MS and multivariate data analysis

MS and multivariate data analysis

E. T. Contis et al. (Editors) Food Flavors: Formation, Analysis and Packaging Influences © 1998 Elsevier Science B.V. All rights reserved 159 Determ...

701KB Sizes 23 Downloads 25 Views

E. T. Contis et al. (Editors) Food Flavors: Formation, Analysis and Packaging Influences © 1998 Elsevier Science B.V. All rights reserved

159

Determination of the Cause of Off-Flavors in Milk by Dynamic Headspace GC/MS and Multivariate Data Analysis R. T. Marsili and N. Miller Research Dept., Dean Foods Technical Center, P.O. Box 7005, Rockford, IL 61125, U.S.A.

Abstract Milk is susceptible to off-flavor development by a variety of mechanisms. Multivariate analysis of GC data can be used to determine the cause of off-flavor development in customer complaint milk samples. Control samples of normal, good-tasting milk were analyzed by dynamic headspace GC/MS and a GC/FID technique that measures free fatty acids. The samples were then subjected to common abuse conditions (exposure to light, copper, and sanitizer) at various levels and re-analyzed. The abused samples provided the basis for KNN and SIMCA classification modeling. Results show that multivariate analysis can accurately predict the type of sample abuse responsible for off-flavors.

1. INTRODUCTION Milk is susceptible to formation of off-flavors by various mechanisms. Shipe et al. [1] have listed seven descriptors of off-flavor in milk based on causes: heated, lipolyzed, microbial, transmitted (from feed and weeds), light-induced, oxidized, and miscellaneous. Light-induced off-flavors, undoubtedly the most common flavor defect in milk, have two distinct components. Initially a burnt, activated sunlight flavor develops and predominates for about two or three days. Degradation of sulfur-containing amino acids of the serum (whey) proteins has been blamed for this reaction. The second component is attributed to lipid oxidation. This off-flavor, often characterized as metallic or cardboardy, usually develops after two days and does not dissipate. It has been estimated that exposure of milk in blow-mold plastic containers to fluorescent lights in supermarket dairy cases is responsible for the development of light-induced off-flavors in some 80% of store samples [2]. Two other common causes of off-flavor in milk are contamination with pro-oxidant metals (especially copper) and contamination by sanitizer (especially peroxyacetic acid-based sanitizer). Copper can be transmitted to milk from feed sources and leached from pipes and valves used in processing equipment. Sanitizer which hasn't been completely rinsed from processing lines after cleaning can contaminate milk. Peroxyacetic acid and other popular new robust sanitizers do an excellent job sanitizing processing equipment between runs, but because of their improved stability, they have a long lifetime in milk and can increase the risk of off-flavor development. These three mechanisms of off-flavor development can potentially generate off-flavors in

160 milk by degrading polyunsaturated fatty acids in milkfat triglycerides and phospholipid fractions and/or by degrading milk proteins. Each of these three types of off-flavor mechanisms can potentially generate similar oxidation byproducts (e.g., hexanal) in milk [3]. Deciding which mechanism is responsible for the off-flavor in a particular sample is impossible to do simply by tasting samples — even when trained organoleptic evaluators are used. The goals of this study were: (a) to determine if multivariate statistical analysis of gas chromatographic data can be used to classify milk with off-flavors by the type of abuse mechanism and by the level of abuse which has occurred (i.e., low, medium, high exposure levels); and (b) to use multivariate analysis to indicate which chemical byproducts are the best indicators of these three types of abuse mechanisms. Classification modeling can be conducted with multivariate analysis techniques and involves the computation and graphical display of class assignments based on multivariate similarity of one sample to others. One example of this technique is identification of bacteria based on fatty acid profiles of lipid extracts from bacterial cell walls [4]. Multivariate analysis has been applied to the study of a wide variety of food and beverage problems. A few examples include milk shelf-life prediction [5], chemotyping of essential oils [6], discriminating aromas of coffee samples [7], classifying wine samples [8], and characterizing peppermint oils [9]. Recently, Horimoto et al. used Principal Component Similarity Analysis (PCSA) for the classification of microbial defects in milk based on dynamic headspace GC data [10].

