spectroscopic techniques

Determination of Omega-3 Fatty Acids in Fish Oil Supplements Using Vibrational Spectroscopy and Chemometric Methods Michael Yemane Bekhit,a,* Bjørn Grung,a Svein Are Mjøsa,b a b

Department of Chemistry, University of Bergen, P.O. Box 7803, Bergen N-5020 Norway Nofima BioLab, Kjerreidviken 16, Fyllingsdalen NO-5141 Norway

The potential of Fourier transform infrared (FT-IR), near-infrared (NIR), and Raman spectroscopic techniques combined with partial least squares (PLS) regression (PLSR) to predict concentrations of eicosapentaenoic acid (EPA), docosahexaenoic acid (DHA), and total omega-3 fatty acids (n-3 FAs) in fish oil supplements was investigated. FT-IR spectroscopy predicted EPA (coefficient of determination (R2) of 0.994, standard error of cross-validation (SECV) of 2.90%, and standard error of prediction (SEP) of 2.49%) and DHA (R2 ¼ 0.983, SECV ¼ 2.89%, and SEP ¼ 2.55%) with six to seven PLS factors, whereas a simpler PLS model with two factors was obtained for total n-3 FAs (R2 ¼ 0.985, SECV ¼ 2.73%, and SEP ¼ 2.75%). Selected regions in the NIR spectra gave models with good performances and predicted EPA (R 2 ¼ 0.979, SECV ¼ 2.43%, and SEP ¼ 3.11%) and DHA (R 2 ¼ 0.972, SECV ¼ 2.34%, and SEP ¼ 2.60%) with four to six PLS factors. Both the whole and selected NIR regions gave simple models (two PLS factors) with similar results (R2 ¼ 0.997, SECV ¼ 2.18%, and SEP ¼ 1.60%) for total n-3 FAs. The whole and selected regions of Raman spectra provided models with comparable results and predicted EPA (R2 ¼ 0.977, SECV ¼ 3.18%, and SEP ¼ 2.73%) and DHA (R2 ¼ 0.966, SECV ¼ 3.31%, and SEP ¼ 2.56%) with seven to eight PLS factors, whereas a simpler model (three PLS factors) with R2 ¼ 0.993, SECV ¼ 2.82%, and SEP ¼ 3.27% was obtained for total n-3 FAs. The results demonstrated that FT-IR, NIR, and Raman spectroscopy combined with PLSR can be used as simple, fast, and nondestructive methods for quantitative analysis of EPA, DHA, and total n-3 FAs. FT-IR and NIR spectroscopy, in particular, have the potential to be applied in process industries during production of fish oil supplements. Index Headings: Fourier transform infrared spectroscopy; FT-IR spectroscopy; Near-infrared spectroscopy; NIR spectroscopy; Raman spectroscopy; Fish oil supplements; Eicosapentaenoic acid; Docosahexaenoic acid; Omega-3 fatty acids; Chemometrics.

INTRODUCTION Omega-3 fatty acids (n-3 FAs) are families of long chain polyunsaturated fatty acids (PUFAs), which are Received 8 July 2013; accepted 19 March 2014. * Author to whom correspondence should be sent. E-mail: mybekhit@ yahoo.com. DOI: 10.1366/13-07210

1190

Volume 68, Number 10, 2014

found predominantly in fish and fish oils.1 The n-3 FAs and foods from which they are derived have long been recognized as beneficial to human health. Eicosapentaenoic acid (EPA, 20:5 n-3) and docosahexaenoic acid (DHA, 22:6 n-3) are two of the most important members of n-3 FAs, which have essential roles in the prevention and treatment of coronary heart disease, hypertension, arthritis, inflammatory and autoimmune disorders, and diabetes.2–5 Current dietary guidelines recommend a daily intake of 0.5–1.0 g of EPA and DHA for primary and secondary prevention of coronary heart disease, whereas higher pharmacological doses are recommended for treatment of hypertriglyceridaemia.6 The increased awareness of the health benefits of n-3 FAs, especially EPA and DHA, has led to increased consumer demand for fish oil supplements and other fortified foods containing these fatty acids.7 Currently, there are many brands of fish oil supplements available for consumers, most of which claim the active ingredients EPA and DHA on their labels. Rapidly increasing public demand for these products requires strict process control and high standards in quality assurance, which in turn requires simple, fast, and accurate analytical techniques for monitoring these fatty acids both during and after production. The official or routinely used method for analysis of n-3 FAs including EPA and DHA in fish oils is gas chromatography (GC).8,9 Although GC gives a detailed picture of the complete fatty acid composition of a sample, it is a relatively slow methodology and inappropriate for on-line control in processing industries.10–12 Analysis of triacylglycerol-based oils also requires that the sample is derivatized to fatty acid methyl esters. The derivatization is laborious and time consuming and may cause oxidation of lipids.8 With complex samples there is also a risk that chromatographic overlaps or incorrect identifications may affect the results.13 Recently, Fourier transform infrared (FT-IR), nearinfrared (NIR), and Raman spectroscopic techniques have become attractive and promising analytical tools for research, process control, and monitoring in indus-

0003-7028/14/6810-1190/0 Q 2014 Society for Applied Spectroscopy

APPLIED SPECTROSCOPY

trial laboratories. The techniques have analytical and instrumental characteristic advantages, which make them good alternatives to the traditional and more time-consuming analytical methods. They are rapid, do not require reagents, and are nondestructive. Limited or no sample preparation is required, and samples can be analyzed in almost any physical state (solid, liquid, gas, pastes, gels) and at any macroscopic or microscopic level.14 The possibility of coupling the instruments with suitable fiber optics is an important instrumental feature, which makes the techniques very attractive for process control and monitoring.15 Moreover, advances in chemical instrumentation, extensive use of computers, and development of appropriate chemometric methods have made spectroscopic methods suitable for quantitative analysis of a broad range of analytes and samples. The feasibility of using FT-IR, NIR, and Raman spectroscopy in combination with multivariate statistical methods for quantitative determination of fatty acids in edible oils, fats, and other food products has been well documented.16–20 FT-IR spectroscopy has been widely used for quantitative analysis of free fatty acids, characterization and authentication of vegetable oils, and determination of cis and trans content of fats and oils.21–24 There are also studies on determination of DHA and docosapentaenoic acid (DPA) in pork fat25 and analysis of total n-3 FAs in vegetable oils using attenuated total reflectance (ATR) sampling technique combined with multivariate statistical methods.26 Near-infrared spectroscopy has many applications in the food industry and is of particular use in the quality control of raw materials and food products.27 It has had its use in the determination of edible oil parameters,28 analysis of fatty acid composition in fats and oils,29,30 and prediction of fatty acid content.31 Free fatty acids and moisture content of fish oil have been determined using NIR reflectance spectroscopy.19 Fourier transform NIR spectroscopy has also been explored as a rapid method to determine free fatty acids in fats and oils.32,33 Recent developments and improvements in laser light sources, detector technology, optical fibers, and optical filters increased the use of Raman spectroscopy as a robust technique.34–36 It has been successfully used for quantitative analysis of omega-3 and omega-6 fatty acids in pork fat.37 Fatty acid content in butter38 and conjugated fatty acids in milk fat39 have also been predicted with Raman spectroscopy. Fourier transform Raman (FTRaman) spectroscopy has been used in the determination of total unsaturation in vegetable oils,40 detection of olive oil adulteration,41 and authentication of various oils and fats in the food industry.42 Although there are several publications on the analysis of fatty acids in dietary oils and fats using spectroscopic techniques, currently there are no reports on the application of these methods for quantification of n-3 FAs in fish oil supplements using multivariate calibration methods. The objective of this study was, therefore, to investigate the feasibility of using FT-IR, NIR, and Raman spectroscopic techniques for quantitative analysis of EPA, DHA, and total n-3 FAs in pharmaceutical fish oil supplements. Spectra of fish oil supplement samples were collected using FT-IR, NIR, and Raman spectrometers, and concentrations of EPA,

DHA, and total n-3 FAs were determined using GC as a reference method. Multivariate calibration models based on partial least squares regression (PLSR) were developed using the spectral data and concentration variables. The models were validated, and their predictive ability was also investigated, using chemometric methods.

