Original Article Ann Clin Biochem 1998; 35: 624-632

Multianalyte serum analysis using mid-infrared spectroscopy

R Anthony Shawl, Steven Kotowich ', Michael Leroux- and Henry H Mantsch!
From the I Institute for Biodiagnostics, National Research Council of Canada, 435 Ellice Avenue,
Winnipeg, Manitoba, R3B 1Y6, Canada; and the 2Department of Clinical Chemistry, Health Sciences
Centre, Winnipeg, Manitoba, Canada

SUMMARY. This study assesses the potential for using mid-infrared (mid-IR)
spectroscopy of dried serum films as the basis for the simultaneous quantitation of
eight serum analytes: total protein, albumin, triglycerides, cholesterol, glucose, urea,
creatinine and uric acid. Infrared transmission spectra were acquired for 300 serum
samples, each analysed independently using accepted reference clinical chemical
methods. Quantitation methods were based upon the infrared spectra and reference
analyses for 200 specimens, and the models validated using the remaining 100
samples. Standard errors in the IR-predicted analyte levels (Sy/x) were 2,8 gil (total
protein), 2·2 gil (albumin), 0·23 mmol/L (triglycerides), 0·28 mmol/L (cholesterol),
0-41 mmol/L (glucose) and 1·1 mmol/L for urea, with correlation coefficients (IR vs
reference analyses) of 0·95 or better. The IR method emerged to be less suited for
creatinine (Sy/x = J.lmol/L) and uric acid (Sy]» = 140 J.lmol/L) due to the relatively low
concentrations typical of these analytes.

Additional key phrases: clinical chemistry analyses. urea, glucose, protein. albumin

Recent years have witnessed an explosion in
analytical applications of infrared (IR) spectro-
scopy. The IR spectral range extends from 780 to
25000nm (I2800cm- 1 to 400cm- l ) and is
commonly subdivided further into regions in-
cluding the near-IR (4000--12 800 em -I) and mid-
JR (40Q--4000cm- I ) . The potential of this
technique has been realized by combining the
ever-improving sensitivity of modern spectro-
meters with powerful multivariate quantitation
and classification methods that permit essentially
all of the relevant information latent in the
spectrum to be usefully extracted. These ad-
vances have spurred a number of clinical and
diagnostic applications including the quantita-
tion of urine constituents, 1 the differential
diagnosis of arthritis based upon the JR
spectrum of either synovial fluid? or the residue
left upon drying synovial fluid to a film,3.4
estimation of faecal fat content.! microspectro-
scopy of osteonal bone," the IR measurement of
the lecithin/sphingomyelin ratio (an indicator of
foetal lung maturity) in amniotic fluid," and the
Issued as NRCC publication No. 3479\.
Correspondence: R Anthony Shaw.
E-mail: shaw@ibd.nrc.ca
differentiation of leukaemic from normal lym-
phoid cells."
The promise of IR-based analysis is instru-
mentation that can rapidly and simultaneously
quantitate several constituents without using
specific reagents. In contrast to common enzym-
atic and colorimetric methods, which rely upon
unique chemical agents to recognize and
quantify dissolved species, the method proposed
here is founded upon the unique absorption
patterns that characterize the analytes them-
selves. While previous studies based upon the
mid-IR 9 and near-IR lo . 11 spectra of native serum
have demonstrated promising results, neither
has been adopted for routine clinical use. This
may be due to the fact that although both have
shown potential in trials carried out on
individual spectrometers, the long-term stability
and transferability of the methods are unclear.
The objective of this work was to evaluate a
new approach to quantitation of serum con-
stituents using mid-IR spectroscopy. Based
upon transmission spectroscopy of dried serum
films, this approach eliminates the very strong
water absorptions that otherwise complicate
transmission mid-IR measurements for native
serum. The potential of the method was assessed

