Biosensors and Bioelectronics 65 (2015) 121–138

Contents lists available at ScienceDirect

Biosensors and Bioelectronics journal homepage: www.elsevier.com/locate/bios

Breath sensors for lung cancer diagnosis Yekbun Adiguzel a,n, Haluk Kulah b,c a b c

Department of Biophysics, School of Medicine, Istanbul Kemerburgaz University, Mahmutbey Dilmenler Caddesi, No. 26, 34217 Bagcilar, Istanbul, Turkey METU-MEMS Research and Application Center, Middle East Technical University (METU), Ankara, Turkey METU BioMEMS, Electrical and Electronics Engineering Department, METU, Universiteler Mah., Dumlupınar Bulv. No. 1, 06800 Çankaya, Ankara, Turkey

art ic l e i nf o

a b s t r a c t

Article history: Received 26 June 2014 Received in revised form 9 October 2014 Accepted 10 October 2014 Available online 19 October 2014

The scope of the applications of breath sensors is abundant in disease diagnosis. Lung cancer diagnosis is a well-fitting health-related application of this technology, which is of utmost importance in the health sector, because lung cancer has the highest death rate among all cancer types, and it brings a high yearly global burden. The aim of this review is first to provide a rational basis for the development of breath sensors for lung cancer diagnostics from a historical perspective, which will facilitate the transfer of the idea into the rapidly evolving sensors field. Following examples with diagnostic applications include colorimetric, composite, carbon nanotube, gold nanoparticle-based, and surface acoustic wave sensor arrays. These select sensor applications are widened by the state-of-the-art developments in the sensors field. Coping with sampling sourced artifacts and cancer staging are among the debated topics, along with the other concerns like proteomics approaches and biomimetic media utilization, feature selection for data classification, and commercialization. & Elsevier B.V. All rights reserved.

Keywords: Breath sensor Breath analysis Electronic nose (e-nose) Lung cancer Disease diagnosis

Contents 1. 2. 3.

4.

5.

n

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Breath analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1. Exhaled breath content . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sensor-based diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1. Colorimetric-sensor array . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2. Carbon-polymer array . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3. Single-walled carbon nanotubes array . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4. Surface acoustic wave sensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5. Gold nanoparticle sensor-based array . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6. Metalloporphyrins-coated quartz microbalance array . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7. Optical fiber sensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Considerations for State of the Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1. Utilization of carbon nanotubes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.1. Utilization of functionalized surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2. Metal oxide semiconductor sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.1. Utilization of functionalized semiconductor metal oxide fibers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.2. Utilization of zeolites. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3. Utilization of gold nanorod-metalloporphyrins and pH sensitive dyes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Future perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1. Proteomics approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2. Utilization of biomimetic media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3. Sampling related issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4. Distinction of the cancer stages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5. Feature selection for data classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Corresponding author. E-mail addresses: [email protected] (Y. Adiguzel), [email protected] (H. Kulah).

http://dx.doi.org/10.1016/j.bios.2014.10.023 0956-5663/& Elsevier B.V. All rights reserved.

122 123 124 124 124 125 126 127 127 128 129 129 129 130 131 131 131 132 132 132 132 133 135 135

122

Y. Adiguzel, H. Kulah / Biosensors and Bioelectronics 65 (2015) 121–138

5.6. Commercialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6. Final remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7. Summary and conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1. Introduction Guidelines for early detection and prevention of cancer are intended to raise the awareness of the society, by providing information about a specific cancer, its symptoms and diagnosis, possible causes and prevention, and screening for early diagnosis. According to the American Cancer Society guidelines for the early detection of cancer (Smith et al., 2003a), screening for early detection of cancers favorably influences the survival of patients with cancer. The five-year survival rate of stage I and stage III lung cancer patients improved by 70% and 20%, respectively, by early detection and treatment (Horváth et al., 2009). Yet a distributable and effective screening method was not available at the beginning of this millennium (Bach et al., 2003; Machado et al., 2005; Saba and Kuhuri, 2005; Ghosal et al., 2009), which is still not the case in 2014. The mortalities due to lung cancer have not changed much in the last decades (Mutti, 2008). Cancer kills more than seven million people every year (Peng et al., 2010) and according to the World Health Organization, lung cancer is the leading cancer-related cause of death, killing about 1.3 million people every year (Horvath, 2010). According to a projection of global mortality in 2009, it is expected to emerge as the third and fifth leading cause of death in high and middle-income countries, respectively, by the year 2030 (). Early diagnosis of lung cancer is desirable, but conventional means like chest radiography, sputum cytology or computed tomography do not support the application of a wide-range population-based screening (Chan et al., 2009) and can have radiationexposure related side effects on health. On the other hand, techniques such as gas chromatography–mass spectrometry are expensive, not portable and require personnel with expertise. However, it was proposed as early as the end of 80s that breath may contain markers of lung cancer (Gordon et al., 1985). Diagnosis of lung cancer or any type of disease through a sensor-based approach would greatly improve the specificity of imaging procedures and other -invasive- confirmatory techniques, when early diagnosis is not particularly accessible by a single test and with the desired level of certainty. Electronic noses, which are aiming for lung cancer detection from the breath of lung cancer patients, are seemingly the best approach among the other sensor-based approaches since their use in lung cancer detection is primarily the adaptation and standardization of an already established practice in medicine. Accordingly, investigating exhaled air was recognized in the ancient times by the Greek physicians and used to diagnose diseases from the characteristic odor of the patients′ breath (Miekisch et al., 2004). Good clinicians knew about the sweet, fruity odor of acetone in patients with uncontrolled diabetes; the musty, fishy smell of advanced liver disease; the urine-like scent in adjunct to failing kidneys; and the rotten stink of a lung abscess (Cao and Duan, 2006). Around 400 BC, Hippocrates described fetor oris and fetor hepaticus in his discourse on breath aroma and disease; Lavoisier and Laplace in 1784 demonstrated respiration oxygen consumption and carbon dioxide elimination (Mashir and Dweik, 2009). Traditional Chinese medicine engaged in olfactory diagnosis as well (Wilson and Baietto, 2011). Nebelthau showed in the mid 1800s that diabetics emit breath acetone, and secluded ethanol was isolated from breath in 1874 by Anstie (Mashir and Dweik, 2009). A major breakthrough in the scientific study of

136 136 137 137

breath started in the 1970s with Linus Pauling (Mashir and Dweik, 2009). It was proposed as early as the end of the 80s that breath may contain markers of lung cancer (Gordon et al., 1985). An intriguing fact is that dogs are able to distinguish between breath samples of people with lung cancer, but dogs cannot be standardized (Canavan, 2013). Yet, this practice in medicine was considered old fashioned due to the advancement in instruments. Electronic noses (e-noses) re-popularize the olfactory diagnosis concept that was becoming old fashioned (Fig. 1) (Machado et al., 2005), by combining the instrumental advancements in the sensors field and the age-old breath analysis concepts. The breath analysis practice is based on olfactory perception. Electronic noses or rather breath sensors, are based on establishing a detection principle, which is analogous to the (human) olfactory system. The human nose does not smell diseases but the electronic noses aim to smell diseases, and they work similar to the biological noses. Biological noses are composed of olfactory receptor cells that each contain one type of odorant receptor and respond to a specific odor through the activation pattern of several receptors, each of which detects limited number of substances (Röck et al., 2008). In other words, a single olfactory neuron responds to different odors, hence is mildly selective; and at the same time, a pattern recognition process produced by multiple olfactory neurons identifies and classifies odors, like electronic noses that are based on analysis with semi-selective electronic sensor array (Oh et al., 2011). Thus, the non-selectivity issue, which is considered problematic in case of analytical chemical sensors, is actually a strategy that resembles natural olfactory perception, by recognizing the fingerprint pattern of odors, through a combination of “not-fully-selective” response patterns (Di Natale et al., 2007). Persaud and Dodd have constructed an electronic nose as early as the 80s, using semiconductor transducers and incorporating design features suggested by their proposal, and reported that their device could reproducibly discriminate between many odors (Persaud and Dodd, 1982). The design features of the model nose that was offered by Persaud and Dodd were analogous with the detection by broadly tuned receptor cells, organized in a convergent neuron pathway (Persaud and Dodd, 1982). By this means, complex odorant mixtures were suggested to be discriminated, without the need for highly specialized receptors. They assumed that there is no need for odor-specific transducers and the ratio of odors can be processed to identify odors. They built an electronic nose with semiconductor transducers, accordingly. Today, breath analysis for disease (lung cancer) diagnostics with electronic noses is considered in a similar manner, which is rather distinct from looking for a single biomarker within the breath of the patient. Electronic noses rely on arrays of chemical vapor sensors for stereochemical detection of odorant molecules that are named volatile organic compounds (VOCs) (Machado et al., 2005). Statistical or structural algorithms are used to discriminate and identify odorant samples (Machado et al., 2005). Several medical applications of electronic noses exist, like in the case of hemodialysis (Fend et al., 2004) liver failure (Wlodzimirow et al., 2014), cancer detection (Di Natale et al., 2003; Xu et al., 2013), bacterial pathogen discrimination (Pavlou et al., 2000), glucose monitoring for diabetics (Ping et al., 1997), and urinary tract infections (Pavlou et al., 2002). Among those, lung cancer detection by means of

Y. Adiguzel, H. Kulah / Biosensors and Bioelectronics 65 (2015) 121–138

123

Fig. 1. Analogy of e-nose and the human olfactory system. In the figure, small cubic dots that are dispersed around the sensor array represent individual volatiles. Semiselective olfactory receptor cells are represented here to form up an array that consists of n number of cells (n¼ 5, here). So, each rectangle in the array is a semi-selective olfactory receptor cell. Each receptor cell, namely the rectangle in the array, detects a limited number of volatiles and a specific response pattern is created by the array, due to the volatile content of the smell. In humans, this is achieved at the olfactory bulb, and in a sensor array, at the preprocessor. This recognized smell is identified in the olfactory cortex, through recognition of the specific pattern by cognitive processes. In case of sensors, the pattern recognition algorithm of choice serves this purpose.

electronic noses will be discussed in detail, following Section 2, which informs briefly about the concept, the disease related volatiles in breath, and the exhaled breath content.

2. Breath analysis Breath analysis is a sub-topic of the electronic nose concept. In addition to the advantages of breath analysis like safety, noninvasiveness, and easy accessibility, it is potentially applicable to a direct and real-time monitoring system, and a breath sample has a less complicated sample matrix than serum or urine (Peng et al., 2008). Similar to the odor-quality relationship in food (D’Amico et al., 2010), specific correlation of bodily odor composition with the disease is the underlying principle in breath sensors. In the seminal work of Pauling and co-workers, 250 substances from breath samples and 280 substances from urine vapor were analyzed from human subjects, by using temperature-programmed gas–liquid partition chromatography, after a defined diet to eliminate intestinal flora (Pauling et al., 1971). In 1999, variation of VOCs in the breath of 50 normal humans was investigated with gas chromatography and mass spectroscopy (GC–MS), in the C4– C20 range (Phillips et al., 1999a). It leads to the observation of 3481 different VOCs. Only 27 of these, which were probably sourced by

general metabolic processes, were common among all subjects. High quantitative variation among individuals persisted. However, a small variation was observed in the total number of breath VOCs, in a fairly narrow range of population (Phillips et al., 1999a). The list of relevant compounds for any disease in the breath sample defines a set of compounds at various concentrations, namely a chemical pattern, which can be detected by pattern recognition methods. Application of pattern recognition algorithm sorts a collection of multivariate data into classes, discriminating those pertaining to specific VOCs that are related to tumor diseases (Pennazza et al., 2011). For instance, exhaled volatiles′ comparison of the blood-sourced volatiles of 10 healthy controls and 10 stage I cancer patients revealed 23 compounds in the breath of lung cancer patients, with high levels of hexanal and heptanal (Deng et al., 2004). The same concept could be valid for issues such as rare cell detection from blood, based on, for instance, their elevated cell-surface receptors or other chemical signatures. In support, genetic mutations of lung cancer cells were profiled through their headspace VOCs, which were emitted from their membranes (Peled et al., 2013). In addition, cancerous cell surface proteins display a shifted pattern, in comparison to non-cancerous cell surface proteins (Hewett, 2001; Bao et al., 2012). Therefore, it is possible to approach cancer cell detection and discrimination

124

Y. Adiguzel, H. Kulah / Biosensors and Bioelectronics 65 (2015) 121–138

Table 1 Classes of volatile organic compounds in the breath, revised from the original table in Mazzone (2008). Additional sources are Miekisch et al. (2004) and Horváth et al. (2009). Class of the VOCs

