Machine Learning Applied to Chemical Analysis: Sensing Multiple

Sep 26, 2012 - Machine Learning Applied to Chemical Analysis: Sensing Multiple. Biomarkers ..... The dashed line separates the training data (left of ...
0 downloads 0 Views 3MB Size
Article pubs.acs.org/ac

Machine Learning Applied to Chemical Analysis: Sensing Multiple Biomarkers in Simulated Breath Using a Temperature-Pulsed Electronic-Nose Phillip H. Rogers, Kurt D. Benkstein, and Steve Semancik* Biochemical Science Division, National Institute of Standards and Technology, 100 Bureau Drive, Gaithersburg, Maryland 20899-8362, United States S Supporting Information *

ABSTRACT: Monitoring of chemical species in breath offers an approach for the detection of disease and other conditions that cause homeostatic imbalance. Here, we demonstrate the use of microsensor-based devices for detecting select biomarkers in simulated exhaled breath as a step toward enabling fast and inexpensive breath-screening technology. Microhotplate elements functionalized with three chemiresistive metal-oxide films (SnO2, In2O3, and CuO) were used to acquire data in simulated breath containing single targets [(5 to 20) μmol/mol ammonia, methanol, and acetone], as well as mixtures of those species. All devices were operated with programmed thermal cycles featuring rapid temperature excursions, during which film resistances were measured. Material-specific temperature programs were optimized to achieve temperature-dependent metal-oxide sensing film conductance levels and target selectivity. A supervised hierarchical machine-learning algorithm using linear discriminant analysis for dimensional reduction of sensing data and discrimination was developed. This algorithm was employed in the classification and quantification of biomarkers. This approach to microsensor data collection and processing was successful in classifying and quantifying the model biomarkers in validation-set mixtures.

V

In contrast, microsensor arrays can be manufactured inexpensively and can enable characterization of a patient’s exhaled breath history. They may utilize technology that can be trained specifically to each patient.7,9,14−22 Assuming that suitable sensitivity and selectivity can be achieved, inexpensive chemical sensor arrays such as electronic-nose (e-nose) devices functionalized with chemiresistive metal-oxide transducers may prove viable for the detection and quantification of volatiles in breath. Metal oxide chemiresistors tend to be broadly selective; that is, they show a change in resistance with a changing amount of many different analytes and interferences. This broad selectivity is the result of the transduction mechanism based upon interaction of the metal oxide surface with adsorbed analyte and background molecules.23 Several approaches may be used to enhance the selectivity of metal oxide chemiresistors including the use of dopants, modulation of operating conditions (e.g., temperature, field effects, light), and use of multielement arrays with varied types of sensing materials (different metal oxides, conductive polymers, etc.). While chemical modification of the chemiresistor can render it more sensitive to one chemical or class of chemical, having broad selectivity can be useful for a sensor when in an array configuration. Additionally, thermally

olatile organic compounds, as well as some volatile inorganic chemicals (collectively referred to as “volatiles” here), are produced by core metabolic reactions and are common in exhaled human breath in concentrations ranging from sub-nmol/mol to tens of μmol/mol. Depending on the individual and his/her health status, the presence and concentrations of volatiles can vary significantly. The composition of the mixture of volatiles can act as a “fingerprint” of an individual’s homeostasis. There has been a significant effort in the past decade to develop methods which can measure, classify, and quantify the components of these mixtures for monitoring human health and diagnosis of disease.1−8 Methods currently exist that are capable of selectively determining the presence and concentration of volatiles in exhaled breath. These methods can be categorized in one of two ways: (1) lab-based instruments and (2) inexpensively manufactured microsensor arrays.9 Lab-based equipment, such as gas chromatographs, can be extremely sensitive and selective to a broad range of analytes, but they tend to be expensive and take several minutes to hours to reach an analytical conclusion.10−13 These approaches are therefore out of reach to individuals interested in rapid diagnostic data (possibly to be used as a sensory input to a diagnostic agent) and for home-based monitoring. Furthermore, lab-based instruments that evaluate the composition of discrete samples offer only snapshots of a patient’s molecular outgassing history. This article not subject to U.S. Copyright. Published 2012 by the American Chemical Society

Received: June 18, 2012 Accepted: September 26, 2012 Published: September 26, 2012 9774

dx.doi.org/10.1021/ac301687j | Anal. Chem. 2012, 84, 9774−9781

Analytical Chemistry

Article

Figure 1. (A) Exploded view schematic of a single microsensor element (approximately 100 μm across), and (B) each material-specific temperature program selected after analyses of preliminary tests on the target analytes in a dry air background.

