-
Full Terms & Conditions of access and use can be found
athttp://www.tandfonline.com/action/journalInformation?journalCode=laps20
Download by: [Australian Catholic University] Date: 15 August
2017, At: 22:15
Applied Spectroscopy Reviews
ISSN: 0570-4928 (Print) 1520-569X (Online) Journal homepage:
http://www.tandfonline.com/loi/laps20
Exploring the Potential of Applying InfraredVibrational
(Micro)Spectroscopy in Ergot AlkaloidsDetermination: Techniques,
Current Status, andChallenges
Haitao Shi & Peiqiang Yu
To cite this article: Haitao Shi & Peiqiang Yu (2017):
Exploring the Potential of Applying InfraredVibrational
(Micro)Spectroscopy in Ergot Alkaloids Determination: Techniques,
Current Status, andChallenges, Applied Spectroscopy Reviews, DOI:
10.1080/05704928.2017.1363771
To link to this article:
http://dx.doi.org/10.1080/05704928.2017.1363771
Accepted author version posted online: 14Aug 2017.
Submit your article to this journal
View related articles
View Crossmark data
http://www.tandfonline.com/action/journalInformation?journalCode=laps20http://www.tandfonline.com/loi/laps20http://www.tandfonline.com/action/showCitFormats?doi=10.1080/05704928.2017.1363771http://dx.doi.org/10.1080/05704928.2017.1363771http://www.tandfonline.com/action/authorSubmission?journalCode=laps20&show=instructionshttp://www.tandfonline.com/action/authorSubmission?journalCode=laps20&show=instructionshttp://www.tandfonline.com/doi/mlt/10.1080/05704928.2017.1363771http://www.tandfonline.com/doi/mlt/10.1080/05704928.2017.1363771http://crossmark.crossref.org/dialog/?doi=10.1080/05704928.2017.1363771&domain=pdf&date_stamp=2017-08-14http://crossmark.crossref.org/dialog/?doi=10.1080/05704928.2017.1363771&domain=pdf&date_stamp=2017-08-14
-
ACCEPTED MANUSCRIPT
1
ACCEPTED MANUSCRIPT
Ergot Alkaloids Research with Infrared Spectroscopy
Exploring the Potential of Applying Infrared Vibrational
(Micro)Spectroscopy in Ergot Alkaloids Determination:
Techniques,
Current Status, and Challenges
Haitao Shi & Peiqiang Yu*
Department of Animal and Poultry Science, College of Agriculture
and Bioresources, the University of
Saskatchewan, 51 Campus Drive, Saskatoon, Canada, S7N 5A8
*Corresponding author, Peiqiang Yu, Ph.D., Professor and
Ministry of Agriculture Strategic Research
Chair, Department of Animal and Poultry Science, College of
Agriculture and Bioresources, University of
Saskatchewan, 6D10 Agriculture Building, 51 Campus Drive,
Saskatoon, Canada, S7░N 5A8, Tel: +1
306 966 4132, Fax: +1 306 966 4150, E-mail:
[email protected]
ABSTRACT
Ergot alkaloids are toxins produced mainly by Claviceps fungi
and are considered as one of the most
important groups of mycotoxins. Rapid and reliable detection
techniques are urgently required by
producers, importers and market regulators. As a promising
alternative to conventional wet chemistry,
infrared (IR) based techniques are non-destructive, rapid, and
cost-effective. However, very limited
studies on the qualitative or quantitative analysis of ergot or
ergot alkaloids in food or feed based on IR
vibrational spectroscopy have been reported so far. Being a
secondary technique, the accuracy of IR
method heavily dependents on the robustness of chemometrics
models. This paper aims to offer a brief
overview of the ergot alkaloids issue in food and feed,
conventional detection methods, theoretical
principles of IR-based techniques, and commonly used
chemometrics for spectral data processing. In
addition, the current application status of IR spectroscopy in
ergot research is also considered.
KEYWORDS
ergot alkaloids; mycotoxin; vibrational spectroscopy;
chemometrics
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
https://crossmark.crossref.org/dialog/?doi=10.1080/05704928.2017.1363771&domain=pdf&date_stamp=2017-08-12
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 2
Introduction
Concept of ergot alkaloids
Ergot alkaloids (EA) are toxic metabolites produced mainly by
Claviceps fungal species, most notably by
Claviceps purpurea, which infect the seed heads of cereals and
wild grasses at the time of flowering (1,
2). Claviceps fusiformis and Claviceps purpurea, which have been
blamed as the causes of some severe
human/animal epidemics, are the two most important species among
them (3). More than 600
monocotyledonous plants, including barley, wheat, rye, rice,
corn, oats, and forage grasses, are vulnerable
to these fungal species (4). When infected by these fungi, the
developing seeds will be replaced by ergot
bodies (also known as sclerotia) which may contain various EA
and are usually harvested together with
uncontaminated grain (5, 6). The types of EA and their
concentrations in ergot bodies can be affected by
the maturity of ergot bodies, types of fungi and plant, weather
conditions, geographical locations, etc (7,
8). More than 80 different types of EA have been discovered and
more than 70 of them are produced by
Claviceps species (4). Clavine EA, lysergic acid amide EA, and
peptide EA are the three most common
categories. Ergocristine, ergosine, ergotamine, ergometrine,
ergocornine, and ergocryptine that belong to
the group of Clavines are regarded as the most dominant and
toxic EA which are frequently detected in
food and feed stuffs (7). Figure 1 shows the chemical structure
of the six major EA that derived from
PubChem Compound database (9).
Contamination status in food and feed
Ergot contamination in feed and food could be observed in almost
all parts of the world. In 2011, up to
20% of the wheat produced in western Canada was contaminated by
ergot to a certain extent, and this
situation is likely to get worse in the future according to the
results of climate-change prediction models
(10). The contamination status of six major EA and their epimers
in cereals and cereal products (803
samples) in 13 European countries has been investigated between
2010 and 2011 (11). Results showed
that 34% of wheat feed, 48% of triticale feed, 52% of rye feed,
76% of food products, 86% of wheat food,
and 95% of rye food were contaminated with EA (concentration of
total EA varies from 1 to 12340
μg/kg). The occurrence of ergometrine, zearalenone,
deoxynivalenol, and alternariol in beer collected
from German market has been investigated and results showed that
ergometrine was detected in 93% of
beer samples with the concentration between 0.07 and 0.47μg/L
(12). Mulder et al. (2015)
(13)investigated the occurrence of EA and tropane alkaloids in
cereal products for young children and
infants in Netherlands (13). Both the major EA and seven minor
EA (i.e. agroclavine, chanoclavine-1,
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 3
elymoclavine, ergine, erginine, festuclavine, and lysergol)
content of 113 samples (i.e. breakfast cereals,
biscuits and cookies) has been determined and results showed
that 54% samples were contaminated with
EA and tropane alkaloids were found in 22% samples. Tittlemier
et al. (2015) developed a method for
analyzing 10 EA in grain samples and studied the contamination
status of these EA in western Canadian
cereal grains (14). They observed a strong linear relationship
between the amount of ergot bodies and the
content of EA (R2 = 0.826 and 0.911 for Canadian western Red
Spring and Canadian western Amber
Durum, respectively), and grain samples collected at harvest had
higher concentrations of total EA than
that collected from shipments. Besides, ergocristinine was
reported as the most dominant alkaloids in
wheat and durum samples from grain shipments. Endophyte-infected
grasses (e.g. fescue and tall fescue)
toxicosis in pasture-based regions have been frequently
associated with EA, and increased occurrences of
EA issues has been observed in regions that typically not
related to grazing-based farming due to
international forage trading (15).
Effects on animals and human
Ergot alkaloids have been recognized as a group of the most
important natural toxins and pharmaceuticals
for a long time (16). The earliest recording of the effects of
ergot that can be authenticated has been
discovered in Chinese ancient documents (originated around
1100░BC) which described the application
of ergot as an obstetric medication (4). The severe pathological
syndromes caused by the ingestion of
food or feed that contaminated with EA are known as ergotism. In
944--945 AD, ergot poisoning caused
the death of about 20000 people in France, and this was probably
the first documented epidemic of
ergotism (4). In Europe, St. Anthony’s Fire, the severe
epidemics caused by intaking of ergot
contaminated grains and grass, was one of the things that made
the Middle Ages a horrible time (17).
Due to the advancement in crop and food science (e.g. crop
management, grain cleaning techniques,
and contaminants detection methods), ergotism is now regarded as
a very rare disease for human and
usually resulted from overdosing of EA-based medicine rather
than ingestion of EA contaminated food
(18). In many countries, ergot-contaminated grain is not allowed
to be used as food for humans and
redirected for feedstuffs of animals (10). However, ergotism is
continuing to be a significant issue for
people living in some underdeveloped regions and animals (2, 3,
18).