MATERIALS AND METHODS Samples and Sample Preparation. Sixty-one fish oil supplements (omega-3 products) claiming to contain EPA and DHA were purchased from pharmacies or retail stores in France, Germany, Italy, Norway, Spain, and the United Kingdom. All of the products were in capsules that were liquid at room temperature. From each of the 61 products, four to eight capsules were pierced, and the fish oil was carefully squeezed out and collected in a clean 4 mL screw-cap glass vial. The samples were stored in a refrigerator at 4 8C without any pretreatment before spectroscopic analyses. All products were within shelf life when analyzed. Gas Chromatographic Analysis. Fatty acid methyl esters were prepared from 32 6 8 mg of the fish oil products (one or two drops, depending on viscosity) using a modified version of the American Oil Chemists’ Society (AOCS) official method Ce 1b-89.43 Two capsules of each product were analyzed using an Agilent 7890 gas chromatograph equipped with flame ionization detector and a BPX-70 column with L = 60 m, i.d. = 0.25 mm, and df = 0.25 lm (SGE, Ringwood, Australia). Blank samples and reference fish oil were prepared daily to check the quality of the reagents, and empirical response factors based on the reference mixture GLC-793 (Nu-Chek prep, Elysian, MN) were applied to correct the chromatographic areas. The chromatographic conditions were based on the methodology described in the paper by Mjøs and Solvang,44 where two different temperature programs were applied. The chromatographic fatty acid analyses used in this study were the same as the data used in a previous study of trans isomers of EPA and DHA,45 and further details on the GC methodology can be found in this paper. The concentrations of all the single fatty acids including EPA and DHA were expressed as percentages relative to the total mass of fatty acids. Total n-3 FAs were the sum of concentrations of eight single n-3 FAs: a-linolenic acid (18:3 n-3), stearidonic acid (18:4 n-3), eicosatrienoic acid (20:3 n-3), eicosatetraenoic acid (20 : 4 n-3), eicosapentaenoic acid (EPA, 20 : 5 n-3), heneicosapentaenoic acid (21:5 n-3), docosapentaenoic acid (DPA, 22:5 n-3), and docosahexaenoic acid (DHA, 22:6 n-3). Fourier Transform Infrared Spectroscopic Measurements. Fourier transform infrared spectra of the 61 fish oil samples were collected using a Nicolet Prote´ge´ 460 E.S.P. spectrometer (Nicolet Instrument Corporation, Madison, WI) equipped with infrared (IR) source, deuterated triglycine sulfate-potassium bromide detector, and KBr beam splitter. Sample measurement was carried out using a single-bounce ATR sampling accessory (ASI SensIR Technologies, Danbury, CT) equipped with a circular diamond crystal (1 mm effective diameter) mounted on a zinc selenide focusing element.

APPLIED SPECTROSCOPY

1191

Sample scan was initiated by carefully placing the fish oil onto the surface of the diamond crystal using a disposable Pasteur pipette until the crystal was completely covered. After each sample measurement the thin oil film was gently wiped off, and the crystal was thoroughly cleaned with hexane followed by methanol and distilled water and wiped dry using lint-free tissue paper. Sample spectra were collected at room temperature in reflectance mode, and each spectrum was a result of 32 co-added scans at a resolution of 4 cm1 over a spectral range of 4000–650 cm1. A background scan of the clean ATR crystal with no sample (i.e., air) was recorded every 30 min. All samples were scanned in duplicate with a two day interval between measurements. Data acquisition and instrumental settings were controlled using OMNIC software version 4.1a (Thermo Fisher Scientific Inc., Madison, WI). Spectra were imported to Sirius software version 8.5 (Pattern Recognition Systems AS, Bergen, Norway) as an ASCII file for data analyses. Raman Spectroscopic Measurements. Raman spectra of the 61 fish oil samples were recorded using a Raman spectrometer (RamanRxn1, Kaiser Optical Systems, Inc., Ann Arbor, MI) equipped with a NIR externalcavity-stabilized diode laser (Invictus, Kaiser Optical Systems, Inc., Ann Arbor, MI). Operating power of the laser was less than 500 mW continuous wave at an excitation wavelength of 785 nm. An air-cooled chargecoupled device (CCD detector) operating at 40 8C was used in the instrument. The instrument was connected to a fiber optic probe attached to a 20 cm long immersion optic with a sapphire window at its tip. The immersion optic had an outside diameter of 6 mm, and the sapphire window had a diameter of 3 mm. Measurements were carried out with the immersion optic in direct contact with the surface of the sample. To avoid cross contamination, after each sample measurement the tip of the immersion optic was thoroughly cleaned with hexane followed by methanol and distilled water and then wiped dry with clean lintfree tissue paper. The samples were measured in duplicate at room temperature with a two-day interval between each measurement. Spectra were collected in the wave number range of 3450–0 cm1, which was covered using two measurement windows. Intensity of the spectra was displayed as detector counts against Raman shift in wave numbers. Exposure time was adjusted to 5 s, and number of accumulation was set to 1. All data acquisition and control of the instrumental parameters were carried out using HoloGRAMS software version 4.0 (Kaiser Optical Systems, Inc., Ann Arbor, MI). The spectral data were exported to HoloReact software version 2.0 (Kaiser Optical Systems, Inc., Ann Arbor, MI) and then transferred to Matlab software version 7.4.0. (The MathWorks, Inc., Natick, MA). The data were imported to Sirius software version 8.5 (Pattern Recognition Systems AS, Bergen, Norway) for preprocessing and statistical analyses. Near-Infrared Spectroscopic Measurements. The 61 fish oil samples were scanned using FOSS NIRSystems 6500 spectrometer (Foss NIRSystems, Inc., Silver Spring, MD). A polystyrene internal wavelength linearization

1192

Volume 68, Number 10, 2014

standard and lead sulfide transmission detector were used in the instrument. The ceramic standard of the spectrometer was used as an internal reference. Samples were measured using the fiber optic probe of the instrument. The tip of the probe was screw-fitted with a circular camlock cell (15 mm diameter) made from an aluminum-plated quartz window. The cell pathlength was adjusted to 1 mm using a 0.5 mm feeler gauge blade. The sample was filled into the cell as a thin oil film using a disposable glass Pasteur pipette. After each measurement the cell was cleaned thoroughly with hexane followed by methanol and distilled water and then wiped dry using lint-free tissue paper. Spectra were collected in transflectance mode in the wavelength range of 1100–2500 nm with 2 nm intervals (700 data points) by co-adding 32 scans at a resolution of 4 cm1. Intensity of the spectra was displayed as absorbance against wavelength. An air-filled blank cell was scanned as a reference, and each sample was measured in duplicate at room temperature with a twoday interval between measurements. Spectral data acquisition and instrument system monitoring was carried out using Vision software version 2.11 (Foss NIR Systems, Inc., Silver Spring, MD). Spectra were exported as an ASCII file and transferred to Sirius software version 8.5 (Pattern Recognition Systems AS, Bergen, Norway) for statistical analyses. Spectral Preprocessing. Raman spectra collected with visible laser from biological samples are often dominated by background signal due to fluorescence from organic molecules and other contaminants. The spectra of the fish oil samples had severe fluorescence background, especially in the lower wave number region (1820–0 cm1), resulting in variable baseline offsets and masking of the Raman bands. Different preprocessing methods were tested on the Raman spectra to remove the fluorescence background. A least squares based polynomial curve fitting (with fourth-order polynomial and 15 iterations) was tested on the replicate Raman spectra with or without standard normal variate (SNV) transformation.46,47 In addition, Savitzky–Golay first derivatives (with second-order polynomial and 11 points) of the polynomial-curve-fitted Raman spectra were also used for calibration. The preprocessed Raman spectra of the two replicate measurements were averaged before multivariate modeling. Minor baseline shifts in the FT-IR spectra and overlapping peaks in the NIR spectra were observed. To remove these effects and investigate whether lower prediction errors could be obtained, the Savitzky–Golay first and second derivatives with second-order polynomial and 11 points were tested on both the FT-IR and NIR spectra. Both the raw and preprocessed FT-IR and NIR spectra were used for multivariate modeling. In this paper, only the preprocessing methods or combinations of methods that provided the lowest prediction error and model complexities are presented. Chemometric Analysis. The FT-IR, NIR, and Raman spectra covering the whole and selected spectral regions were used to develop all multivariate calibration models. Important spectral variables that contribute to modeling a specific response variable, EPA, DHA, or

TABLE I. Mean, range, and standard deviation of concentrations of all omega-3 fatty acids including EPA and DHA in all of the fish oil samples.