624

by developing optimal quantitation models for
eight serum analytes; total protein, albumin,
triglycerides, cholesterol, glucose, urea, creati-
nine and uric acid. The accuracy of the predicted
concentrations confirmed the suitability of the
method for the simultaneous determination of
six (total protein, albumin, triglycerides, choles-
terol, glucose, urea) of these eight analytes.
MATERIALS AND METHODS
Clinical chemistry
A total of 300 serum samples was analysed for
total protein, albumin, triglycerides, cholesterol,
glucose, urea, creatinine and uric acid using
accepted reference clinical chemistry methods.
All samples were selected randomly from a
service laboratory. Analyses for triglycerides,
cholesterol and uric acid were carried out using a
Kodak Ektachem DT-60 analyser (Eastman
Kodak Co., Rochester NY, USA). Analyses
for albumin (bromcresol purple), creatinine
(Jaffe-rate reaction), glucose (hexokinase), total
protein (biuret) and urea (urease) were per-
formed using a Hitachi 717 analyser (Boehringer
Mannheim, Indianapolis IN, USA) with re-
agents from Boehringer Mannheim.
Infrared spectroscopy
Sample preparation
The foundation of the proposed analytical
method is the infrared absorption spectrum of
the film left after evaporation of water from
serum. In principle, the sample could be
prepared by spreading a small volume of serum
onto an IR-transparent material, allowing it to
dry, and measuring the absorption spectrum of
the film. In practice, however, the accuracy of
this method would be compromised by any
variability in the amount of serum successfully
deposited on the window, particularly with
manual sample preparation. In order to com-
pensate for this variability, and to assess its
impact upon the overall accuracy of the method,
we have added an internal standard to each
serum sample. It is worth stressing that this step
could be dispensed with for an automated
sample preparation system, and is only included
here to estimate the accuracy that might be
achieved using a completely automated analyser.
The IR absorption spectrum of the thiocya-
nate ion (SCN -) includes an absorption at
2060 em -I, in a spectral region that contains
no absorptions due to the serum itself. Impreci-
sion in the film preparation was therefore
Infrared serum analysis 625
compensated for by adding equal amounts of
this ion to all of the serum samples, and
subsequently normalizing all of the spectra to
equal intensities for the SCN- absorption at
2060cm- l .
Prior to measuring the infrared spectrum,
I mL of serum was diluted with an equal volume
of 4 g/L aqueous potassium thiocyanate (KSCN)
solution. All spectra were measured within 24 h
of the reference analyses. In order to assess the
precision of the derived analyte concentrations,
two replicate samples were prepared from each
serum specimen by spreading 7/-lL of the diluted
serum evenly on the surface of circular
(13 x 2 mm) BaF 2 windows. All films were air
dried for 30 min prior to measuring the infrared
spectra.
Spectral measurements
Transmission spectra were measured using a
Digilab FTS-40 Fourier transform infrared
spectrometer (Bio-Rad Laboratories, Cam-
bridge MA, USA) equipped with a liquid-
nitrogen -cooled mercury-cadrnium-telluride
(MCT) detector. A total of 256 scans (2 min
collection time) were co-added at a resolution of
4 em -I (no apodization), using a clean BaF 2
window for the background.
Calibration models
Methods
The analytical method used here is a secondary
method. Independent reference analyses were
therefore carried out for all of the serum
specimens, using accepted clinical chemistry
methods and instrumentation. Eight calibration
models were then derived to yield relationships
between the IR spectra and the various analyte
concentrations. These models were derived using
only a subset of the IR spectra-the 'training
set'-in combination with the corresponding
reference analyte levels. Each model was then
validated by predicting analyte levels for a
separate 'test set' of IR spectra and comparing
the predicted values to the reference analyses.
The calibration models were developed for
each of the eight analytes using the partial-least-
squares (PLS) method. A full account of this
method is beyond the scope of this report,
however excellent descriptions are in the litera-
ture. 12•13 The practical aspects of the procedure
have some parallels to multiple point linear
regression, in that both judgement and many
trial calibrations are required to determine:
Ann Clin Biochem 1998: 35
626 Shaw et al.
