ELECTRONIC NOSE CHEMICAL SENSOR VERSUS : A FEASIBILITY STUDY FOR THE DIFFERENTIATION OF APPLE FLAVORS AND ESSENCES

W. N. Marrazzo, P. H. Heinemann, R. A. Saftner, R. E. Crassweller, E. Leblanc

ABSTRACT. The ability of chemical sensors to differentiate and match samples of apple flavors and essences is of interest to the flavor and food industries. This study evaluated the feasibility of using a prototype chemical sensor for the differentiation of apple flavors and essences. Volatile headspace gases from 20 flavors and four essences were measured with a gas chromatograph and the prototype chemical sensor. A principal component analysis (PCA) reduced the chemical sensor data into two principal factors that accounted for 96% of the variance in the sensor measurements. A hierarchical cluster analysis (HCA) of the PCA data put two flavors and essences into one cluster and the remaining 22 into another. The HCA analysis also provided a breakdown of the clusters into subgroups by similarity. The subgroups were used to determine matches with eight unknowns, six flavors, and two essences. The prototype chemical sensor correctly identified all six unknown flavors. Three were absolute matches, and three matched more than one unknown (one of which was the correct flavor). Visual comparisons of gas chromatograms of known and unknown samples led to absolute matches of five of the six unknown flavor samples. Neither measuring method identified the two unknown essences. These results suggest that it is feasible to use the chemical sensor prototype to differentiate between apple flavors or essences. One potential use of chemical sensor technology is in quality control and food safety programs of food ingredients and products. Keywords. Apple, Electronic nose, Essences, Flavors.

pple flavors and essences are food ingredients spectral or temporal separation, and also evaluated by the hu- that emit a vast array of aromatic volatile gases, man nose after separation of components. Newer detectors which contribute to the sensorial quality of foods rely on chemometrics, and vary from solid-state sensors to in which they are incorporated (MacGregor et al., rapid-analysis mass spectrometers (Sarig and Delwiche, A1964). The headspace atmosphere of different flavors and es- 1998). Disadvantages of solid-state sensors include sensor sences are distinct both qualitatively and quantitatively poisoning, no chemical structure information, time-consum- (Morton and MacLeod, 1990; Dimick and Hoskin, 1983). Re- ing calibration, short- and long-term drift, and water and al- liable, inexpensive analytical methods to accurately detect cohol interference (InfoMetrix, 1999). An array-type and differentiate flavors and essences are needed. The human chemical sensor designed to alleviate these disadvantages nose as a sensory analytical tool remains the most widely was developed by Cyrano Sciences (Doleman et al., 1998). used technique for analysis of flavors and essences. Early in- This sensor array is based on the physical characteristic that strumental detectors, categorized at that time as electronic carbon black organic polymer composites swell reversibly noses or e-noses, utilized a traditional analytical approach of upon exposure to volatile compounds. This sensor array is spectroscopy or chromatography. The headspace atmo- fast when compared to the response times of gas chromato- spheres of flavors and essences were evaluated with either graphs and mass spectrometers. Recent studies compare the pros and cons of gas chromatography and mass spectroscopy (Heiden et al., 2002) to measure aromatic volatile headspace Article was submitted for review in November 2004; approved for gases. The need exists to compare a to publication by the Information & Electrical Technologies Division of standard analytical instruments. The chemical sensor array’s ASABE in August 2005. quick response time, instrument portability and ease of use Use of a product name or product by the USDA does not imply approval would facilitate industrial and laboratory applications to or recommendation of the product to the exclusion of others that also may be suitable. characterize the aromatic volatile gases in the headspace at- The authors are William N. Marrazzo, Former Graduate Fellow, and mosphere of foods and food ingredients. Paul H. Heinemann, ASABE Member Engineer, Professor, Department The flavor and essence industry is a significant part of the of Agricultural and Biological Engineering, The Pennsylvania State food industry. Among their product line is a vast array of University, University Park, Pennsylvania; Robert A. Saftner, Plant apple flavors (naturally occurring apple components and Physiologist, USDA-ARS-BARC-PQSL, Beltsville, Maryland; Robert E. Crassweller, Professor, Department of Horticulture, The Pennsylvania other flavor-associated components) and essences (apple State University, University Park, Pennsylvania; and Eric LeBlanc, concentrates). These products, which contain aromatic Statistician, USDA-ARS-BARC Statistics Group, Beltsville, Maryland. volatile compounds, are combined and incorporated into a Corresponding author: Paul H. Heinemann, Department of Agricultural variety of processed foods to impart aromas and flavors and Biological Engineering, 224 Agricultural Engineering Bldg., The Pennsylvania State University, University Park, PA 16802; phone: reminiscent of apples. The primary objective of this study 814-865-2633; fax: 814-863-1031; e-mail: [email protected]. was to determine the feasibility of comparing the prototype

