UAEU Dietary Fiber of Intact Cereal Food Products presentation

Paper presentation:

Make a review of the selected paper

First slide: Name of the paper, Journal and year.

Introduction: from the paper and also from other sources if the paper brings few information about fibers. we are not interested only about the health benefits, if the paper talks only about it, get more information related to food technology. We are working with cereal technology, try to make a nice introduction regarding the role of fiber in food, different quality aspects in food related to fiber.

Materials and Methods: Make it clear for the others to understand. You can also add as a flow chart. Add materials used quantities according to the methodology and the methodology should be complete. If the methodology cites the AOAC method, please search for the method and explain.

Results and discussion: Explain the results regarding the content of fiber. If the paper brings other aspects related to sensory and quality analysis, please add these results. Add results (graphs, tables, figures with legends) in one slide and explain the results and discussion with few texts if necessary.

At the end of Results and Discussion session, summarize the results and discussion as topics, to make it clear for everybody, like a conclusion for the results. It will serve for questions and answers.

Conclusion: What is the authors conclusion of the paper and what did you learn from the experiments, why is this paper relevant for food technology? Do you think the paper has any gap? Comment what do you think they should have added.

J. Agric. Food Chem. 2000, 48, 4477−4486
4477
Determination of Total Dietary Fiber of Intact Cereal Food
Products by Near-Infrared Reflectance
Douglas D. Archibald*,† and Sandra E. Kays
Downloaded via UNITED ARAB EMIRATES UNIV on April 6, 2020 at 12:29:28 (UTC).
See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.
Quality Assessment Research Unit, Agricultural Research Service, U.S. Department of Agriculture,
Russell Research Center, Athens, Georgia 30604-5677
Near-infrared reflectance spectra of cereal food products were acquired with a commercial dualdiode-array (Si, InGaAs) spectrometer customized to allow rapid acquisition of scans of intact
breakfast cereals, snack foods, whole grains, and milled products. Substantial gains in the
performance of multivariate calibration models generated from these data were obtained by a
computational strategy that systematically analyzed the performance of various spectral windows.
The calibration model based on 137 cereal food products determined the total dietary fiber (TDF)
content of a test set of 45 intact diverse cereal food products with root-mean-squared error of crossvalidation of between 1.8 and 2.0% TDF, relative to the laborious enzymatic-gravimetric reference
method. The calibration performance is adequate to estimate TDF over the range of values found
in diverse types of cereal food products (0.7-50.1%). The method requires no sample preparation
and is relatively unaffected by specimen moisture content.
Keywords: Total dietary fiber; cereal food products; nondestructive analysis; near-infrared
reflectance; spectroscopy; InGaAs diode array; multivariate calibration; spectral window
INTRODUCTION
The dietary fiber component of foods is important for
the health of individuals and the public at large (Gorman and Bowman, 1993; Marlett and Slavin, 1997), and
fiber content often is used in the marketing of products.
As a consequence, the Nutrition Labeling and Education
Act of 1990 requires food producers to report total
dietary fiber (TDF) on consumer product labels (United
States Congress, 1995). In the United States, the
current accepted definition for TDF is the sum of lignin
and polysaccharides not digestible by human alimentary
enzymes as estimated by AOAC method 991.43, which
approximates the enzymatic aspects of human digestion
by gravimetric analysis of the residue after a sequence
of timed enzymatic treatments (AOAC, 1990; Lee et al.,
1992; AOAC, 1992).
Fiber assays are time-consuming and often produce
a considerable amount of chemical waste, so it is not
surprising that efforts have been made to develop
instrumental methods. Technology for near-infrared
reflectance (NIRR) analysis of dietary fiber was reviewed by Kays et al. (1999). Briefly, Baker published
the first reports on estimation of neutral detergent fiber
(NDF) in human foods in the early 1980s (Baker, 1983;
Baker, 1985). However, the near-infrared (NIR) analysis
of fiber in animal forage began in the mid-1970s (Norris
et al., 1976; Marten and Templeton, Jr., 1989) and
culminated in a handbook in 1985 (USDA, 1985; USDA,
1989), and an official method for acid-detergent fiber
(ADF) in 1988 (Barton, II, and Windham, 1988). This
laboratory recently has developed spectroscopic methods
for dietary fiber in human foods; these methods apply
to a broad range of cereal food products using milled
specimens (Barton, II, et al., 1995; Kays et al., 1996;
Windham et al., 1997; Kays et al., 1997; Archibald et
al., 1998a; Archibald et al., 1998b; Kays et al., 1998;
Kays and Barton, II, 1998; Kays et al., 2000).