2. EXPERIMENTAL 2.1. Sampling for classification modeling One set of samples, consisting of homogenized whole-fat milk (3.3% fat), 2% fat milk, skim milk (<0.5% fat), and the raw milk from which they were made, was obtained from three different dairies (Chemung, IL; Albuquerque, NM; and Murray, KY). Milk was intentionally abused to provide samples that could be used to calibrate statistical prediction models. Each of the 12 samples was subjected to the following abuse conditions and assigned a classification code corresponding to one of the 10 class assignments indicated (in bold) below: Class A: Light abuse: • Class Al; Low level exposure (150 footcandles fluorescent light for 3 hr at 4°C; analyzed after 24 hr of storage at 4°C). • Class A2; Medium level exposure (150 footcandles fluorescent light for 8 hr at 4°C; analyzed after 24 hr of storage at 4°C). • Class A3; High level exposure (150 footcandles fluorescent light for 24 hr at 4°C and then analyzed). Class B: Copper abuse: • Class Bl; Low level contamination (5 ppm Cu added as cupric chloride, incubated 24 hr at 4°C and analyzed). • Class B2; Medium level contamination (50 ppm Cu added as cupric chloride, incubated 24 hr at 4°C and analyzed).

161 • Class B3; High level contamination (50 ppm Cu added as cupric chloride, incubated 72 hr at 4°C and analyzed). Class C: Sanitizer abuse. (Oxonia™ active sanitizer was used. It is a peroxyacetic acid sanitizer used extensively in the dairy industry and is the trademark of Ecolab, Inc., St. Paul, MN.): • Class CI; Low level contamination (400 ppm Oxonia active sanitizer, incubated at 4°C for 24 hr and analyzed). • Class C2; Medium level contamination (800 ppm Oxonia active sanitizer, incubated at 4°C for 24 hr and analyzed). • Class C3; High level of contamination (1200 ppm Oxonia active sanitizer, incubated at 4°C for 24 hr and analyzed). Class D: No abuse (fresh, control, good-tasting milk). The above sampling scheme generated a total of 120 samples to be used for classification modeling by KNN and SIMCA. GC analysis of these 120 samples generated approximately 80 different peaks for each sample, resulting in approximately 9,600 peak areas that were initially considered for use in classification modeling.

2.2. Gas chromatography methodologies 2.2.1. Dynamic headspace GC/MS methodology A milk sample (20 g) was heated to 40°C and purged with helium at a rate of 15 mL/min for 20 min using the Purge and Trap System from Scientific Instrument Services (S.I.S., Ringoes, NJ). The helium dry-purge line was set at a 10 mL/min flow rate. Volatiles were trapped on a Tenax-filled desorption cartridge. After the 20 min sampling period, the Tenax cartridge was removed from the S.I.S. system and placed in the desorption chamber of a CDS PeakMaster concentrator (CDS Analytical, Oxford, PA). Volatiles from the trap were desorbed by heating the cartridge to 185°C and purging with helium at a rate of 15 mL/min for 10 min. Volatiles were then trapped onto a second Tenax adsorbent trap in the CDS PeakMaster concentrator. After collection on this trap, an additional dry-purge cycle was used to eliminate any water vapor that may have accumulated on the trap during stripping. Dry purging was conducted at a temperature of 40°C for 4 min. The volatile components were then desorbed onto the analytical column from the second trap by heating to 200°C for 10 min and cryofocused at -100°C with liquid nitrogen prior to column injection. The analytical column used was a 30 m x 0.25-mm i.d. DB-5 fused-silica capillary column with a film thickness of 1 |Lim (J&W Scientific, Folsom, CA). The initial column temperature was 50°C for 1 min, then heated to 180°C at a rate of 6°C/min and held at 180°C for 4 min. The column was then heated to 240°C at 6°C/min and held at 240°C for 8 min. The performance of various types of capillary columns for the analysis of volatile chemicals in dairy products by dynamic headspace GC was evaluated by Imhof and Bosset [10]. This study showed best performance was obtained with a capillary column coated with a thick film of polydimethylsiloxane. GC/mass spectrometry was performed with a Varian Saturn 3 system from Varian Analytical Systems (San Fernando, CA).