Many factors, such as host plants, fungal species, growth
environment, size of ergot bodies, and the
interaction with other mycotoxins that might concurrent in
cereals and feed could affect the toxicity of EA
(3, 10). After uptaking food or feed that contaminated with
certain amount EA, clinical symptoms may
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 4
appear in a few hours or several months (3). The diagnosis of
ergot poisoning is usually difficult since
some symptoms are similar with other disease such as respiratory
disorders, frostbite, and foot rot (10,
19).
In general, there are four forms of ergotism in mammals,
including convulsive ergotism, gangrenous
ergotism, enteroergotism, and hyperthermic ergotism (10). Sheep
and horses are usually more vulnerable
to convulsive ergotism than cows; gangrenous ergotism is more
common in pigs and cows that could
cause lameness which may be followed by the partial or entire
loss of hooves, tail, ears, and even limbs;
enteroergotism and hyperthermic ergotism are considered as less
severe and might give rise to vomiting,
fever or endocrine disorders (2, 10).
The health and performance of farm animals can be harmed by
eating EA contaminated feed and the
seriousness can be influenced by other factors such as animal
species, age, environment temperature, etc
(10). Previous studies confirmed that the growth rate,
reproductive performance, pregnancy rate, and
sperm motility of many animals (e.g. cattle, sheep, swine,
poultry, etc.) could be negatively affected by
the consumption of ergot contaminated feed (10, 20, 21).
Due to the limitation of detection technology and other factors,
previous toxicology studies mainly
focus on ergot bodies or total EA rather than single ergot
alkaloid and their interactions. Consequently,
the regulatory limits are usually established based on
concentration of ergot bodies or total EA (2).
However, the concentration and proportion of individual EA
within sclerotia can vary significantly, and
their toxicities in humans and animals still need further
research (10).
Strategies to reduce their harmful effects
Up to 82% ergot bodies in the grain could be cleared away with
proper cleaning procedures at mills (2).
However, high concentration of EA still has been detected in
cereals and cereal products from time to
time according to the results of previous surveys (7).
Furthermore, cleaning procedures might be
ineffective when ergot bodies are produced in dry climatic
environment which usually have similar size
with grain or when the original ergot bodies break into smaller
pieces during harvest, transportation, and
storage (22).
Minimizing the production of ergot bodies in the field is of
great importance for the control of EA
contamination of feed and food. Genetic engineering might be a
potential approach for the developing of
ergot-resistant crop varieties. Some sorghum genotypes that
present high levels of ergot-resistance have
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 5
been discovered and further effort is required to explore the
physiological basis and inheritance for
resistance in them (23). The genetic difference for
ergot-resistance in cytoplasmic-male sterile rye has
been studied and more research on the specific resistant
mechanisms is needed (24). In addition, field
management, environmental conditions, and the application of
biological and chemical agents are also
considered as important factors which can influence the forming
of mycotoxins pre-harvest (25).
It is generally assumed that ergot bodies can only be formed
during pre-harvest period and the
concentration of EA is supposed to be relatively constant after
host crops are harvest (19). However,
another study reported that the aerobic instability resulted
from extended storage favors the growth of
ergot-producing fungi and could lead to an increase in
ergopeptinines concentration in high-moisture
grain (2). Therefore, appropriate storage management is required
to prevent the increase of EA content in
feed.
The toxicity levels of EA could be affected by their inherent
chemical structure and processing
methods. Temperature and pH could influence the C9 = C10 double
bond contained in EA, therefore heat
treatment might be an effective way to reduce the toxicity of EA
(10, 26). The effect of hydrothermal
treatments on EA content in rye has been studied and the
treatment resulted in approximately 10%
reduction in total EA content (27).
A very promising dietary approach to minimize the detrimental
effects of mycotoxins is the application
of mycotoxin binders (28). Some mycotoxin binders are able to
selectively bind EA and other mycotoxins
in feed. In a mycotoxin binding study, five major EA (i.e.
ergosine, ergotamine, ergocornine,
ergocryptine, and ergocristine) and three other mycotoxins have
been effectively absorbed by surfactant
modified zeolites (29). Some mycotoxins binders that are claimed
to be capable of absorbing EA are
commercially available on the market now, while more animal
studies are expected to further confirm
their effectiveness.
Brief introduction of infrared spectroscopy
Infrared spectroscopy is regarded as one of the most important
analytical methods available to
researchers, which is developed based on the vibrations of
chemical bonds within a molecule. (30).
Infrared region (14000-50░cm1) of the electromagnetic spectrum
can be segmented into three major
areas, including near-infrared (NIR; 14000-4000░cm1),
mid-infrared (MIR; 4000--400░cm1), and far-
infrared (FIR; 400--50░cm1) regions (31). The IR light can
induce the excitation of vibrations of bonds,
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 6
which can be influenced by many factors, such as the mass of
atoms, firmness of bonds, as well as other
features of molecules (32). Specific functional group within a
molecule has characteristic IR absorption at
certain frequency ranges, and a spectrum can be generated when
the changes in absorption of IR light by
molecules are recorded by a spectrometer (33). In a spectrum,
the absorption bands are characterized by
wavelength (energy), shape (environment of bonds), and intensity
(polar character or polarizability) (30,
34). As a consequence, information related to the molecular
characteristics of a sample can be revealed by
analyzing the specific spectrum.
Classical wet chemistry methods are well known due to their
precise and accurate determination of
various constituents in food and feed. However, they are usually
time-consuming, expensive, and based
on special skilled operators. IR spectroscopy is
non-destructive, chemical-free, cost-effective, rapid, and
has been suggested as a powerful and promising substitution to
conventional methods (30, 35). Another
distinct advantage easily overlooked is that IR spectroscopy
could predict multiple parameters
simultaneously based on a single spectrum if proper calibration
models were developed (36, 37).
The original chemical composition and structural characteristics
of food and feed materials might be
altered by fungal infection and mycotoxin contamination (38-40).
IR techniques have the potential to
detect such changes and perform qualitative and/or quantitative
analysis by examining differences in
specific bands.
The application of IR based technique relies heavily on the
construction of calibration model by
chemometrics, which needs great efforts, takes long time and
could be regarded as a disadvantage.
However, the analysis time saved during the practical
application of IR based method could easily make
up for the initial time investment required for calibration, and
this advantage will keep growing as the
technique continues to be applied.
Conventional methodology for the determination of ergot
alkaloids
Traditionally, visual inspection has been applied as a simple
method to monitor the ergot contamination
status of grains since the ergot bodies usually have darker
color and larger size than original grain kernels
(10). However, the size of sclerotia might vary greatly
depending on grain types as well as other external
factors (e.g. climate, humidity, fungal strains, etc.). Besides,
it’s impossible to determine the
concentration of EA by visual inspection. To effectively monitor
the contamination status of EA in food
and feed and investigate their toxicodynamics and toxicokinetics
in animal models, analytical techniques
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 7
of high sensitivity and selectivity are highly required (18).
Enzyme-linked immunosorbent assay
(ELISA), thin-layer chromatography (TLC), gas chromatography
(GC), High-performance liquid
chromatography (HPLC), liquid chromatography-tandem mass
spectrometry (LC/MS/MS) are commonly
used wet chemistry methods for the detection of EA in food and
feed (2, 5, 18).
To make EA “available” for determination, these wet chemistry
methods usually need time consuming
extraction and cleanup procedures that rely on hazardous and
expensive organic solvents (39).
Furthermore, the final detection stage may also need well
trained operators, expensive equipments, and a
long testing time.
Rapid screening methods
The spatial structure of the mycotoxin can be distinguished by a
certain antibody, and the ELISA method
was developed based on this principle (41). Compared with other
wet chemistry methods, ELISA seems
to be a sampler and faster technique for EA screening (2). A
competitive inhibition ELISA method has
been used to detect ergot bodies and ergonovine in wheat and
grass seed (42). This method has also been
applied to determine the total EA concentration in different
native Turkish grasses (43). Nevertheless,
cross reaction is often a great concern for ELISA method and
cross-reactivity might vary between
different groups of EA (2, 10).
As a rapid and cost-efficient method for mycotoxin detection,
TLC can analysis multiple mycotoxin-
contaminated samples simultaneously, but it is unable to make
precise or sensitive measurements (44). By
taking the approach, the EA need to be extracted from samples
with proper solvent and then spotted on a
thin-layer plate. A semi-quantitative TLC method has been
developed for the determination of ergovaline
content in tall fescue (21). Nevertheless, this method is now
rarely used in EA determination area.
Quantitative methods
Compared with traditional liquid chromatography, HPLC employs
much higher operational pressures
which can accelerate the separation process effectively. The
HPLC with UV-diode array detection or
fluorescence detection (HPLC-FD) has been a powerful tool for
quantitative mycotoxin analysis (45).