The work presented in this manuscript is the first
effort to determine TDF from spectral properties of
intact cereal food products taken directly from the box
or bag. The study started with the proposal that it would
be possible to create a single TDF calibration from NIRR
spectra of diverse intact cereal food products. One
objective of the analysis was to determine if there are
classes of products that are more difficult to model.
NIRR measurements were made with a commercial
spectrometer that incorporates silicon and InGaAs diode
arrays spanning the range from 400 to 1700 nm. Due
to the high sensitivity of InGaAs diode arrays, these
devices are finding use in applications of NIR spectroscopy that require rapid remote reflectance measurements (Huthfehre et al., 1995). In the study presented
here, the rapid measurement capability was used to
obtain multiple optical samples representing NIRR from
a large surface area of each cereal food product, because
many of the products were highly heterogeneous. The
NIR portion of the spectral range measured by the
silicon and InGaAs arrays (800-1700 nm) is suited to
this application because it has more penetrating ability
than the longer wavelength NIR. The geometry of the
measurement also may be appropriate for at-line or online process measurements for a variety of constituents.
MATERIALS AND METHODS
* E-mail: dda10@psu.edu. Telephone: (814) 865-8449.
Fax: (814) 863-7043.
† Current address: Agronomy Department, 116 Agricultural
Sciences and Industries Building, The Pennsylvania State
University, University Park, PA 16802.
Samples and Sample Treatments. One-hundred
thirty-six cereal food products, including breakfast
products, snack foods, flours, and baking mixes, were
obtained from local retail stores. Few retail products
10.1021/jf000206j This article not subject to U.S. Copyright. Published 2000 by the American Chemical Society
Published on Web 08/31/2000
4478 J. Agric. Food Chem., Vol. 48, No. 10, 2000
Archibald and Kays
Table 1. Cereal Food Product Groups Represented in
the Study after Removal of Five Outliers
I
II
III
IV
V
VI
a
product class
calibr.
set
test
set
full
set
cereals without fruit or nuts
cereals with fruit or nuts
crackers, chips and biscuits
grain flour and baking mixes
whole and crushed grains
predominantly bran
total
44
13
11
7
10
7
92
13
5
10
3
10
4
45
57
18
21
10
20
11
137
abreakfast
abreakfast
Not including breakfast products found in other classes.
Table 2. Composition and Property Variation of
Untreated Cereal Food Products in the Study
minimum
maximum
range
mean
median
# samples
moisture
%a
TDF
%b
protein
%c
sugar
%d
fat
%d
density
g/mLe
2.3
13.7
11.3
6.4
5.1
137
0.7
50.1
49.4
10.2
7.6
137
4.0
22.3
18.3
11.8
11.8
137
0.0
55.6
55.6
13.5
8.8
137
0.0
25.0
25.0
4.9
3.2
137
0.056
0.842
0.786
0.319
0.227
113
a By oven method. b Total dietary fiber by AOAC method as
described in Materials and Methods. c Combustion nitrogen analysis × 6.25. d As determined from the product nutrition labels.
e Bulk product density by mass/volume measurements.
were found in the middle to upper range of TDF.
Therefore, the sample set was supplemented by creating
6 simulated cereal products from 6 real products by
binary mixing, resulting in a total of 142 untreated
intact cereal food products. Five of the 142 were identified as outliers, as described later. As outlined in Table
1, approximately one-third of the 137 untreated samples
was reserved for testing, and the remainder was used
to optimize the parameters of the calibration model.
Sample composition was diverse (Table 2). Therefore,
available untreated samples were distributed between
calibration and test sets to achieve similar sample
histograms for the constituents TDF, protein, sugar, and
fat, based on either reference measurements or the
values obtained from product nutrition labels.
Because the condition of the sample may not be
controllable by the analyst in some applications, calibration development involved specimens that were
treated to simulate varying moisture content levels, and
degrees of breakage of products. The spectra of treated
specimens were used to characterize and stabilize the
model response. Treated specimens were created by
reducing the particle size of products (by crushing in
heavy plastic bags) or by altering the moisture content
by desiccation or contact with moist air. The particle
size and moisture treatments expanded the number of
specimens from 137 to 361, and increased the moisture
content range substantially to 2.2-21.4%, with the
mean and median moisture contents each increasing by
1.1% over the untreated set described in Table 2.
Although the character of many products was altered
substantially by crushing, the range, mean, and median
of product bulk density did not change significantly after
particle size treatments, because breakfast cereal products were the primary products treated, and these had
low-to-intermediate bulk densities. When treated samples
were used to stabilize or test models, they were assigned
to the same set, calibration or test, as their untreated
counterpart.
NIRR Sampling Strategy and Data Collection.