Transactions of the ASAE Vol. 48(5): 2003−2006 2005 American Society of Agricultural Engineers ISSN 0001−2351 2003 Cyrano Sciences chemical sensor analyses of apple flavors ric analysis used for the chemical sensor were the methods of and essences to standard analytical approaches for matching choice of the GC expert and statisticians consulted for the and differentiating different apple flavors and essences. project. Headspace atmospheres of known and unknown flavors and essences were measured with a prototype Cyrano Sciences CYRANO SCIENCES MULTIPLE ARRAY CHEMICAL SENSOR - chemical sensor and with a gas chromatograph. DATA LOGGER PROTOTYPE The multiple array chemical sensor - data logger used in this study was provided by Cyrano Sciences, Inc., Pasadena, MATERIALS AND METHODS Cal. (prototype data logger serial number DL-0027-012999, sensor set number CN-7100A-608). It is a portable polymeric FLAVORS AND ESSENCES resistance change sensor unit that connects to a power supply Twenty apple flavors (generous gifts from Alex Fries, and a laptop computer through a parallel printer port. The Cincinnati, Ohio; McCormick, Hunt Valley, Md.; and Takasago, Rockleigh, N.J.) and four apple essences (gener- instrument is sometimes referred to as a data logger, a multiple array chemical sensor, a chemical sensor, an ous gifts from Flavor Materials International, Avenel, N.J.; “e-nose,” or an “electronic nose.” The chemical sensors and Sun Pure - Mastertaste, Lakeland, Fla.) were stored in the dark at 4°C between uses. After each use, the headspace of (32 total) are carbon elements impregnated with non-con- ducting polymers whose resistance changes when exposed to each storage bottle was flushed with nitrogen gas and different volatile headspace gases. Only 31 of the 32 sensors resealed. While these efforts were made to maintain product stability, some product degradation may have occurred were used in this study since the manufacturer indicated that the output from one of the sensors was unreliable. The during the course of this study. Formulations for these custom-made chemical sensors in the prototype data logger samples are proprietary. Notation referring to specific products was replaced with code numbers 1 through 24. In used in this study are not expected to be as stable, accurate, or precise as those manufactured using refined mass addition, to serve as a repeat “blind” measurement, eight of production techniques. In this regard, subsequent perfor- the same products (six flavors and two essences) were randomly selected and coded with the numbers U1 through mance by production versions of the portable chemical sensor unit should improve due to standardized sensor U8. To avoid bias, the identity of the blind products was not manufacturing methods. revealed until after all of the sample headspace atmospheres from the coded flavors and essences had been measured and the data analyzed. CHEMICAL SENSOR ANALYSES A one-half mL aliquot (0.5 mL) of each flavor and essence was transferred to a 500 mL Erlenmeyer flask and sealed with SAMPLE IDENTITY a Teflon-lined rubber stopper. Samples were allowed to After all headspace volatile measurements were com- equilibrate for 5 min at room temperature (23°C ±1°C). pleted and the data analyzed, the identities of the coded Cyrano Sciences designated the sampling protocol cycle as samples were revealed. Two of the coded samples were follows: a 10 s baseline, room air flush, a 25 s headspace essences, and the other six were flavors. sampling, and then a 60 s, room air flush. Sensor chamber temperature was maintained at 28.0°C, and flow rate through GAS CHROMATOGRAPHIC ANALYSES the sensor chamber was kept at 150 mL min−1 using the Aliquots (0.5 mL) of each coded product were transferred default pump setting (3 V). A single data set consisted of the to individual 10 mL vials and crimp capped with a average of five consecutive samplings of the same sample. Teflon-lined septum. The samples were equilibrated for Since immediate data analysis was not possible, two data sets 5 min at 20°C. For headspace volatile sampling, a solid- were consecutively collected. Only data sets in which the phase microextraction (SPME) fiber coated with polydime- headspace atmosphere maintained an apparent steady state thylsiloxane (PDMS, 1 cm long, 100 mmthick, Supelco Co., throughout sampling were used for analysis. The data were Bellefonte, Pa.) was used to collect and concentrate aromatic normalized for each of the sensors to equalize the concentra- volatile gases in the headspace of the sample vials. These tion effect. fibers act by virtue of their sorption characteristics (Arthur et al., 1992). After 16 min of exposure time, the sorbed volatile headspace gases were desorbed from the fiber for 2 min at 250°C (injector port temperature) into a glass-lined splitless RESULTS AND DISCUSSION injection port of a gas chromatograph (model 5890a, Agilent GAS CHROMATOGRAPHY Technologies, Rockville, Md.) equipped with electronic Composite chromatograms of representative flavors and pressure control and a flame ionization detector (FID). essences are shown in figure 1. These matches were made by Apple flavor and essence volatile headspace gases were visual analysis of gas chromatograms, and by analytical separated using a capillary column (HP-5, 11 m × 0.1 mm ID, evaluations (PCA and HCA) of data sets from the chemical 0.34 mmcoating thickness). The carrier gas was ultra-purified sensors of the prototype data logger and their coded hydrogen (6.0 research) at a flow velocity of 52 cm s−1. The identification. The chromatograms from many of the 26 fla- temperature program was isothermal for 2 min at 40°C and vors (20 known and six unknown coded samples) and six then raised at a rate of 30°C min−1 to 250°C and held for essences (four known and two unknown coded samples) were 3 min. For identification of unknown samples, chromato- unexpectedly similar to each other in peak profiles. As grams of each unknown sample (coded U1 to U8) were expected, the chromatograms of the essences generally had visually compared to chromatograms of each known sample higher peak magnitudes than that of the flavors. However, (coded 1 to 24). This visual examination and the chemomet- there were sufficient qualitative and quantitative differences

