Anal. Chem.
2003, 75, 754-760
De ve lopm e nt a nd V a lida t ion of a N e a r-I nfra re d
M e t hod for t he Qua nt it a t ion of Ca ffe ine in I nt a c t
Single T a ble t s
Magali Laasonen, †,‡ Tuulikki Harmia-Pulkkinen, ‡ Christine Simard, § Markku Ra1 sa1 nen, | and
Heikki Vuorela* ,†
Department of Pharmacy, Division of Pharmacognosy, P.O. Box 56 (Viikinkaari 5E), FIN-00014 University of Helsinki,
Helsinki, Finland, and Pharmia Oy, P.O. Box 387, FIN-00101 Helsinki, Finland
A near-infrared spectroscopic method was developed and
validated for determining the caffeine concentration of
single and intact tablets in a Finnish pharmaceutical
product containing 5 8 .8 2 % (m/ m) caffeine.The spectral
region of interest contained a total of 4 7 4 data points. The
second derivative of Savitsky-Golay, a standard normal
variate, and mean centering were used as spectral pre-
processing options. The feasibility study showed non-
uniformity of caffeine repartition within each tablet. Thus,
spectra were recorded from both faces of the tablets, and
the analysis result for a single tablet was reported as the
average of both face determinations. Precision of the
method was validated because the relative standard
deviations from repeatability and intermediate precision
tests were below 0 .7 5 % (m/ m). Accuracy validation
proved that the NIR results were not significantly different
( P ) 0 .0 9 , n ) 1 2 ) from the results obtained with the
reference HPLC method. The limit of quantification for
caffeine was 1 3 .7 % (m/ m) in the tablets. The method was
found to be unaffected by NIR source replacement, but
the repeatability of the results was affected if the sample
holder was not placed in the correct position in the light
beam. Routine NIR analysis of caffeine in tablet form was
found to be more flexible and much faster than that
performed with the HPLC method.
Caffeine or 1,3,7-trimethylxanthine is a widely used drug
throughout the world. This alkaloid occurs naturally in tea leaves,
coffee beans, cocoa beans, and maté leaves and is traditionally
used for its stimulatory effects.1 The caffeine found in medicines
and food supplements can be of natural origin (e.g., as a byproduct
from caffeine-free product manufacture) or produced by chemical
synthesis.
* Corresponding author. E-mail: heikki.vuorela@ helsinki.fi.
†
Department of Pharmacy, University of Helsinki.
‡
Pharmia Oy.
§
ABB Bomem Inc., 585 Charest Boulevard, East Suite 300, Quebec, PQ,
Canada, G1K 9H4.
|
Department of Chemistry, University of Helsinki.
(1) O’Neil, M. J.; Smith, A.; Heckelman, P. E.; Obenchain, J. R.; Gallipeau, J. A.
R.; D’Arecca, M. A.; Budavari, S. Merck Index, an Encyclopedia of Chemicals,
Drugs and Biological; 13th ed.; Merck & Co., Inc.: Rahway, NJ, 2001; p
275.
754 Analytical Chemistry, Vol. 75, No. 4, February 15, 2003
The most common methods used for analyzing caffeine are
probably HPLC, recommended by the American Pharmacopoeia
USP 25, and the UV method, even though these techniques are
slow and solvent consuming. Researchers have recently shown
interest in near-infrared spectroscopy (NIR) for the analysis of
caffeine in its “natural matrix”, i.e., in green tea leaves2 or in
coffee.3,4 So far, only one study has been reported on the NIR
quantification of caffeine isolated from a pharmaceutical prepara-
tion, and it was published almost thirty years ago.5 This study,
however, did not make use of the main advantage of NIR
spectroscopy, i.e., determining a compound without having to
isolate it from its matrix.