(a) what (if any) data pretreatment is required
(e.g., baseline/offset correction, smoothing,
derivation, etc.)
(b) which spectral region(s) should be included
in the model (in contrast to linear regres-
sion, the analytical information is derived
from patterns over spectral regions rather
than from the intensities at individual
frequencies)
(c) how many PLS factors should be included in
the model (the analogous decision in multi-
ple point linear regression would be how
many frequencies to include in the model)
An PLS calibration models were derived by
systematically exploring these three choices until
no further improvements in accuracy could be
achieved. Spectral manipulations and PLS
calibrations were carried out using the commer-
cial software packages GRAMS/32 (Galactic
Industries, Salem NH, USA) and Win-IR (Bio-
Rad Laboratories, Cambridge MA, USA).
Derivation of calibration models: an overview
The first step in the workup of the IR spectra
was to normalize all of the spectra to a common
absorption intensity for the C - N stretching
mode of the KSCN internal standard. An
interim PLS calibration model was then derived
for each analyte using a training set comprised
of the averaged replicate spectra. The same
model was then used to predict concentrations
for the individual replicate spectra in the
training set; outliers that could thereby be traced
to imprecision in the spectral measurements
were removed and the calibration models
optimized. The eight resulting calibration mod-
els, based upon the averaged replicate spectra
for 200 serum specimens, where then validated
by predicting the analyte concentrations for the
remaining 100 samples.
Finally, in order to address the question of
whether spectral normalization is necessary, a
parallel set of PLS calibration models was
evaluated using the IR spectra without normal-
ization.
RESULTS AND DISCUSSION
Infrared spectra
The infrared spectrum is, in essence, a reflection
of the infrared 'colour pattern' characteristic of
the sample. The basis of quantitation is that each
constituent contributes a unique absorption
pattern to the overall spectrum, governed by
the unique set of molecular vibrations character-
Ann Clin Biochem 1998: 35
istic of each distinct molecular species. The
spectra of creatinine, uric acid, urea, albumin,
cholesterol, triglycerides and glucose are all
distinct from one another, which provides the
basis for the selectivity in IR-based serum
analysis. The quantitative information is carried
by the relative intensities of the various con-
stituent spectra contributing to the unique
absorption profile of each serum specimen.
A representative I R absorption spectrum is
displayed in Fig. I. Not surprisingly, the general
appearance of this and all serum spectra is
dominated by the absorptions of the protein
constituents. For example, the major features at
1540 ern -1 and 1655 em -I correspond to the
protein amide II (N-H bending) and amide I
(C = 0 stretching) vibrations respectively, while
the high frequency band at 3300 cm - \ corre-
sponds to the amide A (N-H stretching) mode.
The feature at 2060 cm -I is the C N stretching
absorption of the KSCN internal standard.
While species of lower concentration also
contribute to the overall absorption profile,
their contributions are much weaker than the
major protein absorptions. In addition, the
absorptions from the various, relatively minor.
constituents may be so closely spaced that they
cannot be distinguished in the absorption
spectra of Fig. I. For this reason, infrared
analytical models often benefit from mathema-
tical pretreatment of the spectra to enhance the
effective resolution of closely spaced bands;
further details of the procedure used here are
outlined in the Appendix.
Initial calibration models
We illustrate here the development of calibration
models for each of the analytes, using glucose as
an example.
The ultimate goal of this procedure is to
derive an optimal relationship (calibration
model) relating the IR spectra in the training
set to the corresponding known reference
analyte concentrations, using the PLS method
as the basis for deriving this relationship. To this
end, a number of provisional models were
tested, varying both the choice of spectral
region(s) and the number of PLS factors. It is
important to note that at this stage the models
were tested using only the spectra/analyte levels
for the samples in the training set. The accuracy
of these provisional models was therefore
estimated using cross-validation, as described
in the Appendix.
 Infrared serum analysis 627

