ± 0.170 0.769 ± 0.124 3.57 ± 1.11 1.43 ± 1.11 0.98 ± 0.85 4.02 ± 0.85 4 (6,10,11,12) 0.782 ± 0.119 0.793 ± 0.144 0.796 ± 0.183 0.780 ± 0.132 3.84 ± 0.91 1.16 ± 0.91 1.02 ± 0.92 3.98 ± 0.92 5 (1,4,5,6,7) 0.809 ± 0.132 0.830 ± 0.153 0.812 ± 0.185 0.805 ± 0.138 4.03 ± 0.92 0.97 ± 0.92 0.94 ± 0.93 4.06 ± 0.93 6 (1,4,5,6,7,10) 0.786 ± 0.119 0.811 ± 0.160 0.786 ± 0.186 0.780 ± 0.134 3.93 ± 0.92 1.07 ± 0.92 1.07 ± 0.93 3.93 ± 0.93 7 (1,2,4,5,6,7,9) 0.785 ± 0.134 0.803 ± 0.157 0.798 ± 0.187 0.784 ± 0.138 3.86 ± 1.01 1.14 ± 1.01 1.01 ± 0.93 3.99 ± 0.93 8 (1,2,4,5,6,10,11,12) 0.781 ± 0.111 0.790 ± 0.134 0.798 ± 0.161 0.782 ± 0.116 3.82 ± 0.89 1.18 ± 0.89 1.01 ± 0.81 3.99 ± 0.81 9 (1,4,5,6,8,9,10,11,12) 0.785 ± 0.120 0.802 ± 0.140 0.786 ± 0.177 0.779 ± 0.134 3.92 ± 0.82 1.08 ± 0.82 1.07 ± 0.89 3.93 ± 0.89 10 (1,2,4,5,6,7,8,9,11,12) 0.774 ± 0.125 0.777 ± 0.145 0.800 ± 0.198 0.772 ± 0.141 3.74 ± 0.95 1.26 ± 0.95 1.00 ± 0.99 4.00 ± 0.99 11 (1,2,3,4,5,6,8,9,10,11,12) 0.768 ± 0.133 0.793 ± 0.163 0.776 ± 0.180 0.767 ± 0.136 3.80 ± 1.08 1.20 ± 1.08 1.12 ± 0.90 3.88 ± 0.90 12 (1,2,3,4,5,6,7,8,9,10,11,12) 0.749 ± 0.120 0.764 ± 0.150 0.774 ± 0.205 0.746 ± 0.142 3.62 ± 1.05 1.38 ± 1.05 1.13 ± 1.03 3.87 ± 1.03a Highest accuracy among all machine learning results with the number of used channelsb Values corresponding to the results of highest accuracyTP, true positive; TN, true negative； FP, false positive; FN, false negative