Most of the HPLC protocols applied for mycotoxins quantification
are similar and many of them have
been adopted as the official AOAC methods (46). HPLC is an
excellent tool for EA detection since it’s
able to quantify trace components at the parts per trillion
levels (10). It has been frequently used for
quantification of the six major EA and their corresponding
isomers (27).
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 8
Gas chromatography is a widely used chromatography technique for
separating and detecting
substances that can be vaporized without decomposition. In GC, a
capillary or packed column is used as
stationary phase, and an inert gas usually works as the moving
phase (47). The separation of components
is based on their relative affinity with the stationary phase.
However, GC is unsuitable for the detection of
ergopeptide alkaloids since the hot injector could cause the
decomposition of them (2).
Taking advantage of both the separation power of liquid
chromatography and excellent sensitivity of
mass-spectrographic techniques, LC-MS/MS has gained more and
more popularity in recent years (48).
Krska et al. (2008) developed and validated an accurate method
for the simultaneous quantification of six
major EA as well as their corresponding epimers in food and
cereal samples with limits of quantification
of 0.17 to 2.78 μg kg1 based on LC-MS/MS technique (49).
Recently, Guo et al. (2016) demonstrated the
feasibility of quantifying up to 25 EA in cereal samples
simultaneously by this technique (5).
Concept and principal of infrared spectroscopic techniques
Near-infrared spectroscopy
As a commonly used vibrational-spectroscopy-based technique, NIR
spectroscopy has been applied to a
wide range of applications from agriculture to pharmaceutical,
as well as petroleum industries (35). With
the optimization of instrument and development of chemometrics,
NIR has become a prominent analytical
technique for both cereal quality monitoring and fundamental
research related to characterizing complex
biological systems (36, 50).
The NIR spectra contain relatively weak overtones and
combinations of fundamental vibrations mainly
related to O-H, N-H, and C-H functional groups (31, 36). The
intensity of absorption bands decreases
with the increase of the rank of overtones. Moreover, absorbing
a single photon might simulate more than
one vibration, and the interaction of these concurrent
vibrations could give rise to combination bands
which also included in the NIR spectrum (30, 51). The structural
selectivity of NIR spectra is much lower
than that of MIR due to the substitution of different
combinations and overtones in NIR region (31).
Therefore, NIR spectra are usually difficult to be interpreted,
and the assignment of specific absorption
bands to certain functional groups is technically more
challenging (36).
However, NIR light can better penetrate into the sample and
allow the directly analysis of strongly light
scattering or highly absorbing materials without complex sample
preparation (52).
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 9
Globar-sourced MIR molecular spectroscopy
MIR spectroscopy measures information from frequencies of
fundamental molecular vibrations, and it is
commonly used to identify the structural characteristics of
organic compounds (53). The MIR spectrum
could be divided into four major regions, including the X-H
stretching region (4000-2500░cm1), the
triple bond region (2500-2000░cm1), the double bond region
(2000-1500░cm1), and the fingerprint
region (1500-400░cm1) (30, 32). In contrast to NIR, most of the
peaks in MIR region are sharp and
narrow due to fundamental vibrations of the matrix compounds
(54). In addition, the absorption bands
could be assigned to specific functional groups due to the
higher degree of resolution of MIR spectra (55).
The fingerprint region (1200-700░cm1) of MIR spectra contains
abundant structural information, and
unique absorption bands of protein, polysaccharides,
carotenoids, and lipids can be found within this
region (56). Fourier transform infrared (FTIR) spectrometer is a
commonly used device to produce such
spectra, and it circumvented obvious defects of traditional
dispersive IR spectroscopy technique (57).
Up to present, most quantitative spectroscopy has been developed
on the basis of NIR spectroscopy,
the application of MIR technique for quantitative determination
is still under development and not as
mature as for NIR (58).
Infrared hyperspectral imaging
As a branch of photography and spectroscopy, spectral imaging
has been applied as a powerful technique
for astrophysics, remote sensing, satellite imaging, and
terrestrial military applications (59, 60). In recent
years, the application of this relatively new and
non-destructive method has expanded into more and more
areas such as agriculture, biology, archaeology, pharmaceuticals
(36, 60).
Generally, spectral imaging can be classified as multispectral
imaging (MSI) and hyperspectral
imaging (HSI) mainly according to the number of spectral bands
and spectral resolution (60, 61). MSI
collects data within only several spaced spectral bands, while
HSI continuously measures reflected energy
in more numerous bands (tens to hundreds). As a consequence, HIS
generates dataset that contains more
abundant spectral information and is highly sensitive to even
subtle variation in the spectra (61).
Hyperspectral imaging collects three dimensional information,
including wavelength (z) and spatial (x,
y) dimensions (36, 61). The acquired data is presented as a
three-way data matrix (x, y, z) which is
typically mentioned as hypercube (36). There are four distinct
strategies for the acquisition of
hyperspectral datacube, including point-scanning spectrometry,
pushbroom spectrometry, wavelength-
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 10
scanning spectrometry, and snapshot imaging spectrometry (61).
The pushbroom strategy is regarded as
the most widely used in food industry since its readiness
facilitates the on-line analysis purpose (62).
Different regions of the electromagnetic spectrum can be
employed by the HSI system, and the visible
(400-700░nm) and NIR (700-2500░nm) spectral range are more
frequently used (60). By using the
conventional NIR technique, only a mean spectrum that contained
absorption values at corresponding
wavelengths can be obtained. As a result of combining digital
imaging with NIR spectroscopy, the near-
infrared HSI can provide both NIR spectral data and spatial
information simultaneously, which allows
visualizing chemical characteristics and determining chemical
constituents of specific samples at the same
time (36, 60). There are also several disadvantages of
hyperspectral imaging, for example, more sensitive
detector and high-performance computer are needed to meet the
requirement of data acquisition and
processing, and the complex hypercube usually need greater
storage capacity and more professional data
analysis techniques (36).
Synchrotron infrared microspectroscopy
The joint application of IR spectroscopy and microscopy are
referred as IR microspectroscopy, which has
become a highly sensitive, label free, and intact analytical
tool for a variety of biological and medical
research since the mid-20th century (63, 64). IR
microspectroscopy allows scientist to explore molecular
information at a cellular or subcellular level, which is
necessary to the study of bio-chemical properties of
materials of interest (64, 65). Up to now, FTIR
microspectroscopy has been widely applied to different
research fields (biology, materialogy, art restoration, and
forensics). However, due to the intrinsic
brightness of conventional thermal IR light source, the
signal-to-noise ratio (S/N) of FTIR
microspectroscopy is still unsatisfied (66). Synchrotron light,
which is characterized by its high brightness
(100-1000 times brighter than a conventional thermal light
source), can deliver the whole IR wavelength
range (from NIR to FIR) and is considered as the only “white”
source (67, 68). By using synchrotron
radiation as light source, the spectral resolution and the S/N
of FTIR microspectroscopy have been greatly
improved, abundant spectral and spatial information could be
obtained, which allowing the in-depth
exploration of individual cells and tissue sub-structures (66,
68). The high quality spectral data obtained
makes it possible for researchers to perform more reliable
analysis (36). In addition, no induced damage
to the samples has been detected since the brightness advantage
of synchrotron light is not because it
generates higher power than other conventional thermal IR source
(66). Despite the above advantages, it
should be noted that the availability of synchrotron light
sources is still limited around the world and they
are mainly for research purposes at present.
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 11
Common chemometrics techniques for qualitative and quantitative
analysis
Due to the large amount of variables contained in a spectrum, IR
spectral data is multivariate in nature. To
effectively extract the useful chemical information from the
complex spectra that would correlate with
parameters of interest, chemometrics techniques are highly
needed (30). A diversity of views on the word
chemometrics and its definition could be found from the
literature (69). The International Chemometrics
Society defined chemometrics as “the science of relating
measurements made on chemical system or
process to the state of the system via application of
mathematical or statistical methods” (70). Designing
and optimizing experimental procedure and extracting maximum
chemical information from analytical
data have been regarded as some of the major areas of
chemometrics (71). To date, chemometric has been
used as routine technique for chemical researchers, and
multivariate analysis of chemical data is always a
critical part of this interdisciplinary field (72). By the
application of chemometrics, scientists are able to
handle typical chemistry and spectroscopy problems more
efficiently, such as monitoring the process
status, exploring the structural characteristics of unknown
organic compound, predicting the activity or
property of a chemical group, quantifying components in complex
compounds, and classifying samples
into corresponding groups, etc (30, 35, 72).
Spectral preprocessing
The spectral data obtained from spectrometer contains not only
sample information, but also noise and
back-ground information. Preprocessing of IR spectra to remove
possible spectral variations unrelated to
chemical composition is regarded as an essential procedure
before multivariate calibration (33, 35, 36).