In2O3, and CuO) was used to generate the data for this study. Sensing films were formed by localized calcification on microhotplate elements from micropipetted metal-hydroxide sols followed by localized annealing to 500 °C in dry air. Prior to the experiments reported here, the sensing array was intermittently aged for over 100 h of thermal cycling to improve reliability. This sensing device was also used in a prior optimization study performed to establish custom temperature programs associated with each of the sensing materials for monitoring the target analytes examined in this study.28 During operation, the three elements of the chemiresistive microsensor array were independently heated with the materialspecific pulsed temperature programs. These three temperature programs are shown in Figure 1 with an exploded-view schematic of a single microhotplate chemiresistive sensing element. Sensing-film resistance values were only collected during the base temperatures indicated in Figure 1 by the highlights. Once collected, the resistance data were normalized using an approach previously developed that involved both a self-normalization and a range-normalization.28 In short, the previous study focused on developing procedures for the optimization of sensor operation based on an exhaustive, multidimensional database. In collecting the sensing data for this previous database, the temperature program was incrementally varied by increasing an intermittent base temperature of a base/ramp temperature program after every 6 h gas exposure cycle to create an extensive database consisting of a broad range of temperatures and temperature histories. The gas exposure cycle consisted of acetone, ammonia, and methanol exposures in varying concentrations in a dry air background. Evaluating this database with linear discriminant analysis (LDA) and principle component analysis (PCA), we determined that certain temperature bases and excursions provided in the temperature programs shown in Figure 1B contribute most to analyte separability and orthogonality within dimensionally reduced data sets. Simulated-Breath Exposures. Testing of the microhotplates involved exposure to a gas stream that alternated between zero-grade dry air humidified to 20% relative humidity

cycled e-noses use rapid temperature pulses to induce a level of analytical orthogonality in microarrays of metal oxide chemiresistors.18,21,24−27 Using advanced signal processing methods to pick out the differences in the data streams caused by the various analytes imparts discrimination capabilities to the sensing system. It has been demonstrated previously that chemiresistive sensors are sensitive to individual biomarkers in simulated breath experiments.16,18 It is a different task, however, to demonstrate cross-selectivity and quantification of a sensor array to more than one biomarker in complex mixtures. Furthermore, this task is difficult to perform with reaction models once multiple targets and mixtures are considered. Instead, the sensor system must be “taught” to recognize the environment in a way analogous to that of a forensics K-9 that receives supervised training to sniff out controlled substances or volatile biological material. Here, multivariate data streams from three metal-oxide microsensors operated in a pulsed-temperature mode were used to track the presence and amount of three potential biomarkers in simulated exhaled breath. The focus of this work was to train the sensing system to recognize variations in individual biomarkers presented to the e-nose while changing the mixture compositions and concentrations. To facilitate the training and testing of the sensor system, a hierarchical machine-learning algorithm was employed, which utilizes linear discriminant analysis (LDA) at each of the nodes for supervised discrimination and quantification. After training, this hierarchical algorithm was able to detect and quantify trained and untrained concentrations and mixtures of the three biomarkers. The approach was found to work well with the specific chemiresistive microsensor array studied here. We believe that the employed hierarchical algorithm represents a more generally applicable method that can be used in the training of a variety of sensor systems.



