A Mobile E‐Nose Prototype for Online Breath Analysis

Exhaled breath provides information of individual's holistic physiological condition, making non‐invasive exhaled breath testing a convenient and effective choice for personalized health management. This work proposes a mobile e‐nose prototype equipped with a miniaturized gas delivering unit, an advanced sensor array, a customized analog to digital conversion circuit board, an online data transmission, and machine learning algorithms that can effectively detect and analyze volatile compounds in exhaled breath, providing a systematic design strategy for non‐invasive disease diagnosis. The gas sensor array, as a core element, is composed of eight graphene‐based chemiresistive materials which are sensitive to trace gas analytes. In an online breath test study with 401 cases, patients with respiratory diseases can be distinguished from the healthy with an overall accuracy higher than 80%. In addition, classification of breath samples between patients with chronic obstructive pulmonary disease and the healthy higher than 85% is achieved. These results indicate the proposed portable e‐nose prototype holds great promise in personalized health state monitoring.


Introduction
[3] During the ongoing pandemics, anyone can get sick with the COVID-19 and become seriously ill or die at any age. [1][4] According to World Health Organization, COVID-19 has caused 6.7 million confirmed deaths reported by 7 January 2023, and COPD is the third leading cause of death worldwide causing 3.23 million deaths in 2019. [5,6]Prevention, diagnosis and treatment are essential to the control of both acute and chronic respiratory diseases, which also meet the needs of the United Nations' Sustainable Development Goals by 2030. [7]Among the strategies, timely and correctly diagnosis remains a priority to slow the progression of the disease.Currently, routine examination methods for respiratory diseases can be divided into physical examination, blood gas analysis, lung imaging, bronchoscopy, and laboratory examinations.All methods require patients to go hospital, and some of the tests are invasive and/or radioactive.The current state of the pandemic and respiratory illnesses has brought to light the urgent need for improved technologies to combat these diseases, and rapid point-of-need testing plays a key role in this endeavor.
Exhaled breath (EB) containing abundant endogenous volatile inorganic and organic compounds provides valuable information into the physiological basis of exchange of gases between blood and air.10][11] Respiratory system is unique for EB practice, providing biological samples directly from the organ with every breath and allowing direct non-invasive measurement of ongoing metabolic processes. [12,13][16] Sniffing volatolomics for auxiliary diagnosis can be dated back to traditional Chinese medicine culture and ancient Greeks. [17]Early association of volatile compounds and human disease formed the foundation for the current use of breathomics in early diagnosis and stratification of lung diseases. [18,19]Apart from diseaseinduced metabolic process, microbial metabolism also produces characteristic gases, as mounting evidences suggest that the airway microbiome and gut microbiota are associated with respiratory disease phenotypes and endotypes, and that dysbiosis contributes to airway inflammation. [20,21]Therefore, a holistic measurement of EB may have potential to enhance the diagnostic process, improve clinical outcomes and facilitate timely start of guided therapy. [22]Some of the EB tests for respiratory diseases have been authorized by FDA, e.g., breath nitric oxide test system (Class II), [8] COVID-19 diagnostic test (emergency use authorization). [23]They are completely noninvasive, but are hardly available for daily use especially for healthcare at home.
Gas chromatography-mass spectrometry is a standard approach to identify individual volatile compounds, which is specific but time-consuming.Throughout the last decades, more technologies are developed for EB analysis, e.g., selected ion flow tube mass spectrometry, [24] ion mobility spectrometry, [25] laser spectrometry. [26]Despite the advancements and precisions of these techniques, they are often suffered from being complex, costly, laboratorial, and not suitable for real-time daily healthcare.Among different techniques for EB analysis, sensors are gaining increasing attention in disposable (e.g., paper-based wearable sensors [13,27] ) or nondisposable form(e.g., Cyranose 320, [28] Aeonose [29] ), especially with the increasing activity in the field of sensing in 21st century. [30]Electronic-noses (e-noses) are usually cost-effective, easy to operate and can be used in real-time to recognize patterns of breathprint, making them potential for pointof-care testing (POCT) practice. [31]Unlike traditional sensors for quantitative measurement of target analytes, e-noses do not focus on identifying individual compound, but rather on recognizing distinct breath profiles inherent in different diseases using pattern recognition algorithms.However, disease caused variation of volatile molecules in EB usually spans over trace amount level (sub-ppm to ppb level).The sensitivity of current commercialized e-nose comprising array of metal oxide semiconductor or carbon doped conductive polymer is usually not sufficient.