Mean centering, baseline offset, derivatives, detrending,
multiplicative scattering correction (MSC),
standard normal variate (SNV), and wavelet transform are
commonly used preprocessing methods for
spectral data (35, 73, 74). Both the individual mathematical
preprocessing methods and proper
combinations of them could be used to reduce various types of
physical variations, chemical interference,
and negative effect of spectral side information (35)
Generally, SNV and MSC are regarded as the most widely used
techniques for reducing light scattering
(commonly found in samples with different particle size
distribution) and correcting baseline offsets;
derivatives (applied with a smoothing step) could reduce
random-noise influence, highlight subtle band
shape, improve the resolution of overlapping bands, and preserve
the relative band-intensity information;
detrending method is usually used for correcting baseline shift
and curvilinearity of densely packed or
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 12
powered samples, and follows SNV to reduce additional variations
in baseline shift and curve linearity
(33, 35, 36). By subtracting the average value from each
variable, mean centering could substantially
enhance differences between samples in terms of both
concentration and spectral response (35, 75).
Wavelet transform utilizes variable window size to analyze
signals at different resolutions or scales,
breaks up the signal into shifted and scaled versions of the
original wavelet function (74). Many previous
studies suggest that wavelet transform could efficiently
eliminate the background noise from IR spectral
data, and the denoising performance could be affected by factors
of wavelet function, decomposition
level, and threshold (74, 76). Figure 2 shows the raw (A) and
second derivative (B) spectrum of ground
wheat sample in the NIR region (680-2500░nm).
All the pretreatment techniques aim to remove un-modeled
variability in spectral data, optimize the
signal from useful information, and facilitate subsequent
calibration and exploratory study (35, 73).
However, it should be noted that the valuable information
related to parameters of interest could be lost
by the application of a wrong or too severe preprocessing method
(73).
Wavelength selection
Infrared spectra usually contain vast quantities of information
related to molecular physicochemical
properties that can be used for analytical purpose (30, 77).
However, considerable amounts of information
that irrelevant to the properties of interest (e. g.
interactions, variabilities, noise, etc) also contained in the
spectra which may bring detrimental effects on the modeling
process. Therefore, the spectral regions
employed might have substantial influence on the performance of
classification and regression models.
Multivariate model constructed with the whole wavelengths does
not always yield good prediction
performance since some of the wavelengths contain unrelated
information which may distort the models
(35). Spectroscopic model developed with selected important
wavelengths might achieve better predictive
performance and helps to speed up the computational speed of
analyzing new samples. Wavelength
selection is especially important for three-way data matrix that
obtained by HSI or microspectroscopy
techniques since the whole wavelength data sets always contain
massive information and need more
computing time (62).
Up to now, it is still challenging to select the most
informative wavelength variables from whole
spectra for calibration. There are many variable selection
algorithms available to researchers. Variable
selection methods could generally be classified into three
groups: filter methods, wrapper methods, and
embedded methods (78-80). In brief, filter approaches will
calculate the variable importance ranking
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 13
criteria and then remove the unwanted variable by applying a
filtering rule (e.g. threshold cut-off);
wrapper methods wrap around an appropriate learning machine,
which is applied as the evaluation
criterion, such as classification or prediction error; embedded
approaches are similar to wrapper methods,
except that variable selection is achieved simultaneously with
the training process. Detailed information
of advantage and disadvantage of these methods could be found in
Liu et al (2014) (79).
Successive projections algorithm (SPA) has been reported as one
of the popular filter methods (81). As
a forward variable selection technique, SPA focuses on the
selection of effective wavelength variables
that containing minimum redundant information to reduce of
detrimental effects of colinearity (82, 83).
This algorithm begins with one variable and merges another new
variable iteratively till a certain number
of wavelengths are obtained. When there are large amount of
variables, SPA is more effective to reduce
the workload of computation than GA technique since it uses
simple vector projection operations (82).
Regression coefficient analysis (RCA), uninformative variable
elimination (UVE) and Genetic
algorithms (GA) has been considered as commonly used wrapper
methods for variables selection and
have been modified for important wavelengths selection purpose
(79, 81, 84). Based on the PLS
algorithm, RCA is one of the popular filter methods for the
determination of sensitive wavelengths (52,
81, 84). Regression coefficients (RC) could be used to reveal
which wavelength variables could affect the
dependent variable in a greater degree (83). A RCA plot for PLS
model to predict crude protein content in
wheat samples is shown in Figure 3.Wavelengths with high
absolute RC values indicated that they were
the most informative wavelengths for the specific models. When
applying this method, the setup of
specific threshold value of RC is suggested to facilitate the
picking up of informative variables (85). UVE
assessing the significance of the variables by comparing their
reliability index, a function of the
regression coefficients, with the ones of artificial random
variables (78). It has been employed in
wavelength selections in IR spectroscopy and achieved good
prediction performance (79). However,
latent variables are still required to be used for modeling
since the quantity of variables selected by UVE
is still too large, therefore UVE has been combined with other
algorithms in some studies for spectral
variable selection (86). Based on guided random search
techniques, GA allow the efficiently exploration
of solution space and facilitate the implementation of parallel
processing (82). Overfitting is usually
considered as the major concern when applying GA for variable
selection (87). Another disadvantage of
GA is that the selection of important variables could be
difficult when the spectra contain too many
wavelengths (large size of search domain) (84). Some researchers
reported that complex spectra (e.g. with
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 14
a large quantity of wavelengths and/or narrow peaks) could be
handled effectively by the sequential
application of interval PLS algorithms and GA techniques
(84).
Support vector machine (SVM) has been reported as one of the
commonly used embedded methods
(79, 88). However, SVM method in standard formulation does not
assess the significance of variables and
is therefore inappropriate for wavelength selection (89). As a
variant of the standard L2-norm SVM, L1-
norm SVM constrains the L1-norm of the fitted coefficients and
has the ability of automatically selecting
variables, not shared by the standard L2-norm SVM (90). The
L1-norm SVM may have obvious advantage
over the L2-norm SVM especially when there are redundant noise
variables. However, it is reported that
L1-norm SVM is unable to select groups of highly correlated
variables and the number of selected
variables could be bounded by the size of the training data
(90).
Regression methods
The relationship between reference values of chemical
ingredients to be evaluated and the specific
absorption bands contained in the spectra could be revealed by
the application of appropriate regression
methods (30, 34, 36). There are many regression algorithms
available for developing quantitative IR
models, among which multiple linear regression (MLR), Principal
component regression (PCR), partial
least squares regression (PLSR), SVM, and artificial neural
networks (ANNs) are most commonly used in
agricultural research (33, 91, 92).
MLR is the easiest linear regression technique which could be
used to relate the variations of a
dependent variable (Y) to the variations of several independent
variables (X) based on ordinary least
squares regression (36). MLR works well only when there are a
limited number of independent variables
(usually less than the number of samples) which are
non-collinear and have a well understood relationship
to the dependent variables (72, 93). However, MLR is usually
inappropriate or inefficient for spectral data
analysis since they usually contain more than hundreds of
wavelength variables which might be highly
collinear (93).
The PCR and PLSR work well when the number of X-variables is
much greater than the number of
samples and they are not affected by colinearity effects (94,
95). In PCR analysis, principal component
analysis (PCA) is performed firstly to decompose the X-matrix to
obtain principal components (PCs), and
then these PCs will be applied as alternatives to original
X-variables for fitting the MLR model (94). In
PCR, the multi-colinearity problem could be solved effectively,
but the selection of optimum subset of
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 15
predictors is still unachievable and some noise will remain in
the PCs that selected for calibration (94,
96).
The PLSR is a well-know statistical technique for exploring the
relation between a large number of
independent variables and dependent variables based on reducing
the dimensionality of data set (95). It is
particularly powerful in constructing linear calibration models
with IR spectral data since the negative
effect of irrelevant spectral variations could be reduced by
this algorithm (97). To a certain extent, PLSR
could be considered as the effective combination of PCA and MLR.
In contrast to PCR, PLSR
decomposes X-matrix and Y-matrix simultaneously, takes both of
predictor variables and response
variables into account, and orders the PCs based on their
relevance for the prediction of dependent
variables (94). As a consequent, PLSR model usually requires
smaller number of PC than PCR model for
achieving comparable predictive ability (75). In non-linear
situation, the modified versions of PLS such as
spline-PLS and polynomial-PLS, which replacing the initial
linear function with spline or polynomial
function, could be applied to develop regression models
(98).
The SVM is a supervised learning model with associated learning
algorithms that can efficiently
conduct multivariate analysis (99). The function kernel in SVM
technique allows implicitly mapping
sample space to a high-dimensional eigenspace, thus the
non-linear problems could be transformed into
linear ones (100, 101). It has been widely employed in different
areas, especially in food quality
monitoring (91, 102). As an optimized version of standard SVM,
least-squares SVM (LS-SVM) uses
least-squares linear system as the loss function and can
simplify the calculations of classical SVM and
investigate both nonlinear and linear relationships between
response variables and spectral characteristics
more rapidly (91, 102, 103).