Example of the VOCs

Saturated hydrocarbons

Ethane, pentane1, aldehydes Lipid peroxidation of fatty acid components of cell membranes—triggered by reactive oxygen species

Unsaturated hydrocarbons Oxygen-containing

Isoprene

Mevalonic pathway of cholesterol synthesis

Markers of oxidative stress, less amounts can be produced by protein oxidation and bacterial colonies, and excreted fast Produced from acetyl-coenzyme A

Acetone

Decarboxylation of acetoacetate from lipolysis or lipid peroxidation Incomplete metabolism of methionine

Elevated in liver failure and allograft rejection

Sulfur-containing Nitrogen-containing

Ethyl mercaptane, dimethylsulfide Dimethylamine, ammonia

Example of the production mechanism

Additional information

Elevated in liver impairment and uremia

issue, in a similar manner with the detection of lung cancer subjects, through their shifted breath VOCs′ patterns. 2.1. Exhaled breath content The majority of exhaled breath content is nitrogen, oxygen, carbon dioxide, water, and inert gases (Mazzone, 2008). There is also minor contribution of trace components. Sources of these volatiles are the atmospheric components that originate from many solvents and petroleum-based products, and the VOCs, which are generated as a result of the biochemical processes of the body. The classes of VOCs, which can be measured in the exhaled breath, are unsaturated and saturated hydrocarbons, together with oxygen-, sulfur-, and nitrogen-containing compounds (Table 1) (Mazzone, 2008). In addition to the information that is given in Table 1, the breath markers in certain diseases and applications can be found as a list in the work by Cao and Duan (2006) and further information is present in other works (Weitz et al., 1991; Philips et al., 1993, 1995, 1999b; Mendis et al., 1995; Narasimhan et al., 2001). Still, there is a more comprehensive review of the biomarkers in exhaled breath condensate by Grob et al. (2008). Besides, Haick and co-workers recently reviewed the origin and implementation of breath volatile cancer biomarkers (Haick et al., 2014), and previously, the same research group assessed the possible biochemical pathways of lung cancer related VOCs as well (Hakim et al., 2012). Former studies involving conventional techniques such as GC– MS are very sensitive and can measure specific VOCs, in addition to their concentrations (Mazzone, 2008). Yet, those are not ideal tests, because they are expensive, require expertise, and they need to collect and transport the breath contents to the devices (Mazzone, 2008). Mashir and Dweik made a review on the exhaled breath analysis, as an interface between medicine and engineering (Mashir and Dweik, 2009). They provided information about the identified markers in the exhaled breath, along with their chemical structure, associated medical conditions, and the techniques that were used to measure them. Among the techniques, GC–MS was presented as the dominating technique, alone or in combination with other techniques, such as automated thermal desorption, laser spectroscopy, selected ion flow tube mass spectrometry (SIFT-MS), and mid infrared spectroscopy (MIR). These other techniques like gas chromatography, H-breath test, and MIR were also utilized either alone or in combination. Among the identified biomarkers, 39 (hexane; heptane; octane; nonane; decane; undecane; 1-methyl-2-pentylcyclopropane; isoprene 2-methyl-1,3butadiene; 2-methylpentane; methylcyclopentane; cyclohexane; 2-methylheptane; 2,2,4,6,6-pentamethyl heptane; 2-methyloctane; 3-methyloctane; 3-methylnonane; 1-hexene; 1-heptene; 1-octene; benzene; trimethylbenzene; ethylbenzene; propylbenzene; toluene; o-toluidine 2-amino-1-methylbenzene; m-toluidine

3-amino-1-methylbenzene; aniline; styrene; xylenes; 2,3-dihy dro-1-phenyl-4(1H)-quinazoline; n-propanol; 2-propanol; formaldehyde; acetaldehyde; hexanal; heptanal; 1-phenyletanone; 2-butanone; and isopropyl myristate/tetradecanoic acid) were indicated to be associated with lung cancer (Mashir and Dweik, 2009). The compounds of interest in the breath are generally found to be in a 1–20 ppb range as distinctive mixture compositions whereas those compounds are elevated to a 10–100 ppb range in the breath of lung cancer patients (Liu et al., 2011).

3. Sensor-based diagnosis Gas sensing devices with sensor arrays detect and identify VOCs. Quartz microbalance sensors coated with metalloporphyrins, metal oxide semiconductors, polymer coated surface acoustic wave devices, conductive polymers, nonconductive polymer/carbon black composites, fluorescent dye/polymer systems, and chemoresponsive dyes are among the sensor arrays that have been utilized for this purpose (Mazzone, 2008). Some are mentioned below, entitled with their respective sensor systems. 3.1. Colorimetric-sensor array The colorimetric sensor arrays differentiate the analyte (s) through their composite responses, rather than one-by-one detection of the analyte(s) (Mazzone et al., 2012). Mazzone and coworkers described the diagnosis of lung cancer through the analysis of exhaled breath by using a colorimetric sensor array (Mazzone et al., 2007). The array was composed of 36 spots, each with a different chemically sensitive compound (Fig. 2), which responded to the reactive compounds with a color change.

Fig. 2. The colorimetric sensor array consisting of 36 chemically sensitive dots, impregnated on a disposable cartridge (Mazzone et al., 2007). Reprinted with permission from BMJ Publishing Group Ltd.

Y. Adiguzel, H. Kulah / Biosensors and Bioelectronics 65 (2015) 121–138

The sensitivity limits of the sensor for specific VOCs was in the ppm range at the lower parts, and in the ppb range at the upper parts. This study was carried out with 143 individuals. Among these, 21 were controls, 49 had non-small cell lung cancer, and the rest had one of the following diseases: chronic obstructive pulmonary disease, idiopatic pulmonary fibrosis, pulmonary arterial hypertension, and sarcodiosis. For the test, exhaled breath was drawn over the sensor array by a pump. The color of sensor spots changed when they responded to specific VOCs in the breath samples. Changes in color were analyzed after being dissected into their RGB components, which were further converted to numerical values. The prediction model was built by observing 70% of the subjects and testing the remaining of the subjects. Lung cancer prediction accuracy was moderate (73.30% sensitivity and 72.40% specificity), and lower than those reported previously (Di Natale et al., 2003; Machado et al., 2005). The authors attributed this to the superiority of the other sensor systems in discriminating the VOCs′ patterns, or to the differences in the sampling techniques or the tested populations (Mazzone et al., 2007). The Suslick group has contributed to the colorimetric sensing for a long time (Rakow and Suslick, 2000; Askim et al., 2013). The approach, which is described above, was improved by a joint effort (Mazzone et al., 2012). Accordingly, breath signatures were combined with clinical risk factors. It was previously expressed that the contribution of the patients′ clinical history to the outcomes of breath analysis, and carrying out studies with wide international subjects′ groups are important (Horvath, 2010). The subjects of this study were not of international origin. However, 92 lung cancer and 137 control subjects were used, and this was the

125

highest number among existing breath sensor studies for lung cancer detection. Four separate logistic prediction models were developed for data analysis. The sensitivity and specificity results of the most accurate models are presented in the footnote of Table 2, within Section 4, “Considerations for State of the Art”. 3.2. Carbon-polymer array Castro and co-workers reported an electronic nose design, consisting of five sensors that were made of hierarchically structured conductive polymer nanocomposites (CPC) (Castro et al., 2011). Five different polymer matrices and carbon nanotubes (CNT) lead to the conductive architecture of CPC, which helped to obtain the desired sensitivity and selectivity. One order of magnitude improvement in LOD could be observed, due to chemical specific interactions, depending on the selected polymer in the CPC (Kang et al., 2010). This CPC transducer was tested with nine VOCs that were chosen among the lung cancer biomarkers in breath (Castro et al., 2011). Principal component analysis (PCA) was used as the pattern recognition tool, to separate vapor clusters. The depiction of the measurement set-up and the chemoresistive response (relative amplitude) of the sensor array in the presence of isopropanol solvent vapor is shown in Fig. 3. Machado and co-workers previously hypothesized that the electronic nose could identify and discriminate between lung diseases, especially the bronchogenic carcinoma (Machado et al., 2005). The sensor was a handheld electronic nose, with 32 polymer composite sensors, and a sampling and data acquisition system, along with a processor. The subjects in the study had

Table 2 The studies of gaseous chemical sensing devices for the evaluation of breath VOCs in lung cancer, revised from the original table in Mazzone (2008) and D’Amico et al. (2010). Author

Yr

Cancer subjects Control subjects Sensor system

Di Natale et al. Chen et al. Machado et al. Mazzone et al. Peng et al. Dragonieri et al.

2003 2005 2005 2007 2008 2009

35 5 14 49 15 10

18 5 45 21 15 10

D’Amico et al. Mazzone et al. Santonico et al.

2010 28 2012 137 2012 20

36 92 10

Quartz crystal microbalance  (QCM) Surface acoustic wave Carbon-polymer array Colorimetric sensor array SWCNTs array Commercial n-composite array, Cyranose 320 Metalloporphyrins-coated QCM sensors Colorimetric sensor array Metalloporphyrins-coated QCM sensors

Model accuracy Validation accuracy or performance NA NA 71.60% 85.90% NA NA

100.0% (sens), 94.00% (sp) 80.00% 71.40% (sens), 91.90% (sp) 73.30% (sens), 72.40% (sp) Relative humidity dependent 85.00% and 90.00%a

NA NA 85.00%, 90.00%d

Three models were builtb Results differed with the modelc Results differed with the sampling approachd

Abbreviations: Yr stands for year, sens stands for sensitivity, sp stands for specificity, and n-composite stands for nano-composite. a The smell prints from non-small cell lung cancer (NSCLC) patients clustered distinctly from those of COPD subjects. Canonical discriminant analysis gave a crossvalidation value of 85% correct. The NSCLC patients were discriminated from healthy controls as well, and canonical discriminant analysis gave a cross-validation value of 90.00% correct. b Three models were built with the data, each aiming separately as follows: the 1st model aimed classification between lung cancer patients and controls; the 2nd model aimed discrimination of lung cancer patients from those with the other lung diseases; and the 3rd model aimed discrimination of lung cancer patients from both the patients with the other lung diseases and the control group. The cross-validated results of the 1st classification model provided 100.0% specificity and 85.00% sensitivity, that of the 2nd model provided 85.70% correct classification and 92.80% sensitivity for lung cancer diagnosis, and that of the 3rd model provided 79.30% correct classification and 89.30% sensitivity for lung cancer diagnosis. c Four separate logistic prediction models were developed for analysis. 72 sensor predictors (24 colorants  3 color values, namely red, blue, and green) were included in the 1st logistic regression model. 72 sensor predictors and 4 clinical predictors (smoking status, age, sex, COPD) were incorporated in the 2nd model; and in the 3rd model, 4 clinical predictors were forced into the model that contained selected variables among the 72 predictors. The 4th model incorporated those 4 clinical predictors. The sensitivity and specificity results of the most accurate models were reported. These are respectively as follows: 70.00% and 86.00%, according to model 2, for non-small cell cancer patients and control groups 80.00% and 86.00%, according to model 3, for adenocarcinoma patients and control groups 91.00% and 73.00%, according to model 3, for squamous cell cancer patients and control groups 90.00% and 83.00%, according to model 1, for adenocarcinoma and squamous cell cancer patients groups 89.00% and 85.00%, according to model 3, for small cell cancer patients and control groups 78.00% and 95.00%, according to model 3, for small cell and non-small cell cancer patients groups 81.00% and 73.00%, according to model 2, for stages I and II and stages III and IV cancer patients groups 70.00% and 86.00%, according to model 3, for patients with survival less than 12 months and survival more than 12 months groups Abbreviations: Yr stands for year, sens stands for sensitivity, sp stands for specificity, and n-composite stands for nano-composite. d The model that was cross-validated with the Leave-One-Out criterion gave the highest correct classification overall for both the endoscopic breath sampling data (90.00%) and the bag breath sampling data (85.00%). The performances of the diagnostic technique were as follows: 97.50% sensitivity and 75.00% specificity for the former, and 85.00% sensitivity and 85.00% specificity for the latter.

126

Y. Adiguzel, H. Kulah / Biosensors and Bioelectronics 65 (2015) 121–138

Fig. 3. Scheme of the chemo-resistive sensing device (a) and the chemo-resistive response (relative amplitude) of the sensor array in the presence of isopropanol solvent vapor (b) (Castro et al., 2011). Reprinted with permission from Elsevier.

relatively advanced disease, which lead to the need for the assessment of the sensor in a diverse population with patients at various stages of the disease. In addition, the predictive values, which were obtained from the validation patient group, may not be that suitable in other populations (Machado et al., 2005).