The advantages6 of NIR spectroscopy over traditional methods
for quantitation in solid form are very attractive due to the physical
properties of the NIR region. The low molar absorptivity of NIR
bands permits the measurement of solid samples with little or no
sample preparation, thus avoiding manipulation errors. NIR
spectroscopy is therefore environment-friendly because there is
usually no need to dilute samples. The speed of this technique is
due to both the minimum amount of sample preparation and the
short time needed to record spectra. Moreover, the NIR signal
contains both physical and chemical information about the
samples. Thus it can be used for qualitative analysis7,8 and for
measuring different physical parameters on the sample (e.g.,
particle size, tablet hardness,9 or thickness of the tablet coating).
NIR can also be used for monitoring on-line processes.10
Despite all these advantages, only a small number of papers
have been published on the NIR quantitation of intact tablets9,11-13
compared to the large number of qualitative NIR applications
(2) Schulz, H.; Engelhardt, U. H.; Wegent, A.; Drews, H. H.; Lapczynski, S. J.
Agric. Food Chem. 1 9 9 9 , 47 , 5064-5067.
(3) Downey, G.; Boussion, J. J. Sci. Food Agric. 1 9 9 6 , 71 , 41-47.
(4) Fabian, Z.; Izvekov, V.; Salgo, A.; Orsi, F. Anal. Proc. 1 9 9 4 , 31 , 261-263.
(5) Allen, L. J. Pharm. Sci. 1 9 7 4 , 63 , 912-916.
(6) Blanco, M.; Coello, J.; Iturriaga, H.; Maspoch, S.; De La Pezuela, C. Analyst
1 9 9 8 , 123 , 135R-150R.
(7) Laasonen, M.; Rantanen, J.; Harmia-Pulkkinen, T.; Michiels, E.; Hiltunen,
R.; Räsänen, M.; Vuorela, H. Analyst 2 0 0 1 , 126 , 1122-1128.
(8) Laasonen, M.; Harmia-Pulkkinen, T.; Simard, C. L.; Michiels, E.; Räsänen,
M.; Vuorela, H. Anal. Chem. 2 0 0 2 , 74 , 2493-2499.
(9) Chen, Y.; Thosar, S. S.; Forbess, R. A.; Kemper, M. S.; Rubinovitz, R. L.;
Shukla, A. J. Drug Dev. Ind. Pharm . 2 0 0 1 , 27 , 623-631.
(10) Rantanen, J.; Räsänen, E.; Tenhunen, J.; Känsäkoski, M.; Mannermaa, J. P.;
Yliruusi, J. Eur. J. Pharm. Biopharm . 2 0 0 0 , 50 , 271-276.
(11) Blanco, M.; Coello, J.; Iturriaga, H.; Maspoch, S.; Pou, N. Analyst 2 0 0 1 ,
126 , 1129-1134.
10.1021/ac026262w CCC: $25.00 © 2003 American Chemical Society
Published on Web 01/17/2003
reported. The main reason for this is probably because Pharma-
copoeias do not give guidelines for the application of NIR in
quantitative analysis, even though there are recommendations14
on how to perform a qualitative NIR analysis. Moreover, the ICH
guideline 15 was not aimed at the validation of nonseparative
procedures, which makes the validation of NIR methods more
difficult to perform.
In this work, we have developed and validated a NIR diffuse
reflectance method for the determination of the caffeine content
in intact single tablets. This method can be used to determine
the caffeine content of a batch, as well as the content uniformity
of caffeine tablets, and to validate the manufacturing process, e.g.,
by controlling content deviations or comparing the caffeine content
of tablets pressed by each punch of the press. Validation of the
method was performed according to the ICH guideline 15 where
applicable and using recommendations from Moffat et al.16
The main points of this study are the following: (i) Caffeine
is quantified for the first time in intact tablets by a validated NIR
method. (ii) Spectral features of synthetic and natural caffeine were
compared. (iii) Spectral differences between the tablet faces were
demonstrated and explained. (iv) Statistical analysis of the model
residuals was found to be an interesting new parameter to include
in the method validation. (v) The method was found to be as
reliable and much faster than the reference HPLC method.