Artificial neural networks are widely used artificial
intelligence techniques that can perform non-linear
modeling (104). ANNs employ algorithms designed to analysis and
process information in a similar way
as the human brain and have the ability to learn from experience
by themselves to solve a variety of
problems (105). The multiple layers contained in ANNs can
perform different functions including
receiving information, processing the information by non-linear
algorithm, and solving specific problems
(104). ANNs have shown great potential in solving complex
problems (e.g. pattern recognition,
categorization, and prediction) in multiple research fields in
recent years, and are getting more and more
attention in food research (106, 107).
Classification methods
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 16
As fundamental chemometric techniques, multivariate
classification methods are designed to construct
models that can be used for recognizing unknown objects to
proper groups based on calculation (108).
For the qualitative analysis by IR spectroscopy, both supervised
and unsupervised classification methods
are available to researchers (36). When applying supervised
methods, the categories/classes of samples
for developing classification model are known in advance (104).
For unsupervised methods, however, the
information regarding to sample categories/classes is
unavailable (36). PCA and hierarchical cluster
analysis (HCA) are widely used unsupervised classification
method and usually applied to investigate the
potential relationships between samples (77, 109). Commonly used
supervised classification methods
including soft independent modeling of class analogy (SIMCA),
linear discriminant analysis (LDA), PLS
discriminant analysis (PLS-DA), SVM, and ANN (108, 110).
Based on feature reduction algorithm, PCA forms the basis of
many multivariate analysis techniques
and is mostly applied to explore the structure of data set and
investigate potential relations between
samples at the beginning of data analysis (36, 110). PCA could
reduce the dimensionality of original data
sets by deriving a new data set which composed of PCs, and then
a score (eigenvector) is assigned to each
spectrum to define its relationship to the specific principal
component (109). A two or three-dimensional
scatter plot could be created using the eigenvector to show the
similarity between the spectra. It is worth
noting that only gross variability can be identified by PCA, and
the variability within a group and among
groups can’t be distinguished by it (111).
HCA aims to build a hierarchy of clusters and reveal relatively
homogeneous groups of variables
according to a predefined metric (112). HCA uses an iterative
procedure to explore the possible structure
of samples: classifying spectra that with high similarity into a
cluster, calculating the distance between
clusters, and incorporating the most similar pairs into new
clusters (109, 113). Either Euclidean distance
or Mahalanobis distance can be used for distance calculation,
and the later is more popular (77, 114).
HCA usually serves as an exploration tool in the early stage of
data mining.
SIMCA is a widely used supervised classification method that
developed based on the modeling
properties of PCA (104, 110). In SIMCA, A PCA is applied for
each class of the samples and distinct
confidence regions can be created around different classes. An
unknown sample is projected in each PCs
space that depicts a specific class to assess if it belongs to
the class or not (115).
Like the PCA, LDA is another widely used dimension reduction and
feature extraction technique (116).
In LDA, the input data is projected onto a vector space with
lower-dimension thus the ratio of the within-
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 17
group distance and between-group distance is maximized (111,
116). The Mahalanobis distances are
calculated for assigning objects to the nearest categories
(77).
Partial least squares algorithm (PLS1 deals with one response
variable situation and PLS2 handles
multiple response problems) is originally developed to deal with
the problem of overdetermined
regression, while scientists in the field of application have
discovered its potentiality to perform as a
discrimination and classification technique and further modified
it for those purpose (111). PLS-DA is the
classification technique based on modeling the differences
between different groups with PLSR (117).
With this technique, spectral data and reference group
membership information of training samples could
be used to construct calibration models. PLS1-DA aims to solve
binary classification problem, while
PLS2-DA could be applied when there are several classes need to
be modeled (77, 111).
The SVM and ANN techniques mentioned in the previous section
could be also used for classification
purpose. SVM technique originally aims to solve binary
classification problems, and has been extended to
deal with multiclass situations in recent years (118). A variety
of ANN learning strategies and
architectures which could handle unsupervised or supervised
classification problems have been developed
(108). The selection of proper ANN strategy should be based on
the nature of problems.
It is worth noting that the hypercubes (three dimensional data)
obtained by hyperspectral imaging
systems could not be applied for qualitative and quantitative
multivariate analysis directly. Appropriate
multivariate image analysis techniques are needed to transform
the data from hypercube form to two
dimensional data before applying aforementioned multivariate
spectral data analysis (36).
Current status of applying infrared spectroscopy in ergot
research
Previous ergot studies with infrared spectroscopy
As one of the prevalent ergopeptine alkaloids, ergovaline is
frequently detected in endophyte-infected
grasses and is associated with some severe livestock maladies
(119). Roberts et al. (1997) established a
spectroscopic method based on NIR technique for the
quantification of ergovaline content in tall fescue
(120). All samples were freeze-dried and ground prior to
analysis. The HPLC method was applied to
obtain reference values of ergovaline and a NIR spectrometer
(1110-2490░nm) was used to collect the
spectra of 99 samples. The prediction model was constructed by
using PLSR technique. Important
wavelength ranges for the quantification of ergovaline content
in tall fescue has been detected. The
determination coefficients and standard errors were 0.93 and 35
μg/kg (DM basis) for calibration and 0.88
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 18
and 44 μg/kg (DM basis) for cross-validation. Although this
model could only be applied to a limited set
of samples, it demonstrated the possibility of using NIR
technique for detecting micro constituents in
samples. However, due to the absence of an independent
prediction sample set for external validation, the
results obtained in this study could be overoptimistic.
Gislum et al. (2007) conducted a study to explore the
possibility of constructing a chemometric model for
classifying Danish grass samples contaminated with ergovaline
(121). Reference values of ergovaline in
167 plant samples were obtained by HPLC method, and the
reflectance spectra were collected by using a
NIR spectrometer (1100-2500░nm). PCA was performed based on the
MSC pre-processed spectra.
Nevertheless, their results showed that PCA technique was unable
to distinguish between contaminated
and uncontaminated grass samples, possibly because the limited
sensitivity of conventional NIR
technique is not enough to detect the small concentrations of
ergovaline in grass. Moreover, improved
discrimination performance between contaminated and
uncontaminated samples might be achieved by
coupling appropriate wavelength selection techniques to PCA
since PCA algorithm models the total
variance in the data.
In another study, Roberts et al. (2005) developed an empirical
prediction equation to measure total
ergot alkaloids content in tall fescue by NIR technique (122). A
diverse population of tall fescue samples
(N = 84) were collected and tested for total ergot alkaloids
content by ELISA method. NIR spectra (1110-
2490░nm) of these samples were generated, preprocessed by
second-derivative transformation, and
regressed with ELISA reference data using modified partial least
squares regression. The final calibration
(omitted toxic endophyte contaminated stockpiled samples) had a
1-variance ratio of 0.89, with a mean
and standard error of cross validation of 0.682 0.11. They
concluded that NIR technique could be used
for quantifying total ergot alkaloid content in tall fescue.
Vermeulen et al. (2012) developed a method based on HSI
technique and chemometrics for the online
determination of ergot bodies in cereals (6). Seven wheat
samples with differing content of ergot bodies
(0-10000░mg/kg) were prepared for the construction of models and
an additional sample was used for
validation of the established models. Hyperspectral images were
collected with a near infrared HSI
system (1100-2400░nm) and then subjected to multivariate image
and multivariate spectral analysis
(PLSDA and SVM). The correlation between reference values and
predicted values yielded by PLSD and
SVM models was higher than 0.99. The limits of quantification
and detection were 341░mg/kg and
145░mg/kg, respectively. This study showed that the HSI
technique has the potential to be used as a rapid
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 19
and accurate inspection tool in processing industry. In
subsequent research, Vermeulen et al. (2013)
investigated the transferability of the established HSI
procedure (1). In this study, an NIR hyperspectral
line scan camera was employed to collect images for developing
the models for determining ergot bodies
in cereals. Discrimination models were constructed by PLSDA
(without spectral preprocessing) and
SIMCA (spectra preprocessed by mean centering and SNV)
techniques and the wavelength regions
between 1573 and 2399░nm were selected. The models were
successfully transferred to commercial HSI
systems and the correlation between reference values that
obtained by official stereo-microscopic method
and predicted values offered by multivariate models was higher
than 0.94. This study showed the
potential of using HSI technique to screen ergot contaminated
cereals automatically at industrial level.
Problems and challenges
Up to now, very few studies on the qualitative or quantitative
analysis of ergot or EA in food or feed
based on IR spectroscopy techniques have been reported as
compared with other wet chemistry method.