Accurate subsampling was required to split intact cereal
food specimens for treatments or reference analyses. To
do this, typically three to six boxes or bags of a product
were mixed thoroughly and then poured out at a
constant rate on a rotating platform. The resulting
doughnut-shaped pile was split along a diameter to form
two subsamples. To obtain a representative optical
subsample for a product, the product was packed onto
the measuring window eight individual times. For each
“repack”, the instrument was set to collect four rescans
of the stationary specimen. In most instances, the
regression models were generated from the average
spectrum for each specimen (eight repacks by four
rescans). The repeat data was used to analyze the
variance, within the calibration model, that was associated with rescans or repacks.
The NIRR spectra were collected with a Perten
Instruments model DA7000 spectrometer (Springfield,
IL). This instrument has silicon and InGaAs diode
arrays and an intense broadband light source, making
it possible to measure reflectance from a large area of
the specimen surface (about 10 cm diameter). The diodes
are centered at 10-nm intervals, but software is used
to spline-interpolate spectra to a data interval of 5 nm.
After acquiring a reference scan of Spectralon (Labsphere, Inc., North Sutton, New Hampshire), the instrument returned the -log[Rsample/Rreference] for the
sample, with the two spectral regions spliced at 950 nm
to cover the range from 400 to 1700 nm. Although the
instrument internally averages 30 spectra per second
of acquisition time, for the purposes of this study, a
spectral scan was defined as the spectrum generated
by the instrument after a 1-s acquisition. To facilitate
the loading and unloading of specimens, the spectrometer was operated in side-view mode with the addition
of a custom-built hopper with a drop hatch and an
adjustable sample thickness (2.5-10 cm) (Figure 1).
With this design, the Spectralon reference disk also can
be inserted in place of the sample. The manufacturer
recommends taking a reference scan every hour, or
sooner if suggested by the automatic instrument diagnostics. In the present study, reference scans were
collected more frequently, typically after every second
sample (two by eight repacks).
Reference Analysis of Total Dietary Fiber. The
TDF of the specimens was determined by AOAC method
991.43 as described previously (Lee et al., 1992; AOAC,
1992; Kays et al., 1998). Not every sample was subjected
to an identical analysis procedure: for products high
in fat (>10%) or sugar (>20%), solvent extraction
procedures were used to either defat or desugar the
sample prior to application of the enzymatic-gravimetric method. An AOAC validation study for method
991.43 yielded a mean repeatability (standard deviation
of blind duplicates) of 0.71% TDF and a mean interlaboratory reproducibility of 1.28% for three cereal
products analyzed by 11 laboratories (Lee et al., 1992).
The standard error of the laboratory (SEL) for application of the AOAC TDF reference method in this study
was 0.39% when calculated from 46 of the cereal food
product specimens with replicates performed on different days. SEL is the pooled standard deviation of the
repeatability (ASTM, 1995).
Calibration Model Development. Data processing
was performed in the MATLAB computing environment
(The MathWorks, Natick, MA) using custom subroutines that incorporated de Jong’s PLS algorithm (de
Jong, 1993) and the Savitzky-Golay derivative (Sav-
Determination of TDF of Cereal Food Products
J. Agric. Food Chem., Vol. 48, No. 10, 2000 4479
Figure 1. InGaAs/silicon dual-diode-array NIR spectrometer and hopper accessory for rapid loading and measurement of the
reflectance of a large surface area of each cereal food product.
itzky and Golay, 1964; Madden, 1978) MATLAB functions available in the PLS Toolbox 2.0 (Eigenvector
Research, Manson, WA). Preliminary analysis of the
data led to the conclusion that interference from color,
hydration effects, or irrelevant chemical variations was
larger than that due to instrument noise or scattering
differences of the various intact products. Furthermore,
because interference was not evenly distributed across
the spectrum, it was found that selecting subsets of
wavelength variables substantially improved the calibration performance. Establishing which variables to
retain or delete, however, was not trivial because of
extensive band overlap and mixing, the insufficient
knowledge about the spectral properties of the various
components of dietary fiber, and the presence of a highly
variable background.
