Journal of the American Oil Chemists' Society

, Volume 89, Issue 12, pp 2143–2154

Evaluating the Transferability of FT-NIR Calibration Models for Fatty Acid Determination of Edible Fats and Oils Among Five Same-make Spectrometers Using Transmission or Transflection Modes with Different Pathlengths

Authors

    • NIR Technologies Inc.
  • John K. G. Kramer
    • NIR Technologies Inc.
  • Magdi M. Mossoba
    • Food and Drug Administration, Center for Food Safety and Applied Nutrition, Office of Regulatory Science
Original Paper

DOI: 10.1007/s11746-012-2116-9

Cite this article as:
Azizian, H., Kramer, J.K.G. & Mossoba, M.M. J Am Oil Chem Soc (2012) 89: 2143. doi:10.1007/s11746-012-2116-9

Abstract

Fourier transform near-infrared (FT-NIR) spectroscopy in conjunction with partial least squares 1 (PLS1) calibration models was previously reported to be an alternative method to GC for the rapid determination of the fatty acid (FA) composition of fats and oils. These calibration models had been developed based on accurate GC data (primary reference method) and observed FT-NIR spectra. In the present three-laboratory limited collaborative study, the transferability of these pre-developed calibration models to four other FT-NIR spectrometers from the same manufacturer was evaluated. Six samples were selected that provided a wide range of FA contents. Our results indicate that these models were successfully transferable to spectrometers operating in the transflection mode with 2- or 4-mm pathlength fiber optic probes or in the transmission mode using 5-mm, but not 8-mm, outer diameter tubes. The predicted FA composition fell within the statistically accepted limits of agreement between FT-NIR and GC. The FT-NIR precision data were consistent with those reported in a published GC collaborative study. The application of FT-NIR to the determination of the total content of SFA, trans FA, MUFA, and PUFA is cost-effective and potentially suitable for the rapid screening of commercial products for compliance verification with labeling regulations.

Keywords

FT-NIRRapid fatty acid determinationTransferabilityCalibration modelstrans Fatty acidsSaturated fatty acids

Introduction

The introduction of labeling regulations in the US, Canada, and many other countries mandated the declaration of the total saturated fat and total trans fat contents on the nutrition facts labels of all food products and dietary supplements [1, 2]. Capillary gas chromatography (GC) has been the industry’s method of choice because it provides detailed information on the fatty acid (FA) composition of foods [3, 4]. However, it should be noted that a thorough GC analysis [4] will take hours, require the use of organic solvents and chemicals, and the need to first derivatize the fat and oil to their fatty acid methyl esters (FAMEs) prior to their separation by GC. A successful GC determination is dependent on correct identification of the peaks, accurate quantification, and sorting out of overlapping peaks, much of which are subject to the analyst’s expertise [47]. Compared to GC, Fourier transform near-infrared (FT-NIR) spectroscopy offers several advantages [810]. There is no requirement for sample preparation of fats and oils, and no need for chemicals and solvents and their subsequent disposal. The method is non-destructive to the test samples and can be measured with either fiber optic probes or by using transmission tubes. However, most importantly FT-NIR is much faster (<5 min per test sample) and less costly than GC. The collection of an FT-NIR spectrum takes less than a minute, and the application of calibration models (data analysis) to determine the FA composition is complete in minutes, and is fully automated and independent of analyst bias [810]. Because calibration models have been previously developed [810], their subsequent application to FA profiling is both rapid and cost-effective. FT-NIR is ideally suited to rapidly determine the total SFA, trans FA, cis MUFA, and cis PUFA of fats and oils and potentially suitable for the rapid screening of commercial products for compliance verification with labeling regulations.

An alternative attenuated total reflection (ATR)-Fourier transform mid-infrared (FTIR) method has also been reported recently, which is also rapid (<5 min) and requires no derivatization of fats and oils [1114]. However, this method provides only the total trans fatty acid (trans FA), but not the total saturated fatty acid (SFA) content, whose declaration is also mandated by the various labeling regulations [1, 2]. ATR-FTIR methods were validated and published by the Association of Analytical Communities (AOAC) [12] and by the American Oil Chemists’ Society (AOCS) as official method AOCS Cd 14e-09 in 2009 [13]. By contrast, FT-NIR in conjunction with partial least squares 1 (PLS1) calibration models is a better alternative because it can be used to rapidly (<5 min) determine not only the total trans FA content, but also the total SFA content, as well as those of all the other FA in fats and oils [810]. For decades FT-NIR technology has been widely used by many industries for numerous applications, but not for the determination of the FA composition of fats and oils because building reliable PLS1 calibration models is time-consuming and requires the input of accurate FA composition data from a primary reference method such as GC [15].