Herein, a graphene-based mobile e-nose prototype for online breath analysis is proposed to address the issue.The whole prototype includes a gas storage unit, a micro-pump, a gas sensor array, an analog to digital conversion (ADC) and control circuit, a Wi-Fimodule, and a software combining test and algorithm functions.Among them, the gas sensor array and the gas storage unit are the core elements.[34][35] Previously, we have demonstrated that homogeneous loading of reduced graphene oxide (rGO) on polyethylene terephthalate (PET) film can detect sub-ppm and ppb level of volatile compounds and the graphene array doped with different type of metal ions can discriminate trace volatile gases. [34,35]However, the previous works is only in a laboratory stage.An e-nose system with a compact design for future practical application is urgently needed.In the present work, the gas sensor array is composed of eight graphene based chemiresistive materials which are sensitive to trace gas analytes.To achieve a more compact prototype, each pair of interdigitated electrodes (IDEs) of the sensor array are uniformly arranged on a circle to make a homogeneous gas flow stream when sample gas is input from the center.In addition, we specially designed the gas storage chamber so that EB can maintain a longer retention time in the chamber at initial concentration.A 401-case study of rapid online breath test for respiratory disease analysis is conducted.Machine learning was applied for analyzing and comparing the breathprints.Through multi-layer perceptron (MLP), patients with respiratory diseases can be distinguished from the healthy.In addition, patients with COPD can be better classified from the healthy.The results of the study demonstrate that the proposed hand-held e-nose prototype has great potential for implementation in daily clinical practice.

Design and Fabrication
Sensor arrays are potentially becoming convenient devices for the detection and monitoring of respiratory diseases.Online breath analysis is designed for real-time analysis and to provide feedback of personal health states according to breath exhaled at that moment.Concept of the e-nose working procedure is depicted in Figure 1a.When the breath is tested through the e-nose immediately, data are analyzed through the server, and the result is feedback to the user through any mobile terminal.The e-nose can connect to online database for online storage of the sensor signals to enrich the breath database, and for online invocation of the models to match analysis requirements.In addition, cloudconnected e-nose is possible for online learning to make smarter decision models.
The e-nose device, see Figure 1b, contains a gas chamber, a micro-pump, a gas sensor array, a circuit for data acquisition and device control, and a Wi-Fi-module.The Wi-Fi-module is a bridge connecting the e-nose and the terminal unit.The terminal unit deploys software to control the e-nose, record data, and analyze the collected data.The size of the e-nose is essential to make a hand-held device.In that case, we specially designed the gas sensor array (Figure 1c,d) and the gas chamber (Figure 1e).
Gas sensor array is an effective solution to construct a device with high selectivity.It produces cross-reactive responses and can be analyzed by pattern recognition to discriminate target gas.The gas sensor array in this work is composed of eight graphene sensing materials HXS001-HXS008 as previously reported in our work (detailed information can be found in Experimental Section).Reduced graphene oxide (rGO) layer were self-assembled by metal species and form crosslinked rGO networks with rougher, porous, and wrinkled nanostructures.The introduction of metal species in rGO complex can enhance the gas adsorption ability and selectivity, producing resistance change upon gas adsorption owing to the charge transfer and swollen effects. [35]The porous structure and homogeneous sensing layer provide large surface-active sites and facilitates diffusion and adsorption of target gases.Meanwhile, the synergetic coupling effect [36,37] between graphene and different metal species help the sensor array sensitively response to a variety of trace gas analytes with different patterns.Although carbon nanotubes (CNTs) and metal oxide semiconductor (MOS) have been used in the construction of e-nose, we did not choose them because the sensitivity of CNTs and MOS is not as high as the rGO-metal species complex.Meanwhile, MOS sensor needs to work at high temperature, which limits its application in a mobile and wireless device due to its high power consumption.In the present work, the microelectrodes were coated with eight types of rGO-metal species complexes.Each sensing site was uniformly arranged on a circle substrate and compacted within 9.8 mm to make a homogeneous gas flow stream when sample gas was input from the center, Figure 1c.