Ql
o
c
ra
z
.c o
o
Ul
Ul
~
.c
«

i I I i I I i I I
BOO 1200 1600 2000 2400 2BOO 3200 3600 4000
Wavenumber (crn")

FIGURE I. Representative infrared absorption spectrum for a film obtained hy drying 7 J.lL of a 50:50 mixture of
serum and aqueous KSCN (4 giL) onto circular BaF] windows (/3 mm diameter). The highlighted KSCN absorption
at 2060cm- 1 lVa.l· used as an intemal standardfor spectral scaling (see text).

Once relationships between the model accu-
racy and choice of spectral range became
apparent, the spectral region was varied system-
atically until it was clear that no further
improvements in the model accuracy were likely.
Occasionally, it became apparent that two (or
more) separate regions were required, thereby
eliminating superfluous spectral information
(i.e., 'noise') contained between the two regions.
Precision and outlier detection
Precision outliers
Manual sample preparation unavoidably leads
to some imprecision in the spectra, and hence
the analytical results. Since replicate spectra
were measured for all specimens, it was
straightforward to identify and remove the
specimens whose spectra were most seriously
affected. Concentrations were predicted for the
replicate spectra and compared in pairs. The five
pairs showing the worst agreement were identi-
fied, and the corresponding specimens removed
from the calibration set.
The replicate spectra were also used to gauge
the role that imprecision in the IR method plays
in governing the overall accuracy. The contribu-
tion to the total variance that may be attributed
to this source was estimated as:
(cr; p,,,,;,;on)2 = ~(C, - C; "vc)2 IN
where 'i' is the sample number, C, the concen-
tration predicted for one of the two replicate
measurements for sample 'i', ciave the average of
the two measurements and N the number of
samples in the training set.
Concentration and spectral outliers
One spectrum was unique in that it gave rise to
particularly severe outliers in seven of the eight
calibrations. In all cases, the large analytical
errors could be attributed to spectral features
that were not modelled by PLS (part of the PLS
regression is a reconstruction of the infrared
spectra; these spectra emerged as clear outliers in
scatterplots of spectral residual vs sample
number). This sample was therefore removed
from all calibration trials. Standard statistical
tests'? identified very few additional samples as
outliers; one such specimen was removed from
the training set for total protein, one for
albumin, one for triglycerides, none for choles-
terol, two for glucose and two for urea. With the
spectra giving no indication of unusual chemical
interference, these outliers may correspond to
atypically large errors in the reference analyses.
Final model optimization and validation
After removing the outliers identified above, the
final glucose calibration model was derived by
fine-tuning the initial model. Small adjustments

Ann cu« Biochem 1998: 35

628 Shaw et al.

to the choices of spectral region were explored, 2·0

and the appropriate number of PLS factors ---- SEP
c::- -0- SEC
determined, as outlined in the Appendix. For llJ
rn 1·6
glucose, the PLS calibration model that proved t5
most accurate used only the frequency range llJ
~ 1·2
925-l250cm-l, which is not surprising since the
mid-IR spectrum of glucose includes several very ~
Cii
strong absorptions in this region. The optimal 'E 0·8
rQ
spectral region(s) for glucose and the seven other "C
C
species are summarized in Table 1. .l!1
The role that the number of PLS factors plays
rn 0-4
in the calibration model is illustrated in Fig. 2. o 2 4 6 8 10 12 14 16
The standard error of calibration (SEC) for
Number of PLS factors
glucose decreased rapidly as the early factors
were added and reached a stable plateau as FIGURE 2. Trends in the standard errors of calibration
additional factors were included in the model; (SEC) and validation (SEP) with increasing factors in
the addition of factors beyond the tenth resulted the partial-least-squares (PLS) model. The SEC and
only in a marginal improvement. In contrast, SEP (see Appendix) are the root mean square (RMS)
while the standard error of prediction (SEP) for deviations between the JR-hased and reference analyses
for the samples in the training set and the samples in the
the test set was nearly identical to the SEC for test set.
models of up to 10 PLS factors, the two began to
diverge at 11 factors. This behaviour was typical
of all of the analytes and reflects a balance
between underfitting the data in the training set,
thereby failing to extract all of the relevant
information that is available from the spectra,
and overfitting, in which case the SEC is
artificially low.
Figure 3 compares the IR-predicted glucose
levels to the corresponding reference analyses
Q) 5
for both the test and training sets. This figure Ul
0
U
graphically illustrates the fact that the lO-factor ::::I