TABLE II. Correlation coefficients between concentrations of EPA, DHA, and total n-3 FAs. Correlation coefficient

Fatty acida

Mean

Range

Standard deviation

18:3 n-3 18:4 n-3 20:3 n-3 20:4 n-3 20:5 n-3 (EPA)b 21:5 n-3 22:5 n-3 22:6 n-3 (DHA)c n-3 FAsd

0.60 1.83 0.13 1.13 24.14 1.25 3.80 20.59 53.47

0.00–2.82 0.00–6.27 0.00–0.32 0.00–1.93 0.01–57.32 0.00–2.80 0.29–8.78 5.68–79.92 19.51–94.24

0.39 1.10 0.07 0.47 12.95 0.66 1.88 12.63 21.49

a

Percentage relative to total mass of fatty acids. b Eicosapentaenoic acid. c Docosahexaenoic acid. d Total omega-3 fatty acids.

total n-3 FAs (referred to as y variables) were selected based on variable selectivity ratio (SR) plots. SR plots provide more localized regions of the spectra and have better ability to focus on the most significant variables due to the higher sensitivity of the ratio compared with correlation coefficients between spectral variables and responses.48 A plot of variable selectivity ratio against the whole spectral range for each y variable was studied, and important spectral regions with high values of selectivity ratio were selected manually.49 Principal component analysis (PCA) was performed to detect outliers, to look for trends or groupings, and to examine relevant and interpretable structures in all the spectral data sets. All the spectral data were mean centered prior to PCA. In addition to PCA scores plot, residual standard deviation (RSD) versus objects or RSD versus leverage plots were also used to identify outlier samples. Partial least squares (PLS) regression was used to develop all the calibration models. Mean centering and a significance level of 0.05 (5% probability) were set as PLS model parameters prior to calibration. For each y variable, calibration outliers were identified and removed from the sample population. Models were developed using all the samples as a calibration set, and the models were validated using the full crossvalidation leave-one-out method. Standard error of cross-validation (SECV) expresses the prediction error of the full cross-validation method and is defined as vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi u N uX u ðy  y^  dÞ2 i u i t i¼1 SECV ¼ ð1Þ N 1 where yi is the measured response value and y^i is the predicted response value of the ith sample, where sample i was left out of the data set when performing the calibration, N is the number of samples in the calibration set, and d is the bias, which is the average of cross-validated prediction residuals. The optimum number of PLS factors for all the models were determined by plotting the SECV against the number of PLS factors and selecting the number of PLS factors that gave a minimum

Fatty acid

EPA

DHAb

n-3 FAsc

EPA DHA n-3 FAs

1.000 0.194 0.812

1.000 0.725

1.000

a b c

a

Eicosapentaenoic acid. Docosahexaenoic acid. Total omega-3 fatty acids.

value of SECV.50 In addition, values of cross-validation standard deviations for each PLS factor and percentage variance of independent and dependent variables explained by each PLS factor were also examined. Model performance was reported as coefficient of determination (R2) and the SECV. To investigate the stability and robustness of all the models, separate models were also developed and validated using independent validation (test) sets. In this study, samples were manually divided into a calibration and validation set in such a way that both sets showed approximately the same values of mean, range, and standard deviation of concentrations of each y variable.51 Models were developed using the calibration sets, and the prediction accuracy was then tested using the test sets. Model performance was reported as R2, standard error of prediction (SEP) and bias. All data analyses were carried out using Sirius software version 8.5 (Pattern Recognition Systems AS, Bergen, Norway).

RESULTS AND DISCUSSION Fatty Acid Composition. Descriptive statistics (mean, range, and standard deviation) of concentrations of the y variables in all fish oil samples used in the regression analysis is presented in Table I. The range of fatty acid concentrations makes the sample set suitable for developing robust calibration models. Correlation coefficients of concentrations of the fatty acids are presented in Table II. Concentrations of EPA and DHA are the least correlated, whereas both EPA and DHA have correlation with levels of total n-3 FAs. Fourier Transform Infrared Spectral Features. Fourier transform infrared spectra of the fish oil samples have distinct bands that can be assigned to several functional groups found in lipids. Figure 1a shows a spectrum of one of the samples, with the peak positions of important functional group frequencies of fatty acids labeled. Detailed descriptions of the different modes of vibrations for important functional groups corresponding to the peak positions of the FT-IR spectrum labeled in Fig. 1a are given in various literature.16,17,26,52–54 The ratio of the band due to the =C–H cis stretching at 3012 cm1 to the bands due to the CH2 asymmetric (2923 cm1) and CH2 symmetric (2852 cm1) stretching can be used to compare concentrations of total PUFAs in the fish oil samples. Higher ratios of band height indicate higher concentrations of total PUFAs.26 The ester carbonyl (–C=O) stretching vibration can be identified at 1743 cm1 in the spectrum. This strong peak shifts to

APPLIED SPECTROSCOPY

1193

FIG. 1. (a) Typical raw FT-IR spectrum and (b) fluorescence background-corrected Raman spectrum of a fish oil sample with labels showing important functional group frequencies.

1735 cm1 for fatty acids in ethyl ester forms when compared with the position for triglyceride forms of fatty acids.52,53 Several absorption bands of different intensities are observed in the fingerprint region of the FT-IR spectrum between 1500 and 650 cm1. Raman Spectral Features. The Raman spectra of the fish oil samples have a significant baseline offset due to fluorescence background, a distinctive feature of Raman spectra measured on biological samples. However, the signal-to-noise ratio is high enough for vibrational bands to be distinguished and frequencies to be easily assigned. A typical fluorescence background-corrected Raman spectrum of a fish oil sample indicating peak positions for important functional group frequencies of fatty acids is shown in Fig. 1b. The different modes of vibrations of important functional groups in fatty acids corresponding to the peak positions of the Raman

FIG. 2. Typical raw NIR spectrum of fish oil sample illustrating important bands and shoulders related to molecular vibrations of fatty acids.