To make variable selection more systematic, a computational strategy called spectral window preprocessing was developed. This procedure extensively tested
the data set to determine which spectral windows to
include and which windows to delete (Archibald and
Akin, 2000). In this study, the spectral window preprocessing technique was implemented as described
below. The calibration set of untreated samples (n )
92) was evaluated first by conventional full-spectral
partial least-squares regression (PLSR) (Martens and
Næs, 1989) to remove obvious outliers and to establish
the optimal spectral transformation. The latter was
determined by comparison of the calibration-set rootmean-squared error of cross-validation (RMSECV) (Martens and Næs, 1989) produced by various types of
spectral preprocessing. Second, a spectral window deletion experiment was performed where model leave-oneout cross-validation statistics were compiled for a
triangular array of spectral window widths (5, 10, 15,
…, 400 nm) and window positions (beginning at 800, 805,
…, 1695 nm). The optimal range to delete was determined from the window parameters of the model with
the lowest RMSECV. The next step was to compute a
spectral window addition experiment, with widths of
400, 405, …, 900 nm and beginning positions of 800, 805,
…, 1300 nm, with all of the data except for the variables
of the deleted window. Again, the global minimum
RMSECV was used to determine the best limits, which
in this case established the spectral range. These
exhaustive computational testing experiments involved
92 samples and 181 spectral variables and may be
calculated in a few days with currently available
desktop computer technology, although the computation
time can be reduced easily by changing the spectral
4480 J. Agric. Food Chem., Vol. 48, No. 10, 2000
Archibald and Kays
Figure 2. Response image for a spectral window deletion
experiment for the prediction of TDF of freshly opened, intact
cereal food products from SNV-transformed second derivative
NIRR calibration spectra (n ) 92). The gray level of each pixel
represents the RMSECV of the calibration set for a 13-factor
PLSR model when a spectral window of a given size (ordinate)
and position (abscissa) was deleted from the matrix of calibration data. The highest error of the color bar (darkest shade)
was set to the value of RMSECV obtained using the full
spectral range (2.63% TDF for 850-1700 nm). The bullet
symbol indicates the optimal region to exclude.
window limits, the window step size, or the number of
cross-validation segments. Predictive models were generated from the optimized set of PLSR model parameters using untreated and treated calibration samples
and were evaluated with the test set.
RESULTS AND DISCUSSION
Development and Testing of Calibration Models
with Untreated Samples. Initial attempts to develop
a TDF calibration model with the use of either underivatized or derivatized -log(R) spectra identified 3
of the 95 original untreated calibration samples that
consistently were outliers. One specimen (toasted wheat
germ) was removed because its protein content was far
higher than any other product. A second was removed
because of its extremely different optical properties
(whole popcorn kernels). A third (buckwheat groats) was
omitted because the TDF predictions were always
extremely high compared to the measured value. Of the
preliminary models evaluated with the untreated calibration set, digital second derivative (D2) spectral
processing followed by standard normal variate transformation (Barnes et al., 1989) (SNV) and mean centering yielded the best PLSR cross-validation error for the
NIR region from 800 to 1700 nm. The derivative was
calculated by the Savitzky-Golay convolution method
(Savitzky and Golay, 1964; Madden, 1978) with a width
of nine data points (45 nm) and a third-order polynomial
fit. To improve these results, two spectral window
computational experiments were conducted to optimize
the selection of the spectral range. First, a PLSR
window deletion experiment on D2-SNV data was
calculated, where the SNV normalization was performed
for each tested set of wavelength variables (Figure 2).
This showed that substantial improvement in RMSECV,
from 2.6 to 2.1% TDF, was obtained by deleting the
spectral range represented by the bullet symbol in
Figure 2. Those spectral window parameters produced
the lowest calibration set RMSECV among all of the
Figure 3. Plots illustrating the transformation of intact cereal
food product NIRR spectra for optimization of the linear
calibration for TDF: (a) unprocessed calibration spectra (n )
92), where R is [Rsample/Rreference]; (b) D2-SNV preprocessed
calibration spectra with vertical dotted lines indicating the
separation between included and excluded regions; and (c)
spectral activity within the model expressed as the standard
deviation across samples of the regression-vector-weighted
spectra (STD(RVWS)). RVWS are computed by element-byelement multiplication of each preprocessed calibration spectrum by the optimized regression coefficients. Abbreviations:
ov ) overtone; comb. ) combination band; str ) overtone of
bond stretching vibration.
possible window deletions that formed the grid of the
computation. The other computational experiment evaluated which spectral window to include, and this procedure established the best spectral limits for the preprocessing (data not shown). The latter step produced
only a slight improvement in error but used two fewer
factors. The effect of the preprocessing parameters
optimized by the spectral window experiments is shown
in Figure 3a,b. The spectral range from 845 to 1685 nm
was selected, except for the window from 1370 to 1480
nm. Second derivative preprocessing alone effectively
extracted the structural bands from optical artifacts.
However, the process of normalizing the relevant bands
and extraction of signal with the PLSR algorithm was
improved by deleting the highly variable spectral region
between 1370 and 1480 nm.
Determination of TDF of Cereal Food Products
Figure 4. Estimates of the uncertainty of PLSR models for
the determination of TDF from NIRR of a diverse set of freshly
opened, intact cereal food products: (a) RMSECV for the
untreated calibration set (ncal ) 92) with D2-SNV preprocessing over the optimal spectral range; (b) test set error (ntst )
45) estimated by prediction of the test set by the calibration
set alone (*), or by cross-validation on the specimens in the
test set combined with the calibration set (×); (c) full set error
(ncal + ntst) estimated as RMSEC (*) or RMSECV (×). The
circles mark the error estimate for the optimal model as
determined from part a, whereas the square symbol indicates
the optimal model by the cross-validation procedure when both
the calibration and test samples were included.