PLS model provides virtually identical accuracy Cl

for the training set and for the test set-a finding "5l
ts
25
common to all of the IR-based methods
presented below. This is perhaps the clearest
¥c. Training set
possible indicator that the underlying calibration a: 20

TABLE 1. PLS models 15

Spectral region( s) No. of
Analyte (cm- I ) . factors·· 10

Albumin 1100-1800 12 5
Cholesterol 1100-1300,1700-1800, II
2800-3000
Glucose 925--1250 10 0
Total protein 900-1800 13 0 5 10 15 20 25
Triglycerides 900-1500, 1700-1800, 7 Reference glucose (mmoI/L)
2800-3000
Urea 1400-1800 13 FIGURE 3. Comparison ofglucose levels as provided by
Total CO2 900-1500 14 the reference analyses and as predictedfrom the infrared
Creatinine 800-1600, 2800-3500 7 (JR) spectra using a Ill-factor partial-least-squares
Uric acid 800-1800, 2800-3500 9 (PLS) model. Separate plots for the training set (i.e.,
the samples used to calibrate the method) and the test
• = Spectral regions included in the optimized PLS set (i.e., the samples used to validate the method)
calibration model; •• = number of PLS factors in the illustrate that the method is equally accurate for both
optimized model. data sets. The line of identity is included for reference.