3.3. Single-walled carbon nanotubes array Peng and co-workers developed arrays of chemically sensitive, resistive vapor detectors (Peng et al., 2008). For this purpose, semiconductive random networks of single-walled carbon

Fig. 4. Principal components score plots of an array of 10 sensor films, upon exposure to simulated “healthy” and “cancerous” pattern at (a) 80% relative humidity (RH), (b) 10% RH, (c) 1% RH, and at (d) 80% RH (Peng et al., 2008). 50 times pre-concentration was applied. Reprinted with permission from Peng et al. (2008) Nano Letters 8:3631. Copyright 2008 American Chemical Society.

Y. Adiguzel, H. Kulah / Biosensors and Bioelectronics 65 (2015) 121–138

nanotubes (SWCNTs) were processed into arbitrary size, by using conventional microfabrication techniques, impinged on insulating nonpolymeric organic materials. Utilizing random networks circumvented the necessity of structure and position control, such as in case of working with single SWCNTs. This was made possible by the averaging of the properties, when employing randomly distributed SWCNTs. Further, applying SWCNTs provided more discriminatory power to construct sensors with desired capabilities because discrimination of closely related alkane species was possible in a mixture of such alkane species as well. The devices described above were used to detect lung cancer and to differentiate between the VOCs in the lung cancer patients′ breath, and in those of the healthy controls′. Briefly, the GC–MS measurements were first carried out for disease markers′ detection. Then, the simulation works were performed with the selected disease-marker VOCs (trimethyl benzene, styrene, decane, octane, and 1-hexene) that were present in most lung cancer (real) samples, with the highest concentrations. Next, an array of highly discriminative adsorption films was prepared with 10 different nonpolymeric organic adsorption materials. These adsorption materials had different chain lengths and branching, aromatic configurations, and functional groups in order to increase the interaction of the VOCs with the adsorption material. The simulated results of pre-concentrated healthy individuals′ and lung-cancer patients′ breath analysis revealed discrimination as pattern clusters among the two groups upon decreasing humidity conditions because water molecules have high affinity towards SWCNTs (Fig. 4). The analysis of real breath sample measurements revealed only a partial overlap within the control groups′ cluster patterns, without any overlap among the results of controls and lung-cancer patients. Variation within the same healthy control groups′ cluster patterns was mostly due to differences in diets, genetic backgrounds, and metabolic states among the individuals of the same group. The differential responses of the specific sensor sections to certain biomarker VOCs were observed. This was attributed to charge transfer from adsorbed species to the SWCNTs, modifications of contact work functions, and carrier scattering by adsorbed species. Different from a “lock-and-key” approach, obtaining a fingerprint of each VOC biomarker with a cross-reactive sensor array brings about the possibility of a number of VOCs in a complex multicomponent (bio)chemical media to be detected at once. 3.4. Surface acoustic wave sensor For artificial olfaction, the most suitable choices among the available gas-sensing technologies are acoustic wave sensors, including quartz crystal microbalance (QCM) gas/odor sensors (Wyszynski and Nakamoto, 2009). Wang and co-workers described a surface acoustic wave (SAW) gas sensor with a gas chromatography interface (Wang et al., 2008). It was used to detect lung cancer specific breath VOCs. SAW sensor was preferred for its higher sensitivity, which was further enhanced by the polymer film technology. Before measurement, respiratory gas was enriched by an adsorption tube, and desorbed at the inlet of gas chromatography capillary by elevated temperature (Wang and Liu, 2011a). VOCs were then moved along with the carrier gas into the capillary, and separated there. These VOCs were detected by means of condensing on the SAW surface. The detection time was related to the VOCs′ physical and chemical characteristics since their passage through the gas chromatography capillary was accelerated or decelerated accordingly. Thus, a gas chromatography capillary coupled SAW sensor became a virtual sensor array with numerous orthogonal (non-overlapping) sensors. Such a system can distinguish hundreds of different gases. During breath testing, data

127

analysis was performed after data collection and signal processing for pattern recognition. Back propagation (BP) artificial neural network (ANN) algorithm combined with imaging was used for this purpose (Wang et al., 2008). 15 lung cancer patients were detected by comparing their breath sample results with those of the 10 healthy individuals and 7 chronic bronchitis patients. A content result was obtained by using 4 select lung cancer patients and healthy controls for BP-ANN validation. Specific VOCs that are exhaled by lung cancer cells in the microenvironment were shown to be the source of lung cancer biomarkers. This early e-nose prototype was relying on a virtual array of SAW sensors (Chen et al., 2005). The earliest version of this sensor was a polymer coated SAW sensor with frequency change response, used along with data processing with ANN algorithm (Yu et al., 2003). The following lung cancer biomarker VOCs were detected: styrene, decane, undecane, isoprene, benzene, 1-hexene, hexanal, propyl benzene, and 1,2,4-trimethyl benzene. 3.5. Gold nanoparticle sensor-based array Peng and co-workers reported that 5 nm gold nanoparticle (GNP) sensor could differentiate between the healthy controls and patients suffering from lung, breast, colorectal, and prostate cancers (Peng et al., 2009, 2010). This was achieved regardless of the variables other than the subjects′ health status, and in a comparable manner with GC–MS. The sensor was a tailor-made array of cross-reactive sensors that were based on organically functionalized GNPs (Peng et al., 2010). The GNPs were serving to connect electrodes and to form paths between them. At the same time, the surfaces of GNPs were the adsorption sites for analytes. Following sample collection, 14 cross-reactive nanosensors with different functionalities were designed and tested for feasibility of getting odor fingerprints instead of detecting individual constituents. This process was rather obtaining a fingerprint spectrum of a complex biomolecule. Hence, the approach was termed as a black-box approach. Typical cancer biomarker detection limits of the GNP sensors were roughly ranging from 1–5 ppb to 10 ppt. During measurement, each sensor of the array underwent a reversible change in electrical resistance upon exposure to the sample breath. PCA plots of the sensor array′s resistance responses of lung cancer patients′ and healthy controls′ measurements revealed a very good separation in the principal component space between the patterns of controls and patients (Fig. 5A). Even, the four distinct cancer types were distinguished in the same principal component plot, with only minimal overlap between the prostate cancer, breast cancer and healthy clusters (Fig. 5E). This overlap could represent limitation of the approach in applicability of breath sensors to all cancer types or several cancer types at once and with similarly pronounced detection capabilities. Still, lung cancer and colon cancer are well distinguished with no overlap of the results of healthy group and cancer patients′ groups (Fig. 5E). Although the investigation was defined as a proof of concept study due to lack of a larger size of population, it was a major step towards the development of a robust breath test for cancer. The authors were suggesting the improvement of sensor array while maintaining the minimum sensitivity to potential confounding effects of common metabolites, against the possible increase in the overlaps in a prospective study with a larger-sized population. GC– MS/solid-phase microextraction analysis was also performed in this study. Contrary to the results of the GNP array, sensitivity to colon cancer was slightly lower in the GC–MS/solid-phase microextraction analysis. In this work, some of those VOCs for diagnosing the disease of interest were common among more than one disease. Namely, toluene was common among the lung and prostate cancer tests; 3,3-dimethyl pentane was common among the lung and breast

128

Y. Adiguzel, H. Kulah / Biosensors and Bioelectronics 65 (2015) 121–138

Fig. 5. The PCA plots of the GNP sensor array's resistance responses for (A) lung cancer (LC) patients and healthy controls, (B) colon cancer (CC) patients and healthy controls, (C) breast cancer (BC) patients and healthy controls, (D) prostate cancer (PC) patients and healthy controls, and (E) all of the four types of cancer patients and healthy controls (Peng et al., 2010). (Original image is slightly modified, for paper formatting.) Reprinted with permission from Nature Publishing Group.

cancer tests; and 1,1′-(1-butenylidene)bis benzene was common among the lung and colon cancer tests. On the other hand, 2-amino-5-isopropyl-8-methyl-L-azulenecarbonitrile was common among the prostate, breast, and colon cancer tests. The rest of the 6 common VOCs among the 33 VOCs for lung cancer, which were selected to be inspected, were 1-methyl-4-(1-methylethyl)benzene, dodecane, and 2,3,4-trimethyl hexane. The rest of the 6 common VOCs among the 39 VOCs for colon cancer, which were selected to be inspected, were 1,3-dimethyl benzene, 1-iodo nonane, [(1,1-dimethylethyl)thio] acetic acid, and 4-(4-propylcyclohexyl)-4′-cyano[1,1′biphenyl]4-yl ester benzoic acid. The rest of the 5 common VOCs among the 54 VOCs for breast cancer, which were selected to be inspected, were 5-(2-methylpropyl)nonane, 2,3,4-trimethyl decane, and 6-ethyl-3-octyl ester 2-trifluoromethyl benzoic acid. The rest of the 4 common VOCs among the 36 VOCs for prostate cancer, which were selected to be inspected, were p-xylene and 2,2-dimethyl decane. Except for 1-methyl-4-(1-methylethyl)benzene and toluene, all these VOCs were used for the first time. It was also indicated that a limited number of and the most suitable VOCs were selected to be inspected during the tests despite considerably higher number of VOCs' being specific for the respective diseases. This was due to the large experimental scatter, and some of these VOCs were occurring in very distinctive concentration ratios. The latter situation

was making those VOCs potential biomarkers that needs to be examined through the overall change in their compositions. 3.6. Metalloporphyrins-coated quartz microbalance array D′Amico and co-workers studied differentiation of lung cancer from other lung diseases and control groups with electronic nose (D’Amico et al., 2010). The gas sensor array of the artificial olfaction system was based on eight quartz microbalance sensors coated with high sensitivity bearing molecular films of metalloporphyrins. In this study, apart from those of lung cancer patients', artificial lung cancer breaths were generated by adding three alleged volatile markers of tumor to the breath samples of controls. For the marker analysis, VOCs were added in mixture to breath samples collected from 2 control individuals, according to a method optimized for halitosis studies. Eventually, eight samples were generated for each individual. Samples were then analyzed with the gas sensor array following the same experimental procedure that was used for real breath samples. The sensor responses from the artificially created patient group drifted significantly towards those obtained in case of the lung cancer patients. Despite the fact that a correlation was demonstrated, it was indicated that alterations in the breath of lung cancer patients are complex and addition of few VOCs would not turn a normal breath to that of a lung cancer patient. In addition, a new method for the segregation of the deepest portion of a single breath was presented with good results in terms of rejection of the exogenous compounds and probably of the compounds sourced from the upper respiratory tract. Yet, it was not possible to ascertain long-term effects of chemotherapy on the discrimination power of the sensor array. The most meaningful results in lung cancer diagnosis with gas sensor arrays for the years between 2003 and 2009 were presented through this study. The same group improved their results by developing the sampling method through the use of an endoscopic probe that enabled sampling near the tumor mass (Santonico et al., 2012). By

Y. Adiguzel, H. Kulah / Biosensors and Bioelectronics 65 (2015) 121–138

129

Fig. 6. The schematic configuration of the analytical apparatus for determination of VOCs in exhaled air (Silva et al., 2011). The inset figures show (a) the SPME sampling procedure used for the GC–MS breath analysis reference method and (b) the optical microscopy (OM) and scanning electron microscopy (SEM) images of the sensor head. (OF – optical fiber, SB – sampling bulb, v – valve, SPME – solid phase microextraction) Reprinted with permission from Elsevier.

this approach, 75% correct classification among the adenocarcinoma and squamous cell carcinoma subtypes of lung cancer was demonstrated for the first time. 3.7. Optical fiber sensor Silva and co-workers developed an optical fiber sensor for clinical diagnosis (Silva et al., 2011). The aim was to determine eight volatile organic compounds in human breath. Schematic configuration of the analytical apparatus is illustrated in Fig. 6. Variations in optical power (in dB) were detected as a result of changes induced by the analyte molecules in the refractive index of the sensitive polymeric film. These changes in the reflected optical power were reversible and proportional to the amount of analyte at the recognition surface. In this study, selected VOCs (ethane, pentane, heptane, octane, decane, benzene, toluene, and styrene) were breath biomarkers of several distinct disorders including lung cancer. The detection limits of the analytical system were ranging from 0.8 pmol/L (for heptane) to 9.5 pmol/L (for decane). The intriguing fact was the comparability of analytical performance of the sensor to breath analysis by GC–MS, the reference method. Apart from the remote measuring capabilities by optical fibers, the other advantages of this sensor system was unmatched sensitivity, non-invasiveness, and multi-analyte detection of chemical, biological, and clinical relevance. Moreover, analysis time was 10 min. The device was exhibiting adequate accuracy and a high statistical degree of linearity for the calibration model and VOCs′ detection. Besides, humidified mixtures were having identical retention times with the dry standard samples, without any peak amplitude variation in the humidified mixture.