162 2.2.2. Free fatty acid GC/FID methodology Milk samples were analyzed for free fatty acids by the method described by Deeth [12]. Approximately 10 g of milk and 5 mL of a C17 carboxylic acid internal standard solution (20 mg diluted to 100 mL with ethyl ether) were subjected to solid phase extraction with an alumina column. After washing with 2 x 5 mL portion of hexane/ethyl ether (1:1, v/v), the free fatty acids were eluted off the alumina with a 6% solution of formic acid in diisopropyl ether. The column used for FFA analysis was a 30 m x 0.25-mm i.d. FFAP fused-silica column with a film thickness of 0.25 |im (J&W Scientific, Folsom, CA). The column was heated to 140°C for 1 min, then heated to 240°C at a rate of 15°C/min and held at 240°C for 20 min. The injection volume was 2.0 |iL, and the split ratio was 30:1. The injector temperature was 250°C, and the flame ionization temperature was 260°C. This test measured acetic acid from the decomposition of peroxyacetic acid. The test can also be used to monitor free fatty off-flavors generated from active lipolytic enzymes reacting with milkfat and free fatty acids produced as psychrotrophic bacterial metabolites. 2.3. Multivariate classification methods Multivariate analysis was conducted with Pirouette™ multivariate data analysis software from Infometrix (Seattle, WA). Two types of classification techniques were investigated: KNearest Neighbors (KNN) and Soft Independent Modeling of Class Analogy (SIMCA). The KNN method is a similarity-based classification method which attempts to categorize unknown samples exclusively on their multivariate proximity to other samples of preassigned categories. In contrast to KNN, which is based simply on Euclidean distances among sample points, SIMCA develops principal component models for each class or category in the training set. When classifications are attempted for unknown samples, a comparison is made between the unknown's data and each class model. The class model which best fits the unknown, if any, is the class assigned to that sample. While KNN and SIMCA are both similarity-based techniques, their calculation approaches are distinctiy different. KNN is more appropriate to use for a sample-poor environment. KNN measures the Euclidean distance between the unknown sample and each of the known samples in the training set. The category assignment for the unknown is made by a plurality vote of the nearest neighbors, with an option of one to k neighbors considered. KNN is tolerant of sample-poor situations and is the only method which works well when categories are strongly subgrouped. Although SIMCA is not well suited for the case where there are only a few samples per category, SIMCA models the location and distribution of a category in the measurement space by constructing a principle component representation of this distribution for each category. Class assignments for unknown samples are based on their proximity to the nearest category model. SIMCA creates factor-based models of each category membership, with three possible outcomes: sample is a member of one category, sample is a member of more than one category, sample is not a member of any category. 2.3.1. Exploratory data analysis Peak area data for 80 different GC peaks from 120 milk samples were combined into a Pirouette spreadsheet. Peak areas were converted to peak area ratios by dividing the area of each peak by the area of an internal standard peak. 4-Methyl-2-pentanone was the internal