EXPERIMENTAL SECTION
Material. The pharmaceutical product studied was in the form
of commercially available 170-mg caffeine tablets produced by
Pharmia Oy (Helsinki, Finland). The active principle content
(anhydrous caffeine) was 58.82% (m/ m), and the excipient mass
consisted of cellulose, lactose, and other minor excipients. Twenty-
two production batches were supplied by Pharmia Oy, and 11
laboratory-made batches were prepared in order to broaden the
caffeine concentration range covered by production samples.
These batches were obtained by overdosing or underdosing the
samples by adding caffeine or excipient mass during the mixing
process prior to tableting. The laboratory-made tablets had a
weight similar to those from the production batches. The caffeine
concentration covered a very wide range: 0 (excipient tablets) -
100%(m/ m) (pure caffeine tablets). Two other laboratory batches
were prepared using the same formula as for the production
batches, except that caffeine was replaced by theobromine (3,7-
trimethylxanthine) or theophylline (1,3-trimethylxanthine).
HPLC Analysis. The active principle concentration in each
batch was obtained as the average of two HPLC determinations.
The isocratic reversed-phase HPLC method used a methanol/
0.05 M NaH2PO 4 (30:70) mobile phase, a Lichrocart 125-4
precolumn, a Lichrospher 5-µm, 100 × 4.60 mm RP-18 Merck
column, a flow rate of 1 mL/ min, and a run time of 12 min. The
UV absorbance was measured at 275 nm. The calibration curve
(12) Broad, N. W.; Jee, R. D.; Moffat, A. C.; Smith, M. R. Analyst 2 0 0 1 , 126 ,
2207-2211.
(13) Thosar, S. S.; Forbess, R. A.; Ebube, N. K.; Chen, Y.; Rubinovitz, R. L.;
Kemper, M. S.; Reier, G. E.; Wheatley, T. A.; Shukla, A. J. Pharm. Dev.
Technol. 2 0 0 1 , 6 , 19-29.
(14) European Pharmacopoeia, 4th ed.; Council of Europe: Strasbourg, 2002;
pp 55-56.
(15) ICH Harmonised Tripartite Guideline: Validation of Analytical Procedures
Methodology, International Conference on Harmonization, 1996.
(16) Moffat, A. C.; Trafford, A. D.; Jee, R. D.; Graham, P. Analyst 2 0 0 0 , 125 ,
1341-1351.
( R2 ) 0.999) was established from duplicate determinations of
standard samples at six different concentrations of caffeine in
methanol/ NaH2PO 4. Ten tablets per batch were ground to a fine
powder, and 50 mg of this mass was dissolved in methanol/ NaH2-
PO 4 (30:70) using sonication. The extracts were filtered and then
transferred into HPLC vials. The precision of the method was
approximated by the standard deviation (eq 1) from the pooled
results of duplicate determinations made on validation samples.
x∑
k)n
2 2
[( xk1 - jxk) + ( xk2 - jxk) ]
k)1
s) (1)
n
SE ) s/ x2 (2)
where n is the number of duplicates, xk1 and xk2 are the individual
duplicate results, and jxk is the mean of the duplicates. The
standard deviation was 0.67% (m/ m) leading to a standard error
(SE) of 0.47% (m/ m) using eq 2.
NIR Analysis. The spectra were recorded on a FT-NIR MB160
spectrometer (ABB Bomem, Inc, Quebec, Canada) fitted with a
Powder Samplir reflectance accessory, a quartz-halogen lamp,
and a cooled InAs detector. The reflectance standard package was
supplied by Labsphere Inc. The software package from ABB
Bomem Inc. included Grams 32 version 4.04, PLSplus/ IQ version
3.03, AIRS (Advance Infrared Software) version 1.54, and SYSTAT
version 10, from SPSS Inc.