The calibration set RMSECV versus the number of
factors for the optimal preprocessing parameters is
shown in Figure 4a. An 11-factor model was optimal for
the 92-member calibration set (RMSECV ) 2.10% TDF,
R2 ) 0.96). On the basis of the residual standard
deviation of the spectral residuals, two test set samples
were determined to be outliers of the model and so were
removed, leaving 45 test samples. One sample was “Low
Fat Apple Cinnamon Muffin Mix” and the other was
“Cinnamon Toast Crunch” breakfast cereal. The fact
that both contain cinnamon is probably not a factor,
because several other products contained cinnamon but
were not spectral outliers. The muffin mix may be
spectrally unique because it contained an unusual
component, artificial fruit particles made of dextrose.
Similarly, the breakfast cereal may be a spectral outlier
as a result of the composition of the crystalline sugar
coating. Since the spectral window technique evaluated
a large number of wavelength sets, it could be argued
that the calibration set cross-validation error estimates
are likely to be too low. Therefore, final conclusions
about model performance were based on analysis of a
large test set that was not used for optimization of the
spectral window ranges. An estimate of the model error
was generated from the test set by two methods (Figure
4b). By simple prediction of the test set samples using
the calibration model based on 92 samples, the rootmean-squared error of prediction (RMSEP) (Martens
and Naes, 1989) was 1.92% TDF for an 11-factor model
(R2 ) 0.96). By leave-one-out cross-validation of each
of the 45 test samples (while always retaining the 92
calibration samples), the RMSECV was 1.92% TDF for
an 11-factor model and 1.84% for a 13-factor model (R2
J. Agric. Food Chem., Vol. 48, No. 10, 2000 4481
Figure 5. Plot of predicted versus measured TDF values for
the 92 untreated calibration samples (circles) and 45 untreated
test samples (triangles) as estimated by full cross-validation
on all 137 samples. The PLSR predictions were generated with
13 factors using the spectral preprocessing parameters established by computational experiments performed on the untreated calibration samples. The thick solid line is the ideal
correlation with an intercept of 0 and slope of 1. The thin lines
are parallel to the ideal correlation at plus or minus two times
the full-set RMSECV.
) 0.96). Unlike the aforementioned RMSEP statistic,
the test-set cross-validation method estimated the performance of a PLSR model produced from 137 samples
(instead of 92). More independent samples generally
yields more modeling power, and hence, the data set
could support a model with two additional factors to
achieve slightly better estimates of the same test
samples. The number of factors added is reasonable, as
ASTM guidelines recommend an infrared multivariate
calibration set contain at least six spectra per regression
factor (ASTM, 1995). The number of samples in this
application probably was critical because the products
have widely varying ranges of composition and optical
properties. By use of all of the calibration and test data
for untreated samples, the error was estimated to be
1.99% or 1.96% TDF by the RMSECV method for 11or 13-factor models, respectively (both with R2 of 0.96)
(Figure 4c). The root-mean-squared error of calibration
(RMSEC) on all samples was considerably lower (1.61
and 1.53% TDF, Figure 4c). The ratio of TDF range to
test set RMSECV was over 25, indicating good discriminating power for the TDF content of intact cereal food
products. This also can be judged visually in Figure 5.
The model testing and stabilization studies described
in the remainder of this report use the preprocessing
and spectral windows determined from the 92 untreated
calibration samples, and estimate model errors by
performing cross-validation on the test set with 13 PLSR
factors.
Analysis of Repeated Spectral Measurements
for Untreated Samples. Repeated measurements of
each specimen were taken in two ways: (1) the specimen
was scanned four times without removal (a rescan), and
(2) the specimen was loaded into the hopper eight times
(a repack). Each of the individual scans was preprocessed with the optimized parameters (D2-SNV), and
the rescan or repack variances in the model for each
specimen were pooled (Figure 6a,b). The noise due to
rescans arose primarily from the instrument and should
be extremely low because of negligible changes in the
4482 J. Agric. Food Chem., Vol. 48, No. 10, 2000
Archibald and Kays
Figure 7. Variation in TDF prediction error due to the
fluctuation in the spectral properties of the untreated subsample loaded into the sample hopper: (a) model error when
individual repacks were used for calculating and testing the
model; and (b) decrease in sampling error as increasing
numbers of repack scans were averaged. The error was
estimated from the 45-element test set by cross-validation
using 92 untreated calibration samples and 13 PLSR factors.