EXPERIMENTAL SECTION Microhotplate Sensing Array. A microhotplate sensing array possessing three chemiresistive sensing elements (SnO2, 9775

dx.doi.org/10.1021/ac301687j | Anal. Chem. 2012, 84, 9774−9781

Analytical Chemistry

Article

Figure 2. Exposure schedule for the three target biomarkers, acetone (top), ammonia (middle), and methanol (bottom) used in the training (blue) and testing (red) of the sensor system. Regions where all three biomarkers are zero are indicative of either 20% relative humidity air or a simulated breath without biomarker exposure.

selective to that analyte alone. The levels of the hierarchy act as a filter for environmental effects, complex backgrounds, and analyte mixtures, turning a sensor array that is sensitive to many volatile organic compounds (VOCs) into a sensing system that detects acetone concentration in breath, for example. The other advantage of a hierarchical approach is apparent when considering an ever-expanding training set. As greater numbers of backgrounds, mixtures, and analytes are added, an all-in-one approach would likely fail, because analyte clusters would likely grow and become less well-defined, having a detrimental effect on analyte classification. A hierarchy can handle a greater number of mixtures, backgrounds, and analytes by simply adding additional hierarchical nodes to filter the larger data set. The hierarchy we used in this study, displayed schematically in Figure 3, shows the five nodes used to determine the presence of breath, any biomarker, and each of the three specific biomarkers examined in this study. To discriminate at each of the nodes, LDA was used to dimensionally reduce the data set to a single dimension which separated the “yes/ present” state from the “null/not present” state. LDA eigenvectors determined with training exposures of varying mixtures and concentrations of each of the biomarkers built categorization and classification properties into the algorithm. These properties and the diminished susceptibility to sensor drift are results that derive from the single dimension, dual state projection of the sensing data at each node. In order to discriminate in this way, the LDA-determined eigenvector was heavily weighted in dimensions that produce high correlation in all sensing data. This stabilizes the system to variations in the analyte and background composition of the synthetic mixture. LDA was applied to normalized resistance data designated as the sensing signal at each of the nodal discrimination points. At each node, a supervision array of length equal to the number of measurements made in the training set had elements valued at either 0 or 1, indicating that the nodal case was not or was satisfied at that measurement. This training was repeated for all nodes. After training, a conditional parameter was defined that had to be satisfied for a measurement to be deemed a “yes”

using a dew-point generator and simulated-breath exposures (77% N2 and 3% to 5% CO2 by volume balanced by O2 at 60% relative humidity) flowed at 1000 cm3/min over the sensor array. Simulated breath exposures containing biomarkers had target concentrations that ranged between 5 μmol/mol and 20 μmol/mol for all of the three biomarkers examined. Training exposures included single and multibiomarker mixtures of (0, 5, 10, and 20) μmol/mol of each of the biomarkers in the simulated breath stream, whereas the test mixtures were randomly varied (but kept within the training concentration range). Figure 2 outlines both the sensor training (blue) and test (light red) regions of the 24 h gas exposure cycle. Areas in Figure 2 that do not show biomarker concentrations are time segments where either a breath exposure was performed without a biomarker present or when the sensor was exposed to air only.



COMPUTATIONAL APPROACH To improve the capabilities of our temperature-modulated sensing method as both a sensitive and a selective means of tracking complex mixtures in gas streams, we have developed a hierarchical approach to data handling for the classification and quantification of target biomarkers. This hierarchy is designed to break the evaluation process into manageable pieces that can be represented by bimodal outputs. The approach is somewhat analogous to playing Twenty Questions with a data set (with potentially fewer questions). It is expected that similar classification results could be achieved using simultaneous multicategory classifiers, i.e., an all-in-one approach. That stated, the hierarchical approach employed here does have advantages over the simultaneous approach in that it is optimized for solving the particular measurement problem of interest (through supervision in the training process), and can be less susceptible to sensor drift related to desensitization. Essentially, a hierarchical approach can be used to tailor a sensor array to act as a selective and deterministic sensing device for a target analyte, even if the sensor is not inherently 9776

dx.doi.org/10.1021/ac301687j | Anal. Chem. 2012, 84, 9774−9781

Analytical Chemistry

Article

compared to background air (20% relative humidity). Simulated breath exposures in this study contained 65% relative humidity to simulate the highly humid conditions found in human breath. This relative humidity condition in the simulated breath exposures facilitated exploration of a wide range of biomarker concentrations and mixtures (Figure 2) from calibrated mixtures of the biomarkers balanced with dry air in compressed gas cylinders. In addition to the analyte concentrations, the CO2 concentration (balanced by O2) was also varied with each breath exposure to simulate fluctuations in respiratory gas exchange. Displayed in Figure 4 are the resistance measurements portrayed as “perturbed isotherms,” denoted as such due to the