In addition, the handful of available studies employ spectra in
the NIR region and mainly focus on the
detection of ergot bodies in grain samples or total EA/
ergovaline in grasses. No published research, to the
authors’ knowledge, focuses on exploring the possibility of
using IR based techniques for the rapid
determination of major six EA or other EA in cereals and forages
so far.
As a secondary analytical technique, the performance of IR
spectroscopy method relies heavily on the
robustness of multivariate models which are developed based on
calibration samples. The performance of
multivariate models could be affected by many factors. For
instance, low accuracy and poor
reproducibility of reference values obtained with wet chemistry
method usually leads to reduced
robustness of IR model; the variability of reference value range
has substantial influence on the predictive
ability of IR model, while building a calibration set with a
broad range of variability could be a tough
task; sample heterogeneity has negative effects on calibration
especially when the spectra are scanned
from intact samples (36, 123).
No study on the limit of direct detection of EA in food and feed
samples by IR based techniques has been
reported yet. The concentrations of EA (at ppb or ppm levels) in
cereals or grasses are generally several
orders than the content of main chemical components (e.g.
protein, fiber, fat, etc.). Therefore, the
difference in IR spectra of contaminated and uncontaminated
samples might be caused mainly by changes
in major chemical constituents rather than by variations of
mycotoxin content (124). The limited
sensitivity of conventional NIR or MIR methods could be
incompetent for the direct determination of EA
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 20
in complex matrixes (40). For instance, the sensitivity limit of
conventional NIR spectroscopy for
analyzing most constituents is about 0.1% (125). Thus, it is
speculated that the discrimination and
quantitation models for EA as well as other mycotoxins in
previous studies might be constructed mainly
based on specific changes in chemical components or other
physical alterations which resulted from
mycotoxigenic fungi infection or mycotoxin contamination.
Reliable calibration results only come from appropriately
constructed calibration sample sets. The
acquisition of proper samples prior to calibration stage could
always be a tough task. The concentration
ranges of EA in feed and food are too large, and the
distribution of calibration samples usually unfit for
developing reliable models. In an ongoing mycotoxin project
conducted in University of Saskatchewan,
more than 600 feed and food samples were collected and analysed
for major EA concentration
(unpublished data). However, it is still unable to obtain a
proper sample set for calibration. Take the
wheat sample set for example, 56 of the 80 samples were
contaminated by EA with the range of total EA
from below 1 to above 21000 ppb. In addition, 45 samples were
co-contaminated with more than one EA.
The histogram plot (Figure 4) of total EA concentration shows
the undesirable distribution of these
samples, in which only a handful of samples at higher
concentration while most of the samples at lower
concentration. With this highly skewed data set, it might be
difficult to obtain reliable calibration results.
Feed and food samples usually could be infected by more than one
mycotoxigenic fungi and
contaminated with multiple EA as well as other mycotoxins (10,
11, 13). Under this situation, changes in
chemical constituents and other physical characteristics that
related to EA content are even more complex
and the calibration will be more challenging. Besides, the
samples itself is another challenging factor
since the small quantity of EA/ergot bodies might spread
unevenly in samples.
The development of HSI and microspectroscopy techniques has the
potential to improve the detection
sensitivity of IR based methods considerably. While the
processing of the complex data hypercube is still
a challenging task (36). Errors related to multiple factors
(e.g. sample shape, camera, etc.) could always
be found in images which need to be removed by multivariate
image analysis techniques. What’s more,
the reference concentration value for each spectrum at each
pixel in an image is almost impossible to be
measured by currently available wet chemistry techniques. In
this case, regression models have to be
constructed based on the average value of whole sample.
Outlook and Conclusions
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 21
As a secondary analytical method, IR spectroscopy relies on
accurate reference values and the subsequent
calibration models. The establishment of appropriate calibration
and validation sample sets is of vital
importance to the chemometrics model construction (126, 127).
Most of the previous mycotoxin studies
focus on demonstrating the feasibility of IR based techniques
and usually conducted with relatively small
calibration sample sizes and narrow ranges of concentration.
Therefore it is important to obtain proper
sample sets that with sufficient sample size, representative
concentration, and proper distributions prior to
the calibration step.
Appropriate techniques for spectral outlier detection,
preprocessing, and wavelength selection are
extremely important for extracting maximum amount of relevant
information for modeling. During the
calibration stage, it is recommended to employ both linear and
non-linear regression/classification
techniques for the exploration of all possibilities since the
relationships between spectral characteristics
and EA content may be non-linear. It is essential to evaluate
the performance of calibration models with a
completely independent validation set (external prediction set)
which has not participated in the
calibration step, otherwise the predictive ability of
established model might be overestimated (36).
Unfortunately, many previous IR-based calibration studies have
been reported without independent
prediction set. To facilitate other researchers to evaluate the
accuracy and repeatability of established
multivariate models objectively, statistics of calibration,
cross-validation, and external prediction should
be reported in a proper way.
The ergot issues will continue to be great challenges to the
health of humans and animals. More
comprehensive efforts should be made to reduce their formation,
decrease their toxicity, and minimize
their detrimental effects. Effective strategies could be made
only based on the accurate and timely
detection of EA. To make the determination of EA faster, safer
and greener, more studies focus on
exploring the feasibility of analyzing EA in food and feed
matrixes by different IR techniques, especially
advance molecular spectroscopy techniques, are highly expected
in the near future.
Acknowledgements
The authors thank Brian Chelack and Warren Schwab at Prairie
Diagnostic Services Inc., College of
Veterinary Medicine, the University of Saskatchewan, Canada, who
provided feed samples to the current
project, and also thank Zhiyuan Niu and Na Liu (University of
Saskatchewan) for their technical support.
The Ministry of Agriculture Strategic Research Chair (PY)
Programs have supported by various grants
from the Natural Sciences and Engineering Research Council of
Canada (NSERC-Individual Discovery
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 22
Grant and NSERC-CRD Grant), Saskatchewan Agriculture Strategic
Research Program Fund,
Agricultural Development Fund (ADF), SaskMilk, Saskatchewan
Forage Network (SNK), Western Grain
Research Foundation (WGRF), SaskPulse Growers, SaskCanola,
Prairie Oat Grower Association
(POGA), etc.
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 23
References
1. Vermeulen, P., Pierna, J. F., Van Egmond, H. P., Zegers, J.,
Dardenne, P., and Baeten, V. (2013) Validation
and transferability study of a method based on near-infrared
hyperspectral imaging for the detection and
quantification of ergot bodies in cereals. Anal. Bioanal. Chem.,
405:7765-7772.
2. Krska, R. and Crews, C. (2008) Significance, chemistry and
determination of ergot alkaloids: A review. Food
Additives and Contaminants, 25:722-731.
3. Belser-Ehrlich, S., Harper, A., Hussey, J., and Hallock, R.
(2013) Human and cattle ergotism since 1900:
Symptoms, outbreaks, and regulations. Toxicol. Ind. Health,
29:307-316.
4. Schiff Jr, P. L. (2006) Ergot and its alkaloids. Am. J.
Pharm. Educ., 70:98.
5. Guo, Q., Shao, B., Du, Z., and Zhang, J. (2016) Simultaneous
determination of 25 ergot alkaloids in cereal
samples by ultraperformance liquid chromatography–tandem mass
spectrometry. J. Agr. Food Chem., 64:7033-
7039.
6. Vermeulen, P., Pierna, J. F., Egmond, H. V., Dardenne, P.,
and Baeten, V. (2012) Online detection and
quantification of ergot bodies in cereals using near infrared
hyperspectral imaging. Food Additives &
Contaminants: Part A, 29:232-240.
7. Di Mavungu, J. D., Malysheva, S. V., Sanders, M., Larionova,
D., Robbens, J., Dubruel, P., Van Peteghem, C.,
and De Saeger, S. (2012) Development and validation of a new
LC–MS/MS method for the simultaneous
determination of six major ergot alkaloids and their
corresponding epimers. Application to some food and feed
commodities. Food Chem., 135:292-303.
8. Storm, I. D., Rasmussen, P. H., Strobel, B. W., and Hansen,
H. C. B. (2008) Ergot alkaloids in rye flour
determined by solid-phase cation-exchange and high-pressure
liquid chromatography with fluorescence
detection. Food Additives and Contaminants, 25:338-346.
9. Kim, S., Thiessen, P. A., Bolton, E. E., Chen, J., Fu, G.,
Gindulyte, A., Han, L., He, J., He, S., and Shoemaker,
B. A. (2015) PubChem substance and compound databases. Nucleic
Acids Res.:gkv951.
10. Coufal-Majewski, S., Stanford, K., McAllister, T., Blakley,
B., McKinnon, J., Chaves, A. V., and Wang, Y.
(2016) Impacts of cereal ergot in food animal production.
Frontiers in veterinary science, 3.