Figure 6. Pooled repeat standard deviation per specimen at
each wavelength as calculated from the four rescans of the
eight repacks obtained for each untreated cereal food product
(n ) 137, both calibration and test samples): (a) 137 pooled
rescan standard deviation spectra create the shaded area, the
line with the square symbols marks the mean repack standard
deviation, and the line with the circle symbols marks the mean
rescan standard deviation; (b) similar to part a, except that
137 pooled repack standard deviation spectra create the
shaded region. Each scan was preprocessed with the second
derivative and SNV applied over the optimal range.
sample condition, instrument condition, and elapsed
time between measurements. However, the rescan noise
varied with wavelength, having many of the features
of the average cereal product spectrum. Rescan noise
was the same magnitude as the repack noise above 1680
nm, which probably accounted for the optimal upper
limit of the spectral range established by the spectral
window experiments. The rescan noise was only marginally smaller than the repack noise in the range
1500-1680 nm, a region that contains at least two
important spectral signals. These spectral features can
be seen in the spectral plot and the plot of spectral
activity within the model, as described later (Figure
3a,c). Rescan noise increased at wavelengths shorter
than 950 nm, where the silicon array operates. Rescan
noise was approximately the same as the repack noise
from 845 to 930 nm, a range that contained one
important spectral peak (see below). However, repack
noise created by the tails of absorption by chromophores
was the main cause of the preferred deletion of the lower
region of the spectral range (800-845 nm). Increased
repack noise also was observed in the deleted window
from 1370 to 1480 nm, primarily as a result of variation
in absorbance bands attributed to the free O-H stretch
of crystalline sugars, which arose from the nonuniform
distribution of added sugar in many cereal food products. There were some similar free O-H bands at 976
nm, but these bands were small enough that they did
not degrade the calibration model appreciably (compare
Figure 6 with Figure 3b). The repack noise increased
well above the background at 942 and 967 nm. These
wavelengths are near the splice point between the
silicon and InGaAs spectral ranges. For cereal foods, the
spectral region of the splice was a relatively featureless
valley between two absorbance bands. Differences in
reflected signal of each repack generated a mismatch
in the slope of each spectrum meeting at the splice point.
In the region between 1100 and 1365 nm, the repack
noise was much larger than the rescan noise (Figure
6). This range contained many spectral features that
were significant to the model (Figure 3c and see below).
When each repack scan was analyzed separately by
the PLSR model, substantially different TDF predictions
were produced, causing the RMSECV for the test set to
vary widely (Figure 7a). This is thought to be a result
of the heterogeneous nature of many of the products.
The fluctuation was minimized by averaging multiple
optical samples to represent a larger area of each
specimen. When six or seven repacks were averaged,
the model error was relatively constant (Figure 7b). An
additional repack yielded no improvement. Moreover,
no correlation was found between the pooled repack
noise for a spectrum and the residual error in predicting
TDF when using the eight-repack average spectra (data
not shown). Since the diameter of the illumination area
was approximately 10 cm, approximately 550 cm2 of
product surface must be measured with the NIR beam.
In contrast to repacks, use of two or more rescans had
no effect on the model performance (data not shown).
Spectral Basis of the Calibration Model Developed from Untreated Samples. Exact spectral interpretation of the NIRR model for TDF was difficult, as
molecular vibration absorbance bands were highly
overlapped. Also, the indigestible fraction of food products consists of several very different components,
further confusing the identification of specific bands.
Determination of TDF of Cereal Food Products
J. Agric. Food Chem., Vol. 48, No. 10, 2000 4483
Table 3. Statistical Analysis of Test Set TDF Prediction Error and Bias for the 6 Product Classes
product
classa
# of samples
(within class)
SECVb
(within
class)
class
biasc
# of samples
(outside
class)
SECVb
(outside
class)
bias
(outside
class)
T-test on bias
(class vs
outside class)d
F-test on SECV
(class vs
outside class)e
F-test on SECV
(class vs
class I)f
I
II
III
IV
V
VI
13
5
10
3
10
4
1.40
1.81
1.28
2.31
2.43
2.00
-0.08
-0.05
-0.06
0.04
-0.07
0.73
32
40
35
42
35
41
1.99
1.84
1.97
1.80
1.62
1.80
0.05
0.02
0.03
0.01
0.03
-0.06
58%
53%
55%
51%
56%
79%
72%
62%
72%
78%
83%
70%
70%
53%
81%
82%
76%
a Described in Table 1. b Standard error of cross-validation (ASTM, 1995) calculated for the 13-factor PLSR D2-SNV model with optimized
spectral range, and using only untreated calibration samples. c Bias is the difference between the average predicted values and the average
measured TDF values. d Probability that the bias is different between the samples in the product class and samples outside the product
class. e Probability that the SECV is different between the samples in the product class and samples outside the product class. f Probability
that the SECV is different between the samples in the product class and samples in product class I.