1194

Volume 68, Number 10, 2014

spectrum labeled in Fig. 1b are reported in several papers.11,20,26,40,53,55,56 The Raman shift at 3014 cm1 (asymmetric =C–H cis stretching) has been described as a useful indicator of the different degrees of unsaturation,20,37,57 and it is correlated with the iodine value.55 This band shifts toward higher wave number as the number of olefinic double bonds increases in the fatty acid chain.26 Raman spectroscopy is a complementary technique to the infrared spectroscopy, and both are based on fundamental transitions. Similar to the infrared bands, it can be expected that the ratio of the band at 3014 cm1 to the bands at 2894 cm1 or 2852 cm1 may be used to compare concentrations of total PUFAs in the fish oil samples.26,58 Near-Infrared Spectral Features. Raw NIR spectrum of a fish oil sample with labels on major absorption bands is shown in Fig. 2. The spectrum shows relatively weaker and broader peaks than the corresponding FT-IR and Raman spectrum because NIR absorptions are based on overtones and combination modes of vibrations of molecules, which are weaker due to lower transition probabilities.59,60 The most important absorption maxima associated with molecular vibrations of lipid functional groups are clearly visible at 1710, 1760, 2144, 2178, and 2308 nm. NIR absorption frequencies, functional groups, and modes of vibrations for bands or shoulders marked in Fig. 2 are thoroughly discussed elsewhere.18,19,27–29,61,62 The NIR regions of the spectrum near 1710 and 1760 nm have been studied by several authors, and it has been reported that increasing values of absorbance in the vicinity of 1710 nm correlates with higher degrees of unsaturation.61 In this study, the NIR spectra of the fish oil samples show significant differences in absorbance around 1710 and 1725 nm. The intensity of the bands near 1710 nm increases for samples with high levels of

TABLE III. PLSR results using all samples and for calibration and validation sets based on the whole and selected regions of raw FT-IR spectra. Cal. set (all samples)a 1

2

Fatty acid

Spectral range (cm )

R

EPA

Wholed 3100–2700 þ 1820–650 3100–2700 1820–650 Whole 3100–2700 þ 1820–650 3100–2700 1820–650 Whole 3090–2800 þ 1790–650 3090–2800 1790–650

0.990 0.984 0.956 0.994 0.978 0.979 0.958 0.983 0.982 0.985 0.979 0.984

DHA

n-3 FAs

c

Cal. set (Nb = 40) 2

Validation set b

SECV

PLS factors

R

SECV

N

3.05 3.29 3.76 2.90 3.09 2.97 3.65 2.89 2.75 2.73 3.26 2.78

7 6 7 6 7 7 7 6 2 2 2 2

0.991 0.991 0.956 0.984 0.985 0.976 0.942 0.979 0.981 0.983 0.974 0.980

3.73 3.73 4.23 3.98 3.16 3.10 4.01 3.20 3.00 2.81 3.46 3.08

18 18 18 18 17 17 17 17 19 19 19 19

R2

SEPc

Bias

PLS factors

0.960 0.963 0.936 0.967 0.950 0.940 0.935 0.948 0.985 0.986 0.976 0.981

2.98 2.78 3.98 2.49 2.66 2.88 3.97 2.55 2.86 2.75 3.68 3.28

1.180 0.999 0.958 0.421 0.462 0.202 0.487 0.253 0.275 0.200 0.205 0.315

7 6 7 6 7 6 7 6 2 2 2 2

a

All samples were used as calibration set for full cross-validation. Number of samples after all samples were split into calibration and validation sets. c Percentage relative to total mass of fatty acids. d Whole wave number range: 4000–650 cm1. b

unsaturation. The peak maxima in this region is shifted to 1725 nm for samples with low levels of total PUFAs, and the peak at 1760 nm becomes more intense for these samples.29 Fourier Transform Infrared Spectral Data Calibration. Regression models based on raw FT-IR spectra gave lower prediction errors than those based on the preprocessed spectra (Table III). The models based on the whole FT-IR spectral range predicted EPA and DHA with SECV values of 3.05% and 3.09%, respectively, and required seven PLS factors. Total n-3 FAs were predicted using the whole spectral range with a coefficient of determination (R2 = 0.982) that is comparable to that of EPA and DHA, and an SECV value of 2.75%, but with reduced model complexity (two PLS factors). Selected regions gave models with slightly reduced prediction errors and model complexity when compared with the models based on the whole FT-IR region for EPA and DHA. However, the spectral region between 3100 and 2700 cm1, for both EPA and DHA, and the region between 3090 and 2800 cm1, for total n-3 FAs, gave lower values of R2 and slightly higher prediction error. For total n-3 FAs, models based on the selected regions (3090–2800 cm1þ 1790–650 cm1 and 1790–650 cm1) gave comparable results to the models based on the whole FT-IR region. All the samples used for the full cross-validation method were split into a calibration and validation set, and separate models were established, and performance of the models was tested using the independent validation sets (Table III). In general, all the results indicate that the whole region of the FT-IR spectrum can be used to predict all the response variables with high values of R2 and low prediction errors. The results also show that for FT-IR spectral data, selection of variables has no significant effect on model performance rather than slightly reducing model complexity for EPA and DHA (six PLS factors). Regression coefficients illustrate major regions in the spectra that are important for the predictions. A plot of regression coefficients of a two PLS factor model for total n-3 FAs based on the FT-IR spectra is shown in Fig. 3a.

Important variables with positive coefficients correspond with peaks at 3012, 1735, and around 696 cm1 (–(CH2)n– and cis =C–H symmetric rocking), which are all related to high levels of unsaturated fatty acids. The highest negative coefficients were observed at 2921, 2852, and 1747 cm1. Higher absorbance near these bands correlated well with low levels of unsaturation in the fish oils. The plot also indicates that the range from 3090 to 2800 cm1 and 1800 to 1600 cm1 and the fingerprint region below 1500 cm1 appear to contribute most to the prediction of total n-3 FAs. Raman Spectral Data Calibration. Different preprocessing methods were tested on the Raman spectra, and the polynomial curve fitting with SNV transformation gave models with the lowest prediction errors. The PLSR results are shown in Table IV. Models based on the whole Raman region and the selected regions of 3150– 2460 cm1 þ 1800–769 cm1, and 1800–769 cm1 gave comparable values of R2, SECV, and number of PLS factors for all the response variables. The region between 3150 and 2460 cm1 gave models with the lowest values of R2 and the highest prediction error. The higher prediction error in this spectral range occurred because most of the bands in this region are not related to olefinic double bonds.11 It is also evident from the Raman spectra of the fish oil samples that most of the bond vibrations related to the level of unsaturation are found in the lower wave number region between 1800 and 769 cm1. In addition to the full cross-validation method, separate models were established, and performance of the models was tested using independent validation sets (Table IV). All the results demonstrate that the models based on the region between 3150 and 2460 cm1 þ 1800 and 769 cm1 or the region between 1800 and 769 cm1 can be used to predict all response variables with comparable degrees of accuracy. These regions can be used to predict the response variables as accurately as the models based on the whole Raman spectral range. Similar to the full cross-validated models based on all the samples, model performance was reduced when the higher wave number region (3150–2460 cm1) was used

APPLIED SPECTROSCOPY

1195

FIG. 3. Regression coefficients plot of total n-3 FAs (a) for a two-PLS factors model based on the whole range of raw FT-IR spectra and (b) for a three-PLS factors model based on the whole range of background-corrected Raman spectra.

for predicting EPA and DHA. This region also gave slightly increased prediction error and model complexity for total n-3 FAs. A plot of regression coefficients against Raman shift covering the whole spectral range for total n-3 FAs is shown in Fig. 3b. The plot indicates that the spectral region between 3450 and 2460 cm1 and the region below 1800 cm1 appear to contribute most to the prediction of total n-3 FAs. Important variables correspond well with peaks at 3014, 2892, 2852, 1657 cm1 (cis C=C stretching), 1440 cm1 (CH2 scissoring), and near 1265 cm1 (cis =C–H symmetric rocking) that are all related to vibrations of different functional groups of fatty acids. The three peaks at 3014, 1657, and 1265 cm1 with positive coefficients are related to level of unsaturation in the fish oils, and these peaks have been used by other

authors for quantification and classification of unsaturated pure fats and oils.11,20 Near-Infrared Spectral Data Calibration. Among the different preprocessing methods tested on the NIR spectra, the Savitzky–Golay first derivative with second-order polynomial and 11 points gave the lowest prediction error. PLSR results are shown in Table V. Models based on the whole NIR spectral region predicted EPA and DHA with comparable values of R2, SECV, and the same number of PLS factors. Total n-3 FAs were predicted with the highest R2 value, the lowest prediction error (SECV = 2.18%), and two PLS factor model. Model performance increased when the NIR spectral variables were reduced. The model based on the spectral range of 1530–1900 nm for EPA, and the region