The FT-NIR calibration models used were all previously developed [810] and tested using one Bruker FT-NIR spectrometer equipped with a fiber optic probe adjusted to a 2-mm pathlength. However, the potential to transfer these PLS1 calibration models to other FT-NIR spectrometers from the same or multiple vendors has not yet been evaluated, and the former is the subject of the present three-laboratory limited collaborative study, which is the first step in establishing a standard procedure. Model transferability is challenging because it must take into consideration all sources of variance between different instruments, experimental conditions, and test sample matrices. A successful transfer involves the application of models developed on one instrument to other FT-NIR spectrometers and must result in similar prediction errors. Calibration models are strongly dependent on the accuracy of the primary GC reference data, and any inherent error in the GC determinations would negatively impact the performance of the FT-NIR calibration models [810, 15]. Calibration transfer also depends on developing robust models with an adequate variety of test samples that cover the desired concentration ranges for each of the FAs expected in unknown test mixtures. The robustness of the models used in the present study has been discussed previously [810].

In the present limited validation study, calibration transfer for the rapid FT-NIR determination of the FA composition of fats and oils was tested in three different laboratories. This study was designed to explore the possibility of applying the calibration models pre-developed on one FT-NIR spectrometer to four other spectrometers from the same vendor under as many experimental conditions as possible, such as different measurement modes and pathlengths. All five FT-NIR spectrometers used were equipped with transmission and/or transflection accessories with different pathlengths, specifically 2- and 4-mm fiber optic probes, and/or 5- and 8-mm outer diameter (OD) transmission tubes that correspond to 3- and 6-mm pathlengths, respectively.

Materials and Methodology

Several edible oils and fats were selected for this study, which included soybean, canola, sunflower, coconut, and palm oils and a partially hydrogenated (PH) soybean oil. All the oils were obtained from local grocery stores except the PH soybean oil, which was a gift from Bunge Canada (Toronto, ON, Canada). The six products chosen for this study were typical oils used to cover the type and range of FA one would find in edibles oils on the market. For example, canola was selected because of the high content of oleic and linolenic acids, soybean oil for the high content of linoleic and linolenic acids, palm and coconut oils for their high content of the saturated FA 12:0, 14:0, and 16:0, and the PH soybean oil for its high content of trans FA.

FT-NIR and Chemometrics (PLS1) Measurements

Spectral measurements were collected in the 12,800–4,000 cm−1 range on five Bruker Optics (Billerica, MA) FT-NIR spectrometers equipped with OPUS (Optics User’s Software) available in two models, namely a Matrix-F model equipped with a fiber optic probe of 2-mm pathlength and four different Multi Purpose Analyzer (MPA) model spectrometers operating in the fiber optic mode (2- or 4-mm pathlength) as well as in the transmission mode (5- and 8-mm OD transmission tubes, equivalent to approximately 3- and 6-mm ID pathlengths, respectively). Replicate measurements were acquired for each of the test samples investigated in three different laboratories (FDA, College Park, MD, USA; NIR Technologies, Oakville, ON, Canada, and Bruker Optics, Milton, Ontario, Canada). All measurements were obtained at room temperature except those for hard fats (coconut, palm, and PH soybean oil), which were recorded at 50 °C. In the case of fiber optic measurements of hard fats, the samples were placed in test tubes that were warmed to 50 °C in a heating block (Fisher Scientific, Ottawa, Ontario, Canada), and then the spectra were acquired. Five replicate FT-NIR spectra were collected for every test sample on each instrument both in the transflection and/or transmission mode, and each spectrum consisted of five co-added scans at 8 cm−1 resolution. Each of these two sets of five replicate spectra were subsequently averaged and analyzed using the pre-developed PLS1 calibration models available from NIR Technologies Inc. (Oakville, Ontario, Canada). The four PLS1 calibration models used in this study were a very low trans FA (<3 % trans FA) and a low to medium trans fat (<20 % trans FA) model suitable for low trans fat oils (soybean, canola, and sunflower oils), a saturated fat model for fats and oils high in SFAs such as coconut and palm oils, and a high trans fat model for PH soybean oil. The decision as to which calibration model was most suitable for a given sample was obtained by a fully automated spectral analysis protocol in which all FT-NIR models are applied consecutively. FT-NIR models in which predicted FA values fall outside the specified FA range for a certain calibration model are flagged by the software and are then ignored for that sample. These hierarchical decisions are undertaken in an automatic process independent of analyst bias. The sums of each of the SFA, trans FA, MUFA, and PUFA were obtained and reported for each oil using the appropriate calibration models from NIR Technologies Inc.

The resulting collaborative study data for test samples were analyzed according to the Statistical Program 2001 developed by the AOAC Statistics Committee (AOAC International, Gaithersburg, MD, USA) [16]. The precision statistical parameters include: s(r), repeatability standard deviation; s(R), reproducibility standard deviation; RSD(r), repeatability relative standard deviation; RSD(R), reproducibility relative standard deviation; and HORRAT value (http://www.aoac.org/dietsupp6/Dietary-Supplement-website/HORRAT_SLV.pdf), which are given in the AOAC Harmonization Guidelines [16, 17].