Gas sensing performance of the sensor array toward prevalent endogenous volatile compounds was studied, including acetone, ammonia, nitric oxide (NO), and hydrogen sulfide (H 2 S).Acetone is a potential biomarker in EB for diabetes mellitus of which 1.71 ppm is usually a critical concentration for anomaly. [38]mmonia is a potential biomarker for patients with liver or renal failure, and breath ammonia higher than 2 ppm can be related to dysfunction. [39,40]NO and H 2 S disorders are highly related to respiratory diseases. [9,41,42]In the respiratory tract, NO is generated via oxidation of L-arginine that is catalyzed by NO synthase (NOS), and NO derived from inducible NOS isoform seems to be a proinflammatory mediator with immunomodulatory effects which has beneficial effects (e.g., host defense) but also several harmful effects (e.g., genotoxicity). [41]H 2 S is the third endogenous gasotransmitter to be discovered after NO and carbon monoxide.H 2 S has been reported to be produced throughout the actions of H 2 S-generating enzymes or redox reactions be-tween the oxidation of glucose and element of sulfur.H 2 S mainly has anti-inflammatory and antioxidant roles, but it also has proinflammatory effects under certain conditions where rapid release of H 2 S in tissues occurs. [42]NO and H 2 S are typically in ppb range in EB.The standard gas testing concentrations were thus regulated according to their concentrations in EB for sensor validation.The dynamic responses of each sensing site (HXS001-HXS008) of the graphene array toward acetone, ammonia, NO, and H 2 S were shown in Figure 2a-d.The discrimination capability of array toward 4 types of model gases was verified by discrimination analysis (Figure 2e).The result showed that different gases at different concentrations were well classified as they were clearly separated.Meanwhile, the concordant dynamic response curves during three cyclic tests (Figure S1a-d, Supporting Information) and long-term stability (Figure S2, Supporting Information) indicated the good reproducibility of the array in gas sensing.Response and recovery time (Figure S1e,f) was calculated which showed a little different, and all of the following gas tests were conducted with non-equilibrium method by controlling the same response and recovery time.In addition to standard gases, simulated EB samples were prepared by adding standard gases into EB samples collected from one healthy individual.Different samples including EB, EB added with NO, and EB added with NO and NH 3 were analyzed.The sensor array also showed good classification results toward EB samples spiked with different types of standard gases (Figure S3, Supporting Information).The high sensitivity and good classification results indicated that the array is qualified for EB test.