TABLE IV. PLSR results using all samples and for calibration and validation sets based on the whole and selected regions of polynomialcurve-fitted and SNV-transformed Raman spectra. Cal. set (all samples)a Fatty acid EPA

DHA

n-3 FAs

Spectral range (cm1) d

Whole 3150–2460 þ 1800–769 3150–2460 1800–769 Whole 3150–2460 þ 1800–769 31502460 1800–769 Whole 3150–2460 þ 1800–769 3150–2460 1800–769

Cal. set (Nb = 40)

R2

SECVc

PLS factors

R2

SECV

Nb

R2

SEPc

Bias

PLS factors

0.973 0.970 0.965 0.977 0.965 0.965 0.959 0.966 0.993 0.989 0.980 0.987

3.42 3.52 4.32 3.18 3.32 3.33 4.35 3.31 2.82 2.92 3.31 3.08

8 7 8 8 8 8 8 8 3 3 3 3

0.968 0.969 0.966 0.969 0.947 0.951 0.917 0.976 0.979 0.980 0.978 0.980

3.19 3.11 4.29 2.95 4.44 4.36 5.14 3.95 3.33 3.16 3.51 3.21

18 18 18 18 17 17 17 17 19 19 19 19

0.959 0.958 0.935 0.961 0.943 0.945 0.796 0.947 0.981 0.980 0.972 0.979

2.85 2.86 3.52 2.73 2.66 2.59 5.33 2.56 3.27 3.33 3.66 3.48

0.327 0.347 0.560 0.324 0.064 0.121 1.024 0.023 0.077 0.055 0.111 0.184

8 8 7 8 7 8 7 8 3 3 3 3

a

All samples were used as calibration set for full cross-validation. Number of samples after all samples were split into calibration and validation sets. c Percentage relative to total mass of fatty acids. d Whole wave number range: 3450–769 cm1. b

1196

Validation set

Volume 68, Number 10, 2014

TABLE V. PLSR results using all samples and for calibration and validation sets based on the whole and selected regions of Savitzky– Golay first derivative of NIR spectra. Cal. set (all samples)a Fatty acid EPA

DHA

n-3 FAs

2

c

Cal. set (Nb = 40) 2

Spectral range (nm)

R

SECV

PLS factors

R

Wholed 1530–1900 þ 2070–2500 1530–1900 2070–2500 Whole 1530–1940 þ 1960–2500 1530–1940 1960–2500 Whole 1630–1870 þ 2100–2500 1630–1870 2100–2500

0.971 0.956 0.979 0.947 0.969 0.969 0.972 0.922 0.997 0.995 0.997 0.988

3.43 3.45 2.43 3.80 3.20 3.24 2.34 3.81 2.18 2.20 2.18 2.42

6 5 5 5 6 6 6 4 2 2 2 2

0.963 0.964 0.950 0.961 0.946 0.945 0.942 0.916 0.987 0.987 0.987 0.985

Validation set b

SECV

N

R2

SEPc

Bias

PLS factors

3.38 3.26 3.23 3.53 3.72 3.76 3.21 4.33 2.45 2.47 2.47 2.67

18 18 18 18 17 17 17 17 19 19 19 19

0.900 0.902 0.935 0.875 0.967 0.968 0.952 0.939 0.995 0.995 0.996 0.994

3.92 3.89 3.11 4.38 2.96 2.98 2.60 3.39 1.68 1.68 1.60 2.00

0.203 0.255 0.091 0.337 0.009 0.064 0.187 0.088 0.042 0.029 0.172 0.236

4 4 4 4 5 5 5 4 2 2 2 2

a

All samples were used as calibration set for full cross-validation. Number of samples after all samples were split into calibration and validation sets. c Percentage relative to total mass of fatty acids. d Whole wavelength range: 1100–2500 nm. b

between 1530 and 1940 nm for DHA gave the lowest prediction error. For total n-3 FAs, all models based on the selected regions gave comparable prediction errors to the model based on the whole NIR region. To investigate the robustness of the NIR predictions, separate models were developed and their predictive ability was tested using validation sets. The results from model validation based on the test sets (Table V) also indicate that models based on the lower NIR region could give the lowest prediction error for both the single fatty acids. For total n-3 FAs, all of the selected NIR regions gave simple and reliable models with high values of R2 and low prediction errors. Similar to the full crossvalidated models, the region between 1630 and 1870 nm gave the lowest prediction error for total n-3 FAs. All the results demonstrate that model performance increased when selected variables in the NIR region were used for calibration, and the models predicted all the studied response variables with the lowest prediction errors. Regression vectors based on the Savitzky–Golay first derivative of NIR spectra were difficult to interpret. Thus,

FIG. 4. Regression coefficients versus wavelength plot of a two-PLS factors model for total n-3 FAs based on the whole range of raw NIR spectra.

a plot of regression coefficients against wavelengths for a two PLS factor model for total n-3 FAs developed using the whole range of raw NIR spectra is shown in Fig. 4. The plot illustrates that important variables correspond with peaks near 1708 and 1728 nm (first overtone of C–H stretching from CH3, –CH2–, and –HC=CH– groups). The band around 2142 nm (=C–H and C=O stretching) and the peaks at 2302 nm and 2348 nm, which are related to C–H combination and deformation vibrations of CH3 and CH2 groups, also contribute most to the prediction of total n-3 FAs. The regression plot also indicates that the NIR region between 1110 and 1450 nm has no significant contribution in predicting the fatty acids. The PLSR model results in Tables III and IV revealed that both FT-IR and Raman spectroscopy are able to predict the contents of EPA, DHA, and total n-3 FAs with good correlations and low prediction errors, with FT-IR yielding slightly lower prediction error and reduced model complexity than Raman. The lower FT-IR region (1820–650 cm1) gave models with six PLS factors, and with the highest R2 value and the lowest prediction error for EPA (R2 = 0.994, SECV = 2.90%, and SEP = 2.49%) and DHA (R2 = 0.983, SECV = 2.89%, and SEP = 2.55%). Both Raman and IR spectroscopy are based on fundamental transitions where most of the bond vibrations related to double bonds occur below 1800 cm1. Similar to the FT-IR spectra, the optimal models for both single fatty acids using Raman spectra were obtained when the lower region (1800–769 cm1) was used for calibration, and this region predicted EPA (R2 = 0.977, SECV = 3.18%, and SEP = 2.73%) and DHA (R2 = 0.966, SECV = 3.31%, and SEP = 2.56%) with eight PLS factors. The results also suggested that models based on FTIR spectra predicted total n-3 FAs with slightly reduced prediction error and model complexity when compared with the corresponding models based on Raman spectroscopy. FT-IR spectroscopy predicted total n-3 FAs with high values of R2 (0.985), the lowest model error (SECV = 2.73% and SEP = 2.75%), and reduced model complexity (two PLS factors), whereas Raman spectroscopy required a three PLS factor model with R2 value of