GC Analysis

The GC analyses of these fats and oils were determined using previously described methods [7, 18, 19] and are briefly summarized below. All fat and oil products (about 20 mg each) were methylated using two separate procedures: anhydrous 5 % HCl/methanol (by weight) for 1 h at 80 °C and 0.5 % NaOCH3 in methanol (#33080; Supelco Inc., Bellefonte, PA, USA) for 15 min at 50 °C. Each of the two resultant FAME products were extracted with hexane after addition of water (5 %, v/v). Hexane was removed, and the FAMEs were purified by TLC on silica gel G plates (Fisher Scientific, Ottawa, ON, Canada) using the developing solvent hexane/diethyl ether/acetic acid (85:15:1). The FAME band on the TLC plate was identified after spraying the plates with 2′,7′-dichlorofluorescein in methanol, and the band was visualized under ultraviolet light. The FAME band on the TLC plate was scraped off, transferred into a Pasteur pipette (5.75 inches) that contained a previously cleaned glass wool plug (glass wool washed with chloroform/methanol, 1:1), and the FAME products were eluted with hexane. A concentration of about 1–2 mg/ml FAME in hexane was used for GC analysis.

All the FAME preparations were analyzed by GC (Hewlett-Packard Model 5890 Series II, Palo Alto, CA, USA) equipped with a splitless injection port flushed after 0.3 min, a flame ionization detector, an autosampler (Hewlett-Packard, Model 7673), a 100-m CP-Sil 88 fused capillary column (100 m × 0.25 mm i.d. × 0.2 μm film thickness; Varian Inc., Mississauga, ON, Canada), and a Hewlett-Packard ChemStation software program (version A.10). The operating conditions were: injector and detector temperatures both set at 250 °C; H2 was used as carrier gas (1 ml/min) and for the flame ionization detector (FID) (30 ml/min); the other FID gases were N2 makeup gas (30 ml/min), and air (300 ml/min). All gases were of the highest purity. Two temperature programs were used: (1) initial temperature of 45 °C and held for 4 min, programmed at 13 °C/min to 175 °C and held for 27 min, then programmed at 4 °C/min to 215 °C and held for 35 min, and (2) initial temperature of 45 °C and held for 4 min, programmed at 13 °C/min to 150 °C and held for 47 min, then programmed at 4 °C/min to 215 °C and held for 35 min) [19]. The results of the two GC temperature programs complemented each other and served to resolve all the cis- and trans-18:1 isomers and to resolve any cis-20:1 from cis/cis/trans-18:3 isomers. The FAMEs were identified by comparison with a GC reference FAME standard (GLC #463) spiked with a mixture of four positional conjugated linoleic acid (CLA) isomers (#UC-59 M) and the long-chain saturated FAME 21:0, 23:0, and 26:0; all lipid standards were obtained from Nu-Chek Prep Inc (Elysian, MN). The trans isomers of linoleic and linolenic acid were prepared as described previously [20]. All GC results are presented as relative percent of total FAME based on the flame ionization response. The GC results of the FAME prepared using the acid- and base-catalyzed procedure were compared. The results of the two methylation procedures were generally similar, except for differences in the CLA isomer composition; acid conditions resulted in a higher content of the trans-,trans-CLA isomers because of acid-catalyzed isomerization [21]. The final FA composition was obtained by averaging the results of both methylation procedures, but the CLA isomer distribution was taken from the results of the base-catalyzed methylation. Identification of the CLA isomers was confirmed using silver ion HPLC separation [19]. Because the GC results were used as the primary reference data for developing the FT-NIR calibration models, every attempt was made to separate and report all GC peaks, since FT-NIR is matrix dependent [15].

Results and Discussion

The sums (as percent of total FA) for each of the total SFA, trans FA, MUFA, and PUFA for the oils determined by FT-NIR in the present study are given in Table 1. As detailed in this table, between seven and nine independent sets of replicate FT-NIR data (totaling 14–18 replicates) were collected in each of the three laboratories on five different spectrometers from the same manufacturer using both fiber optic probes (2- or 4-mm pathlength) and/or transmission tubes (5 mm OD). The data obtained with the 8 mm OD transmission tubes was not included in Table 1 and will be discussed separately below. Each of the six oils investigated in the present transferability study had previously been analyzed by GC and included in the development of the quantitative PLS1 calibration models [8]. These models have been developed and optimized by using leave-one-out cross validation (LOOCV) and PLS1 analysis [8]. In order to determine accuracy, FT-NIR results presented in Table 1 were compared to the corresponding GC determinations of the sums of each of total SFA, trans FA, MUFA, and PUFA for each of the six test samples, expressed as percent of total FA. FT-NIR data precision will be presented and evaluated first, and a discussion on the accuracy of FT-NIR determinations will be subsequently detailed.
Table 1

Fatty acid content (as percent of total fat) determined by GC and FT-NIR

Methodology:

GC

FT-NIR

Laboratory:

 

1

2

3

Mean

STDEV

Instrument model:

 

Matrix

MPA 1

MPA 2

MPA 3

MPA 4

 

Transflection/transmission modea:

F2

F4

T5

F2

T5

F2

T5

F2

T5

Independent sets of FT-NIR measurements:

Set 1

Set 2

Set 3

Set 4

Set 5

Set 6

Set 7

Set 8

Set 9

Duplicate measurementsb:

M1

M2

M1

M2

M1

M2

M1

M2

M1

M2

M1

M2

M1

M2

M1

M2

M1

M2

PH Soybean Oil 

 Σ SFA

18.7

18.1

18.3

18.5

18.5

18.2

18.3

15.7

15.65c

19.1

19.1

18.7

18.8

18.5

18.5

17.1

17.1

17.8

17.9

18.1

0.8

 Σ trans FA

28.6

25.8

25.3

27.6

27.6

27.8

27.9

29.3

29.4

26.9

27.0

31.0

31.2

28.4

28.4

27.6

27.2

28.0

28.0

28.0

1.5

 Σ MUFA

27.9

31.9

31.7

28.5

28.5

28.6

28.2

31.9

32.3

28.7

28.7

27.3

27.0

28.0

27.9

32.4

32.6

29.5

29.3

29.6

1.9

 Σ PUFA

20.4

22.78d

21.9

22.1

22.0

22.5

22.3

21.6

21.5

23.1

23.1

22.5

21.8

21.5

21.4

19.5e

19.4e

22.1

22.0

22.1

0.5

Soybean Oil

 Σ SFA

16.0

16.3

16.2

17.1

17.2

17.8

17.8

15.3

15.3

17.2

17.3

17.0

17.3

16.6

16.6

16.3

16.6

16.3

16.2

16.7

0.7

 Σ trans FA

1.6

1.3

1.3

1.7

1.7

1.6

1.6

1.7

1.8

1.7

1.7

1.2

1.4

1.7

1.6

1.1

1.1

1.5

1.5

1.5

0.2

 Σ MUFA

21.5

23.5

23.0

21.4

21.5

20.3

20.3

25.2

25.6

21.4

21.4

21.5

22.1

21.2

21.2

24.9

24.6

24.4

24.3

22.7

1.8

 Σ PUFA

61.6

59.7

59.5

61.0

60.9

61.2

61.2

60.2

60.0

61.2

61.2

59.8

59.5

61.7

61.7

59.6

60.0

60.2

60.3

60.5

0.8

Canola Oil

 Σ SFA

8.0

7.6

7.1

8.3

8.3

8.5

8.7

7.7

7.1

9.9

9.9

9.3

9.2

8.1

8.1

8.7

8.3

8.8

8.2

8.4

0.8

 Σ trans FA

1.7

0.76

0.76

1.29

1.30

1.31

1.30

1.54

1.50

1.67

1.67

0.94

0.90

1.28

1.29

0.96

0.97

1.15

1.16

1.2

0.3

 Σ MUFA

63.3

61.4

62.6

63.3

63.2

61.5

61.5

67.0

67.2

62.4

62.4

62.0

62.3

62.5

62.6

64.8

64.3

65.1

63.9

63.3

1.7

 Σ PUFA

25.8

27.1

26.9

25.9

26.0

26.8

26.7

24.6

24.2

25.0

25.0

24.6

24.5

26.1

26.1

23.8

24.1

25.6

25.6

25.5

1.0

Sunflower Oil 

 Σ SFA

11.7

11.7

11.8

11.6

11.3

11.8

11.7

10.0e

10.1e

11.4

11.4

11.7

11.7

11.3

11.4

11.6

11.5

11.0

11.0

11.5

0.3

 Σ trans FA

0.8

1.0

1.0

1.3

1.3

1.2

1.2

1.2

1.2

1.2

1.2

0.8

0.8

1.2

1.2

1.0

0.9

1.1

1.1

1.1

0.2

 Σ MUFA

24.5

24.0

24.1

24.3

24.3

22.8

22.8

29.3e

28.9e

24.1

24.1

25.1

25.2

23.9

23.8

24.1

24.3

26.3

26.2

24.3

1.0

 Σ PUFA

62.7

64.3

63.6

63.8

63.9

64.7

64.6

59.9

60.4

63.6

63.5

60.1

60.7

63.8

63.9

62.6

62.5

62.3

62.2

62.8

1.6

Coconut Oil

 Σ SFA

89.1

89.9

89.6

90.1

90.1

90.1

90.4

88.7

88.7

89.4

89.4

88.3

88.4

89.0

89.1

 

 

 

 

89.4

0.7

 Σ trans FA

1.5

2.2

2.2

2.1

2.1

1.2

1.4

2.0

2.0

1.3

1.5

1.8

1.9

0.8

0.9

 

 

 

 

1.7

0.5

 Σ MUFA

7.3

7.3

7.3

7.3

7.4

7.2

7.0

8.1

7.9

7.1

7.1

6.5

6.7

7.7

7.5

 

 

 

 

7.3

0.4

 Σ PUFA

1.8

1.8

1.9

1.6

1.6

1.8

1.8

1.3

1.3

1.7

1.8

0.74e

0.55e

2.0

2.0

 

 

 

 

1.7

0.2

Palm Oil

 Σ SFA

52.9

51.5

51.2

53.7

53.7

53.5

53.4

53.7

53.6

53.1

53.1

52.0

52.6

53.1

53.2

 

 

 

 

53.0

0.8

 Σ trans FA

2.1

3.1

3.2

3.4

3.5

2.6

2.6

4.0

3.8

2.8

2.8

3.8

4.0

2.8

2.9

 