reported for sensor-based analyte detection. In clinical studies of lung cancer diagnosis, sensitivity is the probability that a lung cancer patient will be tested as positive and specificity is the probability that a healthy person will be tested as negative. Deviations from the gold standards of 100.0% sensitivity and 100.0% specificity values are errors. These include the lung cancer patients who were tested as negative (false negatives) and the healthy individuals who were tested as lung cancer (false positives). The 100.0% sensitivity and 94.00% specificity values (Table 2) obtained by Di Natale et al. (2003) demonstrate the potential of lung cancer diagnosis by breath sensors. This result is even better than a recent lung screening trial that resulted in 93.80% sensitivity and 73.40% specificity for low-dose computed tomography (Moyer, 2014). Still, 94.00% specificity indicates that 6 of the healthy individuals among a total number of 100 healthy subjects were tested as positive (false positive), namely diagnosed as lung cancer. Normally, when there is a positive test result with conventional diagnostic techniques, a physician asks for further confirmation of the disease by additional tests especially when the symptoms are insufficient and the therapy for the diagnosed disease requires burdensome surgical operations, radiotherapy, or chemotherapy. Besides, in case of studies on developing diagnostic tests, false positive results and false negative results bear the risk of unrevealing the true potential of the approach. Repetition tests with false negative and false positive results could be a means to overcome this problem since consistently obtaining false negative and false positive results for a subject would be pointing at some unconventional situations that may require him/her to be excluded from the study group. To the best of our knowledge, there are currently no studies that are applying such a strategy. 4.1. Utilization of carbon nanotubes

4. Considerations for State of the Art State of the Art of studies for lung cancer diagnosis with breath sensors is summarized in Table 2. Future trends are outlined afterwards, starting with the utilization of carbon nanotubes. Sensitivity and specificity values are important for evaluating the results of clinical studies. In addition, these terms need to be differentiated from the sensitivity and specificity values that are

Chen and co-workers reported a breath sensor with carbon nanotubes, operated by field effect of polarization and ionization (Chen et al., 2010). The sensor utilized the properties such as small tip ratio, high aspect ratio and electrical conductivity of the carbon nanotubes. It had high sensitivity in detecting faint breath, without any recovery issue like in case of adsorption type sensors. It was fabricated with micro-machining technologies in order to deal with the stochastic nature of the multi-walled carbon nanotubes′

130

Y. Adiguzel, H. Kulah / Biosensors and Bioelectronics 65 (2015) 121–138

distribution and inefficient bonding to the microelectrodes. It functioned under alternative current. The information collected from the sensor was frequency and strength of the exhaled breath since health condition is reflected also by the physical characteristics such as respiration frequency and the tidal volume of the exhaled breath. The sensor was operating steadily despite the interference of ambient air flow and temperature. The fabrication of the devices with multi-walled carbon nanotube (MWCNTs) is briefly as follows: After generating the main device structure with classic microfabrication techniques, electrophoresis was employed as a novel processes in order to coat the interdigitated electrodes of the device with MWCNTs. This was carried out in acetone that contained Mg(NO3)2 (as an additive) and MWCNTs (5 mg/L, 495% pure), and under a constant electric field of 10 V/cm for a duration of 5 min. This was followed by electroplating a thinner nickel film on the MWCNTs to strengthen the bonding of MWCNTs on the underlying nickel, lying on top of Cr/Cu. According to a previous impedance spectroscopy study that utilized MWCNTs as the sensing elements of capacitance and resistance based sensors, reducing gases are chemisorbed on the surface of MWCNTs (Varghese et al., 2001). Moreover, increasing impedances with elevated humidity levels and partial pressures of carbon dioxide, carbon monoxide, and ammonia were attributed to p-type conductivity in semiconducting MWCNTs. It was also attributed to the formation of Schottky barriers in between the metallic and semi-conducting nanotubes. Humidity, carbon dioxide, and carbon monoxide responses were reversible. Methanol, chloroform, and tetrahydrofuran could be distinguished from the lung cancer related VOC, toluene, by using MWCNT transducers (Feller et al., 2014). The amplitudes of MWCNT-based transducers were changing with regard to the type of VOCs and the sensitivities were increasing with the polarity of the target molecules. Last, the carbon nanotube chemoresistive sensors do not need to operate at several hundred degrees Celsius, like in case of metal oxide sensors.

4.1.1. Utilization of functionalized surfaces The selectivity of carbon nanotubes (CNTs) for specific gaseous molecules can be attained or improved by functionalization of the CNTs. The sensitivity of functionalized CNTs with respect to the type of functional groups and length of the CNTs was reported as follows: COOH4 NH2 4OH 4short CNTs 4long CNTs (Feller et al., 2014). This trend was more pronounced with increase in the polarity of the molecules, and vice versa. Polymer dispersions of CNTs lead to enhancement in the amplitudes. This increase was ranging by a factor of 5 to 100 respectively for methanol and chloroform. However, functionalization of the CNTs with the COOH group was found to counteract this by penalizing the CNTs' dispersion. Additionally, the average gap between the CNTs can be used to tune selectivity. Layer-by-layer deposition of the CNT-polymer composites can be achieved by spraying. As a result, random networks of CNTs that are prepared by this way become less conductive but more sensitive than those prepared by membrane filtration under vacuum. In case of CNTs, sensor resistances change dramatically by just a small number of VOCs′ binding to the surface. This is observed when CNT/CNT junctions' gaps do not differ much more than 10 nm in average. It is due to perturbation of electrons' motion in the CNT random network through the changes in the gaps of CNT/CNT junctions. The network disconnections promote tunneling conduction and loss of ohmic conduction. This is the basis of so-called Quantum Resistive Sensors. The sensors were prepared by changing the polymer matrix, in which the CNTs were dispersed. As a result, 7 polar (water, ethanol, methanol, acetone, propanol, isopropanol, and 2-butanone) and 11 non-polar (chloroform, toluene, benzene, styrene, cyclohexane, o-xylene, n-propane, n-decane, 1,2,4-trimethyl benzene, isoprene, and 1-hexene) VOCs of lung cancer were detected (Chatterjee et al., 2013). The response was at the ppm level and the response time was a couple of seconds. The information given above was obtained through the studies with MWCNTs. The functionalization adds considerably to the responses of SWCNTs as well (Liu et al., 2011). Liu and co-workers studied detection of both polar and nonpolar VOCs by using SWCNTs functionalized with organic materials (Liu et al., 2011). The schematic of the test device and illustration of the

Fig. 7. The schematic of the test device (a, b, and c) and functionalization process of SWCNTs (denoted as SWNTs, on the figure) are shown along with the molecular structures of tricosane and pentadecane (on the left) (Liu et al., 2011). The interdigitated electrode coated with SWCNTs is illustrated in (a), cross-section view of the electrode is shown in (b), and ichnography of the test device is shown in (c). The original version of this image is slightly modified in order to improve the legibility. DMF stands for dimethylformamide and ID stands for interdigitated. Reprinted with permission from Elsevier.

Y. Adiguzel, H. Kulah / Biosensors and Bioelectronics 65 (2015) 121–138

functionalization process are shown in Fig. 7. Two nonpolymeric organic materials (tricosane and pentadecane) were coated on the SWCNTs to prepare a conductometric type sensor. The device was then exposed separately to 1,2,4-trimethybenzene and decane respectively as the polar and nonpolar VOCs. The resistance changes of the biosensor elevated by about 5–7% for 1,2,4-trimethybenzene and 2–3% for decane as a result of the SWCNTs′ surface modifications. The resistance of functionalized SWCNTs had a noticeable change in the first 10 minutes after exposure to the VOCs and stabilized in 30 mins. The results revealed pronounced sensitivity to polar VOCs, by tricosane-functionalized SWCNTs.

4.2. Metal oxide semiconductor sensors An electronic nose array named CN e-Nose II was designed in the Zhejiang University (Wang and Liu, 2011b). This gas sensor array included 5 TGS and 3 MQ metal oxide semiconductor sensors, and an NE-NH3 electrochemical sensor for detecting the digestive tract sourced VOCs and trace amounts of ammonia. The detection range for the lung cancer biomarker toluene was 1– 30 ppm when the TGS2602 model sensor was used. The other VOCs that were detected are H2S, NH3, ethanol, and H2. The other sensor models that were utilized are TGS813 (used for H2, CO, CH4, isobutene, and ethanol detection) TGS822 (used for CH4, CO, isobutene, ethanol, acetone, n-hexane, and benzene detection) TGS2600 (used for H2, CO, CH4, isobutene, and ethanol detection) TGS2620 (used for H2, CO, methane, isobutene, and ethanol detection) MQ-2 (used for H2, CH4, ethanol, and butane detection) MQ-3 (used for CH4, ethanol, n-hexane, benzene, and liquefied petroleum gas detection) MQ-2 (used for butane, propane, liquefied petroleum gas, and liquefied natural gas detection)

131

The lowest detection range among all the VOCs was 0.1–3 ppm, obtained for H2S. As indicated above, the 813 and 2600 model TGS metal oxide sensors were both used to detect the same VOCs. Their detection ranges were different. It was 500–10,000 ppm for the model TGS813 and 1–100 ppm for the model TGS2600. The latter model was probably more accurate at the given range. 4.2.1. Utilization of functionalized semiconductor metal oxide fibers Structures with high crystallinity such as nanowires exhibit greater stability and can overcome problems like irreversibility (Hill and Binions, 2012). Irreversibility is commonly associated with elevated working temperatures and observed therefore in case of metal oxide sensors. Still, semiconductor metal oxide sensors are advantageous among most of the other breath sensor types, due to their high gas responses, response speeds, stabilities, and lower costs (Kim et al., 2014). Kim and co-workers obtained superior toluene response (Rair/Rgas ¼5.5 at 1 ppm) and crosssensitivity against H2S (Rair/Rgas ¼1.36 at 1 ppm) by using a functionalized semiconductor metal oxide sensor (Kim et al., 2014). It was based on Pd catalysts inside and/or outside of semiconducting metal oxide WO3 nanofibers. Nanofibers were formed by electrospinning. Decoration with Pd was achieved by treating the fibers with Pd nanoparticles after obtaining the nanofibers. The alternative means of decoration was adding PdCl2 to the WO3 nanofiber precursor solution. Applying both types of Pd decoration simultaneously revealed to be the best with a 20 ppb detection limit (at 350 °C) of the lung cancer biomarker toluene (Rair/Rgas ¼ 1.32) (Fig. 8). 4.2.2. Utilization of zeolites The use of zeolites is another means to modify metal oxide sensors for superior responses. Zeolites are a framework of tetrahedral TO4 building units (T ¼Si, Al, etc.) that link with each other by sharing oxygen atoms to form 3D crystalline porous skeletons (Xu et al., 2006). Zeolites form cages and channels of different window sizes due to the T–O–T links that have a variety of rings. This results in a molecular sieve effect since some molecules are allowed to pass through and the others are excluded. In addition,

Fig. 8. The cyclic response levels of pristine WO3 nanofibers (NFs) and Pd-NPs/Pd-embedded WO3 NFs for low concentration of toluene at 350 °C and relative humidity 90% atmosphere (Kim et al., 2014). The inset highlights the concentration dependent response of pristine WO3 NFs and Pd-NPs/Pd-embedded WO3 NFs. Detection limit is 20 ppb with a response (Rair/Rgas) of 1.32. Reprinted with permission from Elsevier.

132

Y. Adiguzel, H. Kulah / Biosensors and Bioelectronics 65 (2015) 121–138

zeolites can be modified in terms of pore size, ion conductivity, ion-exchange with different cations, and adsorption and catalytic selectivity. Based on their earlier studies (Binions et al., 2009a), Hill and Binions stated that incorporation of zeolites onto the surface of a porous metal-oxide sensor (MOS) improves the selectivity and sensitivity (Hill and Binions, 2012). Enhanced response and discriminating power were achieved with a solid state MOS sensor composed of the hydrogen variants of zeolites as transformation element, overlaid on top of the porous chromium titanium oxide sensing element (Binions et al., 2009b). 4.3. Utilization of gold nanorod-metalloporphyrins and pH sensitive dyes Huo and co-workers designed an optical chemical sensor with gold nanorod-metalloporphyrins and pH sensitive dyes, spotted on porous hydrophobic membranes (Huo et al., 2014). The use of gold nanorods as ligands prevented photochemical degradation of VOCidentifier metalloporphyrins. This sensor was aimed to be used for identifying lung cancer biomarkers in breath. Color difference maps of six VOCs (decane, undecane, hexanal, heptanal, benzene, and 1,2,4-trimethylbenzene) were obtained through generating RGB difference maps of the digital images of the sensor responses to these VOCs. Each VOC was ending up in a unique color difference map and different color change patterns were observed at varying concentrations of VOCs (Fig. 9). Final analysis was performed through the PCA and HCA.