2004 TRANSACTIONS OF THE ASAE Table 1. Results of flavor and essence tests, showing Essence matches found for unknown samples.

400 Gas Cyrano Chroma- Sciences Source tography Chemical Bottle Designation 300 Match Sensor Match Type Sample set 1 U1 20[a] 20[a] 20 Flavor 200 U2 12[a] 12[a] and 16 12 Flavor U3 2 3 and 8 1 Essence U4 5[a] 5[a] 5 Flavor 100 Sample set 2 U5 14[a] 14[a] 14 Flavor 0 U6 21[a] 7, 10, 13, 21[a], 22, 23 21 Flavor

Flavor U7 15 20 3 Essence U8 16 7, 10, 13, 21, 22[a], 23 22 Flavor 400 [a] Indicates a correct match with the source bottle.

300 shown in figure 2. The first of the clusters contains flavor 19. Relative volatile abundance (pA) The other cluster is split into two subclusters: flavors U5 and

200 14, and all remaining coded flavors and essences. Within each designated cluster, HCA indicated sample similarity groupings that were used to match and differentiate unknown 100 coded samples to known coded samples. The Cyrano Sciences chemical sensor correctly identified all six unknown flavors. Three were absolute matches, and 0 0246 three matched more than one unknown (one of which was the correct flavor). The two unknown essences were not Time (min) identified correctly. Figure 1. Composite gas chromatograms of flavors and essences. Com- posite gas chromatograms were prepared by merging gas chromato- COMPARISON OF UNKNOWN DETERMINATIONS graphic data from three representative flavors and from three Visual comparison of gas chromatograms and HCA representative essences. Slight peak broadening resulted from minor variations in peak elution times among the three individual gas chromato- sample groupings was used to identify and discriminate grams associated with each plot. unknown and known coded samples and source bottle designations (table 1). Table 1 also lists the sample type, among the chromatograms of the flavor samples to find either flavor or essence. A comparison of the results showed absolute matches for five of the six unknown flavors by visual that in sample set 1 (U1, U2, U3, and U4), the gas inspection (table 1). The chromatograms of the two unknown chromatographic and the Cyrano Sciences chemical sensor essences were too similar to one another and to several of the data analyses correctly matched the three unknown flavors flavors to correctly identify, even though many of the known (U1, U2, and U4). The analysis of the chemical sensor data flavor samples could be excluded due to their relatively low for U2 did not exclude incorrect sample 16. Sample set 2 (U5, volatile abundance. From the gas chromatographs, four of the U6, U7, and U8) was measured after an additional one-month flavors had high levels of ethanol, which may have been storage period and may have undergone some storage-related added as a flavor enhancer (data not shown). This segregated degradation before measurement. For set 2, the chemical those samples from those without high levels of ethanol and sensor analyses correctly identified the three unknown increased the chance of obtaining a correct unknown sample flavors (U5, U6, and U8) but did not exclude incorrect identification. matches for samples U6 and U8. The gas chromatographic technique led to two correct unknown flavor matches (U5 and CYRANO SCIENCES CHEMICAL SENSOR MEASUREMENTS U6) and one incorrect match (U8). Neither analytical STATISTICAL ANALYSIS technique correctly identified the unknown essences (U3 and Principal component analysis (PCA), multivariate hierar- U7). Both techniques incorrectly identified the unknown chical cluster analysis (HCA) and essences as flavors having relatively high total volatile analyses were run with SAS software (SAS, 1997). These abundances. analyses were used to discriminate among various samples. These results indicate that the chemical sensor’s capabili- The mean of five volatile samplings for each flavor, essence, ty to perform this kind of separation shows promise but is not or unknown coded sample was used for data analysis. The consistent (two absolute matches for sample set 1, and only data sets from 31 sensors of the data logger were analyzed one absolute match in sample set 2). This compares to three using PCA. The PCA successfully reduced the data sets from absolute matches in set 1 and two in set 2 for the GC. The the sensors to two principal factors, which accounted for accuracy and discriminatory ability of the chemical sensor 96.4% of the variance in the sensor measurements. The data technique would be improved by identifying those sensors or from the two principal factors were analyzed using HCA. sensor combinations that respond differentially to flavor- and The results of the HCA analysis using average linkage for essence-related aromatic volatile headspace gases and are cluster formation suggest two clusters. The cluster output is little affected by interfering gases such as CO2 and H2O.