However, the relative importance of different spectral
ranges can be demonstrated by calculating the standard
deviation across samples of the set of preprocessed
calibration spectra whose variables have been weighted
by the regression vector (STD(RVWS), Figure 3c). The
STD(RVWS) spectrum indicates spectral bands that
were weighted heavily in calculation of the predictions.
The full-width at half-maximum of each of these bands
fall in the range 20-35 nm, as would be expected for
solid-state vibrational bands. However, the band positions are only approximate, because the D2-SNV preprocessing could cause signal variance to shift from the
peaks of absorbance bands to nearby spectral features
such as valleys. Correlation of individual bands to TDF
was low (typically no greater than 0.5). One plausible
explanation is that even if a band measures an individual component of the food product, that component
may only indirectly be related to TDF. For example, lowlipid-content products did not always have a high fiber
content, but high-fiber-content products often had low
lipid content. Another possibility is that an active band
could be sensitive to more than one component, and
consequently, its value is only significant in the context
of other bands. For example, one band might be sensitive to both digestible and indigestible polysaccharides,
while another was sensitive to lipid and digestible
polysaccharides, and a third to lipid alone.
On the basis of published correlation tables (Murray
and Williams, 1987; Osborne et al., 1993), the molecular
vibrations of various groups of bands were identified in
the plot in Figure 3a. Two sharp bands were assigned
as the first and second overtone of O-H stretching of
free hydroxyl in a crystalline sugar (mainly sucrose). A
sharp feature due to the first overtone of aromatic C-H
of bran was present at 1685 nm, which was an important analytical band for TDF. Various C-H stretch
second overtones and C-H combination bands were
used by the calibration model, as were the C-H stretch
third overtones. The series of C-H stretching bands
suggests that the model was sensitive to a wide range
of functional groups in different types of molecules
(aromatic, methyl, methylene, or methine; lipids, proteins, polysaccharides, and lignins). First and second
overtones of O-H stretch vibrations also were important for the calibration model and could arise from
multiple components of the products (starch, cellulose,
hemicellulose, simple sugars). The absorption of protein
N-H stretch first and second overtones also could
contribute to the model.
Effect of Product Class and Composition on
Model Performance. Cereal food products differ in
their morphology, structure, and composition, and any
one of these characteristics could create bias or error
in the NIRR prediction model. The effect of membership
in the product classes of Table 1 on test set prediction
error and bias was evaluated statistically (Table 3). The
bias-corrected standard error of cross-validation (ASTM,
1995) (SECV) was used in the table because test-set
RMSECV was found to be the same as SECV except for
type VI products (predominantly bran), where it was
slightly higher because of positive bias (2.13 vs 2.00).
The T-test, on the bias between a particular class and
all other data in the test set, indicated that type VI
products had the only product-class-related bias problem
(columns 4, 7, and 8 of Table 3), suggesting that type
VI products were different from the others and were not
well represented in the data set. Product class SECVs
in column 3 of Table 3 were evaluated in two ways: (1)
by comparison of the SECVs produced for the class
against the remainder of the samples in the test set
(columns 6 and 9 of Table 3) and (2) comparison between
the class SECV and the SECV of class I (column 10 of
Table 3). The lowest SECVs were for breakfast products
of class I, and snack products (class III), and these
SECVs were not found to be statistically different from
each other. The breakfast products with fruit or nuts
(class II) had a higher SECV, which was only marginally
statistically significant even by direct comparison to
class I. Classes IV-VI had higher SECVs that were
significant by either statistical comparison. The most
extreme was the whole grain class (V), perhaps because
of relatively less accurate optical sampling caused by
differences in the light penetration through the surface
layers of these products.
Because the cereal food product samples were very
diverse, there may have been spectral interference
between the various constituents of the products. To
examine the magnitude of this potential problem, the
TDF prediction-residuals, or the absolute value of those
prediction residuals, from the untreated-sample test set
were plotted versus the six major sample attributes in
Table 2 (data not shown). Though relatively minor, the
largest observed effect was that elevated sugar content
was correlated with higher bias, but without any correlation to the amplitude of the error. By contrast, as
fat content increased, the bias did not change, but the
error magnitude decreased significantly. As moisture
content increased, within the range found in untreated
samples, bias became slightly more negative, and error
magnitude increased slightly. Among the samples of the
data set, densities and moisture contents were slightly
inversely correlated to sugar content, so the slight
negative bias attributed to moisture may in fact have
been a result of the sugar content.