Figure 3. Schematic depiction of the hierarchical machine-learning algorithm used to classify and quantify biomarkers in simulated breath exposures.

response. This condition was that the measurement in question falls within a number of standard distributions of the mean of the “yes” training distribution. If the result of a nodal calculation resulted in a “yes” outcome, the child node would be calculated; otherwise the current node and all subsequent nodes would be deemed a “null.” In the case that a measurement was valued as a “yes” for a target biomarker, the concentration was calculated. This calculation was made using a similar approach to the bimodal node training outlined above though the aim was to define a regression and not a classification scheme. Instead, training of the quantification nodes was performed by assigning each of the four concentrations used for the biomarker a separate LDA axis distribution. After calculating the eigenvectors, a polynomial regression (order 3) was fit to the concentrations versus the first LDA axis for each of the three biomarkers. When a biomarker classification was satisfied, the concentration was calculated using these regressions. All training and test measurements were calculated with these quantification nodes if a biomarker was detected by the parent node.

Figure 4. Plots of the synchronized isothermal resistance values produced at the base temperatures using the individual temperature programs shown in Figure 1B that are associated with each of the three sensing elements. The dashed line separates the training data (left of the line) and testing data (right of the line). See Figure 2 for the exposure schedule.

effect that the average temperature and thermal history has on the kinetics and chemical equilibria (for the interfacial interactions) outside of what would be traditionally considered a typical isothermal experiment. These perturbed isotherms were assembled from measurements at each of the base temperatures in the temperature programs for the three sensing materials (Figure 1B). Shown primarily to depict the lack of obvious selectivity of the three sensing films to the biomarkers, Figure 4 also shows a dashed line indicating the separation of the training and the test simulated-breath exposures. Instead of working directly with the raw sensing data, the resistance values of each base temperature were normalized to produce calculated sensing signals and these sensing signals were then used as the sensory inputs to the breath analysis hierarchy for training and testing. Upon training each of the nodes, adequate separation between the “yes” and “null” states was observed for each of the classification types: breath, biomarker, and each of the three individual biomarkers (distributions provided in Figure 1S, Supporting Information). Binning conditions can be implemented to minimize the number of false positives or false negatives or for the optimization of minimal contribution of both (the effects of which are provided in the Supporting Information). LDA attempts to find projection axes that minimize within-class scatter and maximize between-class



RESULTS AND DISCUSSION For this work, we used SnO2, In2O3, and CuO chemiresistive metal-oxide sensing films on microhotplate sensing elements that were thermally cycled by embedded poly-silicon thermoresistive heating elements. Acetone, ammonia, and methanol were chosen as target analytes because of reports that link variation in the concentrations of these molecules to various impaired conditions.3−5,13,18,29 Furthermore, the chemiresistive sensing materials used in this study exhibit similar sensitivity to each of the biomarkers so the analytes provide a challenging data set for the signal analysis approach. The primary focus of this study was the demonstration of discrimination and quantification of select biomarkers in a simulated-breath background, though we note that concentrations explored in this report may differ somewhat from those that are associated with a particular disease state. Exhaled breath was simulated as having elevated carbon dioxide and humidity levels and depleted oxygen content as 9777

dx.doi.org/10.1021/ac301687j | Anal. Chem. 2012, 84, 9774−9781

Analytical Chemistry

Article

Figure 5. Classification results for bin size = 4σ (or μ ± 2σ). The top plot shows the classification results for breath and breath-with-biomarker detection. The middle plot shows the classification results for the individual biomarkers. The bottom plot displays temporally magnified results for the test data only. Shading indicates regions where gas exposure conditions should result in a positive or “yes” hit for the respective node, with blue shading indicating the test region and red shading indicating the test data. The dashed line separates the training data (left of the line) and testing data (right of the line).