A “tablet holder” was constructed in order to prevent light
scattering from the beam due to the small nominal diameter (7
mm) of the tablets. The tablet holder consisted of a 3-mm-thick
piece of metal, with a 6-mm-diameter aperture in its center. Tablets
were scanned on both faces, on the basis of the feasibility results,
using the diffuse reflectance mode. Each spectrum was an average
of 64 scans coadded at 16-cm-1 resolution, over the range of
10 000-4000 cm-1. Routine analysis of the caffeine was performed
using the AIRS interface. The total procedure required to obtain
the average result of six tablets per batch took ∼12 min.
System Suitability Testing. System suitability tests, including
frequency, spectral quality, spectrophotometric noise, photometric
linearity, and precision testing, are performed on a regular basis
on the spectrometer.
RESULTS AND DISCUSSION
Feasibility Study. NIR reflectance spectra and their second-
derivative spectra were examined to identify spectral features that
could be correlated with the caffeine concentration. A diffuse
reflectance spectrum of a laboratory-made tablet containing 100%
synthetic anhydrous caffeine was examined. The main absorption
peaks in this spectrum (4128, 4299, 4431, 4673, 5194, 5836, 5980,
7300, and 8587 cm-1) were found to be very similar to those
reported by Downey and Boussion3 for dried coffee extracts and
for the spectrum of a synthetic caffeine. Vibrational assignments 17
of these bands are given in Table 1.
(17) Osborne, B. G.; Fearn, T.; Hindle, P. H. Practical NIR Spectroscopy With
Applications in Food and Beverage Analysis, 2nd ed.; Longman Scientific &
Technical: Essex, England, 1993; pp 30-33.
Analytical Chemistry, Vol. 75, No. 4, February 15, 2003 755
T a ble 1 . V ibra t iona l Assignm e nt s of Absorpt ion Ba nds for Ca ffe ine T a ble t s in t he Ra nge of 1 0 0 0 0 -4 0 0 0 c m -1

wavenumbers (cm-1) assignments

8587 C-H stretching second overtone

7300 combination band: 2C-H stretching and C-H deformation
5836 and 5980 C-H stretching, first overtone
5194 CdO stretching, second overtone
4673 combination band: dCsH stretching and CdC stretching
4431 combination band: O sH stretching and O sH deformation
4299 combination band: CsH stretching and CsH deformation
4128 combination band: CsH stretching and CsC stretching

The second-derivative spectra of laboratory-made tablets

containing 100 and 0% caffeine (i.e., excipient tablets) and of
production tablets containing 58.82% caffeine were compared
(Figure 1). This plot confirmed that the above absorption peaks
were correlated with the caffeine concentration and did not
interfere with the excipient peaks.
To choose the acquisition mode for our study, we analyzed
both faces of a single production tablet and compared the two
spectra. Six spectra per face were recorded and the tablet was
rotated between each recording. The Savitsky-Golay second
derivative was applied to the spectra. Hierarchical clustering was
used to classify the spectral differences between the two tablet
faces. Similarity between pairs of second-derivative spectra was
evaluated by the Pearson distance, which measures the strength
of the linear relationship between two spectra. The Pearson
distance is calculated by 1 - r, where r is the product-moment
coefficient of correlation as shown in eq 3.

∑x∑y
∑xy - N

x(∑ )(∑ )
r) (3)
( ∑x) ∑y)
2 2
2 2
(
x - y -
N N

where x and y are the second-derivative absorbance values from

the two spectra to be compared and N is number of paired
observations, i.e., number of data points in the spectra. The linkage
distances between clusters were measured by the average linkage-
clustering algorithm. This algorithm measures the distance
between two clusters as the average of the distances between all
the points in the clusters. The process is repeated until all the
spectra are linked in one hierarchical classification system,
represented by the dendrogram (Figure 2). As the two clusters
corresponded respectively to the spectra for the front and the back
face of the tablet, the two faces of caffeine tablets have different
spectral features. This result can be explained by considering the
transmission of force in the die cavity of a press during compres-
sion. Figure 1 . Comparison of the spectral features between PLS factor
In a rotary tablet press, the compressional force of the upper 1 loading (A) and second-derivative spectra tablets containing 100
punch is greater than the force derived from the lower punch (B), 58.82 (C), and 0% caffeine (D), over the (a) 9000-8200- and
because the initial pressure caused by the upper punch decreases (b) 6400-5600-cm-1 regions. PLS factor 1 y-values were multiplied
progressively through the powder bed, leading to variations in by a suitable constant to bring the absorbances within approximately
the same range as the other spectra, thus facilitating visual compari-
density.18 son.