 

 

 

3.2

0.5

 Σ MUFA

36.4

38.1

38.2

36.2

36.2

36.4

36.5

39.7

39.9

35.7

35.7

37.2

35.8d

36.4

36.4

 

 

 

 

37.1

1.4

 Σ PUFA

8.9

9.0

9.0

8.8

8.9

9.0

9.0

9.0

8.9d

8.7

8.7

7.7e

7.8e

8.8

8.9

 

 

 

 

8.9

0.1

aModes: F2 fiber optic probe 2-mm pathlength, F4 fiber optic probe 4-mm pathlength, T5 transmission 5-mm OD tubes

bDuplicate measurements: M1 and M2

cOutlier, single Grubb’s test

dOutlier, single Cochran’s test

eOutlier, double Grubb’s test

Precision

Precision statistical parameters for each oil or fat determined by FT-NIR are given in Table 2 for each of the sums of SFA, trans FA, MUFA, and PUFA. The repeatability Cochran outlier test [16] removes extreme values because they show significantly greater variability among duplicate analyses than those reported in other data sets. Three Cochran outliers found in Table 1 were removed. These were: (1) the sum of PUFA (22.78 %) for PH soybean oil from data set 1 with an overall mean of 22.1 %; (2) the sum of PUFA (8.9 %) for palm oil from data set 4 with an overall mean of 8.9 %; (3) the sum of MUFA (35.8 %) for palm oil from data set 6 with an overall mean of 37.1 %. The reproducibility Grubbs’ outlier test [16] removes laboratories or data sets with extreme averages. Grubbs’outliers were found in six sets of data in Table 1 and were removed. These were: (1) the sum of SFA (15.65 %) for PH soybean oil from data set 6 with an overall mean of 18.1 %; (2) the sums of PUFAs (19.5 and 19.4 %) for PH soybean oil from data set 8 with an overall mean 22.1 %; (3) the sums of SFAs in data set 4 for sunflower oil with an extreme averages of 10.0 and 10.1 % and an overall average of 11.5 %; (4) the sums of MUFA in data set 4 for sunflower oil with extreme averages of 29.3 and 28.9 % and an overall average of 24.3 %; (5) the sums of PUFAs (0.74 and 0.55 %) for coconut oil from data set 6 with an overall mean 1.7 %; and the sums of PUFAs (7.7, 7.8 %) for palm oil from data set 6 with an overall mean of 8.9 % (Table 1).
Table 2

Precision data

 

Test sample used in the present FT-NIR study

Selected data taken from the AOCS GC collaborative study [4]

PHa Soy

Soy

Canola

Sunflower

Coconut

Palm

PH Canola

Coconut

Sunflower

PHa Lard

Total saturated fatty acids (SFA) 

 Total number of laboratories

p

9

9

9

8

7

7

11

10

10

10

 Total number of replicates

Sum(n(l))

17

18

18

16

14

14

22

20

20

20

 Overall mean of all data (grand mean)

XBARBAR

18.13

16.68

8.43

11.48

89.37

52.96

16.06

84.19

7.51

40.27

 Repeatability standard deviation

s(r)

0.07

0.12

0.25

0.08

0.11

0.17

0.25

1.47

0.13

0.17

 Reproducibility standard deviation

s(R)

0.87

0.74

0.83

0.26

0.73

0.86

0.83

6.06

0.25

1.68

 Repeatability relative standard deviation

RSD(r)

0.38

0.69

2.98

0.72

0.13

0.32

1.56

1.74

1.70

0.41

 Reproducibility relative standard deviation

RSD(R)

4.81

4.45

9.82

2.27

0.81

1.63

5.14

7.20

3.28

4.18

 HORRAT value

 

1.86

1.70

3.38

0.82

0.40

0.74

1.95

3.51

1.11

1.82

Total trans fatty acids (trans-FA)

 Total number of laboratories

p

9.00

9

9

9

7

7

11

9

10

11

 Total number of replicates

Sum(n(l))

18.00

18

18

18

14

14

22

18

20

22

 Overall mean of all data (grand mean)

XBARBAR

28.02

1.51

1.21

1.10

1.66

3.23

26.27

0.10

0.17

0.99

 Repeatability standard deviation

s(r)

0.16

0.05

0.01

0.02

0.08

0.08

0.51

0.04

0.03

0.10

 Reproducibility standard deviation

s(R)

1.55

0.23

0.29

0.17

0.51

0.54

0.78

0.04

0.10

0.22

 Repeatability relative standard deviation

RSD(r)

0.58

3.22

1.19

2.13

4.54

2.43

1.94

35.35

15.94

10.06

 Reproducibility relative standard deviation

RSD(R)

5.54

15.45

24.13

15.15

30.51

16.76

2.97

35.88

60.34

21.64

 HORRAT value

 

2.29

4.11

6.21

3.84

8.23

5.00

1.21

6.37

11.50

5.41

Total monounsaturated fatty acids (MUFA)

 Total number of laboratories

p

8

9

9

8

7

7

11

10

11

10

 Total number of replicates

Sum(n(l))