scatter, resulting in “Gaussian-like” distributions for the data collected in this study. For this reason, bins are defined in numbers of standard deviations (centered at the mean) or ± a number of standard deviations about the mean. A bin size of μ ± 2σ, where μ is the mean and σ is the standard deviation, was chosen for the classification plots to minimize false positive hits while keeping the number of false negatives low enough to demonstrate the capabilities of the sensing system. The scale of the binning condition can be tuned to optimize for specific application cases. For example, one would likely favor bin sizes of ±2σ or smaller while increasing the integration time of sample collection in cases that are more concerned with false positives. Whereas a larger binning condition would improve sensing speed, it might be more susceptible to misclassifica-

tions. Figure 2S in the Supporting Information outlines the effects of bin size scaling for data collected in this study. Figure 5 displays the classification results for both training (for verification of adequate LDA separability) and test-gas exposures. Figure 5 also includes a temporally magnified plot of the individual biomarker classification results for the test-gas exposures to show that correct classification hits occurred during most biomarker exposures while very few false positive classifications occurred. In every case of the test data, the concentrations of the synthetic breath components and of the biomarker mixtures were different than those of the training conditions. In the few cases where an entire 15 min gas exposure band (in which the concentrations were not varied) did not produce hits from a particular exposed biomarker, there 9778

dx.doi.org/10.1021/ac301687j | Anal. Chem. 2012, 84, 9774−9781

Analytical Chemistry

Article

data to linear functions, as shown in Figure 3S, Supporting Information. There are varying degrees of sensitivity (i.e., the slopes of the linear fits) of the sensing system to each of the target biomarkers. In each of the three cases, the linear fits are adequate, but only in the case of acetone does the extrapolated line intercept the y-axis near the zero-concentration point. For the other two analytes and most obviously for methanol, the nonzero intercepts suggest that the sensor system is nearing response saturation. This response saturation suggests that the sensor system may generate a better calibration for lower concentrations of these analytes. That is, the system as currently configured and operated may be better suited for quantification of ammonia and methanol in lower ranges of concentration. Because of the nonlinearity of the system for ammonia and methanol, we fit the cluster position versus concentration data with a third-order polynomial function (Figure 6B) to better estimate the analyte concentrations. The results of the concentration-dependent LDA training and regression fitting procedure are displayed in Figure 7, where the quantification from both the training and test sensing data are shown. The calculated concentrations fit the training data quite well, as expected for a supervised machine-learning algorithm like the one used here. The concentration calculations for the test data, however, are not as closely matched to the exposure concentrations for all biomarkers and mixtures. Only those data points that were classified as having the specific biomarker present (from the parent node, see also Figure 5) were quantified, which results in the gaps in the quantification plots for various biomarker exposures (displayed as a zero concentration). Ignoring those gaps in quantification, the quantification of acetone was most accurately performed for the unknown concentrations and mixtures of biomarkers. Ammonia and methanol biomarkers displayed a mix of accurate and inaccurate quantification, reflecting their lower sensitivity over the training concentration range (Figure 6 and Figure 3S, Supporting Information). Quantification for these two biomarkers tended to be more accurate for the lowest exposure concentrations. In the misquantified exposures that did occur, the biomarkers were calculated to have concentrations lower than the actual levels present in the mixture. This “sensor system” sensitivity is a convolution of the fundamental sensing mechanisms and sensitivity of the individual sensing elements to the target analytes and the supervised linear operations performed on the sensing data. The result is a higher order consideration of sensitivity where selectivity takes precedence over the system performance, which is akin to the olfactory system in animals. In a real sensing system where complex mixtures are a requirement, this higher order sensitivity is of much greater concern than the sensitivity of one sensing material to one analyte (though they are linked). We have reported on controlled experiments that use one approach to classification and hierarchical discrimination. Actual human breath contains hundreds of volatile compounds. Whether one has access to the most advanced laboratory equipment or a microhotplate arraybased electronic-nose, it will be necessary to utilize a gamut of computational techniques to measure all or even several of the target compounds in breath. Employing hierarchical algorithms that use dimensional reduction techniques, such as LDA for nodal decision making, is an expandable approach (e.g., by adding further analyte identification and quantification nodes to the scheme shown in Figure 3) well suited for this task. If correctly designed, an expandable algorithm such as this one