5. Future perspectives 5.1. Proteomics approaches Exhaled breath condensate (EBC) contains nonvolatile substances and proteins as well (Conrad et al., 2008). Specific EBC proteins were studied as potential biomarkers in lung cancer. The application of EBC proteomics to lung cancer diagnostics is missing

(Conrad et al., 2008; Chan et al., 2009). Nonetheless, the EBC biomarker proteins have been established. According to a study by Cheng and co-workers, endothelin-1, tumor necrosis factor-α, and interleukin-6 were elevated in the EBC of non-small cell lung cancer patients (Cheng et al., 2011). However, sensitivity and specificity were needed to be improved for clinical use such as early diagnosis of lung cancer. The major obstacle is the scarcity of the EBC proteins (Conrad et al., 2008). This can prevent reliable assessment of the patient′s condition. Still, ongoing studies are promising. The comparative proteomics analysis of EBC was performed and 29 unique proteins were specific to lung cancer, compared to those found in the control samples (Cheng et al., 2011). This data needs to be confirmed by a study with higher number of patients. 5.2. Utilization of biomimetic media The human sense of smell has an olfactory stimuli discrimination power over 1 trillion (Bushdid et al., 2014). This is enabled by the mammalian olfactory system that has approximately 1000 different olfactory genes (Albert et al., 2000). However, olfactory skills differ among the species to a great extent. For instance, olfactory skill of a well-trained dog has 106 times superior sense of smell than that enabled by the human nose (Wilson and Baietto, 2011). This is partly because a canine has over 100 million olfactory cells (Albert et al., 2000). Sniffer dogs were used to test the presence of lung cancer from the breath samples of volunteers including healthy individuals, lung cancer patients, and COPD patients (Ehmann et al., 2011). Lung cancer was identified with 71.00% sensitivity and 93.00% specificity. In an earlier study for lung cancer diagnosis with sniffer dogs, overall sensitivity was 0.99 (95% confidence interval, 0.99, 1.00) in comparison to biopsyconfirmed conventional diagnosis and specificity was 0.99 (95% confidence interval, 0.96, 1.00) (McCulloch et al., 2006). Heterogeneity of performances across studies and within studies are probably due to the distinctions between the capabilities that are somewhat sourced by genetic characteristics of the dogs and the training methodologies (Lippi and Cervellin, 2011). The route to

Fig. 9. The color difference maps of six different VOCs at 0.75 ml, 1.5 ml and 3 ml saturated vapors of those VOCs at 20 °C (Huo et al., 2014). The RGB values are shown four times larger in these images than the initial difference values. Reprinted with permission from Elsevier. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Y. Adiguzel, H. Kulah / Biosensors and Bioelectronics 65 (2015) 121–138

combine this concept with the State of the Art sensor studies involves exploring biomimicry materials for gas sensing (Wu et al., 2001; Du et al., 2013a). This covers expression of olfactory receptors in cell systems (Ko and Park, 2005), establishing olfactory sensory neuron-based biosensors (Du et al., 2013b), immobilizing odorant-binding proteins on biosensors (Di Pietrantonio et al., 2013) and on nanotubes as the sensor surfaces with improved sensitivities (McEuen, 1998). The field is promising and many interdisciplinary groups have been putting an invaluable effort in this field with a considerable progress (McEuen, 1998). The utilization of biomimicry materials in breath sensors is well represented by the immobilization of olfactory receptors of different species on the sensing region of the devices to serve as odorant binding elements. Different methods like Quartz Crystal Microbalance and Surface Plasmon Resonance can be used to recognize the binding events of odorants with the receptors on the sensor surface. To be more specific, odorant receptors can for example be expressed in cells (Ko and Park, 2005; Sung et al., 2006) and membrane fraction of these cells can be applied to the sensor surface (Sung et al., 2006). This approach will be less cumbersome than isolating the olfactory receptor proteins that normally reside on the membrane without significant loss in the functionality of the protein of interest in order to use it for odor detection afterwards. The detection range of the odorant of interest can be as low as 10–12 M with this approach (Sung et al., 2006). In accordance, the detection limit was found to be 10–12 M when the same receptor protein (ODR-10) was expressed in MCF-7 cancer cell line and odorant binding was detected by Surface Acoustic Wave measurement (Wu et al., 2011). Odorant receptor cells can be preferred to be immobilized on the sensor surface instead of the receptor proteins. In this case, calcium imaging can be used to detect odorant binding since binding of odorants to their receptors on the odorant receptor cells results in calcium influx and membrane polarization (Lee et al., 2009). Accordingly, microfabricated planar electrodes were used as the cell immobilization media in one study (Lee et al., 2009). The field potential after odorant binding to the receptor channels was 10 mV. Engineered cells with increased number of odorant receptor channels proteins were produced in this work and the measurement was essentially electrophysiological. One adverse effect was the observation of diminished sensitivity in cell systems, probably due to cells′ intrinsic transduction cascades (Dacres et al., 2011). Besides, sustaining culture conditions can be required or using special buffer media can be necessitated. These will be problematic in some applications for the requirement of extra work and the addition of delicacy to the system. In addition, odorants would need to be dissolved in such systems. This will also bring complications when dealing with hydrophobic target volatiles. Moreover, attaining proper orientation of the protein on the sensor surface can be another issue of concern (Wu et al., 2011). Still, other than due to its possible superior functionality in living beings, this approach is obviously advantageous to the nonbiomimetic sensor based approaches due to being non-toxic and potentially biocompatible. Olfactory receptors are promising for further advancements in the bio-sensing field as already being crucial parts of olfactory sensory systems of living beings. One last interesting application that will be mentioned exploited protein engineering. The odorant binding protein ODR10 protein was engineered to constitute both the green fluorescent protein (GFP) and the Renilla luciferase protein (RLuc) at close vicinity and away from the odorant binding site (Dacres et al., 2011). Bioluminescence Energy Transfer (BRET) occurred by transfer of the energy emitted by RLuc to GFP. The binding of the odorant prevented BRET and thus recognized. The principle is briefly as follows: contrary to common approaches, energy emission by RLuc was not originated from optical excitation but was

133

initiated by the addition of the bioluminescent substrate of RLuc. The addition of substrate thus initiated BRET. Yet, BRET was ceased by odorant binding since ODR-10 underwent conformational change by odorant binding that lead to separation of GFP and RLuc. Accordingly, the response of bioengineered ODR-10 to the odorant (diacetyl) was linear in a range spanning 10–19 M to 10–10 M (Dacres et al., 2011). This work is developed by CSIRO as a commercial product named Cybernoses. 5.3. Sampling related issues Sampling is a critical issue in terms of the results′ quality and measurements′ success. Atmospheric contaminants, patient preparation, and sample collection are among the issues of concern, for disease diagnosis by breath testing. According to a review by Miekisch and co-workers, endogeneous VOCs in alveolar air are 2–3 folds higher than those in expiratory breath sample (Miekisch et al., 2004). The “dead space gas” is used to depict the excess volume of air that is inhaled but not mixed with blood. So, “alveolar breath” is the exhaled air that is in equilibrium with systemic blood and “end-tidal air” is the last fraction of the expired air (Lourenço and Turner, 2014). The term “end-expired air” is also preferred in order to overrule the expectations that they are identical. Miekisch and co-workers called the dead space gas as the nonalveolar (airways) part of breath (Miekisch et al., 2008). It dilutes the VOCs of interest. In order to overcome the adverse effects of the dead volume gas, one approach is to make the patients breath pure air for a certain amount of time before sampling (Miekisch et al., 2004). This is suggested to be effective but is also burdensome and impractical. An alternative method is measuring the VOCs that persist in the ambient atmosphere and subtracting the result of that measurement from the result of sample measurement result. Concentration of the VOCs in breath minus concentration of the VOCs in air is termed as alveolar gradient (Philips, 1997). This approach is actually complex and not straightforward since alveolar and dead space gas ratios are not constant due to variations such as altering breathing patterns (Miekisch et al., 2004). The exhaled VOCs are prone to change due to variations in the cardiac output, ventilation rate, and ventilation/perfusion ratio, especially when the subject has severe hemodynamic or pulmonary dysfunction. On the other hand, the magnitude of such variations are negligible under normal physiological conditions. It is intriguing that exposure to environmental VOCs may have sustained effect on the exhaled VOCs' pattern since the volatiles that are already absorbed from the ambient air only enter the blood at the time of sampling (Horváth et al., 2009). Therefore, it will take time for those volatiles to be released back within the expired air. However, none of the current sampling approaches consider it. This issue can be addressed by the follow-up studies of serial measurements with subjects inhaling VOC-free gases. This is not a demanding task since sampling of the breath can be performed with high frequencies. Standardizations are required at all level. A relevant concern is normalization. Accordingly, data normalization by taking the ratio of pCO2 (partial pressure of carbon dioxide) in the sample and the end tidal pCO2 should be performed especially when the mixed expiratory samples that include the dead volume gas are used for the measurement instead of alveolar air. The dilution level of the alveolar volume can be judged by this means. The data can be corrected afterwards if there is a patho-physiological situation because the end tidal pCO2 would be changing instead of remaining constant as expected. Additionally, to eliminate the interindividual variation, Miekisch and co-workers measured pCO2 in ten samples that were collected from a single subject in one of their later study (Miekisch et al., 2008). In this work, sampling

134

Y. Adiguzel, H. Kulah / Biosensors and Bioelectronics 65 (2015) 121–138

Fig. 10. The schematic of the breath gas sampling device that was used by Miekisch et al. (2008). The image was re-drawn in accordance with the original.

procedures′ effect on the breath analysis was studied. Alveolar, mixed expiratory and time-controlled samples were collected from the subjects. Then, several VOCs and the pCO2 were analyzed. A device for CO2-controlled sampling was built for this purpose. Alveolar sampling was carried out by two means that are named syringe sampling and bag sampling (Fig. 10). The subject first inspired the room air and then the exhaled gas of the respiratory circuit was collected in a 10 ml gas-tight glass syringe (syringe sampling) or a 1000 ml Tedlar bag (bag sampling). The alveolar phase of expiration was simultaneously identified through CO2 tracing by using a mainstream CO2 analyzing system. The sampling process ended prior to the end tidal CO2 concentration. The ‘alveolar syringe sampling’ was achieved then by withdrawing 10 ml of the expired gas into the syringe when plateau of the CO2 curve was attained. The syringes were then transferred to the CO2 analyzer and kept at body temperature during this process. The ‘alveolar bag sampling’ was performed by filling the bag with the exhaled air repetitively at the alveolar phases of the breathing cycles. The ‘non-alveolar bag sampling’ was carried out either during the complete exhalation duration or started 1 s after the beginning of exhalation. Other than the Tedlar bags and glass syringes, the breath sample can be collected and/or stored in the flexfoil bags, nalophan bags, thermal desorption tubes, micropacked sorbent traps, and metal canisters (Lourenço and Turner, 2014). The sampling by collecting and measuring a single breath or multiple breaths are both possible but the breath should be a representative one if the single breath sampling is preferred (Miekisch et al., 2004). Averaging is offered for the spontaneously breathing subjects. When the breath is collected for the forthcoming steps like preconcentration or analysis, inert tedlar bags or stainless steel canisters are used if the breath is not adsorbed directly onto special materials (Miekisch et al., 2004). The precooling of the adsorbent material to  174 °C can also be done (Knutson and Viteri, 1996). The adsorbent materials serve additionally for concentrating the sample. In this case, the effects like breakthrough and memory due to the different boiling points of the VOCs should be avoided through careful selection of the adsorbent material (Miekisch et al., 2004). High breakthrough volumes are preferred but it can lead to the memory effects when such materials are the only adsorption material. The multibed sorbent traps are therefore better. After being trapped there, the volatiles can be released by heating the trap or applying microwave radiation. During sampling, the breath is offered to be taken from the nose in order to eliminate any possible contamination from the mouth flora (Lourenço and Turner, 2014). Besides, the test subjects can be placed on a defined diet for several days (Pauling et al., 1971). In support, Peng and co-workers previously found variations within the same healthy control groups′ cluster patterns in