Ann cu« Biochem 1998: 3S


models are soundly based, and therefore that the
methods would be equally accurate for routine
analysis.
Summary of PLS analytical models
The accuracy of the IR method is illustrated by
the difference plots in Fig. 4, comparing the
IR-predicted analyte levels for the 100 specimens
in the test set to the corresponding reference
analyses.
The overall performance of the mid-IR/PLS
quantitation models for albumin, cholesterol,
glucose, total protein, triglycerides and urea
suggests that all six of these analytes may
potentially be assayed using rnid-IR spectro-
scopy. For uric acid and creatinine, the IR-based
method is markedly less accurate relative to the
comparatively low mean concentrations for
these analytes.
Precision of IR-based methods compared to
reference methods
Both the calibration and the validation of the
IR-based analyses have been carried out with
reference to the same accepted reference clinical
chemistry methods. Scatter in the plots pre-
sented in Fig. 4 therefore reflects both uncer-
tainties in the IR method and uncertainties in
the reference analyses; even if the IR-based
method were to be perfect, there would remain
scatter in the difference plots reflecting differ-
ences between the reference analyses and the
true analyte levels. We term the corresponding
standard deviation in the reference analyses
'SDD ref' .
There is no way to determine the absolute
accuracy of the IR method without resorting to
more fundamental reference analyses. We can,
however, estimate the role that imprecision in
the IR measurements plays in the overall
performance of the IR-based analyses. This
imprecision might arise from non-uniformity in
the thickness of the dried serum films, impreci-
sion in the dilution step, focal precipitation of
non-protein solutes as the film dries, incomplete
drying of the films or from noise in the IR
spectra. The contribution of these factors was
gauged by the term 'uprccision" defined earlier,
which reflects the extent to which analytical
results differ for replicate films.
Comparison of the SDD (SD of the IR
method relative to the reference method) to
uprecision (SD of IR-predicted values for indivi-
dual measurements relative to the mean values
for replicate measurements) is sufficient to
Infrared serum analysis 629
confirm that imprecision in the IR measure-
ments is not the limiting factor in the overall
performance of any of the IR methods (see
Table 2).
Internal standard and normalization
It was not clear a priori whether the overall
accuracy of the serum analysis would be
improved by normalization to the KSCN
internal standard. Having based the optimiza-
tion of calibrations on spectra that were normal-
ized to the C N stretching mode of the KSCN
internal standard, a second set of calibrations
was derived following exactly the same proce-
dure but using spectra that were not normalized.
While the normalized spectra provided equal
or better agreement in all cases, the differences
were often insignificant (see Table 3). However,
in certain cases, notably for glucose, the normal-
ization step provided a marked improvement. It
seems likely that the need for normalization
would be obviated if the film preparation itself
were better 'normalized', i.e., through auto-
mated rather than manual sample preparation.
SUMMARY AND CONCLUSIONS
This study has demonstrated the potential of a
novel infrared spectroscopic method for serum
analysis. For six of the eight analytes studied
(albumin, cholesterol, glucose, total protein,
triglycerides and urea), the analytical results
are probably precise and accurate enough to
confirm the potential of the method for use in
routine clinical analyses.
The effective implementation of this method-
ology into the clinical laboratory would require
the development of automated sample prepara-
tion methods, the deposition of 5-10 J.lL sample
aliq uots onto infrared-transparent substrates,
and drying to a film. Automation of these steps
would improve the rate of sample throughput as
well as improving the overall precision of the
method. It is also likely that standardized,
reproducible film preparation would lead to
uniformity of film drying for native serum; the
internal standard/dilution approach demon-
strated here is largely a means to compensate
for the inevitable imprecision that arises in
manual sample preparation. Calibration to more
accurate reference methods would undoubtedly
improve the accuracy of at least some of the IR-
based analyses. The results reported here there-
fore represent upper limit estimates of the
Ann Clin Biochem 1998: 3S
630 Shaw et al.

600 5·0
400 __________ -::-__0- _ a
2·5 ------o-----..Q'O,",o,,""---- - - - - - - - - - - - - - - - ~- - - - --
200 a a a oooQlCP ~ ° ~~ a
8 qJ0go fJ °
Oi-~~::----"",",,:--l!f.o'-=;P----­
Q)

~ 0 ~~
8, g~o 0~'O 00
°° ° a
-200 ____ ~ 000 0 0
~__~ ?_~~ 9 _ -- ----------------------- - - -aa - - - - -e- - - - ---
-2·5
-400
-600 -I----.---,------r------r---, -5·0 ~---.,.-,.---._____._-_r___,-_,
o 200 400 600 800 1000 o 5 10 15 20 25 30 35
Uric acid (~moI/L) Urea (mmoI/L)

1·5 10
1·0
--------·----------------0-0---------------- °
5 ------0------0------------------
a ------------
0·5
° 0 0 0 a 0
0 0 0°
000
°ooog a
° 0° gOoOgoo
o
0.0 J-----<>-.....;r--;p\lQ..,.q~_;o__=:--:~
0°0 00 0 a 08S8 0
______________________~ oo__~.QB~~o a _
-0·5
-5 Boo a
-1·0
-1·5 -'---,------.----.,....-----, -1 0 +--.,....--,.---..---..---~~
2 4 6 8 15 20 25 30 35 40 45
Cholesterol (mmoI/L) Albumin (gIL)
1·5 300

1·0 ~------------------------------- 200

~
0
0.5 00 0 0 100 a
o 0
0·0 00
o
o
-0·5 -100
-1·0 a -200
-1·5 +----.-----.,.--,.-----r----, -300 +---.---__,r---r--~-~
o 5 10 15 20 25 o 100 200 300 400 500
Glucose (mmol/L) Creatinine (~moI/L)