16

18

18

16

14

13

22

20

22

20

 Overall mean of all data (grand mean)

XBARBAR

29.92

22.65

63.33

24.34

7.28

37.12

39.23

5.32

81.00

38.75

 Repeatability standard deviation

s(r)

0.17

0.22

0.44

0.07

0.11

0.08

0.82

0.06

1.10

0.36

 Reproducibility standard deviation

s(R)

1.89

1.83

1.80

1.01

0.45

1.47

1.38

0.14

4.04

1.42

 Repeatability relative standard deviation

RSD(r)

0.57

0.96

0.70

0.28

1.53

0.22

2.08

1.13

1.36

0.93

 Reproducibility relative standard deviation

RSD(R)

6.32

8.08

2.84

4.15

6.15

3.97

3.51

2.59

4.99

3.65

 HORRAT value

 

2.63

3.23

1.33

1.68

2.07

1.71

1.53

0.83

2.42

1.58

Total polyunsaturated fatty acids (PUFA)

 Total number of laboratories

p

7

9

9

9

6

6

10

10

11

10

 Total number of replicates

Sum(n(l))

13

18

18

18

12

11

20

20

22

20

 Overall mean of all data (grand mean)

XBARBAR

22.08

60.49

25.51

62.80

1.71

8.89

10.70

1.30

5.41

12.85

 Repeatability standard deviation

s(r)

0.08

0.14

0.19

0.25

0.02

0.01

0.19

0.01

0.08

0.17

 Reproducibility standard deviation

s(R)

0.59

0.79

1.03

1.62

0.23

0.14

0.32

0.03

0.14

0.50

 Repeatability relative standard deviation

RSD(r)

0.36

0.23

0.76

0.40

1.06

0.14

1.76

0.93

1.38

1.34

 Reproducibility relative standard deviation

RSD(R)

2.65

1.31

4.05

2.58

13.41

1.57

3.02

2.05

2.63

3.89

 HORRAT value

 

1.06

0.61

1.65

1.20

3.63

0.54

1.08

0.53

0.85

1.43

PH partially hydrogenated

Repeatability

All test samples had repeatability relative standard deviation, RSD(r), values calculated from intralaboratory data in the range 0.14–1.5 (Table 2), except those for the total SFAs for canola oil, which gave a slightly higher RSD(r) value of 2.98. However, for the totals of low trans FA, the RSD(r) values fell in the higher range of 1.2–4.0. It is noted that test samples with low trans fat levels near 1 % of total fat had superior RSD(r) values, which were found in the present FT-NIR study, relative to those reported in the official GC collaborative study [4]. Specifically, RSD(r) values of 10 and 35 were reported for two fat test samples with GC with overall means of 0.99 and 0.17, respectively [4].

Reproducibility

The reproducibility relative standard deviation, RSD(R), values calculated from interlaboratory data fell in the range between 0.82 and 9.82 for overall mean values from 89.5 to 7.2 (Table 2). For mean values in the lower range between 2.9 and 1.1, the RSD(R) values were significantly higher, ranging from 12.8 to 29.1. These high RSD(R) values were as unsatisfactory as those reported for low trans FA determinations in the official GC collaborative study [4], namely, RSD(R) values of 21.6 and 60.3 for means of 0.99 and 0.17, respectively.

HORRAT Value

The HORRAT value [12] is a useful index of method performance with regard to precision of interlaboratory data. It is the ratio of RSD(R) to PRSD(R), where the latter parameter is the predicted reproducibility relative standard deviation. PRSD(R) is calculated by using the Horwitz formula: PRSD(R) = 2C−0.15, where C is defined as concentration expressed as a mass fraction. For a concentration of 100 %, C = 1.0 and PRSD(R) = 2. Based on empirical data developed from over 10,000 interlaboratory studies published over the last century, the HORRAT value was assigned a magnitude of 1.0, with limits of acceptability between 0.5 and 2.0 [16]. Most of the FT-NIR data in this study gave HORRAT values within this range (Table 2). The two main areas that showed higher HORRAT values were for oils with low concentrations of trans FA from 1.1 to 2.9 %, which gave HORRAT values ranging from 7.8 to 3.8. These values are comparable to HORRAT values of 5.4 and 11.5 reported for fats and oils with total trans FA values of 0.99 and 0.17, respectively, in the official GC collaborative study [4]. HORRAT values greater than 2 may indicate several sources of errors including operating below the limit of determination for the measurement in question [16].