163 standard used for purge-and-trap analysis, and heptadecanoic acid was the internal standard used for free fatty acid analysis. A second internal standard, 2-ethylhexyl acetate, was also employed for dynamic headspace testing but was only utilized as a tool to help locate identical chemical peak components from chromatogram to chromatogram. Three-dimensional Principal Component Analysis (PCA) scores plots and PCA loadings plots were created with the Pirouette Exploratory algorithm and examined to see how well class groupings were accomplished and which independent variables (GC peak area ratios) contributed most to class differentiation. By using this technique to eliminate peaks that contributed little to sample classification assignments and by discarding peaks that were identified by mass spectrometry to be background noise peaks (e.g., chemicals from GC septa), the original number of independent variables was reduced from 80 to only 11. Also, inspection of dendograms created with the Hierarchical Cluster Analysis (HCA) algorithm helped to visualize sample class grouping and was useful in deciding which independent variables to eliminate in order to improve clustering of similarly abused samples. Class clustering in the PCA scores plot was significantly improved after performing two minor data transformations. Specifically, all peak area ratios for dimethyl disulfide were multiplied by 10, and all peak area ratios for acetic acid were multiplied by 100. These transformations increased the influence of dimethyl disulfide and acetic acid peak area ratios in determining class groupings.

V c s X C2,C3 • • - - - " > • 1C2

A2 X /

X

'^-

A3 ,

••••'">'

V*iii>5 C1 X ^ * - - - ' ^ ^ \ • IM ,.-••'

B2,B3f.i

\

)

\

Figure 1. PCA scores plot for all 120 samples using 11 independent variables (i.e., chromatographic peak area ratios). Plot shows clustering of samples according to abuse classification category where: Al, A2, and A3 represent low, medium, and high light abuse levels, respectively; Bl, B2, and B3 represent low, medium, and high copper abuse levels, respectively; CI, C2, and C3 represent low, medium, and high sanitizer abuse levels, respectively; and D represents control (non-abused) milk samples. Sample scores tend to fall primarily on principal component axes.

164 Performing these data transformations (and thereby giving dimethyl disulfide and acetic acid more significance) appeared justified for two reasons: (a) Acetic acid is a known degradation product of peroxyacetic acid, and dimethyl disulfide is a known photoxidation product of methionine, a sulfur-containing amino acid in milk proteins [13]; and (b) class clustering was significantly improved as a result of making these data transformations. When exploratory analysis was performed on the transformed data set (120 samples and 11 independent variables), the PCA scores plot showed that sample scores fall primarily on principal component axes (Figure 1). In fact, the three factors are largely associated with only one original variable, where the acetic acid peak is the only indicator of Oxonia sanitizer, the dimethyl disulfide peak is the most significant indicator of light abuse, and the hexanal peak is the primary indicator of copper abuse. 2.3.2. K-Nearest Neighbors After data reduction and transformation, a KNN model was created using no preprocessing and setting the k value at 5. The model was then saved to allow for predictions of abuse class assignments for unknown off-flavor samples based on peak area ratios for the 11 peaks used in the model. The optional value of k was set at 5 neighbors because the fewest number of misses occurred when k = 5. This means that to achieve the best prediction rate, the "votes" from the 5 closest samples to an unknown should be polled. Because only three independent variables (acetic acid, dimethyl disulfide, and hexanal) were shown to be the primary indicators of abuse type and level, a second KNN model was created using the 120 samples and only these three independent variables. The purpose for this was to determine if milk samples could be as accurately classified with only three chemical indicators as with 11 chemical indicators. 2.3.3. SIMCA modeling A SIMCA model was created using the 11 peak-ratio variables for each of the 120 samples. For SIMCA modeling. Preprocessing was set at None, and the Number of Components was set at 5. SIMCA classification modeling provided diagnostic information about which variables to use. For example, the Distance Object for diagnosing outliers is a plot of Mahalanobis distance vs. sample residual for each class assignment. Using this plot, six of the 120 samples were identified as outliers. Two SIMCA models were created: one model with all 120 samples (including Mahalanobis outliers) and the other with Mahalanobis outliers excluded. In addition, a third SIMCA model was created to see how few variables could be used for class prediction and how accurate predictions were with significandy less independent variables.