Breathprint is highly influenced by environmental factors and water vapor.In this work, online EB test is required to be conducted at ventilated environment with clean air and follows a continuous three-step process, containing baseline, EB sample and recovery line tests.Among the steps, baseline test reduces the influence from the ambient air.A gas chamber is applied for EB storage.Disposable sampling items are used for EB sampling, containing a piece of Polytetrafluoroethylene (PTFE) ingress protection film leveled IP65 to prevent water droplet entering the  chamber.The gas storage chamber was designed to be small and easy to use.Three types of gas chamber structures with 10 volume were designed and optimized according to flow field simulation and experiment test.The three kinds of gas chambers structures include box structure (Figure 3a), box structure with baffles (Figure 3b), and tubular structure (Figure 1e).Box structure is the simplest.During the breath test, the air inlet is kept open and the pump works continuously to extract gas from the sampling port.In the breathing process, EB inlet is opened and breath is exhaled to fill the gas chamber while the excess gas is discharged from EB outlet, and the EB inlet is sealed immediately after exhaling.However, the simulation result show that such structure can only maintain EB mass fraction of higher than 80% for 22 s after breathing when setting the pump rate at 10 sccm.In order to reduce the mixing rate of fresh air and EB, baffles were added near the air inlet (Figure 3b) which turned out only a little longer retention time (27 s).Tubular structure is a little more complicated than box structure, but it turned out much better result.Aiming at easier operation, a symmetric tubular structure was designed, which contains two equivalent pores kept open at both sides of the chamber for gas inlet/outlet, and a sampling pore in the middle of the channel, Figure 1e.EB or air can enter the chamber from either of the two gas inlet/outlet pores, and in the breathing process excessed breath can be discharged from the other side pore.Compared with the former structures, EB mass fraction maintaining performance of the tubular structure is almost two times that of the box structures, which can maintain EB mass fraction higher than 80% for 46 s after breathing when setting the pump rate at 10 sccm, Figure 3c.Optimized gas chamber structure with tubular structure was finally applied.When the baseline reaches to a stable state in the test, the test software instructs the participants to blow continuously for 5 s from either of the two gas inlet/outlet pores, and the excessed breath can be discharged from the other.In the symmetric structure, micro-pump pumps gas analyte from the sampling pore located in the middle of the gas chamber.

Result and Discussion
It is essential to treat respiratory disease timely and correctly when it occurs.A common cold if not properly treated may lead to chronic respiratory disease, other systemic diseases or even death.Chronic respiratory disease when occurred in exacerbation state if not symptomatically treated may result in a worse condition.Preliminary application of the e-nose potential in healthcare management is demonstrated in respiratory diseases where EB is produced.Different volatile metabolites in EB can be products from host and pathogen metabolism, and they originate from 1) endogenous volatile compounds from conducting airways and alveoli, 2) exogenous volatile compounds inhaled and subsequently exhaled or originating from the resident microbiome, and 3) systemic volatile compounds generated elsewhere in the body and transported to the lungs via the blood circulation. [43]Analysis of volatile compounds profiles could potentially improve the diagnostic process or classification of respiratory diseases. [4,44,45]Apart from compounds discussed in Section 2, accumulative data suggest that volatile organic compounds (VOCs) differs in respiratory diseases, including alcohol, aldehyde, alkane, alkene, aromatic, ester, ketone, nitrile, organosulfur, etc. [4,45] A total number of 401 participants enrolled in this practice, including healthy controls and participants with seven different respiratory diseases, Table 1.
Eligible participants who had fasted for at least 2 h were asked to rinse mouth with purified water before breath test to prevent influence from oral gases.Breathprints were compared among three times parallel tests, and a typical sensor output of the parallel tests can be found in Figure S4 (Supporting Information).Variations in resistance of the sensor array were detected.Sensor array responses toward one case of patient with COPD and one case of healthy individual in parallel tests were compared in Figure S5 (Supporting Information), which showed good reproducibility though not exactly overlapped.The responses of 401 participants were compared in Figure S6 (Supporting Information) according to eight clinical states: healthy, COPD, acute upper respiratory infection, pneumonia, asthma, pharyngitis, bronchiectasis, and bronchitis.There were obvious individual differences in response, nevertheless, we can find overall distribution differences in response among different health states.
Three situations were analyzed to demonstrate the possible applications of the e-nose prototype.They were classifications between 1) healthy controls and patients with respiratory diseases, 2) healthy controls and patients with COPD, 3) patients with uninfected and infected respiratory diseases.Responses of the graphene array showed overlaps but also differences in radar plots, Figure 4a-c.We performed nonparametric tests of independent samples of the three binary classifications through SPSS software.Hypothesis test summaries of nonparametric tests output indicated that more than 50% of the sensing sites in the array had significant differences in classified groups, thus demonstrated the feasibility of the proposed prototype, Figure S7 (Supporting Information).As shown in Figure 4c-1-c-8, the responses of graphene array from different EB samples varied among individuals, but there were overall differences between two groups in some sensing sites (HXS001-HXS004, HXS008).The differentiated response values might be resulted from the decrease of reduced gas species (e.g., VOCs) and the increase of NO x levels caused by inflammation in EB samples.