Accuracy

When the overall mean for each total FA value for each oil obtained by FT-NIR was compared to the corresponding GC results in the present study (Table 1), satisfactory Pearson’s correlation coefficients (R2) were found (see Fig. 1a–d). These values were R2 = 1.00 (p < 0.001) for SFA, R2 = 0.99 (p < 0.001) for trans FA; R2 = 0.99 (p < 0.001) for MUFA, and R2 = 0.99 (p < 0.001) for PUFA. Values of p < 0.05 were considered statistically significant. These Pearson’s correlation coefficients indicate that there is a strong positive correlation between the FT-NIR and GC results. The paired t test indicated that there was no statistical difference between the results obtained from these two methodologies (p = 0.434 for SFA, p = 0.606 for trans FA, p = 0.133 for MUFA, p = 0.561 for PUFA); all p values were p > 0.05.
https://static-content.springer.com/image/art%3A10.1007%2Fs11746-012-2116-9/MediaObjects/11746_2012_2116_Fig1_HTML.gif
Fig. 1

Pearson correlations of the fatty acid content (as percent of total fat) determined by FT-NIR and GC for total SFA (a), trans FA (b), MUFA (c), and PUFA (d) of each of the six fats and oils analyzed

An additional comparison of the new FT-NIR spectroscopic procedure with the established GC technique was carried out in order to determine whether there is sufficient agreement between the two sets of quantitative results and to establish whether the new procedure can be recommended as a rapid alternative method for routine analysis. A unique approach was developed by Bland and Altman [22, 23] that involved a graphic presentation of the analysis of variance. A plot of the difference between GC and FT-NIR against their mean was generated for each of the sums of SFA, trans-FA, MUFA, and PUFA for the oils investigated (Fig. 2a–d). The limits of agreement between the two methods (GC and FT-NIR) are represented by the mean difference ± 2 SD of the differences [22, 23]. The Bland-Altman plots in Fig. 2a–d were satisfactory and indicated that all values were within the ±2 SD limits of agreement.
https://static-content.springer.com/image/art%3A10.1007%2Fs11746-012-2116-9/MediaObjects/11746_2012_2116_Fig2_HTML.gif
Fig. 2

The mean difference between GC and FT-NIR fatty acid content for total SFA (a), trans FA (b), MUFA (c), and PUFA (d) of all six fats and oils investigated are presented as described by Bland and Altman [22, 23]. Values outside the mean difference of ±2 standard deviations are considered significant [22, 23]

The data obtained using the 8-mm OD (6-mm pathlength) transmission tubes were not included in Table 1 because the results led to a large standard deviation relative to those obtained with pathlengths of 2, 3, or 4 mm under our experimental conditions. Figure 3 shows significant discrepancies in the quantification of total SFAs and total PUFAs using the 8-mm OD tubes under our experimental conditions. It was not possible to modify the FT-NIR calibration models to process the FT-NIR spectral data collected with the 8-mm OD transmission tubes because the variability in the observed absorption spectra was excessively large (Fig. 4a–b). Figure 4a shows a comparison of the spectra for five replicate measurements of canola oil obtained using 5- and 8-mm OD transmission tubes. The absorption values for the 8-mm OD tubes were larger than 3.0 absorbance units, which is significantly outside the linear range of Beer-Lambert’s law. Figure 4b shows an expanded view of the five replicate measurements obtained with both the 5- and 8-mm OD transmission tubes, and indicated that only the latter exhibited a large scatter. Therefore, it is recommended that these calibration models be applied to FT-NIR spectra collected only with fiber optic probes having a 2- or 4-mm pathlength, or with 5-mm OD (approximately 3-mm ID or 3-mm pathlength) transmission tubes. The use of long-pathlength transmission tubes would have dictated the exclusion from the calibration models of an important wavelength range (near 5,800 cm−1) where saturation occurred. This range was critically needed for the development of some of our models for accurate FA profiling. By opting to use smaller pathlengths, we eliminated the need to exclude such an important spectral range from our calibration models.
https://static-content.springer.com/image/art%3A10.1007%2Fs11746-012-2116-9/MediaObjects/11746_2012_2116_Fig3_HTML.gif
Fig. 3

Comparison of FA content (as percent of total fat) determined by FT-NIR and GC for the two major saturated FA (16:0 and 18:0), total SFA, selected PUFA (18:2n-6 and 18:3n-3), and total PUFAs for canola oil. The following FT-NIR measurement modes were compared: 2- and 4-mm fiber optic probes, and 5- and 8-mm OD transmission tubes

https://static-content.springer.com/image/art%3A10.1007%2Fs11746-012-2116-9/MediaObjects/11746_2012_2116_Fig4_HTML.gif
Fig. 4

Comparison of five replicate absorption spectra of canola oil in the range of 4,500–9,000 cm−1 using 5- and 8-mm OD transmission tubes (a). The expanded region of the same spectra from 5,700 to 5,900 cm−1 is shown in b