3. RESULTS AND DISCUSSION 3.1. Accuracy of KNN and SIMCA models in predicting abuse class 3.1.1. KNN prediction results The 120 abuse samples were treated as unknowns and tested by the KNN model to see how accurately abuse class assignments could be predicted. Two KNN models were tested:

165 one using 11 independent variables and one using only three independent variables (area ratio peaks for acetic acid, dimethyl disulfide, and hexanal). The model with the 11 variables correctly classified 103 of the 120 samples (86%) according to the 10 classification assignments. Examination of misclassified samples showed that when misclassification occurs, frequently it is not because samples are assigned to the wrong class based on type of abuse (none, light, copper, sanitizer); rather, the level of abuse is not properly estimated. Therefore, when class predictions were made based on only four categories (A = light abuse; B = copper abuse; C = sanitizer abuse; and D = no abuse) which ignore level of abuse, the 11-variable KNN model is able to correctly classify 93% of the 120 samples. When the samples in the 120-sample data set were treated as unknowns and analyzed for class assignments by the 3-variable KNN model, 101 of the 120 samples (84%) were correctly classified by both type and level of abuse, and 112 of the 120 samples (93%) were correctly classified by type of abuse but not necessarily level of abuse. Also, examination of the Misclassification Matrix revealed that the KNN model tends to classify class Al samples (low level of light abuse) as control (non-abused) samples. This type of misclassification is not unusual, since low level light exposure (150 footcandles for only 3 hrs) does not generate significant off-flavors in milk. For example, in one sensory taste panel experiment, only 2 of 12 people were able to detect a perceivable off-flavor in class Al homogenized whole milk, 2% fat milk, and skim milk samples. Furthermore, chromatograms of most of the class Al milk samples were nearly identical (both quantitatively and qualitatively) to class D chromatograms. With KNN modeling, accurate classification of samples can be accomplished with only three variables. However, the accuracy of classification indicated by the KNN models is misleading, since classifications were performed on the same samples used to develop the KNN model. In the future, abused samples not used for KNN modeling will be classified with the model to evaluate how well classification accuracy is performed. 3.1.2. SIMCA prediction results When the 120 abuse samples were examined by the SIMCA model, class predictions were less accurate than with the KNN model. The SIMCA model with all 120 abuse samples and 11 variables correctly classified (i.e., accurately predicted both type of abuse and extent of abuse) only 63% of the samples. With Pirouette software, the actual class assignments are presented in a tabular format. The first column provides the best ("Best") estimate of the class membership for the test samples, and the next column provides the next best ("NxtBstl") estimate of the class membership. When the next best estimates were examined, the model correctly classified 83% of the samples as the best or next best estimate of the class membership. This SIMCA model accurately classified 114 of the 120 samples (95%) by correct abuse type but not necessarily by correct abuse level. When the SIMCA model with the six sample outliers removed was used, the accuracy of predicted classes was slightly improved. The SIMCA model with the 114 samples correctly classified 67% of the samples as the best estimate, and 89% were correctly classed as the best or next best class estimate. This SIMCA model accurately classified 106 of the 114 samples (93%) by correct abuse type but not necessarily by correct abuse level. A SIMCA model was created with as few variables as possible. With this data set, the

166 Table 1 Accuracy of abuse class predictions made by KNN and SIMCA modeling I. KNN model: Correct Type But Correct Class Not Necessarily Level Sample Set Predicted of Abuse Predicted 86% 120 Samples, 11 Variables 93% 84% 120 Samples, 3 Variables 93% II. SIMCA model: Correct Type But Correct Predicted Not Necessarily Level Best or NxtBst Sample Set of Abuse Predicted 95% 83% 63% 120 Samples, 11 Variables 93% 89% 67% 114 Samples^, 11 Variables 86% 75% 48% 114 Samples^ 5 Variables^ ^Six Mahalanobis outiiers were identified in the data set and discarded in this model. ^The five variables were acetic acid, dimethyl disulfide, hexanal, heptanal, and nonanal. Correct Predicted as Best

Figure 2. PCA loadings plot for 120 milk samples using 11 variables (i.e., chromatographic peak area ratios). Hexanal (1), dimethyl disulfide (2), and acetic acid (3) are the key chemical indicators of sample abuse.