12 1·0

8 -------------------o----------------oP------- 0·5 ____________o. .o .q,_ Cf---- - - - - --

00 ° 00.0'0
4 ~ oeD OJ a 00 ° 8 6 ' 00 8 CO 0

o o 0 ° '" a
o
o
000°0' _.Q
° "0"0 0 a
a
a
!7& 0 eoCO 00 00
-4 ---------------------O"----------o-a.--o------ ____o__8.._J:J.. 0 __ -- - - CP - '!
o -0·5 ° 0 °
-8
-12 .l...--.-_ _.---_---.--_ _, . - - _ - - , -1·0 +---.---__,r---r--~-~
40 50 60 70 80 o 2 3 4 5
Protein (gIL) Triglycerides (mmoI/L)

Average (reference + IR predicted analyte level)/2

FIGURE 4. Difference plots comparing the infrared (IR)-based analyses to the reference analytical results. Solid
horizontal lines indicate the average difference between the two methods, and dashed lines indicate 95% confidence
intervals.

Ann Clin Biochem 1998: 3S

 Infrared serum analysis 631

TABLE 2. Summary of the comparison of reference to infrared (IR)-based analyses

Median Average
concentration" difference" SOD aprecision

Albumin (g/L) 32 0.3 (0'22) 2.2 0.75

Cholesterol (rnmol/L) 4.8 0.08 (0'03) 0.28 0.10
Glucose (rnmol/L) 6.3 0.06 (0'04) 0.41 0.20
Total protein (giL) 64 I (0'3) 2.8 1.1
Triglycerides (rnrnol/L) 1·8 0.02 (0'03) 0.23 0.05
Urea (mmoI/L) 7·9 -0.1 (0'1) 1.1 0.39
Creatinine (JlmoI/L) 100 II (7) 70 12
Uric acid (JlmoI/L) 380 -8 (15) 140 3S

"Median concentration for samples in the test set; "mean difference between reference and IR-based analyses
(reference method-IR method). The standard error of the mean difference is in parentheses; SDD=standard
deviation of differences (reference method -I R method); up"ci'i"n = estimated contribution of imprecision in
replicate measurements to the overall standard errors (see text).