The lower density zones are located at the periphery, near the
upper punch, and also at the bottom corners.19 As a consequence,
the spectral differences are most probably due to density differ-
(18) Train, D. Trans. Inst. Chem. Eng. 1 9 5 7 , 35 , 258-266. ences between the two faces of the tablet. Spectra were therefore

756 Analytical Chemistry, Vol. 75, No. 4, February 15, 2003

T a ble 2 . Ca libra t ion M ode l De sc ript ion a nd Pe rform a nc e Re sult s
NIR Calibration Model for Quantitation of Caffeine in Tablets
composition range of HPLC values no. of PLS factors
6 batches 48.2-65.2%(m/ m) caffeine
Calibration Model Performance
PRESS SEC SEP
39.94 1.08%(m/ m) 1.05%(m/ m)
Figure 2 . Dendrogram for the hierarchical analysis of second-
derivative spectra of the front (case with odd numbers) and back face
(case with even numbers) of a 58.82% caffeine tablet.
recorded on both faces of each tablet and averaged to give a
representative spectrum for the tablet.
Calibration and Design of the Validation Sets. The produc-
tion and laboratory batches were split into a calibration set, used
to develop the regression equation, and into a validation set, used
to evaluate the model performance. The calibration set (Table 2)
consisted of four laboratory batches and two production batches.
Six tablets per batch were scanned, and the average spectra for
both faces were used in the regression model. HPLC results from
duplicate determinations were assigned to each spectrum of the
calibration set. The validation set consisted of 11 laboratory-made
batches and 20 production batches, spanning the range from 0 to
100% caffeine in the tablets, and 2 laboratory batches containing
theobromine or theophylline instead of caffeine.
Data Preprocessing. The absorbance spectra were treated
mathematically by applying a Savitsky-Golay second derivative
to enhance the resolution by removing the overlapping peaks and
correcting the baseline. Mean centering was then performed to
remove any offset from the data, and standard normal variate
(SNV) correction was applied to remove the major effects of light
scattering. Wavenumber selection was performed in order to
include characteristic spectral features of caffeine identified in the
feasibility study and to exclude regions exhibiting a high noise
level (e.g. 10 000-9000 cm-1). The water absorption band around
5155 cm-1, corresponding to O -H stretching + O-H deformation,
was also excluded. The selected region contained a total of 474
data points.
(19) Aulton, M. E. In Pharmaceutics: The Science of Dosage Form Design; Aulton,
M. E., Ed.; Churchill Livingstone: Edinburgh, 1988; pp 660-661.
preprocessing options
1 second derivative, mean centering, standard
normal variate and region selection
bias outliers R2
0.0033%(m/ m) 0 0.972
Model Development and Performance. The calibration
model was developed using a PLS algorithm and constructed by
cross-validation. The number of significant PLS factors was chosen
as defined by Haaland and Thomas.20 The resulting number of
factors was one (Table 2), which could indicate an underfit model.
The quality of this model was checked by calculating the
prediction bias (average value of residuals), standard error of
calibration (SEC), and standard error of prediction (SEP) (Table
2). The F-test ( R ) 0.01) was applied to determine the statistical
significance of outliers. No outliers were found in our calibration
set. The relevancy of PLS factor 1 information was confirmed by
plotting the first PLS loading factor together with a 100%caffeine
spectrum, a production tablet spectrum, and an excipient tablet
spectrum. (Figure 1) This plot showed that the first factor was
well correlated with the caffeine concentration in the tablet
spectrum and did not interfere with the spectral features of the
excipients.