their study and attributed this observation mostly to differences in the diets, genetic backgrounds, and metabolic states of those individuals (Peng et al., 2008). Their follow-up study for variability in the VOC signatures due to a change in the diet matrix involved VOCs′ monitoring of the breath gas of non-an anesthetized and non-restrained mice by using a proton transfer reaction mass spectrometer with time of flight detection (Kistler et al., 2014). Before the measurements, the mice were fed a standard laboratory chow and then exposed to four semi-purified low- or high-fat diets for four weeks. The change in methanol, methyl acetate or propionate and dimethyl sulfonate were observed. The dimethyl sulfonate is a biomarker of fatty liver disease. The authors suggested that the diet should be considered due to the fact that the mice were unlikely to be affected by this disease. Although the disease of concern was not lung cancer in this case, it is possible that the different dietary preferences could be influential on the breath VOCs profiles and end up in similarities of the subjects' breath VOCs to those of the lung cancer patients'. The diet impacts both the VOCs' signatures and the underlying metabolic functions that influence the exhaled volatiles. This highlights the importance of standardizing the diets for a considerable period before the tests. This diet should be composed of small molecules in order to be thoroughly absorbed and there would be lack of nutrition for the intestinal flora as a result (Pauling et al., 1971). The VOCs from the systemic circulation are contributed by specific trace compounds in the breath that are sourced by airways, saliva, mucus, aerosols in the respiratory tract, bacterial infections at the oral cavity and those in the gut (Lourenço and Turner, 2014). Beauchamp and Pleil carried this concern previously to an upper level by asking if the subject should provide sample on an empty stomach or after a certain period of time following eating (Beauchamp and Pleil, 2013). Broadly speaking, factors like age, gender, smoking, and medical history of the patient like medicaments' use can all influence the sampling and the results together with the possible role of the inter-population differences. Therefore, keeping such variations at minimum would improve the outcomes while generating databases for the representative subjects′ groups of each study group would be of use. A study with the populations from two distinct geographical regions (China and Latvia) revealed separate VOCs′ profile of the patients from different geographical origin (Amal et al., 2013). This is a limitation for a globally applicable disease diagnostics through breath. It is required that changes in VOCs′ as a disease pattern need to be universally applicable. Even, the position of the body during sampling is a concern (Beauchamp and Pleil, 2013). This is yet to be investigated. The results may improve by monitoring the individuals over time in order to eliminate the inter-personal variability (Beauchamp and Pleil, 2013). However, intra-personal changes in the breath sample are out of concern but those can end up being comparably high. Eventually, there are guidelines for patient preparation

Y. Adiguzel, H. Kulah / Biosensors and Bioelectronics 65 (2015) 121–138

before breath testing. Those are not aimed for the lung cancer testing but should have much in common since the ultimate aim is eliminating any possible interference. The recommendations of the American Thoracic Society and European Respiratory Society can be inspected for this purpose. The issue that is highly probable to draw attention in the upcoming years is the need for synchronizing the sampling of breath with the circadian rhythm since oxidative damage to the fluid lining of the lung and the other body parts that affect the VOCs has a circadian rhythm (Miekisch et al., 2004). This has potential to personalize medicine. The other concerns in addition to the contributions by the microbial flora of the gut and mouth are the volatile byproducts that are generated by the personal diet, medications, drugs, and the toxins that one is exposed to Dweik (2011). Furthermore, distinctive approaches such as using endoscopic probes for the breath sampling are developing in order to improve the technique at the expense of non-invasiveness (Santonico et al., 2012). A bronchoscopic apparatus was adapted to collect the breath sample nearby the site that was investigated by the endoscopic probe. The excess air was pumped out into a modified apparatus that allowed the air to be used for the breath testing with an electronic nose. More accurate fingerprint of the disease VOCs could be obtained. Such results can be decisive in a preliminary phase study. In addition, this approach has implications as a breathomic diagnostic validation method. The results were also leading to the arousal of questions about the mechanisms in the generation of the lung cancer specific VOCs. Beyond all these discussions, it should be noted that storage of the breath samples is limited. The breath sample cannot be archived like a histological sample or DNA. This may become problematic under certain circumstances. For instance, re-examination and testing of a sample would not be possible but can be required if complications occur during disease treatments that are decided primarily on the basis of breath test results. Solutions to such situations may be required if the ultimate goal is clinical use. As a result, sampling related issues of the breath testing for lung cancer diagnosis is full of debates and is a fruitful field of research. 5.4. Distinction of the cancer stages The distinction of the cancer stages is a tricky issue. It may need to be handled separately for the patients that have undergone therapeutic intervention and for those others who are yet to receive medication. The breath analysis approach can serve for this purpose as well. It is even suitable for the longitudinal studies on the disease progression such as continuous disease monitoring (Miekisch et al., 2004). Yet, not the case of preliminary therapeutic intervention, but the case of discrimination among various stages of cancer can be tough. This view is based on the study, which reported that there were no significant differences in the sensitivity and specificity of the breath VOCs of the early and advanced lung cancer patients (Phillips et al., 1999b). This finding with the breath analysis is in contrast with the tumor markers' being quantitatively in direct relation with the tumor mass (Phillips et al., 1999b). As a proposed explanation by Phillips and coworkers (Phillips et al., 1999b), CYP enzymes are induced in all the regions of the body and lead to the enhanced catabolism of several VOCs during carcinogenesis (Horváth et al., 2009). As a result, VOCs' profile become similar in all stages of lung cancer and does not change after the tumor resection. This was later supported by the data from their larger scale multicenter study (Phillips et al., 2007) that resulted in scoring of the exhaled VOCs' profile of the patients with resected lung cancer as cancer (Horváth et al., 2009). The early detection of any disease and especially cancer is the hurdle for the early disease intervention that can eventually determine the success of the treatment. The observed discrepancy

135

with the breath analysis (Phillips et al., 1999b) can be affected by variation in the tumor mass/body mass ratio of each patient in adjunct with the differences in the capacities of their alveolar volume in each inhalation. These variations would not be fully selective due not having a linear relationship with the cancer stages of the patients. Yet, developing the sampling approaches (Santonico et al., 2012) and introducing the patients' clinical history as variables (Mazzone et al., 2012) are underway to overcome these hurdles. Probably the most and the first systemic approach up to now is the study on the classification of the lung cancer histology by measuring breath samples with the gold nanoparticle sensors (Barash et al., 2012a). Accordingly, the head space gases were analyzed. Moreover, the lung cancer and healthy cells, the small cell lung cancer and non-small cell lung cancer, the two subtypes of non-small cell lung cancer, and the adenocarcinoma and squamous cell carcinoma were distinguished. 5.5. Feature selection for data classification The multivariate calibration methods aid to resolve the problem of components' overlap in multicomponent mixtures like in case of the volatiles in a breath sample. The measurement data of such samples is pre-processed for multivariate analysis. The principal component analysis is a multivariate analysis method that is used to build a calibration model with the linear relationship between the sensor signal intensities and VOCs in a sample. The multivariate datasets are transformed into a coordinate space wherein vectors in this coordinate set are principal components of the data stream (Feller et al., 2014). The measurement results of the VOCs' mixtures are projected onto this coordinate space. The principal component analysis is used for feature extraction (Jain et al., 2000). Features are the dimensionality of pattern representation and pattern of interest here is the sensor signal profile for the lung cancer specific VOCs' fingerprint. The dimensionality of pattern representation (the number of features) is aimed to be optimized as well as being reduced. However, this needs to be handled with care since reduction in the number of features may end up in a reduced discriminatory power with lowered accuracy (Jain et al., 2000) in disease distinction. The number and type of sensors to be used, sampling and preprocessing manners of the raw data are among the features to be determined, optimized, and reduced wherever possible for the best possible classification performance (Wang et al., 2014) of the selected pattern recognition algorithm. The feature selection serves for this purpose. Wang and co-workers recently examined the feature selection for classifying the chemicals that are measured by a sensor array in an electronic nose (Wang et al., 2014). They used an information-theoretic approach (a filter method) for feature selection (named ‘mutual information’ maximization) on a data set of numerous chemicals that were measured by an electronic nose. The feature selection and subsequent classification were performed on the data measured by 12 metal oxide sensors that were used to perform 10 replicate measurements of 20 chemical analytes with fixed concentrations. The features consisted of 12 sensors and 6 candidate time points, taken from 2 Hz data stream. Each sensor and the time point combination was called as a ‘feature’. A maximum ‘mutual information’ criterion was employed to find the feature subsets. For this purpose, the mutual information among the selected features and the class (identity of chemicals) was maximized with a filter approach. Then, the classification performances of the resulting feature sets were compared. The earlier work of this research group that was performed on feature selection with the same data set searched for the feature set with the best classification performance by a wrapper method (Nowotny et al., 2013). It was an exhaustive search of all the

136

Y. Adiguzel, H. Kulah / Biosensors and Bioelectronics 65 (2015) 121–138

permitted feature combinations (the combinations of sensors and time points). The comparison of the performances of this previous and the ‘mutual information’ methods revealed the achievement of not necessarily the maximum classification performances but the optimum classification performances of the previous method by the latter ‘mutual information’ method (Wang et al., 2014). This was achieved by three orders of magnitude lower computational burden than that was required in case of the former method. In addition, with the selected features, Bayesian Networks was found to be giving the best performance among the common classifiers. The classifiers were found be performing the best with the selected features. Yet, the authors suggested that the approach that they have proposed requires validation with more complex and/or problematic data sets' tests. Examples could be the studies with different analyte concentrations or the studies with analytes that would be collected over longer time periods. This is important in the sense that data noise, sensor drift, different analyte concentrations that actually determine a breath VOCs' signature are among those critical hurdles for a successful sensor performance. In relevance, Guo and co-workers proposed a linear discriminant analysis based sensor selection for an array to choose an optimal configuration of sensors for a particular application when there is a whole set of available sensors (Guo et al., 2011a). The sensors in the array can be cross-reactive to the other diseases even if each is optimized for specific diseases. Therefore, sensor configuration needs to be optimal. This is expected to eliminate the redundant information in detection of a disease by using a sensor array. Shortly before publishing this sensor selection approach, Guo and co-workers suggested improved breath sample identification by a classification based on sparse representation (Guo et al., 2011b). The data evaluation algorithms other than the same share factors that are the sensor array itself and the reference data set were traced as a means of further improvement to the authors' former studies' results with the sensor arrays. The motivation of this approach was based on the fact that the prevailing approaches (the classical classifiers like K nearest neighbor, linear discriminant analysis, artificial neural network, support vector machine) rarely consider the specific character of the data. The number of samples to train the classifier is generally very small in comparison to the high dimensionality of the data in each sample. Therefore, the authors claimed that those classical classifiers would not be working good because their performances depend on the interrelation of the sample sizes, features' number, and classifiers' structure. The results were outperforming classical classifiers regardless of the selected features but improvements were still required and more samples were needed to be tested. 5.6. Commercialization There is an existing market for the gas sensing instruments. The commercial potential of the breath sensing devices can easily find a place among the present gas sensing devices. As an example, Owlstone's Field Assisted Ion Mobility Spectrometer (FAIMS) instrument is a dime-sized product that serves for the gas detection and obtaining chemical signatures in the ppb range within seconds. It is serving for chemical and explosives detection with a series of parallel-connected electrodes that improve the sensitivity with respect to the conventional FAIMS technologies. The conventional FAIMS technologies has rather diminished sensitivities due to the requirement of applying high voltages. The Owlstone's instrument can be programmed to filter out the chemicals that do not have specific mobilities of interest with a parallel gap filter structure. Its flexible nature makes the use this instrument possible for various applications including the breath sensing for lung cancer detection.

There are also developing commercial systems as the breath sensors for lung cancer diagnosis. The most prominent example is probably the Nanoscale Artificial NOSE (™NA-NOSE). It was initially developed under an EU funded project by Prof. Haick and his group. It is a nanomaterial-based artificial olfactory system with increased sensitivity to volatiles in breath including those of the lung cancer. On the other hand, Dr. Kenneth S. Suslick initiated Chemsensing, Inc. to commercialize the product of their research on metalloporphyrin dye arrays that change color upon reaction with the specific VOCs. These existing and prospective products all decrease the rate of false positives through the detection of the VOCs' fingerprint.