Figure 3. PCA loadings plot for control (nonabused) samples and all copper abused samples. Hexanal (1), heptanal (2), octanal (3), nonanal (4), oct-l-en-3-one (5), pentanal (6), and isopentanal (7) are the key copper abuse indicators.

Pirouette software would allow no fewer than five independent variables to be used for modeling. The independent variables included acetic acid, dimethyl disulfide, hexanal, heptanal, and nonanal. As indicated in Table 1, unlike KNN modeling, the accuracy of prediction significantly suffered when fewer independent variables were used for modeling. As in the case with KNN predictions, the accuracy of SIMCA predictions in Table 1 is

167 probably favorably biased because the same samples used to calibrate models were also used as "unknowns" to estimate the prediction accuracy. 3.2. Chemical markers as indicators of abuse mechanisms The PCA loadings plot shown in Figure 2 reveals which chemicals were the most significant contributors to determining class assignments when all 120 abuse samples and 11 variables were considered. These chemicals were dimethyl disulfide, hexanal, and acetic acid. By examining PCA loadings plots for each abuse class separately with all control (nonabused) samples, it was possible to determine the specific reaction byproducts that are the most important indicators for each type of abuse. For example, as determined by examination of PCA exploratory results, including the loadings plot (Figure 3), loadings values table, and eigenvalue table, the key chemical indicators of copper abuse are hexanal»>heptanal, octanal, nonanal>oct-l-en-3-one>pentanal>isopentanal. Of these chemicals, hexanal is produced in highest concentration, but oct-l-en-3-one is likely the chemical most responsible for the metallic off-flavor of the copper-abused samples [14]. The only significant marker for sanitizer abuse was acetic acid, a decomposition product of peroxyacetic acid, and the most significant indicators of light abuse were dimethyl disulfide and hexanal. While acetic acid was the only chemical marker revealed in this study, it is not the chemical responsible for the typical off-flavor noted with samples contaminated with peroxyacetic acid. These samples have a peroxide flavor, and the specific chemicals responsible could not be detected by the tests used in this study. Sample chromatograms of a fresh raw milk (non-abused) sample and a raw milk sample abused by light and copper are shown in Figure 4. This figure shows the 10 dynamic headspace chemicals (marked with an asterisk) that were used for classification modeling. An additional chemical, acetic acid, was quantitated by the free fatty acid GC test, and results for acetic acid were also included in the creation of the KNN and SIMCA models. 3.3. Unexpected results that impact flavor: ester degradation by exposure to light, copper, sanitizer, and heat During the early stages of data analysis, the number of independent variables (peak area ratios) was reduced firom 80 to 12. During the process involved in data reduction, a threedimensional PCA scores plot and a PCA loadings plot were created with 12 independent variables. This PCA loadings plot (Figure 5) shows that methyl butyrate has a significant influence on how sample groupings were made. Elimination of methyl butyrate from the data set significantly improved class groupings in the PCA scores plot, so it was not included in the KNN or SIMCA modeling. However, because the 12-variable PCA loadings plots showed that methyl butyrate demonstrated significant variance between samples, this peak was more closely scrutinized in sample chromatograms. Aqueous standard solutions of methyl butyrate, methyl caproate (observed and identified in several control raw milk samples), and ethyl butyrate along with 4-methyl-2-pentanone internal standard were analyzed to allow quantitation of these peaks in the samples. (Note: Although ethyl butyrate was not detected in any of the samples tested, it was included in the experiment in order to be sure the peak identified as methyl butyrate was methyl butyrate and not ethyl butyrate.) Quantitative results for these esters in some of the samples are shown in Table 2. These results show that ester concentrations are highest in fresh raw milk samples but are lost after