Vol. 48(5): 2003−2006 2005 3.0

2.5

2.0

1.5

1.0 Average Distance Between Clusters Average

0.5

0.0 124620U1U7 38U31115 9 18 12 16 U2 17 24 10 13 23 7 21 22 U6 U8 5 U4 14 U5 19 Sample

Figure 2. Cyrano Sciences chemical sensor HCA cluster analysis of flavors and essence samples 1 to 24 and eight unknowns (U1 to U8). The HCA input was the first two principal factor results from a PCA analysis that includes 96.4% of the variance. All sensors except 31 were used in the analysis.

Improved performance of chemical sensors may also occur REFERENCES following detailed statistical analyses of data sets from Arthur, C., L. Killam, S. Motlagh, M. Lim, D. Potter, and J. different sensors or sensor sets (Marrazzo, 1999). In addition, Pawliszyn. 1992. Analysis of substituted benzene compounds in alternate types of classification techniques such as neural groundwater using solid-phase microextraction. Environ Sci. networks (that are trained with known samples in order to and Tech. 26(5): 979-983. classify unknowns) may provide better results. Dimick, P. S., and J. C. Hoskin. 1983. Review of apple flavor - State of the art. CRC Critical Reviews in Food Science and Nutrition 18(4): 387-409. Doleman, B. J., R. D. Sanner, E. J. Severin, R. H. Grubbs, and N. S. CONCLUSIONS Lewis. 1998. Use of compatible polymer blends to fabricate Data sets from 31 sensors of the data logger were analyzed arrays of carbon black-polymer composite vapor detectors. Anal. using PCA. The PCA analysis of the data sets identified two Chem. 70(13): 2560-2564. principal factors. Further analysis of the chemical sensor data Heiden, A. C., C. Gil, and L. S. Ramos. 2002. Rapid screening of sets using HCA with average cluster formation indicated that headspace samples: Pros and cons of using MS-based electronic 22 of the 24 flavors and essences clustered together and thus noses versus fast chromatography. Application note. Mulheim an der Ruhr, Germany: Gerstel GmbH. could not be distinguished from one another. The same InfoMetrix. 1999. Electronic nose instrumentation: Comparing analysis provided a breakdown of the clusters into subgroups evaluation technologies. Infometrix application note. by similarity. These were used to determine “matches” with Woodinville, Wash.: InfoMetrix, Inc. six unknown flavor and two unknown essence samples. The MacGregor, D. R., H. Sugisawa, and J. S. Matthews. 1964. Apple chemical sensor and clustering technique correctly identified juice volatiles. J. Food Sci. 29: 448-455. all six unknown flavors (three were absolute matches). By Marrazzo, W. N. 1999. Feasibility study for the utilization of comparison, a gas chromatographic technique was less Cyrano Sciences chemical sensor data logger technology for ambiguous (all matches were absolute) but also less accurate comparing apple, Malus domestica Borkh, headspace gases. (five unknowns were correctly identified) at identifying the PhD diss. University Park, Pa.: The Pennsylvania State unknown flavor samples. Neither technique identified the University. Morton, I. D., and A. J. MacLeod. 1990. Chapter 1: The flavour of unknown essences. apples, pears, and quinces. In Food Flavours: Part C. The Flavour of Fruits, 1-41. Developments in Food Science, vol. 3. ACKNOWLEDGEMENTS Amsterdam, The Netherlands: Elsevier. A grant from the USDA and a grant from Cyrano Sciences Sarig, Y., and M. Delwiche. 1998. Utilization of the olfactory supported this research. Additional support was provided by characteristics of fruits and vegetables as a potential method for The Pennsylvania State University, the USDA-ARS Appalachi- determining their ripeness and readiness for harvest: A review. an Fruit Research Center (Kearneysville, W.V.), the USDA- NE-179. Davis, Cal.: University of California at Davis. ARS-BARC Produce Quality and Safety Laboratory SAS. 1997. SAS/STAT Software: Changes and Enhancements (Beltsville, Md.), and Agilent Technologies (Rockville, Md.). through Release 6.12. Cary, N.C.: SAS Institute, Inc.

2006 TRANSACTIONS OF THE ASAE