The calibration equation in the form of NIR value (%m/ m) )
y-intercept ( (std error) + slope ( (std error) × HPLC value (%
m/ m) was found to be Y ) 1.47( (1.60) + 0.97( (0.03) X. The 95%
confidence interval for the slope (0.92-1.03) included one,
suggesting that there was no evidence for a relative systematic
error in the calibration equation. As the confidence interval for
the intercept (from -1.79 to 4.73) included zero, there was
therefore no evidence to suggest a nonzero intercept.
Graphical analysis of the model residuals was performed in
order to check whether the assumptions about the model were
correct. The assumptions made about the model21 were that the
residuals are normally distributed random variables with a mean
of 0 and variance σ2, that they are independent, that they have
constant variance over the concentration range of interest, and
that the variance is independent of the concentration.
The first investigated figure (Figure 3a) was a probability plot
of the expected value for a normal distribution versus the
residuals. The points fell on a straight line, suggesting that the
data followed a normal distribution. Studentized residuals were
then plotted against the NIR values (Figure 3b) and showed that
the residuals were randomly scattered above and below the zero
horizontal of the Studentized values. Therefore, the residuals had
a constant variance and were independent of the caffeine concen-
tration.
Precision Validation. Repeatability. The repeatability was
demonstrated by performing 6 times the determination of a single
(20) Haaland, D. M.; Thomas, E. V. Anal. Chem. 1 9 8 8 , 60 , 1193-1202.
(21) Massart, D. L.; Vandeginste, B. G. M.; Deming, S. N.; Michotte, Y.; Kaufman,
L. In Chemometrics: a Textbook; Vandeginste, B. G. M., Kaufman, L., Eds.;
Elservier Science Publishing Co. Inc.: New York, 1988; pp 76-80.
Analytical Chemistry, Vol. 75, No. 4, February 15, 2003 757

Figure 3 . Residual analysis for the statistical evaluation of the
calibration equation: (a) probability plot of the expected values for a
normal distribution versus the residuals and (b) Studentized residuals
plotted against the NIR results.
tablet from two production batches: V1 and V2. Only one tablet
per batch was analyzed in order to avoid recording error due to
tablet-to-tablet manufacturing deviation. The mean and relative
standard deviations (RSDs) were 58.47 (m/ m) ( 0.74%and 60.01
(m/ m) ( 0.55% for V1 and V2, respectively. As the RSDs were
well below the usual acceptable criterion of 1%, the repeatability
was validated. In addition, paired Student t-tests were used to
compare the caffeine concentration determined from the front and
back face of the tablets. The results for the two faces were
significantly different ( P < 0.005, n ) 6, for V1 and V2), which
confirmed the feasibility results. The concentration difference
between the two faces of the tablet was 1.4 and 1.0% (m/ m)
caffeine for V1 and V2, respectively.
Intermediate Precision. The aim of this test is to establish the
effects of random events on the precision of the procedure. In
our study, the variable parameters were days and operators. This
test was performed on two batches (V2 and V3) using six tablets
per batch. An experimental design with three factors was applied
to assess the intermediate precision: factor A ) operator (3 levels
) 3 different operators), factor B ) day of analysis (3 levels ) 3
days), and factor C ) batch used (2 levels). The results obtained
by three analysts on three different days ( n ) 9) for the two
batches were as follows: mean (95%confidence interval ) 60.63
( 0.22% and 59.62 ( 0.28%, and RSD ) 0.61 and 0.48% (m/ m),
respectively. These results are well below the usual accepted RSD
of 2%.
The variability of these two parameters was examined jointly
by means of two-way analysis of the variance (ANOVA). Neither
the parameters nor the interaction produced had any effect on
758 Analytical Chemistry, Vol. 75, No. 4, February 15, 2003
the precision of the result ( P ) 0.64 and 0.90, respectively, for
the interaction effect on the results of the two batches).