168

Control Raw Milk. No Abuse. Class P

-TIC

1. acetone

IS1

2. dimethylsulfide

3. 2-butanone 4. chloroform 5. dichloroethane

6. 3-methyl-2-butanone

13. limonene 14. nonanal* 15. decanal 151 « 4-methyl-2-pentanone 152 = 2-ethlyhexyl acetate B = chemical from GC septum

7. methyl isobutyrate 8. 2-pentanone

10

IS2

9. pentanal* 10. methyl butyrate 11. hexanal* 12. methyl caproate* B

6 5. 7

11

12 14 13

15

660

Light Abused Raw Milk. Class A3 16. dimethyl disulfide*

V JUJ

16

wUww T" T r Copper Abused Raw Milk. Class B2| 21. heptanal* 17. isopentanal* 22. 2-heptanone 18.1-p6ntanol(T)* 19. 4-methyl-1 -pentanol (T) 23. 2-heptenal (T) 24. oct-1-en-3-one* 20.1-hexanol(T) 25. octanal* 24 26. 2-oct6nal (T) 27. l-heptanol(T) 28. 2-nonenal (T) (T)=Tentative i.d. 25 by mass spec.

w Figure 4. Examples of dynamic headspace chromatograms of raw milk samples showing chemicals used for KNN and SIMCA modeling (* = chemicals used for modeling).

169 Table 2 Concentration of methyl butyrate (ppb), showing degradation effects of light, copper, sanitizer, and pasteurization.

Figure 5. PCA loadings plot for all 120 milk samples using 12 independent variables; methyl butyrate (1) included along with the 11 independent variables used in Figure 2. Hexanal (3) and acetic acid (2) also demonstrate significant betweensample variance.

No. 1^ No. 2 No. 3 Class Type 72 D Raw 71 7.3 49 Raw Al 77 7.1 — — 21 A2 Raw — Raw 4.3 5.9 A3 — 6.2 Raw Bl 50 12 Raw B2 5.5 5.5 1.4 8.2 Raw 4.6 B3 3.2 Raw 7.3 3.7 CI 3.2 3.4 Raw C2 7.3 4.1 2.3 Raw 7.3 C3 <0.2 <0.2 D 2.5 Homo 2% D 1.8 0.7 1.9 Skim 1.2 D <0.2 1.7 ^^o. 1, No. 2, and No. 3 refer to milk from three different processing plants.

exposure to light, copper, and sanitizer. Furthermore, when ester concentrations in homogenized whole fat, 2% fat, and skim milk samples made from these raw milk samples are examined, the data show that ester loss is even more significant. More than likely, heating during pasteurization is responsible for this loss. The implications of these findings are significant and may explain why raw milk is generally regarded as having a more desirable flavor than processed milk. A study by Moio et al. [15] concluded that esters of butanoic and hexanoic acid were the most important contributors to the flavor of bovine raw milk. In this work, the researchers used the Charm (Combined Hedonic Response Measurements) method to identify the odor-active components in bovine, ovine, caprine, and water buffalo milk. Dairy processors may be able to improve the flavor and market appeal of their dairy products by devising some way of maintaining/stabilizing ester levels during heat processing. Adding esters in an encapsulated form or adding them back to milk after pasteurization — but in such a way as not to compromise the microbiological integrity of products — could also be considered.