11. Mavungu, J. D. D., Larionova, D. A., Malysheva, S. V., Van
Peteghem, C., and De Saeger, S. (2011) Survey
on ergot alkaloids in cereals intended for human consumption and
animal feeding. EFSA Supporting
Publications, 8.
12. Bauer, J. I., Gross, M., Gottschalk, C., and Usleber, E.
(2016) Investigations on the occurrence of mycotoxins
in beer. Food Control, 63:135-139.
13. Mulder, P. P., Pereboom-de Fauw, D. P., Hoogenboom, R. L.,
de Stoppelaar, J., and de Nijs, M. (2015)
Tropane and ergot alkaloids in grain-based products for infants
and young children in the Netherlands in 2011–
2014. Food Additives & Contaminants: Part B, 8:284-290.
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 24
14. Tittlemier, S. A., Drul, D., Roscoe, M., and McKendry, T.
(2015) Occurrence of ergot and ergot alkaloids in
western Canadian wheat and other cereals. J. Agr. Food Chem.,
63:6644-6650.
15. Klotz, J. L. and Smith, D. L. (2015) Recent investigations
of ergot alkaloids incorporated into plant and/or
animal systems. Frontiers in chemistry, 3.
16. Schardl, C. L., Panaccione, D. G., and Tudzynski, P. (2006)
Ergot alkaloids–biology and molecular biology.
The alkaloids: chemistry and biology, 63:45-86.
17. Wallwey, C. and Li, S. (2011) Ergot alkaloids: Structure
diversity, biosynthetic gene clusters and functional
proof of biosynthetic genes. Nat. Prod. Rep., 28:496-510.
18. Strickland, J. R., Looper, M. L., Matthews, J. C.,
Rosenkrans, C. F., Flythe, M. D., and Brown, K. R. (2011)
Board-invited review: St. Anthony’s Fire in livestock: Causes,
mechanisms, and potential solutions. J. Anim.
Sci., 89:1603-1626.
19. Cowan, V. E. and Blakley, B. R. (2014) Ergot contamination
in livestock feeds. Anim Health Perspect, 10:1-2.
20. Bandyopadhyay, R., Frederickson, D. E., McLaren, N. W.,
Odvody, G. N., and Ryley, M. J. (1998) Ergot: A
new disease threat to sorghum in the Americas and Australia.
Plant Dis., 82:356-367.
21. Salvat, A. E. and Godoy, H. M. (2001) A simple thin-layer
chromatographic method for the detection of
ergovaline in leaf sheaths of tall fescue (Festuca arundinacea)
infected with Neotyphodium coenophialum. J.
Vet. Diagn. Invest., 13:446-449.
22. Lauber, U., Schnaufer, R., Gredziak, M., and Kiesswetter, Y.
(2005) Analysis of rye grains and rye meals for
ergot alkaloids. Mycotoxin Research, 21:258-262.
23. Reed, J. D., Ramundo, B. A., Claflin, L. E., and Tuinstra,
M. R. (2002) Analysis of resistance to ergot in
sorghum and potential alternate hosts. Crop Sci.,
42:1135-1138.
24. Miedaner, T., Dänicke, S., Schmiedchen, B., Wilde, P.,
Wortmann, H., Dhillon, B. S., Geiger, H. H., and
Mirdita, V. (2010) Genetic variation for ergot (Claviceps
purpurea) resistance and alkaloid concentrations in
cytoplasmic-male sterile winter rye under pollen isolation.
Euphytica, 173:299-306.
25. Kabak, B., Dobson, A. D., and Var, I. I. L. (2006)
Strategies to prevent mycotoxin contamination of food and
animal feed: A review. Crit. Rev. Food Sci., 46:593-619.
26. Müller, I. (2007) A history of thermodynamics: The doctrine
of energy and entropy. Springer Science &
Business Media:
27. Mainka, S., Dänicke, S., and Ueberschär, K. H. (2005) Effect
of a hydrothermal treatment on ergot alkaloid
content in ergot contaminated rye. Mycotoxin research,
21:116-119.
28. Galvano, F., Piva, A., Ritieni, A., and Galvano, G. (2001)
Dietary strategies to counteract the effects of
mycotoxins: A review. J. Food Protect., 64:120-131.
29. Tomašević-Čanović, M., Daković, A., Rottinghaus, G.,
Matijašević, S., and Đuričić, M. (2003) Surfactant
modified zeolites––new efficient adsorbents for mycotoxins.
Micropor. Mesopor. Mat., 61:173-180.
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 25
30. Suart, B. (2004) Infrared spectroscopy: Fundamental and
applications. John Wiley & Sons, Ltd: Chichester,
UK.
31. Rodriguez-Saona, L. E. and Allendorf, M. E. (2011) Use of
FTIR for rapid authentication and detection of
adulteration of food. Annu. Rev. Food Sci. T., 2:467-483.
32. Karoui, R., Downey, G., and Blecker, C. (2010) Mid-infrared
spectroscopy coupled with chemometrics: A tool
for the analysis of intact food systems and the exploration of
their molecular structure− Quality relationships− a
review. Chem. Rev., 110:6144-6168.
33. Gauglitz, G. and Vo-Dinh, T. (2003) Handbook of
spectroscopy. Wiley-VCH: Weinheim, Germany.
34. Larkin, P. (2011) Infrared and Raman spectroscopy:
Principles and spectral interpretation. Elsevier:
Waltham, USA.
35. Moros, J., Garrigues, S., and de la Guardia, M. (2010)
Vibrational spectroscopy provides a green tool for multi-
component analysis. TrAC Trends in Analytical Chemistry,
29:578-591.
36. Manley, M. (2014) Near-infrared spectroscopy and
hyperspectral imaging: Non-destructive analysis of
biological materials. Chem. Soc. Rev., 43:8200-8214.
37. Sivakesava, S., Irudayaraj, J., and Ali, D. (2001)
Simultaneous determination of multiple components in lactic
acid fermentation using FT-MIR, NIR, and FT-Raman spectroscopic
techniques. Process Biochem., 37:371-
378.
38. Kos, G., Sieger, M., McMullin, D., Zahradnik, C., Sulyok,
M., Öner, T., Mizaikoff, B., and Krska, R. (2016) A
novel chemometric classification for FTIR spectra of
mycotoxin-contaminated maize and peanuts at regulatory
limits. Food Additives & Contaminants: Part A,
33:1596-1607.
39. Senthilkumar, T., Jayas, D. S., White, N., Fields, P. G.,
and Gräfenhan, T. (2016) Near-Infrared (NIR)
hyperspectral imaging: Theory and applications to detect fungal
infection and mycotoxin contamination in food
products. Indian Journal of Entomology, 78:91-99.
40. McMullin, D., Mizaikoff, B., and Krska, R. (2015)
Advancements in IR spectroscopic approaches for the
determination of fungal derived contaminations in food crops.
Anal. Bioanal. Chem., 407:653-660.
41. Zheng, M. Z., Richard, J. L., and Binder, J. (2006) A review
of rapid methods for the analysis of mycotoxins.
Mycopathologia, 161:261-273.
42. Shelby, R. A. and Kelley, V. C. (1992) Detection of ergot
alkaloids from Claviceps species in agricultural
products by competitive ELISA using a monoclonal antibody. J.
Agr. Food Chem., 40:1090-1092.
43. Tunali, B., Shelby, R. A., Morgan-Jones, G., and Kodan, M.
(2000) Endophytic fungi and ergot alkaloids in
native Turkish grasses. Phytoparasitica, 28:375-377.
44. Bueno, D., Istamboulie, G., Muñoz, R., and Marty, J. L.
(2015) Determination of mycotoxins in food: A review
of bioanalytical to analytical methods. Appl. Spectrosc. Rev.,
50:728-774.
45. Trucksess, M. W. (1998) Mycotoxins. J. Aoac Int.,
81:128-137.
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 26
46. Kralj Cigić, I. and Prosen, H. (2009) An overview of
conventional and emerging analytical methods for the
determination of mycotoxins. Int. J. Mol. Sci., 10:62-115.
47. Grob, R. L. and Barry, E. F. (2004) Modern practice of gas
chromatography. John Wiley & Sons: New York,
USA.
48. Gros, M., Petrović, M., and Barceló, D. (2006) Development
of a multi-residue analytical methodology based
on liquid chromatography–tandem mass spectrometry (LC–MS/MS) for
screening and trace level
determination of pharmaceuticals in surface and wastewaters.
Talanta, 70:678-690.
49. Krska, R., Stubbings, G., Macarthur, R., and Crews, C.
(2008) Simultaneous determination of six major ergot
alkaloids and their epimers in cereals and foodstuffs by
LC–MS–MS. Anal. Bioanal. Chem., 391:563.
50. Wang, W. and Paliwal, J. (2007) Near-infrared spectroscopy
and imaging in food quality and safety. Sensing
and instrumentation for food quality and safety, 1:193-207.