6. Final remarks The significance of breath sensors is not only differentiating the cancerous samples from healthy ones but discriminating lung cancer from other lung diseases as well (Peng et al., 2010; D’Amico et al., 2010). Several tumor cell lines including lung cancer cell lines could be distinguished by using commercially available electronic noses that contain a signal-processing system and an array of VOCs-reactive conductive polymers (Gendron et al., 2007). However, distinction of the cancer stages is equally important as the detection of cancer for a prompt therapeutic intervention. In one study, VOCs' profile was similar in all the stages of lung cancer and not changing after the tumor resection (Phillips et al., 2007). This was supported with the data of a larger scale multicentre study that was carried out later (Phillips et al., 2007). In addition, Poli and co-workers suggested that the breath analysis should not to be used as a short-term follow-up procedure in the resectionencountered patients (Poli et al., 2005). Yet, dissimilar results on relevant issues could be obtained such as that reported later by Poli et al. (2008). The source of the specific smell that is attributed to cancer needs to be delineated. The recent studies reveal significant progress in this by profiling the genetic mutations of lung cancer cells through their headspace VOCs (Peled et al., 2013). The classification of the lung cancer histology and studies on headspace VOCs of various tumor cell types by using gold nanoparticle sensors contributed as much (Barash et al., 2012b). This is achieved despite common VOCs among different cancer cell lines. For instance, Mazzone (Mazzone, 2008) mentioned that cell-concentration specific production of acetaldehyde is the head-space gas of two cancer cell lines in vitro (Smith et al., 2003b). Furthermore, VOCs from the culture medium of lung cancer cells were differing from the virgin culture medium and from that of the control cell cultures (Chen et al., 2007). The model systems are powerful but complexity of the multicellular organisms cannot be fully represented by the model cell systems. Therefore, model systems are known to provide only conceptual supports. Another issue is that the exhaled VOCs that are appearing along with the lung cancer development are possibly due to the chronic load and burden of the VOCs in the lungs rather than being specific lung cancer biomarkers (Mutti, 2008). A recent review nicely put effort on the “assessment, origin, and implementation of breath volatile cancer markers” by classifying 115 validated cancer-related VOCs based on their fat-to-blood and blood-to-fat partition coefficients (Haick et al., 2014). In this study, the relative concentrations of VOCs in fat, blood, and alveolar breath were estimated. Approaches like ‘combined volatolomics’ may provide sufficient insight to reveal the complete profile of the bodily volatiles. The results of a recent exemplary application enabled obtaining an uncorrelated but complementary information from the VOCs of the subjects' breaths and skins (Broza et al., 2014). Such studies can boost stimuli to the upcoming research in this field.

Y. Adiguzel, H. Kulah / Biosensors and Bioelectronics 65 (2015) 121–138

Further studies on discriminating the other lung diseases from lung cancer, distinguishing the stages of cancer, classifying the different lung diseases, studying the lung cancer specific markers, working with discrete and larger-sized populations, and optimizing the sampling through standardizing the ambient VOCs are needed to be performed with breath sensors (Mazzone, 2008; Dragonieri et al., 2009; Horváth et al., 2009; D’Amico et al., 2010). The ongoing debate on the major source of cancer smell (Horváth et al., 2008) and the variations in the results can be resolved by working on the sampling settings and statistical analytical procedures (Horváth et al., 2009; Broza et al., 2013). Those will hopefully evolve into standardized procedures (Miekisch et al., 2004) that can introduce the clinical history of the patients as analysis parameters (Mazzone et al., 2012). Wireless gas sensing (Potyrailo et al., 2011) may combine advances in the sensors and clinical practice fields together with real time analysis (Miekisch et al., 2004). In addition, application-specific e-nose systems with restricted identification libraries/capabilities may be developed. They can be preferred to restrict the instrument expenses, prevent cross-reactions between diseases, and attain portability (Wilson and Baietto, 2011; Guo et al., 2011a, 2011b). One drawback is the possible presence of high concentrations of a single VOC component in case of a disease-related breath sample. This can result in sensor drifts and prevent the exact calibration of the sensor (Horváth et al., 2009). Additionally, relatively short life-times of the sensors, the necessity of performing adequate work on the method development for each application, and the scarcity of quantitative are among the limitations (Horváth et al., 2009). Yet, it is possible that pure volatile samples and their mixtures can be used to develop pattern recognition libraries once the VOC biomarkers are identified (Wilson and Baietto, 2011). As a result, these non-invasive biosensors (Guilbault and Palleschi, 1995) are optimal for the population-wide screening purposes (Horváth et al., 2009). Reliable tests would be life-saving at the point-of-care but a common approach is necessary and validity of the tests need to be proven through various population-based studies together with the models that consider factors like the long-term effects of drugs and the environmental and non-environmental factors. Despite the shortcomings, the reported success rates on disease identification from breath samples imply robustness of the approach. Moreover, the increased incidence of a disease in the patient group of another disease points also at the common VOCs among their fingerprint VOCs' patterns (Horváth et al., 2009). This can be used for developing risk assessment tests for preventive medicine.

7. Summary and conclusions The diagnosis of lung cancer with breath analysis is promising. Its superiority is the possibility of collecting breath samples safely, with high frequencies, and easily, even from critically ill patients (Horváth et al., 2009) or from those in conditions that bear risk for the patient to provide blood samples (i.e., hemophilia). Hence, the breath sensors are optimal for population-wide screening purposes and for testing at the point-of-care. The evolving sensors field brought about the e-nose applications that utilized colorimetric, composite, carbon nanotube, gold nanoparticle-based, and surface acoustic wave sensor arrays. Still, there are many underway. However, comparing the performances is based on their disease discriminatory powers by creating the fingerprint patterns of diseases rather than usual terms associated with the classical detection schemes. The results rely heavily on the sampling settings and statistical analytical procedures. This leads to significant deviations among the results of studies (Horváth et al., 2009; Broza et al., 2013). There are also other limitations such as the

137

decrease in the sensor sensitivities in presence of water vapor. In addition, the drift of the sensor and inability to make an exact calibration of the sensor can be problematic when there is high concentration of a single VOC component (Horváth et al., 2009). Further drawbacks are the relatively short life-times, necessity of performing adequate work on the method development for each application, and the quantitative data scarcity (Horváth et al., 2009). Additional problems to be tackled or processes to be optimized include but not limited to discriminating multiple diseases at once (1), cancer staging (2), carrying out multicentre studies with larger populations (3), understanding the source of cancer related VOCs (4), developing and standardizing sampling and patient preparation together with registering new sampling approaches (5), proper evaluation of the individual differences that can and cannot be controlled including the evaluation of possible influences of the utilized data analysis tools on the test results (6), and developing pattern recognition libraries (7). This list is critical in the sense that a huge remediation is yet to be done in transferring the breath sensors based lung cancer testing to the clinical applications. However, there is a wide scope of potential breath diagnostics field and all of those will be benefiting by the advancements that share the similar grounds.

References Albert, K.J., Lewis, N.S., Schauer, C.L., Sotzing, G.A., Stitzel, S.E., Vaid, T.P., Walt, D.R., 2000. Chem. Rev. 100, 2595–2626. Amal, H., Leja, M., Broza, Y.Y., Tisch, U., Funka, K., Liepniece-Karele, I., Skapars, R., Xu, Z., Liu, H., Haick, H., 2013. J. Breath Res 7, 047102. Askim, J.R., Mahmoudi, M., Suslick, K.S., 2013. Chem. Soc. Rev. 42, 8649–8682. Bach, P.B., Kelley, M.J., Tate, R.C., McCrory, D.C., 2003. Chest 123, 72S–82S. Bao, L., Hazari, S., Mehra, S., Kaushal, D., Moroz, K., Dash, S., 2012. Am. J. Pathol. 180, 2490–2503. Barash, O., Peled, N., Tisch, U., Bunn Jr., P.A., Hirsch, F.R., Haick, H., 2012a. Nanomedicine 8, 580–589. Barash, O., Peled, N., Tisch, U., Bunn Jr., P.A., Hirsch, F.R., Haick, H., 2012b. Nanomed. Nanotechnol 8, 580–589. Beauchamp, J.D., Pleil, J.D., 2013. J. Breath Res 7, 042001. Binions, R., Davies, H., Afonja, A., Dungey, S., Lewis, D.W., Williams, D.E., Parkin, I.P., 2009a. J. Electrochem. Soc. 156, J46–J51. Binions, R., Afonja, A., Dungey, S., Lewis, D.W., Parkin, I.P., Williams, D.E., 2009b. Sensors, IEEE 2009 Proceedings, pp. 1090–1095. Broza, Y.Y., Kremer, R., Tisch, U., Gevorkyan, A., Shiban, A., Best, L.A., Haich, H., 2013. Nanomedicine 9, 15–21. Broza, Y.Y., Zuri, L., Haick, H., 2014. Sci. Rep. 4, 4611. Bushdid, C., Magnasco, M.O., Vosshall, L.B., Keller, A., 2014. Science 343, 1370–1372. Canavan, N., 2013. Electronic nose sniffs out lung cancer. Medscape http://www. medscape.com/viewarticle/810971. Cao, W., Duan, Y., 2006. Clin. Chem. 52, 800–811. Castro, M., Kumar, B., Feller, J.F., Haddi, Z., Amari, A., Bouchikhi, B., 2011. Sens. Actuator B—Chem 159, 213–219. Chan, H.P., Lewis, C., Thomas, P.S., 2009. Lung Cancer—J. IASLC 63, 164–168. Chatterjee, S., Castro, M., Feller, J.F., 2013. J. Mater. Chem. B 1, 4563–4575. Chen, X., Cao, M., Li, Y., Hu, W., Wang, P., Ying, K., Pan, H., 2005. Meas. Sci. Technol. 16, 1535–1546. Chen, X., Xu, F., Wang, Y., Pan, Y., Lu, D., Wang, P., Ying, K., Chen, E., Zhang, W., 2007. Cancer 110, 835–844. Chen, X., Wang, Y., Wang, Y., Hou, Z., Xu, D., Yang, Z., Zhang, Y., 2010. Sens. Actuator A–Phys 158, 328–334. Cheng, Z., Chan, A.K., Lewis, C.R., Thomas, P.S., Raftery, M.J., 2011. J. Cancer Ther 2, 4103. Conrad, D.H., Goyette, J., Thomas, P.S., 2008. J. Gen. Intern. Med 23 (Suppl. 1), 78–84. D’Amico,, A., Pennazza, G., Santonico, M., Martinelli, E., Roscioni, C., Galluccio, G., Paolesse, R., Di Natale, C., 2010. Lung Cancer—J. IASLC 68, 170–176. Dacres, H., Wang, J., Leitch, V., Horne, I., Anderson, A.R., Trowell, S.C., 2011. Biosens. Bioelectron 29, 119–124. Deng, C., Zhang, X., Li, N., 2004. J. Chromatogr. B: Anal. Technol. Biomed. Life Sci. 808, 269–277. Di Natale, C., Macagnano, A., Martinelli, E., Paolesse, R., D’Arcangelo, G., Roscioni, C., Finazzi-Agrò, A., D’Amico, A., 2003. Biosens. Bioelectron. 18, 1209––1218. Di Natale, C., Paolesse, R., D’Amico, A., 2007. Sensor. Actuator B—Chem. 121, 238–246. Di Pietrantonio, F., Cannatà, D., Benetti, M., Verona, E., Varriale, A., Staiano, M., D’Auria, S., 2013. Biosens. Bioelectron. 41, 328–334. Dragonieri, S., Annema, J.T., Schot, R., van der Schee, M.P.C., Spanevello, A., Carratú, P., Resta, O., Rabe, K.F., Sterk, P.J., 2009. Lung Cancer—J. IASLC 64, 166–170. Du, L., Wu, C., Liu, Q., Huang, L., Wang, P., 2013a. Biosens. Bioelectron 42, 570–580.