was always at least one other biomarker present in the simulated breath mixture that was recognized. In addition to classification of biomarkers in complex mixtures, LDA was also used to facilitate the quantification of biomarker targets. With four concentrations being explored for each biomarker during the training set, three LDA eigenvectors were produced. As a result, three axes were available for developing single- or multivariable regressions with exposure concentration. For this work, only the first LDA axis versus concentration was used to define regressions for the three biomarkers. The LDA axis 1 scores calculated for the individual biomarker quantifications are shown in Figure 6A for the training exposures. The results of these regressions are displayed in Figure 6B. To examine the sensitivity of the system to the three analytes, we also fit the LDA axis 1 score versus analyte concentration (5 μmol/mol to 20 μmol/mol)

Figure 6. (A) LDA axis 1 scores for the concentration-separated sensing data, for each of the three biomarkers used to train for regression-based quantification of biomarkers. The shaded area in the top plot is the region magnified in the bottom plot, which more clearly displays the training exposures where all three biomarkers were present in the training concentrations. (B) LDA axis 1 concentrationbased regressions used in biomarker quantification. The dashed red line indicates the polynomial (order 3) fit to concentration versus cluster position. Error bars displayed in the LDA axis 1 position span ± one standard deviation about the mean position of the cluster associated with each of the target analyte concentrations. 9779

dx.doi.org/10.1021/ac301687j | Anal. Chem. 2012, 84, 9774−9781

Analytical Chemistry

Article

Figure 7. The top plot shows biomarker quantification results (black dots) over the actual exposure concentrations (replotted from Figure 2), indicating that trained and untrained concentrations can be quantified for each of the three biomarkers. The dashed line separates the training data (left of the line) and testing data (right of the line). The bottom plot shows a temporally magnified plot displaying only the quantification of the test data.

would only be limited by the stability, sensitivity, and selectivity of the sensing hardware.

mixtures of acetone, methanol, and ammonia in simulated breath pulses. To facilitate the task of biomarker classification and quantification, a hierarchical machine-learning algorithm was employed for the training and testing of the sensor system. The results demonstrate a tunable system that is capable of classifying and quantifying the individual components of biomarker mixtures in simulated breath. It is likely that increasing sensing signal integration time would lower the number of false negatives observed. In future studies, we plan to explore performance training with exposures of greater variability, range, and variety and to extend this modular approach of chemical sensing via machine-learning algorithm enabled e-noses to other applications.



CONCLUSIONS Enabling e-nose systems with machine-learning algorithms capable of discerning biomarker components in complex mixtures and backgrounds in human breath could extend the utility of gas-phase chemical sensors to a range of biomedical monitoring applications. This work was aimed at the demonstration of biomarker classification and quantification in simulated human breath by a chemiresistive microsensor array. The microsensor array utilized for this study was functionalized with SnO2, In2O3, and CuO chemiresistive sensing films. With these three types of sensing elements, which were thermally cycled, chemical sensing data were acquired for 9780

dx.doi.org/10.1021/ac301687j | Anal. Chem. 2012, 84, 9774−9781

Analytical Chemistry



Article

(24) Afridi, M. Y.; Hefnera, A.; Berninga, D.; Ellenwooda, C.; Varmab, A.; Jacobb, B.; Semancik, S. Solid-State Electron. 2004, 48, 1777−1781. (25) Raman, B.; Hertz, J. L.; Benkstein, K. D.; Semancik, S. Anal. Chem. 2008, 80, 8364−8371. (26) Hierlemann, A.; Gutierrez-Osuna, R. Chem. Rev. 2008, 108, 563−613. (27) Rogers, P. H.; Semancik, S. Sens. Actuators, B 2011, 158, 111− 116. (28) Rogers, P. H.; Semancik, S. Sens. Actuators, B 2012, 163, 8−19. (29) Davies, S.; Spanel, P.; Smith, D.; et al. Kidney Int. 1997, 52, 223.

ASSOCIATED CONTENT

S Supporting Information *

Figures showing the normalized distributions of the LDAseparated data at each of the five classification nodes (S-1), the effect of bin size on false positives/negatives (S-2), and linear fits to the concentration-separated sensing data (S-3). This material is available free of charge via the Internet at http:// pubs.acs.org.