Pattern recognition by MLP was applied for binary analysis.MLP is a kind of feedforward artificial neural network model, which maps multiple input data sets to a single output data set.In the present work, we randomly assigned 70% cases for training and 30% cases for testing, and the result is shown in Figure 5 and Figure S8 (Supporting Information).In detail, in the first situation, classification between healthy controls (174 cases) and patients with respiratory diseases (227 cases), the testing set showed 84.6% specificity, 77.4% sensitivity and overall 80.7% accuracy.Specificity is the percentage of true negative cases account for all actual negative trial cases.Whereas the sensitivity means the percentage of true positive cases in whole actual positive trial cases.Accuracy is the percentage of correctly identified cases account for all trial cases.Receiver operating characteristic (ROC) curve analysis for the model showed that area under curve (AUC) was 0.902 (Figure 5a,b), indicating a high classification ability.In the second situation, classification between healthy controls (174 cases) and patients with COPD (100 cases), the testing set showed 93.5% specificity, 80.0% sensitivity, overall 89.1% accuracy, and 0.969 AUC in ROC curve (Figure 5b,e).In the third situation, classification between patients with uninfected (89 cases) and infected (117 cases) respiratory diseases, the testing set showed 79.2% specificity, 90.9% sensitivity, overall 86.0% accuracy, and 0.950 AUC in ROC curve (Figure 5c,f).MLP estimations (Figure 5a-c) demonstrated that different respiratory disease conditions can be well classified though some scattered dots were observed.One reason accounting for these scattered dots is the variation of EB gas concentration among individuals as mentioned previously.In some cases, underlying medical conditions may coexist especially for the patient group.The above  observations indicate that the proposed portable e-nose prototype have great potential in noninvasive personal health state monitoring and disease risk warning.At the same time, it is possible to make a powerful screening tool for chronic respiratory disease.The above classification results are comparable with previously reported work using commercially available e-noses, which showed accuracy between 68% and 91% by Cyranose 320 [28] or AUC between 0.79 and 0.90 by Aeonose [29] in similar application toward respiratory diseases.However, it is not real-time analysis in case of Cyranose 320, which collected EB samples by Tedlar bags and then analyzed.While in case of the Aeonose, participants were asked to breathe through the device for 5 min while wearing a nose clip.The e-nose prototype proposed in this work is more practical and operable as it only requires 5 s for breath sampling and 4 min for a full test.Further functions can be deployed for such a mobile e-nose, e.g., with establishing a wider range of breath fingerprint database related to clinical data, several specific applications for personalized healthcare management can be added.We believe that e-nose technology has the potential to facilitate personalized medicine through establishing early, accurate diagnosis and monitoring disease course and therapeutic effects.

Conclusion
In brief, a mobile e-nose prototype for online breath analysis is proposed in this work.The online test use ambient air for baseline recording to reduce the influence from environment background.In the real-time test, original state of the EB is retained for testing, and the feedback time can be greatly reduced.The specially designed structures of the sensor array and the gas chamber make a smaller but functional hand-held device possible for more practicable usage.MLP results showed overall 80.7% accuracy and 0.902 AUC in classification between healthy controls and patients with respiratory diseases, overall 89.1% accuracy and 0.969 AUC in classification between healthy controls and patients with COPD, and overall 86.0% accuracy and 0.950 AUC in classification between patients with uninfected and infected respiratory diseases.EB analysis possesses an inherent appeal in real-world clinical practice for respiratory diseases since its non-invasive nature, low patient burden and ability to directly sample from respiratory organ.The preliminary application of the graphene sensor array implanted e-nose prototype demonstrated its possibility for online breath profiling to inform individualized health state related to respiratory diseases, and will have great practical potential in noninvasive personalized healthcare management, which may potentially revolutionize healthcare and point-of-care diagnostics.We are still on our way to improve the proposed e-nose prototype, aiming at developing more sensitive materials, screening and expanding the sites of the sensor array for high discrimination ability in complex gas samples, so as to achieve a much higher accuracy and wider applications.