APPLIED SPECTROSCOPY

1197

0.993 and slightly increased prediction error (SECV = 2.82% and SEP = 3.27%) to predict total n-3 FAs. Compared with FT-IR and Raman spectroscopy, model error and complexity was decreased when NIR spectra were used for predicting the response variables (Table V). For both EPA and DHA, the optimal model using NIR spectroscopy was obtained when selected regions in the NIR spectra were used for calibration. The model based on the region between 1530 and 1900 nm required four to five PLS factors and predicted EPA with the highest R2 value of 0.979, SECV = 2.43%, and SEP = 3.11%. Similarly, this region (1530–1940 nm) gave the best model for DHA (R2 = 0.972, SECV = 2.34%, and SEP = 2.60%) with slightly increased model complexity (five to six PLS factors). Selected regions in the NIR spectra also gave simple models (two PLS factors) with good performances and predicted total n-3 FAs with R2 = 0.997, SECV = 2.18%, and SEP = 1.60%. These results indicate that compared with the models based on FT-IR and Raman spectra, the PLSR models based on NIR spectra are more robust and reliable, since models with few PLS factors are typically more stable.20,50 Studies using FT-IR, NIR, and Raman spectroscopy together with PLS regression to quantitatively determine EPA, DHA, and total omega-3 fatty acids in pharmaceutical fish oil supplements could not be found. However, related works using spectroscopy and multivariate calibration to predict total PUFAs in various samples have been documented. It has been reported that there are absorption bands that can be used to determine total unsaturation in an oil sample with high accuracy.22,63 There are also well-established methods for quantifying specific groups such as total trans double bonds22,23 or free fatty acids19 using FT-IR. Quantifying single fatty acids or groups such as omega-3 polyunsaturated acids is more challenging. There are examples of quantification of single fatty acids based on vibrational spectroscopy in samples with limited matrix variation. These include oleic and linoleic acid in olive oil,64 fatty acids in butter,38 omega-3 fatty acids in cod liver oil,65 and EPA and DHA in samples from a single process and manufacturer.66 Olsen et al. used FT-IR and Raman spectroscopy to determine total omega-3 fatty acids in melted fat from pork adipose tissue.37 Fla˚tten et al. also used FT-IR spectroscopy and PLSR to quantify DPA and DHA in pork fat.25 It should be noted that the variations of the contents of the single fatty acids in the fish oil samples are almost independent in this study and that they are predicted in the presence of other fatty acids. Because of the nonselectivity of spectroscopic measurements, it is difficult to know whether the regression models are based on specific signals in the spectra caused by the compounds of interest or by signals from other compounds that correlate with the quantified compound. Although the correlation between EPA and DHA in the current study is low (Table II), it is emphasized that the samples are randomly selected from the market and the full chemical compositions of the samples are not known. Most samples contain a large number of other fatty acids and other compounds that to some degree correlate with EPA and DHA. The possibility that the calibrations are based on indirect relationships is therefore present.

1198

Volume 68, Number 10, 2014

It has been shown that methylene-interrupted double bonds in polyunsaturated fatty acids give rise to a different absorption pattern than the isolated double bonds in monounsaturated fatty acids in the FT-IR spectra.67 It has also been claimed that the position of the double bond system relative to the carbonyl group (D-position) and the methyl end (x-position) in polyunsaturated fatty acids has some importance for the IR spectra of polyunsaturated fatty acids.26 EPA and DHA have the first bond relative to the carboxyl groups in the D5 and D4 positions, respectively. A subset of the samples applied in the present study was recently characterized using gas chromatography–mass spectrometry.68 Thirty-five peaks were found in levels above 1% of the total fatty acids. Nineteen of these were polyunsaturated fatty acids (two or more methyleneinterrupted double bonds). In addition to EPA and DHA, there was only one of these compounds, 20:4 n-6, that had the first double bond in the D5 position, and there was only one compound, 22:5 n-6, that had the first double bond in the D4 position. These two compounds may be interferents if the D position is of importance, but their levels are low, and the maximal weight percents relative to total fatty acids were only 3.3% and 6.2% for 20:4 n-6 and 22:5 n-6, respectively. They are also concentrated using the same techniques that are applied to increase the levels of EPA and DHA and are correlated to these. The coefficients of determination (R2) were 0.83 for the relationship between 20:4 n-6 and EPA and 0.80 for the relationship between 22:5 n-6 and DHA.

CONCLUSIONS The study demonstrated the feasibility of using FT-IR, NIR, and Raman spectroscopic techniques combined with the PLSR method to predict concentrations of EPA, DHA, and total n-3 FAs in fish oil supplements. The whole and selected regions of FT-IR and NIR spectra provided models with good performances and predicted the fatty acids with the lowest prediction errors. Raman spectroscopy predicted the fatty acids with slightly increased prediction error and model complexity when compared with the corresponding models based on the FT-IR and NIR spectra. Compared with EPA and DHA, total n-3 FAs were predicted with simple models using all the spectroscopic techniques. The spectroscopic techniques together with PLSR were simple, fast, and accurate for quantitative analysis of the studied fatty acids. The fish oil samples were analyzed directly and nondestructively without reagents, requiring a total analysis time of less than 1 min per sample measurement. The techniques presented promising opportunities for useful implementations in the process industries as cost-effective alternatives to the tedious, destructive, and time-consuming chromatographic and wet analytical methods. FT-IR and NIR spectroscopy, in particular, have demonstrated their potential to be used as reliable quantitative methods in quality control laboratories and industrial process monitoring during production of fish oil supplements. 1. C. Wang, W.S. Harris, M. Chung, A.H. Lichtenstein, E.M. Balk, B. Kupelnick, H.S. Jordan, J. Lau. ‘‘n-3 Fatty Acids from Fish or Fish-

2.

3.

4.

5.

6.

7.

8.

9.

10.

11.

12.

13.

14.

15. 16.

17.

18.

19.

20.

Oil Supplements, but Not a-Linolenic Acid, Benefit Cardiovascular Disease Outcomes in Primary- and Secondary-Prevention Studies: A Systematic Review’’. Am. J. Clin. Nutr. 2006. 84(1): 5-17. F. Shahidi, U.N. Wanasundara. ‘‘Omega-3 Fatty Acid Concentrates: Nutritional Aspects and Production Technologies. A Review’’. Trends Food Sci. Technol. 1998. 9(6): 230-240. W.S. Harris, D. Mozaffarian, M. Lefevre, C.D. Toner, J. Colombo, S.C. Cunnane, J.M. Holden, D.M. Klurfeld, M.C. Morris, J. Whelan. ‘‘Towards Establishing Dietary Reference Intakes for Eicosapentaenoic and Docosahexaenoic Acids’’. J. Nutr. 2009. 139(4): 804S819S. X.-J. Cao, B.-K. Hur. ‘‘Separation of EPA and DHA from Fish Oil Using Modified Zeolite 13X and Supercritical CO2’’. J. Ind. Eng. Chem. 2005. 11(5): 762-768. G.G. Dumancas, M. Muriuki, N. Purdie, L. Reilly. ‘‘Chemometric Algorithms for the Direct Determination of Lipids in Synthetic Mixtures and Human Serum’’. J. Biotech. Res. 2010. 2: 34-43. M.K. Duda, K.M. O’Shea, W.C. Stanley. ‘‘x-3 Polyunsaturated Fatty Acid Supplementation for the Treatment of Heart Failure: Mechanisms and Clinical Potential. A Review’’. Cardiovasc. Res. 2009. 84(1): 33-41. V.M. Moretti, F. Caprino. ‘‘Analysis of n-3 and n-6 Fatty Acids’’. In: L.M.L. Nollet, F. Toldra´, editors. Handbook of Seafood and Seafood Products Analysis. Boca Raton, FL, USA: CRC Press, 2009. Pp. 377392. W.K.T. Coltro, M.M.C. Ferreira, F.A.F. Macedo, C.C. Oliveira, J.V. Visentainer, N.E. Souza, M. Matsushita. ‘‘Correlation of Animal Diet and Fatty Acid Content in Young Goat Meat by Gas Chromatography and Chemometrics’’. Meat Sci. 2005. 71(2): 358-363. T. Kajishima, S. Aoki, Y. Nishimoto, E. Hasegawa, E. Moriyoshi, H. Nagayama. ‘‘Quantitative Determination of IPA and DHA in the Refined Fish Oil by Capillary Gas Chromatography: Collaborative Study and Standardization’’. J. Jpn. Oil Chem. Soc. 1996. 45(4): 335339. M. Herna´ndez-Martinez, T. Gallardo-Vela´zquez, G. Osorio-Revilla. ‘‘Rapid Characterization and Identification of Fatty Acids in Margarines Using Horizontal Attenuated Total Reflectance Fourier Transform Infrared Spectroscopy (HATR-FTIR)’’. Eur. Food Res. Technol. 2010. 231(2): 321-329. N.K. Afseth, V.H. Segtnan, B.J. Marquardt, J.P. Wold. ‘‘Raman and Near-Infrared Spectroscopy for Quantification of Fat Composition in a Complex Food Model System’’. Appl. Spectrosc. 2005. 59(11): 1324-1332. D. Cozzolino, M.J. Kwiatkowski, M. Parker, W.U. Cynkar, R.G. Dambergs, M. Gishen, M.J. Herderich. ‘‘Prediction of Phenolic Compounds in Red Wine Fermentations by Visible and Near Infrared Spectroscopy’’. Anal. Chim. Acta. 2004. 513(1): 73-80. I. Medina, S. Aubourg, J.M. Gallardo, R. Perez-Martin. ‘‘Comparison of Six Methylation Methods for Analysis of the Fatty Acid Composition of Albacore Lipid’’. Int. J. Food Sci. Technol. 1992. 27(5): 597-601. K.P.J. Williams, D.N. Batchelder. ‘‘Raman Microscopy-Spectroscopy and Direct Two Dimensional Imaging’’. Spectrosc. Eur. 1994. 6(3): 19-26. V. Baeten, P. Dardenne. ‘‘Spectroscopy: Developments in Instrumentation and Analysis’’. Grasas y Aceites. 2002. 53(1): 45-63. A. Rohman, Y.B. Che Man. ‘‘FTIR Spectroscopy Combined with Chemometrics for Analysis of Lard in the Mixtures with Body Fats of Lamb, Cow and Chicken’’. Int. Food Res. J. 2010. 17(3): 519-526. A. Rohman, Y.B. Che Man. ‘‘Application of Fourier Transform Infrared (FT-IR) Spectroscopy Combined with Chemometrics for Authentication of Cod-Liver Oil’’. Vib. Spectrosc. 2011. 55(2): 141145. T. Woodcock, G. Downey, C.P. O’Donnell. ‘‘Confirmation of Declared Provenance of European Extra Virgin Olive Oil Samples by NIR Spectroscopy’’. J. Agric. Food Chem. 2008. 56(23): 1152011525. D. Cozzolino, I. Murray, A. Chree, J.R. Scaife. ‘‘Multivariate Determination of Free Fatty Acids and Moisture in Fish Oils by Partial Least-Squares Regression and Near-Infrared Spectroscopy’’. Food Sci. Technol. 2005. 38(8): 821-828. E.F. Olsen, E.-O. Rukke, A. Fla˚tten, T. Isaksson. ‘‘Quantitative Determination of Saturated, Monounsaturated, and Polyunsaturated Fatty Acids in Pork Adipose Tissue with Non-destructive Raman Spectroscopy’’. Meat Sci. 2007. 76(4): 628-634.