4. CONCLUSION Since all three abuse types studied involve oxidation reactions, prior to initiation of this study it was thought that each abuse mechanism would generate similar byproducts (predominantly aldehydes) and that the relationship between chemical peak areas generated would be multivariate in nature. However, results indicate the data is strongly univariate. Multivariate analysis is probably unwarranted for the purpose of distinguishing the three abuse types stud-

170 ied here, since the information can be generated by analysis of the three primary indicator peaks. Nonetheless, multivariate analysis with Pirouette proved to be a valuable tool for understanding the relationships between sample abuse treatments and peak area ratios. Advantages of using multivariate analysis for this work included: (a) PCA loadings plots and loadings values proved important in finding secondary reaction byproducts created by copper abuse and were useful in pointing out the ester degradation problem. (b) KNN and SIMCA classification techniques allow chemists with little experience in the interpretation of abuse sample chromatograms to accurately predict the cause of off-flavors in customer complaint samples when the cause is either Hght abuse, copper abuse, or sanitizer abuse. (c) As more types of abuse conditions are investigated, independent variable data may become more multivariate in nature, and results can be easily incorporated into the KNN and SIMCA models generated in this work to allow accurate predictions of abuse types. (d) Viewing sample data with PCA scores plots provides an excellent tool for comparing data in an historical database of customer complaint samples to see which types of abuses are occurring most often and from which dairy processing plants. (e) When unknown complaint samples are submitted and analyzed in the future, peak area ratios can be included in the Pirouette spreadsheet with the 120 samples, and residual plots can be examined with various Pirouette algorithms to see how the chemical profile of unknowns differs from the chemical profiles of the 120 samples in the database. Future studies will include investigation of additional mechanisms of abuse (e.g., heat, microbiological, etc.). Also, as data from more samples are collected, we will investigate making separate KNN and SIMCA models for each type of milk sample (i.e., a model for skim, a model for 2% fat milk, a model for homogenized whole fat milk, and a model for raw milk). This could significantly improve prediction rates, since peak areas for the apolar peaks tend to be lower as the fat content increases in the sample when samples are tested by dynamic headspace GC. When a KNN model was created with only skim samples, the model correctly classified 100% of the samples according to the appropriate abuse type and level when the skim samples were analyzed as unknowns. Similar results were obtained for raw, homogenized whole fat milk, and 2% fat milk samples when separate KNN models were made for these milk types.

5. REFERENCES 1 W.F. Shipe, R. Bassette, D.D. Deane, W.L. Dunkley, E.G. Hammond, W.J. Harper, D.H. Kleyn, M.E. Morgan, J.H. Nelson and R.A. Scanlan, J. Dairy Sci., 61 (1978) 855. 2 S.E. Bamard, J. Dairy Sci., 55 (1973) 134. 3 R.T. Marsili (ed.). Techniques for Analyzing Food Aroma, Marcel Dekker, Inc., New York, NY, p. 245. 4 L.S. Ramos, J. Chromatographic Sci., 32 (1994) 219. 5 B. Vallejo-Cordoba and S. Nakai, J. Agric. Food Chem., 42 (1994) 989 . 6 P. Ramanoelina, J. Viano, J. Bianchini and E.M. Gaydou, J. Agric. Food Chem. 42 (1994) 1177. 7 T. Aishima, J. Agric. Food Chem., 39 (1991) 753.

171 8 I. Moret, G. Scarponi and P. Cescon, J. Agric. Food Chem., 42 (1994) 1143. 9 F. Chialva, A. Ariozzi, D. Decastri, P. Manitto, S. Clementi and D. Bonelli, J. Agric. Food Chem., 41 (1993) 2028. 10 Y. Horimoto, K. Lee, and S. Nakai, J. Agric. Food Chem., 45 (1997) 733. 11 Rcn6 Imhof and Jacques O. Bosset, J. High Resolut. Chromatogr., 17 (1994) 25 . 12 H.C. Deeth, N. J. Dairy Sci. and Tech., 18 (1983) 13. 13 R.T. Marsili (ed.), Techniques for Analyzing Food Aroma, Marcel Dekker, Inc., New York, NY, p. 249. 14 W. Stark, D.A. Forss, J. Dairy Sci., 29 (1962) 173. 15 Luigi Moio, Dominique Langlois, Patrick Etievant, and Francesco Addeo, J. Dairy Research, 60 (1993) 215.