Experimental Section
Fabrication of the Graphene Gas Sensor Array: IDEs of the gas sensor array were patterned from flexible conductive PET film by photolithography and wet etching method to remove redundant ITO.As depicted in Figure 1c, the array consists of eight channels at 9.8 mm size, and pitch widths for each channel were 200 μm (channel 1), 250 μm (channels 2-4),125 μm (channels 5-7), and 75 μm (channel 8).Graphene-based gas sensors with cross-reactive responses on each channel were labeled as HXS001-HXS008 respectively.The sensing materials were synthesized similarly as previously described. [35]Firstly, eight kinds of graphene oxide (GO)-based complexes as sensor precursors were prepared, including GO-Co 2+ with 0.25 mg mL −1 : 7.5 mm ratio, GO-Fe 3+ with 0.25 mg mL −1 : 2.5 or 5 or 7.5 mm ratio, GO-Cu 2+ with 0.25 mg mL −1 : 5 or 7.5 or 10 mm ratio, and GO-Ce 3+ with 0.25 mg mL −1 : 5 mm ratio.Then, 0.5 μL of the precursors were drop-casted onto the array and dried naturally.Followed by step-by-step in situ reduction in hydrazine atmosphere at 70 °C for 15 min for HXS008 and 10 min for others, so that resistances of HXS001-HXS008 can reach 5 kΩ-2 MΩ.
Gas Sensing Performance Test: After fabrication, each sensor array went through gas sensing performance test.The instrumentation setup for resistance measurement was the same as described in the group previously. [34,35]The concentrations of analytes were adjusted by mixing standard gas and carrier gas via an automatic gas mixing system and then delivered into a home-built chamber.All tests were operated at room temperature (≈22 °C).Keithley 6487 (Keithley Instrument, USA) coupled with a multiple relay control board was developed as a multi-channel picoammeter for voltage supply and current recording.Real-time current of the sensor array was recorded by LabVIEW 2020 Community software.
Design of E-Nose Prototype: Overall size of the prototype was 9.3*14.0*5.0 cm.It was powered by two 3.7 V lithium ion batteries, and contained a gas storage unit, a micro-pump (mp6-gas, Bartels, Germany), a graphene gas sensor array, a self-designed 8-channel data acquisition and control circuit, a Wi-Fi-module (WIFI232-B2, USR, China), and a selfdesigned software combining test and algorithm functions.For the gas chamber optimization, box gas chambers (box structure, Figure 2a; box structure with baffles, Figure 2b) with 80*31*4 mm size and 10 mL volume were simulated, and the tubular gas storage chamber (Figure 1e) with 11.2*3.2*0.7 cm size and 10 mL volume was simulated.The tubular gas chamber contains three pores: the equivalent pores at both side of the chamber are used for gas inlet/outlet (inner diameter 3.8 mm); the smaller pore in the middle of the channel is gas sampling port (inner diameter 1.2 mm).Ambient air or sample gas can fill the chamber through gas inlet/outlet pores, and micro-pump pumps the gas in the chamber from the sampling port.EB mass fraction in the gas chamber was calculated by Ansys simulation software (2021 R2).Volume rendering is shown in Figure 1e and Supporting Information 2, when micro-pump keeps pumping analyte gas at 10 sccm after the chamber is fully filled with EB.In addition to the gas chamber, the self-designed data and acquisition and control circuit is composed of eight 24-bit ADCs, so that real-time resistance changes from the eight sensing sites in the sensor array can be recorded synchronously with high precision.