Specificity Validation. The feasibility study demonstrated
(Figure 1) that the absorption peaks correlating with the caffeine
concentration did not interfere with the excipient peaks. This
result was verified by a comparison of the spectral residuals of
three production batches (V4, V5, V6) and one excipient batch
(VL1) when challenged with the method. Six tablets per batch
were analyzed. The acceptance criterion (mean spectral residual
+ 3 standard deviations) was calculated from the results of
production batches from the calibration set. One-sample t-tests
confirmed that the mean residual of the excipient batch was
significantly different ( texp ) 24.8, t ) P < 2 × 10-5, n ) 6) and
35 times higher than the acceptance criterion. The mean spectral
residuals from the production batches were within the acceptance
criterion.
The ability of the method to discriminate caffeine from
theobromine and theophylline, two alkaloids with structures
closely related to caffeine, was also evaluated. Due to manufactur-
ing problems the theophylline and theobromine tablets (VL12 and
VL13) were ∼3 times thinner and lighter than the caffeine tablets.
Nevertheless, their spectra and second derivatives (Figure 4)
showed significant differences between caffeine and theobromine
or theophylline tablets. Thus, this method is able to discriminate
caffeine from excipients and closely related compounds and the
specificity is therefore validated.
Linearity Validation. The linearity was already established
in the calibration stage, on the basis of the evaluation of the
calibration equation. The linearity was also established in the
validation step by predicting different batches over the range of
approximately 60-130% of the nominal caffeine concentration
(about 35-75% caffeine in the tablets) and by comparing the
results to HPLC reference values. To increase the number of
batches, and because of the problems in producing several
laboratory batches, the four laboratory batches from the calibration
set were added to the linearity set. Linearity was thus established
at 10 different concentrations (VL2-VL10, V7), using three tablets
per concentration level. The regression line was calculated by the
method of least squares. The equation for the NIR value ( ) Y) of
the caffeine concentration in percent (m/ m) versus the HPLC
value ( ) X in % (m/ m)) was Y ) 0.15( ( 1.32) + 0.99( ( 0.02) X.
( R2 ) 0.986, SEC ) 1.38%, SEP )1.37%, and prediction bias
-0.32%). The confidence interval for the slope (0.95-1.04) and
for the y-intercept ( -2.55-2.85) included 1 and 0, respectively.
Moreover, the t-test proved that the y-intercept did not differ
significantly from 0 ( texp ) 0.11, P ) 0.91, n ) 30), and analysis
of the variance proved that the slope also did not differ significantly
from 1 ( P ) 0.91). Figure 5 shows the correlation between the
NIR and HPLC values for both the calibration and validation
samples. The results obtained for excipient tablets (VL1) and pure
caffeine tablets (VL11) were included in the plot for comparison
purposes but not taken into account when the regression line was
calculated. The NIR results seem to become nonlinear at very
high concentrations of caffeine ( >80%) but to remain linear over
the lower range (0-30%). This phenomenon could be due to
variations in the physical properties of the tablets pressed without
excipient compared to those pressed with excipients. We can,
however, conclude that the NIR results are linear with the HPLC

Figure 4 . Comparison of the spectral features between second-
derivative spectra from laboratory-made tablets containing 58.82%
theophylline (A), production tablets containing 58.82% caffeine (B),
and laboratory-made tablets containing 58.82% theobromine (C) over
the (a) 6200-5600- and (b) 4600-4000-cm-1 regions.
results within the range of 35-75% (m/ m) caffeine.
Accuracy Validation. The accuracy of our method can be
expressed as the closeness of agreement between the HPLC and
NIR values. The ICH guideline16 recommends assessing accuracy
using a minimum of nine determinations over a minimum of three
concentration levels. However, as the results could be affected
by the physical differences of laboratory-made batches, accuracy
was evaluated on both of the linearity set samples (consisting of
laboratory-made and production samples) covering 10 concentra-
tion levels and on a set of 12 production batches (V7, V10-V20).