21. F.R. van de Voort, J. Sedman, T. Russin. ‘‘Lipid Analysis by Vibrational Spectroscopy’’. Eur. J. Lipid Sci. Technol. 2001. 103(12): 815-840. 22. F.R. van de Voort, A.A. Ismail, J. Sedman. ‘‘A Rapid, Automated Method for the Determination of cis and trans Content of Fats and Oils by Fourier Transform Infrared Spectroscopy’’. J. Am. Oil Chem. Soc. 1995. 72(8): 873-880. 23. M.M. Mossoba, M.P. Yurawecz, R.E. McDonald. ‘‘Rapid Determination of Total trans Content of Neat Hydrogenated Oils by Attenuated Total Reflection Spectroscopy’’. J. Am. Oil Chem. Soc. 1996. 73(8): 1003-1009. 24. A. Al-Alawi, F.R. van de Voort, J. Sedman. ‘‘New FTIR Method for the Determination of FFA in Oils’’. J. Am. Oil Chem. Soc. 2004. 81(5): 441-446. 25. A. Fla˚tten, E.A. Bryhni, A. Kokler, B. Egelandsdal, T. Isaksson. ‘‘Determination of C22:5 and C22:6 Marine Fatty Acids in Pork Fat with Fourier Transform Mid-Infrared Spectroscopy’’. Meat Sci. 2005. 69(3): 433-440. 26. S. Yoshida, H. Yoshida. ‘‘Nondestructive Analyses of Unsaturated Fatty Acid Species in Dietary Oils by Attenuated Total Reflectance with Fourier Transform IR Spectroscopy’’. Biopolymers. 2003. 70(4): 604-613. 27. C. Shiroma, L. Rodriguez-Saona. ‘‘Application of NIR and MIR Spectroscopy in Quality Control of Potato Chips’’. J. Food Comp. Anal. 2009. 22(6): 596-605. 28. S. Armenta, S. Garrigues, M. de la Guardia. ‘‘Determination of Edible Oil Parameters by Near Infrared Spectrometry’’. Anal. Chim. Acta. 2007. 596(2): 330-337. 29. T. Sato, S. Kawano, M. Iwamoto. ‘‘Near Infrared Spectral Patterns of Fatty Acids Analysis from Fats and Oils’’. J. Am. Oil Chem. Soc. 1991. 68(11): 827-833. 30. T. Sato. ‘‘New Estimation Method for Fatty Acid Composition in Oil Using Near-Infrared Spectroscopy’’. Biosci., Biotechnol., Biochem. 2002. 66(12): 2543-2548. ˜ J.A. Ramı´ rez, I. Dı´ az. ‘‘Prediction of 31. M. Pla, P. Herna´ndez, B. Arino, Fatty Acid Content in Rabbit Meat and Discrimination Between Conventional and Organic Production Systems by NIRS Methodology’’. Food Chem. 2007. 100(1): 165-170. 32. H. Li, F.R. van de Voort, J. Sedman, A.A. Ismail. ‘‘Rapid Determination of cis and trans Content, Iodine Value, and Saponification Number of Edible Oils by Fourier Transform NearInfrared Spectroscopy’’. J. Am. Oil Chem. Soc. 1999. 76(4): 491-497. 33. H. Li, F.R. van de Voort, A.A. Ismail, J. Sedman, R. Cox, C. Simard, H. Buijs. ‘‘Discrimination of Edible Oil Products and Quantitative Determination of their Iodine Value by Fourier Transform NearInfrared Spectroscopy’’. J. Am. Oil Chem. Soc. 2000. 77(1): 29-36. 34. E.D. Lipp, M.A. Leugers. ‘‘Applications of Raman Spectroscopy in the Chemical Industry’’. In: M.J. Pelletier, editor. Analytical Applications of Raman Spectroscopy. London, UK: Blackwell Science, 1999. Pp. 106-126. 35. E.C.Y. Li-Chan. ‘‘The Applications of Raman Spectroscopy in Food Science’’. Trends Food Sci. Technol. 1996. 7(11): 361-370. 36. S. Keller, T. Lo¨chte, B. Dippel, B. Schrader. ‘‘Quality Control of Food with Near-Infrared Excited Raman Spectroscopy’’. Fresenius’ J. Anal. Chem. 1993. 346(6-9): 863-867. 37. E.F. Olsen, E.-O. Rukke, B. Egelandsdal, T. Isaksson. ‘‘Determination of Omega-6 and Omega-3 Fatty Acids in Pork Adipose Tissue with Non-destructive Raman and Fourier Transform Infrared Spectroscopy’’. Appl. Spectrosc. 2008. 62(9): 968-974. 38. J.R. Beattie, S.E.J. Bell, C. Borgaad, A.M. Fearon, B.W. Moss. ‘‘Multivariate Prediction of Clarified Butter Composition Using Raman Spectroscopy’’. Lipid. 2004. 39(9): 897-906. 39. M. Meurens, V. Baeten, S.H. Yan, E. Mignolet, Y. Larondelle. ‘‘Determination of the Conjugated Linoleic Acids in Cow’s Milk Fat by Fourier Transform Raman Spectroscopy’’. J. Agric. Food Chem. 2005. 53(15): 5831-5835. 40. R.C. Barthus, R.J. Poppi. ‘‘Determination of the Total Unsaturation in Vegetable Oils by Fourier Transform Raman Spectroscopy and Multivariate Calibration’’. Vib. Spectrosc. 2001. 26(1): 99-105. 41. V. Baeten, M. Meurens, M.T. Morales, R. Aparicio. ‘‘Detection of Virgin Olive Oil Adulteration by Fourier Transform Raman Spectroscopy’’. J. Agric. Food Chem. 1996. 44(8): 2225-2230. 42. V. Baeten, R. Aparicio. ‘‘Edible Oils and Fats Authentication by Fourier Transform Raman Spectrometry’’. Biotechnol. Agron. Soc. Environ. 2000. 4(4): 196-203.