Breath Analysis with the E-Nose Prototype: Online breath test contains assembly of disposable sampling items and gas test.The sampling items contain a mouth piece of spirometry, a piece of PTFE ingress protection film leveled IP65, a Teflon tube connected to the gas chamber.During a breath, typical ventilation volume per respiratory cycle contains 150 mL of dead space section and 350 mL of a section with lung gas exchange. [46]articipants were asked to breath continuously for 5 s which is an acceptable time span of a breath, so that dead space section of a breath can be discharged and alveolar air can be collected in the 10 mL gas chamber.Non-equilibrium method was adopted for dynamic sensing of the breath test.It is a continuous process setting flow rate at 10 sccm and can be divided into three steps: 1) recording of baseline, ambient air is pumped by the micro-pump through gas chamber and the resistance of the sensor was recorded until reaching a stable state; 2) breathing and recording of sampled gas, the test software instructs the participants to blow continuously for 5 s, and the sample gas test lasts for around 1 min; 3) recording of the recovery process, ambient air is pumped again for around 3 min for recovery which is sufficient to reach the stable state at that moment.A complete breath test takes 4 min.
Online breath tests by e-nose prototype were conducted between March 2021 and June 2021.Eligible individuals were those who had not consumed food or medications at least 2 h and had not taken alcohol, tobacco or coffee at least 24 h before sampling.All the breath tests were conducted in ventilated rooms and the participants were asked to rinse mouth with purified water before sampling.A total number of 401 volunteers (adults) participated in this study.Participants were mainly from collaborators of the research group, and internal medicine/respiratory medicine/thoracic surgery department of hospitals.The study population was separated into 2 groups: 174 healthy individuals for control, and 227 participants with respiratory diseases (100 participants with COPD, 46 participants with acute upper respiratory disease, 56 patients with pneumonia, 4 participants with asthma, 4 participants with pharyngitis, 2 patients with bronchiectasis, and 15 patients with bronchitis).Detailed information is listed in Table 1 Statistical Analysis: 1) Pre-processing of data.Electrical response can be detected in term of variation in current or conductance or resistance.In this work, the measured sensor response was converted to relative conductance change: Here, I 0 /G 0 /R 0 are the current/conductance/resistance measured in carrier gas stream, and I/G/R are the current/conductance/resistance measured during analytes exposure.In EB detection, to reduce the data variation in batch to batch analysis, the original responses (R%) on each sensing elements of the sensor array were calibrated using standard gas before inputting them to statistical algorithm.2) Data presentation: typical data in parallel (usually three times) tests.3) Sample size: a total number of 401 cases (Table 1) gas sensor responses were analyzed, in which 401 datasets were applied for classification between healthy controls (174 cases) and patients with respiratory diseases (227 cases), 274 datasets were applied for classification between healthy controls (174 cases) and patients with COPD (100 cases), and 206 datasets were applied for classification between patients with uninfected (89 cases) and infected (117 cases) respiratory diseases.4) Statistical methods used to assess significant differences: significances (p-values) are calculated using Mann-Whitney U test. 5) Software used for statistical analysis: statistical analysis was carried out using IBM SPSS statistics 20.0 to estimate the response pattern of the sensor array.

Figure 1 .
Figure 1.Schematic illustration of the mobile e-nose prototype for online breath analysis.a) An overview of the working procedure of the e-nose prototype, b) schematic diagram of the e-nose, c) circuit configuration of the 8-site sensor array, d) testing chamber of the array, e) structure of the gas chamber and its volume rendering of the mass fraction of EB to air at 22 s after exhaling.

Figure 3 .
Figure 3.Comparison of the EB mass fraction maintaining performance of the gas chambers.Volume rendering of the mass fraction of EB to air of a) box structure, b) box structure with baffles gas chambers at 22 s after exhaling.c) EB mass fraction change of the three structure gas chambers (box, box with baffles, and tubular (Figure 1e) after exhaling.
. The study was approved by the Ethics Committee of the Affiliated Hospital of Hangzhou Normal University (Certification No. 2021(E2)-KS-055), Sir Run Run Shaw Hospital (Certification No. 20201231-53), and the Second Affiliated Hospital of Zhejiang University (Certification No. 2021-0507).

Table 1 .
Enrollment information of participants.