TABLE 3. Comparison of standard deviations (IR 3 Eysel HH, Jackson M, Mantsch HH. Method for
versu.\' reference methods) for models using I R spectra diagnosing arthritic disorders by infrared spectro-
with and without normalization scopy. United States Patent 5,473,160.
4 Eysel HH, Jackson M, Nikulin A, Somorjai RL,
SOD Thomson GTO, Mantsch HH. A novel diagnostic
test for arthritis: Multivariate analysis of infrared
Analyte Spectra Spectra not spectra of synovial fluid. Biospectroscopy 1997; 3:
normalized" normalized 161-7
5 Franck P, Sallerin JL, Schroeder H, Gelot MA,
Albumin (giL) 2·2 2·7 Nabet P. Rapid determination of fecal fat by
Cholesterol (rnrnol/L) 0·28 0·43 Fourier transform infrared analysis (FTIR) with
Glucose (rnmol/L) 0-41 0·65 partial least-squares regression and an attenuated
Total protein (giL) 2·8 ]A
total reflectance accessory. Clin Chem 1996; 42:
Triglycerides (rnmol/L) 0·23 0·24 2015-20
Urea (mmol/l.) 1·1 1·4 6 Paschalis EP, DiCarlo E, Betts F, Sherman P,
Creatinine (JlmoI/L) 70 110 Mendelsohn R, Boskey AL. FTIR microspectro-
Uric acid (umol/L) 130 160 scopic analysis of human osteonal bone. Calcif
Tissue Int 1996; 59: 480-7
SOD = Standard deviation of differences (lR method
7 Liu K-Z, Ahmed MK, Dembinski TC, Mantsch
- reference method); "normalized to an absorption of
HH. Prediction of fetal lung maturity from near-
the internal standard (KSCN absorption at 2060cm -I).
infrared spectra of amniotic fluid. Int J Gynecol
Obstetrics 1997; 57: 161-8
8 Schultz CP, Liu K-Z, Johnston JB, Mantsch HH.
Study of chronic lymphocytic leukemia cells by FT-
IR spectroscopy and cluster analysis. Leukemia
performance of this general approach to serum Research 1996; 20: 649-55
analysis. 9 Heise HM, Marbach R, Koschinsky Th, Gries FA.
Multicomponent assay for blood substrates in
human plasma by mid-infrared spectroscopy and
Acknowledgement its evaluation for clinical analysis. Appl Spectrosc
We gratefully acknowledge the role of Jim Graf, 1994; 48: 85-95
10 Hall JW, Pollard A. Near-infrared spectrophoto-
who carried out the reference analyses.
metry: A new dimension in clinical chemistry. Clin
Chem 1992; 38: 1623-31
II Hall JW, Pollard A. Near-infrared spectroscopic
REFERENCES determination of serum total proteins, albumin,
globulins, and urea. Clin Biochem 1993; 26: 483-90
Shaw RA, Kotowich S, Mantsch HH, Leroux M. 12 Martens H, Naes T. Multivariate Calibration. New
Quantitation of protein, creatinine, and urea in York: John Wiley and Sons, 1989
urine by near-infrared spectroscopy. Clin Biochem 13 Beebe K, Kowalski BR. An introduction to
1996; 29: 11-9 multivariate calibration and analysis. Anal Chem
2 Shaw RA, Kotowich S, Eysel HH, Jackson M, 1987; 59: 1007a-17a
Thomson GTO, Mantsch HH. Arthritis diagnosis
based upon the near-infrared spectrum of synovial
fluid. Rheumatology International 1995; 15: 159-66 A ccepted for publication 27 April 1998

Ann cu« Biochem 1998: 3S

632 Shaw et al.
APPENDIX
Pretreatment of the IR spectra
A number of pretreatment options are available
to convert the spectra to a form that is optimal
for quantitation. For example, fluctuations in
the baseline and slope are easily removed by
evaluating the second derivatives. This proce-
dure also effectively enhances the resolution of
closely spaced absorptions, which in turn
permits better discrimination of features due to
the various species. The second derivative
spectra were adopted initially for all quantita-
tion models, and trial and error testing of a
number of alternatives (deconvolution, first
derivatives, no pretreatment) verified this to be
the optimal choice.
Partial-least squares. Definitions and model
optimization
Cross-validation, standard error of cross-
validation
Cross-validation is a method for evaluating the
accuracy of a PLS model without resorting to an
independent set of test spectra. The predictive
accuracy of a model derived using 'N' specimens
(the spectra and corresponding reference ana-
lyses) is estimated by evaluating N-I new
models, leaving out one of the spectra in each
case. For each of these subsidiary models, the
Ann Clin Biochem 1998: 35
analyte level is predicted for the sample left out.
The process is repeated N times, and the root
mean square (RMS) error in the predicted
analyte levels taken as an estimate of the model
accuracy. This is referred to as the standard
error of cross-validation, or SECV.
Standard error of calibration and validation
The standard error of calibration is the RMS
error in analyte levels predicted for the speci-
mens in the training set, i.e., for the set of
specimens used as the basis for deriving the
calibration model. The SEP is the RMS error in
analyte levels predicted for the specimens in the
test set, i.e., for the independent set of specimens
that was not used in deriving the calibration
model.
Optimization of the number of PLSfactors
Typically, the SECV decreases regularly as the
first few PLS factors are included in the
calibration model. As additional factors are
included, the SECV reaches a minimum, signal-
ling that the inclusion of additional factors will
be of no benefit to the model. The number of
PLS factors is therefore limited to the number
required to bring the SECV to, or close to, a
minimum value.