AUTHOR INFORMATION

Corresponding Author

*E-mail: [email protected]. Notes

The authors declare no competing financial interest.



ACKNOWLEDGMENTS The authors thank Eric Dattoli, Chip Montgomery, and Dean Ripple for valuable comments and assistance relevant to this work. P.H.R. was supported by a NIST Postdoctoral Research Associateship Award administered through the National Research Council.



REFERENCES

(1) Mazzone, P. J. J. Thorac. Oncol. 2008, 3, 774−80. (2) Phillips, M.; Cataneo, R. N.; Saunders, C.; Hope, P.; Schmitt, P.; Wai, J. J. Breath Res. 2010, 4, 026003. (3) Risby, T. H.; Solga, S. F. Appl. Phys. B: Laser Opt. 2006, 85, 421− 426. (4) Smith, D.; Spanel, P.; Fryer, A. A.; Hanna, F.; Ferns, G. A. A. J. Breath Res. 2011, 5, 022001. (5) Barker, M.; Hengst, M.; Schmid, J.; Buers, H.-J.; Mittermaier, B.; Klemp, D.; Koppmann, R. Eur. Respir. J. 2006, 27, 929−936. (6) Cao, W.; Duan, Y. Clin. Chem. 2006, 52, 800. (7) Peng, G.; Tisch, U.; Adams, O.; Hakim, M.; Shehada, N.; Broza, Y. Y.; Billan, S.; Abdah-Bortnyak, R.; Kuten, A.; Haick, H. Nat. Nanotechnol. 2009, 4, 669−673. (8) Turner, C.; Spanel, P.; Smith, D. Physiol. Meas. 2006, 27, 637− 648. (9) Di Francesco, F.; Fuoco, R.; Trivella, M.; Ceccarini, A. Microchem. J. 2005, 79, 405−410. (10) Sanchez, J. M.; Sacks, R. D. Anal. Chem. 2003, 75, 2231−2236. (11) Van den Velde, S.; Nevens, F.; van Steenberghe, D.; Quirynen, M.; et al. J. Chromatogr., B 2008, 875, 344−348. (12) Wang, C. Sensors 2009, 9, 8230−8262. (13) Deng, C.; Zhang, J.; Yu, X.; Zhang, W.; Zhang, X. J. Chromatogr., B 2004, 810, 269−275. (14) Gelperin, A.; Johnson, A. J. Breath Res. 2008, 2, 037015. (15) Chan, H. P.; Lewis, C.; Thomas, P. S. Lung Cancer 2009, 63, 164−168. (16) Righettoni, M.; Tricoli, A.; Pratsinis, S. E. Anal. Chem. 2010, 82, 3581−3587. (17) Machado, R. F.; Laskowski, D.; Deffenderfer, O.; Burch, T.; Zheng, S.; Mazzone, P. J.; Mekhail, T.; Jennings, C.; Stoller, J. K.; Pyle, J.; Duncan, J.; Dweik, R. a; Erzurum, S. C. Am. J. Respir. Crit. Care Med. 2005, 171, 1286−1291. (18) Benkstein, K. D.; Raman, B.; Montgomery, C. B.; Martinez, C. J.; Semancik, S. Sen. J., IEEE 2010, 10, 137−144. (19) Aguilar, A. D.; Forzani, E. S.; Nagahara, L. A.; Amlani, I.; Tsui, R.; Tao, N. J. Sen. J., IEEE 2008, 8, 269−273. (20) D’Amico, A.; Di Natale, C.; Paolesse, R.; Macagnano, A.; Martinelli, E.; Pennazza, G.; Santonico, M.; Bernabei, M.; Roscioni, C.; Galluccio, G.; et al. Sens. Actuators, B 2008, 130, 458−465. (21) Ohira, S. I.; Toda, K. Anal. Chim. Acta 2008, 619, 143−156. (22) Pinheiro, C.; Schäfer, T.; Crespo, J. G. Anal. Chem. 2005, 77, 4927−4935. (23) Barsan, N.; Weimar, U. J. Electroceram. 2001, 7, 143−167. 9781

dx.doi.org/10.1021/ac301687j | Anal. Chem. 2012, 84, 9774−9781