The accuracy was approximated by the mean recovery between
the NIR and HPLC results. Values of 99.4 and 98.9%were obtained
for the two sets of samples, respectively. A further estimate was
given by the SEP (1.37%) and the prediction bias ( -0.32%)
calculated from the linearity data set. Therefore, these estimations
Figure 5 . Correlation statistics between the caffeine content
measured by NIR and HPLC. Three tablets per batch were deter-
mined from the calibration set (+) and validation set (O) samples.
confirm that the accuracy lies within the usual accepted limits of
98-102%for the recovery percentage. In addition, a paired Student
t-test was applied to the results of the 12 production batches and
confirmed that there was no significant difference between the
NIR and the HPLC results ( texp ) -1.84, P ) 0.09, n ) 12).
The accuracy was also established because precision, linearity,
and specificity were demonstrated.
Range Validation. The range was validated because linearity,
accuracy, and precision were validated.
Quantification Limit Validation. Although the quantification
limit (QL) is not required for the validation of assays, we found
this parameter to be of interest and easy to approximate.
Therefore, it was evaluated here, according to the ICH guideline,15
as QL ) 10σ/ S, where σ is the standard deviation of the response,
estimated by the residual standard deviation of the linearity
regression line and S is the slope of the linearity regression line.
The quantification limit was therefore 13.7% (m/ m) caffeine
in the tablets, which confirms the validity of the range.
Robustness Validation. Even if the evaluation of robustness
is not required for a marketing authorization application, it remains
an interesting parameter for evaluating the method performance.
For the current study, environmental conditions (temperature,
humidity, direction of sunlight, and the presence of dust and
vibrations) are always controlled to prevent them from affecting
the results. Moffat et al.17 suggested study of the effects on the
results when changing the sample presentation. This was very
applicable in the current study because the tablet holder could
be easily moved in the beam of the spectrometer.
The robustness was therefore tested by performing six NIR
determinations of the caffeine concentration in one tablet (V8)
using the correct location of the sample holder, i.e., on the center
of the beam (test 1), and six determinations of the same tablet
with the sample holder located off-center (tests 2). One determi-
nation was the average of the determinations on two tablet faces.
The background spectrum was acquired once before each test.
The paired Student t-test showed no significant difference ( texp )
1.14, P ) 0.30, n ) 6) between the mean results of the two tests.
However, the repeatability of the results was acceptable ( <1%) in
test 1 (RSD ) 0.78%) but out of specifications (RSD ) 1.1%) in
test 2. Therefore, the repeatability of the results was affected by
the position of the sample holder, confirming that the sample
holder has to be fixed precisely in the center of the beam.
Analytical Chemistry, Vol. 75, No. 4, February 15, 2003 759
 Robustness was evaluated on a second parameter: replacement suggest that statistical analysis of the model residuals should
of the NIR source. Two batches (V3, V9) were measured both always be included in NIR quantitative method validation.
before and after replacement of the NIR source (approximately
one-year delay between the measurements). The paired Student ACKNOWLEDGMENT
t-test proved that the results were not significantly different before The authors thank Niina Laihanen and Maria Lindblad (Pharmia
and after the lamp replacement ( texp ) - 0.35, P ) 0.73, n ) 6 Oy) for their technical assistance during production of the
and texp ) -1.15, P ) 0.19, n ) 6, respectively). laboratory batches. We are also grateful to Erik Michiels (ABB
Bomem) for his advice during method development.
CONCLUSIONS
A rapid and precise near-infrared diffuse reflectance method
was developed and validated for the determination of the caffeine
Received for review October 28, 2002. Accepted
concentration of intact single tablets. The performance of the December 16, 2002.
method was equal to that of the reference HPLC method and was
furthermore much faster and easier to carry out. Finally, we AC026262W

760 Analytical Chemistry, Vol. 75, No. 4, February 15, 2003