The limits of prediction of any FA determined using these calibration models need to be addressed, since this impacts the usefulness of FT-NIR as a screening tool to assess for example low levels of total trans FA or total PUFA in a product. Since FT-NIR is a secondary method based on primary reference GC data, it is subject to the same limits of quantification as GC [15, 18]. Unfortunately, such limits were not reported for any FA in the published official AOCS GC method Ce 1 h-05 [4], possibly because the precision parameters RSD(r) and RSD(R) found for some FAs generally present at low concentrations were not encouraging; see the results of selected fats and oils taken from the GC official method AOCS Ce 1 h-05 [4] in Table 2. Determining prediction limits of FT-NIR data was addressed by Williams [24] who outlined a statistical approach to develop quantitative calibration models for FT-NIR spectral components present at low concentrations in a matrix. He defined the ratio of SD to the standard error of prediction (SEP) as the residual prediction deviation, RPD, and recommended that RPD values greater than 5, but no less than 3, should be useful to predict reliability of results. Table 3 provides a summary of the RPD values previously obtained for the sums of SFA, trans FA, MUFA, and PUFA values generated using the pre-developed calibration models [8] employed in the present study. As shown in Table 3, the majority of the RPD values ranged well above 3 (and most above 5) with no exception, indicating a satisfactory measure of confidence in the results. The model used for very low trans FA levels (<3 %), with a range between 0.07 % for total trans FA in triolein and 3.06 % for total trans FA in a test sample of lard, had an RPD of 3.2 with a root mean square error of cross validation (RMSECV) of 0.2. The corresponding RPD obtained with the low to medium range trans FA model was 5.3 with an RMSECV equal to 1.3. While the FT-NIR prediction limit for trans FA determinations may be as low as 0.2 % as percent of total fat, the precision of FT-NIR determinations at these low levels is as unsatisfactory as those reported in the official GC method for the same total trans FAs [4]; see Table 2.
Table 3

RPD and RMSECV calibration statistics obtained using pre-developed calibration models [8]

 

RMSECV

RPD

r2

Calibration range

No. of samples in prediction set

Very low trans FA model

 Σ trans (<3 % trans FA)

0.2

3.2

0.90

0–3

96

 Σ MUFA

2.2

10.1

0.99

16.85–99.71

114

Low to medium trans FA model

 Σ SFA

1.0

8.0

0.98

0.13–40.97

95

 Σ trans (<20 % trans FA)

1.3

5.3

0.96

3–20

81

 Σ MUFA

2.3

9.8

0.99

16.85–99.71

103

 Σ PUFA

2.0

10.6

0.99

0.08–71.33

100

High saturated FA model

 Σ SFA

1.2

22.4

1.00

7.29–89.07

47

 Σ trans

0.6

5.1

0.96

0–15.23

54

 Σ MUFA

2.3

7.6

0.98

7.1–71.51

48

 Σ PUFA

0.6

13.8

0.99

1.78–25.98

29

High trans FA model

 Σ SFA

1.6

5.1

0.96

6.41–40.97

81

 Σ trans (> 20 % trans FA)

1.6

12.0

0.99

0.13–58.73

70

 Σ MUFA

3.0

5.4

0.96

10.27–74.52

77

 Σ PUFA

1.4

7.7

0.98

0.05–31.43

62

RPD is defined as the ratio of standard deviation (SD) to standard error of prediction (SEP) (see ref. 24)

RMSECV root mean square error of cross validation

For labeling purposes the total trans FA content does not need to be reported unless it exceeds the legal definition of 0 g trans fat/serving, which is <0.5 g/serving in the US [25] and <0.2 g/serving in Canada [26]. Due to differences in serving size declared on labels for edible fats and oils, these limits correspond to means of <3.6 % trans fat (as percent of total fat) in the US and <2.2 % trans fat (as percent of total fat) in Canada. Both the FT-NIR and GC results of all the unhydrogenated processed edible fats and oils in this study met the legal definition of 0 g trans fat/serving in both countries. Since these values are equal to or higher than the FT-NIR prediction limit of 1.3 % (Table 3) obtained with the low to medium range trans FA model, this methodology can potentially be used to rapidly screen and verify whether a product is in compliance with the declaration of 0 g trans fat on the nutrition facts panel in both the US and Canada [25, 26].

Should there be a need to further improve the reliability of low trans FA determinations below 1.3 % of total fat by FT-NIR, the availability of edible oils having a zero trans fat content to serve as reference standard would be required. FT-NIR calibration models are matrix dependent, and the FA composition of unknown test samples should match those used to generate the low trans fat calibration model. This remains a challenge because even commercially available cold-pressed edible oils used as zero-trans reference material generally contain trace amounts of trans-FAs.

Conclusions

Calibration transfer among three laboratories using five different FT-NIR spectrometers equipped with transmission tubes and/or fiber optic probes acquired from a single vendor proved to be successful. Therefore, the need to develop new and costly calibration models for FA determination on different instruments from the same vendor was eliminated. The application of these pre-developed calibration models yielded satisfactory precision and accuracy under our experimental conditions. Pathlengths of 2 and 4 mm (fiber optic probes) and approximately 3-mm pathlength transmission tubes (5-mm OD) are recommended. Unlike GC, FT-NIR spectroscopy requires no sample preparation for the FA determination of edible fats and oils. The FT-NIR protocol for calculating FA composition is fully automated and independent of analyst skill or bias because calibration models cannot be subjectively modified. FT-NIR is a rapid and viable alternative method for the time-consuming GC methodology. The potential of applying FT-NIR to the routine screening of edible fats and oils for their total SFA, trans FA, MUFA, and PUFA contents for regulatory compliance monitoring in both the US and Canada will be further investigated.

Acknowledgments

We thank Drs. Peter Krygsman and Hui Li of Bruker Optics from Canada and USA for their support and for providing the opportunity to use Bruker MPA FT-NIR instruments in this study.

Copyright information

© AOCS 2012