Stabilization of Models to Moisture and ParticleSize Variation. The variation of coating thickness,
4484 J. Agric. Food Chem., Vol. 48, No. 10, 2000
Archibald and Kays
Table 4. Variation in Prediction Error Due to Particle Size or Moisture Treated Specimens in Either the Calibration or
Test Sets
RMSECVa difference between treated and
untreated pairs test set treatment & MCb range
calibration set description
untreated products
untreated and particle size treated
untreated, particle size, and
moisture treated (MCb< 10.7%) untreated, particle size, and moisture treated (MC < 16.0%) untreated, particle size, and moisture treated (MC < 21.5%) number of test samples # calibr. samples untreated products R2 RMSECVa desiccated (MC < 5.25%) moistened (MC < 10.7%) 137 219 276 1.83 1.80 2.01 0.96 0.96 0.96 -0.28 -0.38 -0.33 0.70 0.76 -1.06c 341 1.80 0.97 -0.73 361 1.87 0.96 -0.76 45 5 moistened (MC < 16.0%) moistened (MC < 21.5%) 0.27 0.26 0.35 3.11 3.10 1.79 -0.56 -0.10 0.87 -0.56 -0.04 0.33 7 13 17 a Calculated for the 13-factor PLSR D2-SNV model with optimized spectral range. b MC is moisture content by oven method. c Bolded italicized numbers indicate the best errors achieved for each treated test set. particle size, and moisture content (MC) of a product can create bias and error in the predictions of NIRR calibration models if the models do not sufficiently account for these effects. However, if the calibration samples have wide variation in these attributes, model error may needlessly be increased for future samples to be analyzed. To evaluate these effects, the data for treated samples were analyzed by observing the change in model error between pairs of treated and untreated specimens in the test set, as predicted by calibration sets containing different sets of treated specimens (Table 4). Note that only 17 of the 45 test samples were treated, so this subset of the test set did not represent very accurately all sample types found in the study. Therefore, the arguments about the effects of treatments are based on differences observed with the same (small) sets of samples. Although some of the calibration sets contained treatments, the calculations were implemented such that the calibration sets for each crossvalidation segment did not contain any treated versions of the sample being predicted by that cross-validation segment. By use of a calibration model developed from untreated products (row 2 of Table 4), the prediction error for a set of five samples improved after desiccation (-0.28 in column 5 of Table 4). In fact, the data suggested that the NIRR models produced better results with drier samples, no matter which set of treatments was included in the calibration. Likewise, moistening products at the highest levels studied (up to 21.5% MC) always created poorer predictions. For example, in column 8 of Table 4, RMSECV for 17 test samples was 3.11 larger using spectra collected after the samples received moisture treatments. The increase in RMSECV was primarily due to a large positive bias for the wet samples, in contrast to the aforementioned small negative bias associated with the inherent moisture content of untreated samples. The highest level of moisture content was extreme for many of these products, resulting in friable saturated products and cereal glazes that were sticky or syrup-like. In the more realistic lower ranges of test-sample moisture content, the error increases due to moisture treatment were not nearly as large (maximum of 0.70). Furthermore, including moisture treatments in the calibration dramatically improved the error for the moist samples but did not greatly alter the RMSECV values for the untreatedsample test set, which fell in the range 1.8-2.0% TDF for the 13-factor models built from various calibration sets. In contrast, including only particle size treatments in the calibration did not yield substantial improvement in the ability to predict wet samples. In the three cases where the maximum moisture content of the calibration set could be matched to that of the test set, the best error was obtained for that test set. This effect was largest for the drier test set (MC < 10.7%). For the test set with MC < 16.0%, there was only a small level of degraded performance of the models as a result of adding calibration samples with moisture contents exceeding 16%. Surprisingly, in many cases the moistened products (except for the highest moisture level) were predicted better than their untreated counterparts. It is speculated that regression performance is improved because moderate levels of hydration convert various types of crystalline sugars to their amorphous forms, which are more spectrally similar to one another. Reeves has previously demonstrated this effect in model studies (Reeves, 1995). Reducing the particle size of a subset of test samples (n ) 20) by crushing treatments improved the RMSECV by 0.09, and furthermore, a slightly better result (0.13 improvement) was obtained when particle size treatments were included in the calibration set (data not shown). However, in regard to predicting the test set of intact cereal food products, the benefits of including particle size treatments in the calibration were marginal, though positive (Table 4). Moisture content variation among untreated products appeared to have a more profound effect on model performance than the effects of particle size or structure. When the particle size and inherent moisture content of cereal food product specimens cannot be adjusted prior to making an NIRR determination, it is recommended that the calibration set include all levels of crushing treatments and moisture treatments up to 16.0% MC. With this calibration set, the RMSECV was 1.80% for the 45-sample untreated test set, and slightly lower residual errors were obtained for 13 test specimens moistened to a moderate level (