138

Y. Adiguzel, H. Kulah / Biosensors and Bioelectronics 65 (2015) 121–138

Du, L., Wu, C., Peng, H., Zhao, L., Huang, L., Wang, P., 2013b. Biosens. Bioelectron 40, 401–406. Dweik, R.A., 2011. J. Breath Res 5, 030201. Ehmann, R., Boedeker, E., Friedrich, U., Sagert, J., Dippon, J., Friedel, G., Walles, T., 2011. Eur. Respir. J. 39, 669–676. Feller, J.-F., Gatt, N., Kumar, B., Castro, M., 2014. Chemosensors 2, 26–40. Fend, R., Bessant, C., Williams, A.J., Woodman, A.C., 2004. Biosens. Bioelectron 19, 1581–1590. Gendron, K.B., Hockstein, N.G., Thaler, E.R., Vachani, A., Hanson, C.W., 2007. Otolaryngol. Head Neck Surg. 137, 269–273. Ghosal, R., Kloer, P., Lewis, K.E., 2009. Postgrad. Med. J. 85, 358–363. Gordon, S.M., Szidon, J.P., Krotoszynski, B.K., Gibbons, R.D., O’Neill, H.J., 1985. Clin. Chem. 31, 1278–1282. Guilbault, G.G., Palleschi, G., 1995. Biosens. Bioelectron 10, 379–392. Guo, D., Zhang, D., Zhang, L., 2011a. Sensor. Actuator B–Chem 157, 265–274. Guo, D., Zhang, D., Zhang, L., 2011b. Sensor. Actuator B—Chem 158, 43–53. Grob, N.M., Aytekin, M., Dweik, R.A., 2008. J. Breath Res 2, 037004. Haick, H., Broza, Y.Y., Mochalski, P., Ruzsanyi, V., Amann, A., 2014. Chem. Soc. Rev. 43, 1423–1449. Hakim, M., Broza, Y.Y., Barash, O., Peled, N., Phillips, N., Amann, A., Haick, H., 2012. Chem. Rev. 112, 5949–5966. Hewett, P.W., 2001. Int. J. Biochem. Cell Biol. 33, 325–335. Hill, D., Binions, R., 2012. Int. J. Smart Sens. Intell. Sys 5, 401–440. Horváth, G., Järverud, G.A., Järverud, S., Horváth, I., 2008. Integr. Cancer Ther. 7, 76–80. Horváth, I., Lázár, Z., Gyulai, N., Kollai, M., Losonczy, G., 2009. Eur. Respir. J. 34, 261–275. Horvath, I., 2010. Lung Cancer—J. IASLC 68, 127–128. Huo, D., Xu, Y., Hou, C., Yang, M., Fa, H., 2014. Sens. Actuator B—Chem 199, 446–456. Jain, A.K., Duin, R.P.W., Mao, J., 2000. IEEE Trans. Pattern Anal 22, 5–37. Kang, N.K., Jun, T.S., La, D.-D., Oh, J.H., Cho, Y.W., Kim, Y.S., 2010. Sens. Actuator B— Chem 147, 55–60. Kim, N.-H., Choi, S.-J., Yang, D.-J., Bae, J., Park, J., Kim, I.-D., 2014. Sens. Actuator B— Chem 193, 574–581. Kistler, M., Szymczak, W., Fedrigo, M., Fiamoncini, J., Höllriegl, V., Hoeschen, C., Klingenspor, M., Hrabě de Angelis, M., Rozman, J., 2014. J. Breath Res 8, 016004. Knutson, M.D., Viteri, F.E., 1996. Anal. Biochem. 242, 129–135. Ko, H.J., Park, T.H., 2005. Biosens. Bioelectron. 20, 1327–1332. Lee, S.H., Jun, S.B., Ko, H.J., Kim, S.J., Park, T.H., 2009. Biosens. Bioelectron. 24, 2659–2664. Lippi, G., Cervellin, G., 2011. Clin. Chem. Lab. Med. 50, 435–439. Liu, F.L., Xiao, P., Fang, H.L., Dai, H.F., Qiao, L., Zhang, Y.H., 2011. Physica E 44, 367–372. Lourenço, C., Turner, C., 2014. Metabolites 4, 465–498. 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., 2005. Am. J. Respir. Crit. Care Med. 171, 1286–1291. Mashir, A., Dweik, R.A., 2009. Adv. Powder Technol. 20, 420–425. Mazzone, P.J., Hammel, J., Dweik, R., Na, J., Czich, C., Laskowski, D., Mekhail, T., 2007. Thorax 62, 565–568. Mazzone, P.J., 2008. J. Thorac. Oncol. 3, 774–780. Mazzone, P.J., Wang, X.-F., Xu, Y., Mekhail, T., Beukemann, M.C., Na, J., Kemling, J.W., Suslick, K.S., Sasidhar, M., 2012. J. Thorac. Oncol. 7, 137–142. Mendis, S., Sobotka, P.A., Euler, D.E., 1995. Free Radic. Res. 23, 117–122. McEuen, P.L., 1998. Nature 393, 15–17. McCulloch, M., Jezierski, T., Broffman, M., Hubbard, A., Turner, K., Janecki., T., 2006. Integr. Cancer Ther. 5, 30–39. Miekisch, W., Schubert, J.K., Noeldge-Schomburg, G.F.E., 2004. Clin. Chim. Acta 347, 25–39. Miekisch, W., Kischkel, S., Sawacki, A., Liebau, T., Mieth, M., Schubert, J.K., 2008. J. Breath Res 2, 026007. Moyer, V.A., 2014. Ann. Intern. Med. 160, 330–338. Mutti, A., 2008. Acta Biomed 79 (Suppl. 1), S11–S23. Narasimhan, L.R., Goodman, W., Patel, C.K.N., 2001. Proc. Natl. Acad. Sci. USA 98, 4617–4621. Nowotny, T., Berna, A.Z., Binions, R., Trowell, S., 2013. Sens. Actuator B—Chem 187, 471–480. Oh, E.H., Song, H.S., Park, T.H., 2011. Enzyme Microb. Technol. 48, 427–437. Pauling, L., Robinson, A.R., Teranishi, R., Cary, P., 1971. Proc. Natl. Acad. Sci. USA 68, 2374–2376. Pavlou, A.K., Magan, N., Sharp, D., Brown, J., Barr, H., Turner, A.P.F., 2000. Biosens. Bioelectron. 15, 333–342. Pavlou, A.K., Magan, N., McNulty, C., Jones, J.M., Sharp, D., Brown, J., Turner, A.P.F., 2002. Biosens. Bioelectron. 17, 893–899.

Peled, N., Barash, O., Tisch, U., Ionescu, R., Broza, Y.Y., Ilouze, M., Mattei, J., Bunn Jr., P. A., Hirsch, F.R., Haick, H., 2013. Nanomed. Nanotechnol 9, 758–766. Persaud, K., Dodd, G., 1982. Nature 299, 352–355. Peng, G., Trock, E., Haick, H., 2008. Nano Lett. 8, 3631–3635. Peng, G., Tisch, U., Adams, O., Hakim, M., Shehada, N., Broza, Y.Y., Billan, S., AbdahDortnyak, R., Kuten, A., Haick, H., 2009. Nat. Nanotechnol 4, 669–673. Peng, G., Hakim, M., Broza, Y.Y., Billan, S., Abdah-Dortnyak, R., Kuten, A., Tisch, U., Haick, H., 2010. Br. J. Cancer 103, 542–551. Pennazza, G., Santonico, M., Martinelli, E., Paolesse, R., Tamburrelli, V., Cristina, S., D ′Amico, A., Di Natale, C., Bartolazzi, A., 2011. Sens. Actuator B—Chem. 154, 288––294. Phillips, M., Erickson, G.A., Sabas, M., Smith, J.P., Greenberg, J., 1995. J. Clin. Pathol. 48, 466–469. Phillips, M., Sabas, M., Greenberg, J., 1993. J. Clin. Pathol. 46, 861–864. Philips, M., 1997. Anal. Biochem. 247, 272–278. Phillips, M., Herrera, J., Krishnan, S., Zain, M., Greenberg, J., Cataneo, R.N., 1999a. J. Chromatogr. B: Biomed. Sci. Appl. 729, 75–88. Phillips, M., Gleeson, K., Hughes, J.M.B., Greenberg, J., Cataneo, R.N., Baker, L., McVay, W.P., 1999b. Lancet 353, 1930–1933. Phillips, M., Altorki, N., Austin, J.H., Cameron, R.B., Cataneo, R.N., Greenberg, J., Kloss, R., Maxfield, R.A., Munawar, M.I., Pass, H.I., Rashid, A., Rom, W.N., Schmitt, P., 2007. Cancer Biomark. 3, 95–109. Ping, W., Yi, T., Haibao, X., Farong, S., 1997. Biosens. Bioelectron. 12, 1031–1036. Poli, D., Carbognani, P., Corradi, M., Goldoni, M., Acampa, O., Balbi, B., Bianchi, L., Rusca, M., Mutti, A., 2005. Respir. Res. 6, 71. Poli, D., Goldoni, M., Caglieri, A., Ceresa, G., Acampa, O., Carbognani, P., Rusca, M., Corradi, M., 2008. Acta Biomed 79 (Suppl. 1), S64–S72. Potyrailo, R.A., Surman, C., Nagraj, N., Burns, A., 2011. Chem. Rev. 111, 7315–7354. Rakow, N.A., Suslick, K.S., 2000. Nature 406, 710–713. Röck, F., Barsan, N., Weimar, U., 2008. Chem. Rev. 108, 705–725. Saba, N.F., Kuhuri, F.R., 2005. Clin. Lung Cancer 7, 92–99. Santonico, M., Lucantoni, G., Pennazza, G., Capuano, R., Galluccio, G., Roscioni, C., La Delfa, G., Consoli, D., Martinelli, E., Paolesse, R., Di Natale, C., D′Amico, A., 2012. Lung Cancer—J. IASLC 77, 46–50. Silva, L.I.B., Freitas, A.C., Rocha-Santos, T.A.P., Pereira, M.E., Duarte, A.C., 2011. Talanta 83, 1586–1594. Smith, R.A., Cokkinides, V., Eyre, H.J., 2003a. CA Cancer J. Clin. 53, 27–43. Smith, D., Wang, T., Sulé-Suso, J., Španĕl, P., El Haj, A., 2003b. Mass Spectrom 17, 845–850. Sung, J.H., Ko, H.J., Park, T.H., 2006. Biosens. Bioelectron. 21, 1981–1986. Varghese, O.K., Kichambre, P.D., Gong, D., Ong, K.G., Dickey, E.C., Grimes, C.A., 2001. Sens. Actuator B–Chem. 81, 32–41. Wang, P., Liu, Q., 2011a. Chapter 5: Chemical Sensors and Measurement. In: Ping Wang, Qingjun Liu (Eds.), Biomedical Sensors and Measurement, pp. 163–164. Wang, P., Liu, Q., 2011b. Chapter 5: Chemical Sensors and Measurement. In: Ping Wang, Qingjun Liu (Eds.), Biomedical Sensors and Measurement, pp. 162–163. Wang, X.R., Lizier, J.T., Nowotny, T., Berna, A.Z., Prokopenko, M., Trowell, S.C., 2014. Plos One 9, e89840. Wang, P., Chen, X., Xu, F., Lu, D., Cai, W., Ying, K., Wang, Y., Hu, Y., 2008. Development of electronic nose for diagnosis of lung cancer at early stage. In: Proceedings of the 5th International Conference on Information Technology and Applications in Biomedicine, in conjunction with the 2nd International Symposium and Summer School on Biomedical and Health Engineering. Shenzhen, China, May 30–31, pp. 588–591. Weitz, Z.E.W., Birnbaum, A.J., Sobotka, P.A., Zarling, E.J., Skosey, J.L., 1991. Lancet 337, 933–935. Wilson, A.D., Baietto, M., 2011. Sensors 11, 1105–1176. Wlodzimirow, K.A., Abu-Hanna, A., Schultz, M.J., Maas, M.A.W., Bos, L.D.J., Sterk, P.J., Knobel, H.H., Soers, R.J.T., Chamuleau, R.A.F.M., 2014. Biosens. Bioelectron. 53, 129–134. Wu, T.-Z., Lo, Y.-R., Chan, E.-C., 2001. Biosens. Bioelectron. 16, 945–953. Wu, C., Du, L., Wang, D., Wang, L., Zhao, L., Wang, P., 2011. Biochem. Biophys. Res. Commun 407, 18–22. Wyszynski, B., Nakamoto, T., 2009. IEEJ Trans. Electr. Electron 4, 334–338. Xu, X., Wang, J., Long, Y., 2006. Sensors 6, 1751–1764. Xu, Z.-q, Broza, Y.Y., Ionsecu, R., Tisch, U., Ding, L., Liu, H., Song, Q., Pan, Y.-y, Xiong, F.-x, Gu, K.-s, Sun, G.-p, Chen, Z.-d, Leja, M., Haick, H., 2013. Br. J. Cancer 108, 941–950. Yu, H., Xu, L., Cao, M., Chen, X., Wang, P., Jiao, J., Wang, Y., 2003. Proc. IEEE Sens. 2003, 1333–1337.

Breath sensors for lung cancer diagnosis.

The scope of the applications of breath sensors is abundant in disease diagnosis. Lung cancer diagnosis is a well-fitting health-related application o...
4MB Sizes 5 Downloads 13 Views