51. Barton, F. E. (2002) Theory and principles of near infrared
spectroscopy. Spectroscopy Europe, 14:12-19.
52. de Oliveira, G. A., de Castilhos, F., Renard, C. M. C., and
Bureau, S. (2014) Comparison of NIR and MIR
spectroscopic methods for determination of individual sugars,
organic acids and carotenoids in passion fruit.
Food Res. Int., 60:154-162.
53. Guillen, M. D. and Cabo, N. (1997) Infrared spectroscopy in
the study of edible oils and fats. J. Sci. Food Agr.,
75:1-11.
54. Yang, H. and Irudayaraj, J. (2001) Comparison of
near-infrared, Fourier transform-infrared, and Fourier
transform-Raman methods for determining olive pomace oil
adulteration in extra virgin olive oil. Journal of
the American Oil Chemists' Society, 78:889.
55. Roychoudhury, P., Harvey, L. M., and McNeil, B. (2006) The
potential of mid infrared spectroscopy (MIRS)
for real time bioprocess monitoring. Anal. Chim. Acta,
571:159-166.
56. Paré, J. and Bélanger, J. (1997) Instrumental methods in
food analysis. Elsevier: New York.
57. Griffiths, P. R. and De Haseth, J. A. (2007) Fourier
transform infrared spectrometry. John Wiley & Sons:
58. Richardson, A. D. and Reeves Iii, J. B. (2005) Quantitative
reflectance spectroscopy as an alternative to
traditional wet lab analysis of foliar chemistry: Near-infrared
and mid-infrared calibrations compared.
Canadian journal of forest research, 35:1122-1130.
59. Zimmermann, T., Rietdorf, J., and Pepperkok, R. (2003)
Spectral imaging and its applications in live cell
microscopy. FEBS Lett., 546:87-92.
60. Fischer, C. and Kakoulli, I. (2006) Multispectral and
hyperspectral imaging technologies in conservation:
Current research and potential applications. Stud. Conserv.,
51:3-16.
61. Gao, L. and Smith, R. T. (2015) Optical hyperspectral
imaging in microscopy and spectroscopy–a review of
data acquisition. J. Biophotonics, 8:441-456.
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 27
62. Feng, Y. and Sun, D. (2012) Application of hyperspectral
imaging in food safety inspection and control: A
review. Crit. Rev. Food Sci., 52:1039-1058.
63. Bellisola, G. and Sorio, C. (2012) Infrared spectroscopy and
microscopy in cancer research and diagnosis. Am.
J. Cancer Res., 2:1-21.
64. Wetzel, D. L. and LeVine, S. M. (1999) Imaging molecular
chemistry with infrared microscopy. Science,
285:1224-1225.
65. Yu, P. (2010) Plant‐based food and feed protein structure
changes induced by gene‐transformation, heating
and bio‐ethanol processing: A synchrotron‐based molecular
structure and nutrition research program. Mol.
Nutr. Food Res., 54:1535-1545.
66. Dumas, P. and Miller, L. (2003) The use of synchrotron
infrared microspectroscopy in biological and
biomedical investigations. Vib. Spectrosc., 32:3-21.
67. Dumas, P., Sockalingum, G. D., and Sule-Suso, J. (2007)
Adding synchrotron radiation to infrared
microspectroscopy: What’s new in biomedical applications? Trends
Biotechnol., 25:40-44.
68. Yu, P., McKinnon, J. J., Christensen, C. R., and
Christensen, D. A. (2004) Using synchrotron transmission
FTIR microspectroscopy as a rapid, direct, and nondestructive
analytical technique to reveal molecular
microstructural− chemical features within tissue in grain
barley. J. Agr. Food Chem., 52:1484-1494.
69. Kiralj, R. and Ferreira, M. (2006) The past, present, and
future of chemometrics worldwide: Some
etymological, linguistic, and bibliometric investigations. J.
Chemometr., 20:247-272.
70. Hibbert, D. B., Minkkinen, P., Faber, N. M., and Wise, B. M.
(2009) IUPAC project: A glossary of concepts
and terms in chemometrics. Anal. Chim. Acta, 642:3-5.
71. Gemperline, P. (2006) Practical guide to chemometrics. CRC
Press, Taylor and Francis Group: Boca Ranton,
FL.
72. Varmuza, K. and Filzmoser, P. (2016) Introduction to
multivariate statistical analysis in chemometrics. CRC
press:
73. Rinnan, Å., van den Berg, F., and Engelsen, S. B. (2009)
Review of the most common pre-processing
techniques for near-infrared spectra. TrAC Trends in Analytical
Chemistry, 28:1201-1222.
74. Singh, C. B., Choudhary, R., Jayas, D. S., and Paliwal, J.
(2010) Wavelet analysis of signals in agriculture and
food quality inspection. Food Bioprocess Tech., 3:2.
75. Nicolai, B. M., Beullens, K., Bobelyn, E., Peirs, A., Saeys,
W., Theron, K. I., and Lammertyn, J. (2007)
Nondestructive measurement of fruit and vegetable quality by
means of NIR spectroscopy: A review.
Postharvest Biol. Tec., 46:99-118.
76. Zhu, D., Ji, B., Meng, C., Shi, B., Tu, Z., and Qing, Z.
(2007) Study of wavelet denoising in apple’s charge-
coupled device near-infrared spectroscopy. J. Agr. Food Chem.,
55:5423-5428.
77. Sun, D. (2008) Modern techniques for food authentication.
Academic Press:
Dow
nloa
ded
by [
Aus
tral
ian
Cat
holic
Uni
vers
ity]
at 2
2:15
15
Aug
ust 2
017
-
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT 28
78. Wang, Z. X., He, Q. P., and Wang, J. (2015) Comparison of
variable selection methods for PLS-based soft
sensor modeling. J. Process Contr., 26:56-72.
79. Liu, D., Sun, D., and Zeng, X. (2014) Recent advances in
wavelength selection techniques for hyperspectral
image processing in the food industry. Food Bioprocess Tech.,
7:307-323.
80. Lu, B., Castillo, I., Chiang, L., and Edgar, T. F. (2014)
Industrial PLS model variable selection using moving
window variable importance in projection. Chemometr. Intell.
Lab., 135:90-109.
81. Dai, Q., Sun, D. W., Cheng, J. H., Pu, H., Zeng, X. A., and
Xiong, Z. (2014) Recent advances in de‐noising
methods and their applications in hyperspectral image processing
for the food industry. Compr. Rev. Food Sci.
F., 13:1207-1218.
82. Araújo, M. C. U., Saldanha, T. C. B., Galvao, R. K. H.,
Yoneyama, T., Chame, H. C., and Visani, V. (2001)
The successive projections algorithm for variable selection in
spectroscopic multicomponent analysis.
Chemometr. Intell. Lab., 57:65-73.
83. Liu, F., Jiang, Y., and He, Y. (2009) Variable selection in
visible/near infrared spectra for linear and nonlinear
calibrations: A case study to determine soluble solids content
of beer. Anal. Chim. Acta, 635:45-52.
84. Leardi, R. and Nørgaard, L. (2004) Sequential application of
backward interval partial least squares and genetic
algorithms for the selection of relevant spectral regions. J.
Chemometr., 18:486-497.
85. Zhang, C., Xu, N., Luo, L., Liu, F., Kong, W., Feng, L., and
He, Y. (2014) Detection of aspartic acid in
fermented cordyceps powder using near infrared spectroscopy
based on variable selection algorithms and
multivariate calibration methods. Food Bioprocess Tech.,
7:598-604.
86. Ye, S., Wang, D., and Min, S. (2008) Successive projections
algorithm combined with uninformative variable
elimination for spectral variable selection. Chemometr. Intell.
Lab., 91:194-199.
87. Leardi, R., Seasholtz, M. B., and Pell, R. J. (2002)
Variable selection for multivariate calibration using a
genetic algorithm: Prediction of additive concentrations in
polymer films from Fourier transform-infrared
spectral data. Anal. Chim. Acta, 461:189-200.
88. Bradley, P. S. and Mangasarian, O. L. (1998) Feature
selection via concave minimization and support vector
machines. ICML. 82--90.
89. Maldonado, S., Weber, R., and Basak, J. (2011) Simultaneous
feature selection and classification using kernel-
penalized support vector machines. Inform. Sciences,
181:115-128.
90. Wang, L., Zhu, J., and Zou, H. (2006) The doubly regularized
support vector machine. Stat. Sinica:589-615.
91. He, H., Wu, D., and Sun, D. (2014) Potential of
hyperspectral imaging combined with chemometric analysis
for assessing and visualising tenderness distribution in raw
farmed salmon fillets. J. Food Eng., 126:156-164.
92. Almeida, J. S. (2002) Predictive non-linear modeling of
complex data by artificial neural networks. Curr. Opin.
Biotech., 13:72-76.