APPLIED SPECTROSCOPY

1199

43. AOCS. Official Methods and Recommended Practices of the American Oil Chemists’ Society. Champaign, IL, USA: American Oil Chemists’ Society, 1998. 5th ed. 44. S.A. Mjøs, M. Solvang. ‘‘Geometrical Isomerisation of Eicosapentaenoic Acid and Docosahexaenoic Acid at High Temperatures’’. Eur. J. Lipid Sci. Technol. 2006. 108(7): 589-597. 45. C. Sciotto, S.A. Mjøs. ‘‘Trans Isomers of EPA and DHA in Omega-3 Products on the European Market’’. Lipids. 2012. 47(7): 659-667. 46. C.A. Lieber, A. Mahadevan-Jansen. ‘‘Automated Method for Subtraction of Fluorescence from Biological Raman Spectra’’. Appl. Spectrosc. 2003. 57(11): 1363-1367. 47. R.J. Barnes, M.S. Dhanoa, S.J. Lister. ‘‘Standard Normal Variate Transformation and De-trending of Near-Infrared Diffuse Reflectance Spectra’’. Appl. Spectrosc. 1989. 43(5): 772-777. 48. O.M. Kvalheim. ‘‘Interpretation of Partial Least Squares Regression Models by Means of Target Projection and Selectivity Ratio Plots’’. J. Chemom. 2010. 24(7-8): 496-504. 49. T. Rajalahti, R. Arneberg, F.S. Berven, K.-M. Myhr, R.J. Ulvik, O.M. Kvalheim. ‘‘Biomarker Discovery in Mass Spectral Profiles by Means of Selectivity Ratio Plot’’. Chemom. Intell. Lab. Syst. 2009. 95(1): 35-48. 50. T. Næs, T. Isaakson, T. Fearn, T. Davies. A User Friendly Guide to Multivariate Calibration and Classification. Chichester, UK: NIR Publications, 2002. 51. CAMO. The Unscrambler User Manual: The Unscrambler Methods. Oslo, Norway: CAMO Software AS, 2006. 52. J. Zhang. Analysis and Characterization of Consumer Products by FTIR, Raman, Chemometrics and Two Dimensional ATR-FTIR Correlation Spectroscopy. [Ph.D. Dissertation]. New Brunswick, NJ, USA: Rutgers, 2009. 53. M.D. Guille´n, N. Cabo. ‘‘Infrared Spectroscopy in the Study of Edible Oils and Fats’’. J. Sci. Food Agric. 1997. 75(1): 1-11. 54. Y. Kim, D.S. Himmelsbach, S.E. Kays. ‘‘ATR-Fourier Transform MidInfrared Spectroscopy for Determination of Trans Fatty Acids in Ground Cereal Products Without Oil Extraction’’. J. Agric. Food Chem. 2007. 55(11): 4327-4333. 55. V. Baeten, P. Hourant, M.T. Morales, R. Aparicio. ‘‘Oil and Fat Classification by FT-Raman Spectroscopy’’. J. Agric. Food Chem. 1998. 46(7): 2638-2646. 56. R.M. El-Abassy, P. Donfack. ‘‘Rapid Determination of Free Fatty Acid in Extra Virgin Olive Oil by Raman Spectroscopy and Multivariate Analysis’’. J. Am. Oil Chem. Soc. 2009. 86(6): 507-511. 57. E.C.Y. Li-Chan. ‘‘Developments in the Detection of Adulteration of Olive Oil’’. Trends Food Sci. Technol. 1994. 5(1): 3-11. 58. H. Sadeghi-Jorabchi, P.J. Hendra, R.H. Wilson, P.S. Belton. ‘‘Determination of the Total Unsaturation in Oils and Margarines

1200

Volume 68, Number 10, 2014

59.

60.

61.

62.

63.

64.

65.

66.

67.

68.

by Fourier Transform Raman Spectroscopy’’. J. Am. Oil Chem. Soc. 1990. 67(8): 483-486. C. Sandorfy, R. Buchet, G. Lachenal. ‘‘Principles of Molecular Vibrations for Near-Infrared Spectroscopy’’. In: Y. Ozaki, W.F. McClure, A.A. Christy, editors. Near-Infrared Spectroscopy in Food Science and Technology. Hoboken, NJ: John Wiley and Sons, 2007. Pp. 11-14. S. Kawano. ‘‘Sampling and Sample Presentation’’. In: H.W. Siesler, Y. Ozaki, S. Kawata, H.M. Heise, editors. Near-infrared Spectroscopy: Principles, Instruments, Applications. Weinheim, Germany: Wiley-VCH, 2002. Pp. 115-124. P. Hourant, V. Baeten, M.T. Morales, M. Meurens, R. Aparicio. ‘‘Oil and Fat Classification by Selected Bands of Near-Infrared Spectroscopy’’. Appl. Spectrosc. 2000. 54(8): 1168-1174. N. Sinelli, L. Cerretani, V. Di Egidio, A. Bendini, E. Casiraghi. ‘‘Application of Near (NIR) Infrared and Mid (MIR) Infrared Spectroscopy as a Rapid Tool to Classify Extra Virgin Olive Oil on the Basis of Fruity Attribute Intensity’’. Food Res. Int. 2010. 43(1): 369-375. F.R. van de Voort, J. Sedman, G. Emo, A.A. Ismail. ‘‘Rapid and Direct Iodine Value and Saponification Number Determination of Fats and Oils by Attenuated Total Reflectance/Fourier Transform Infrared Spectroscopy’’. J. Am. Oil Chem. Soc. 1992. 69(11): 11181123. R.M. Maggio, T.S. Kaufman, M. Del Carlo, L. Cerretani, A. Bendini, A. Cichelli, D. Compagnone. ‘‘Monitoring of Fatty Acid Composition in Virgin Olive Oil by Fourier Transformed Infrared Spectroscopy coupled with Partial Least Squares’’. Food Chem. 2009. 114(4): 1549-1554. M.D. Guille´n, I. Carton, E. Goicoechea, P. Uriarte. ‘‘Characterization of Cod Liver Oil by Spectroscopic Techniques. New Approaches for the Determination of Compositional Parameters, Acyl Groups, and Cholesterol from 1H Nuclear Magnetic Resonance and Fourier Transform Infrared Spectral Data’’. J. Agric. Food Chem. 2008. 56(19): 9072-9079. H. Azizian, J.K.G. Kramer, S. Ehler, J.M. Curtis. ‘‘Rapid Quantitation of Fish Oil Fatty Acids and their Ethyl Esters by FT-NIR Models’’. Eur. J. Lipid Sci. Technol. 2010. 112(4): 452-462. J. Lundbom, M. Kolehmainen, L. Pulkkinen, P. Soininen, M. Tiainen, U. Schwab, M. Uusitupa, M. Tammi. ‘‘Mid-Infrared Spectroscopy and Multivariate Curve Resolution for Analyzing Human Adipose Tissue Triacylglycerols’’. Eur. J. Lipid Sci. Technol. 2010. 112(12): 1308-1314. Z. Wasta, S.A. Mjøs. ‘‘A Database of Chromatographic Properties and Mass Spectra of Fatty Acid Methyl Esters from Omega-3 Products’’. J. Chromatogr. A. 2013. 1299: 94-102.