Introduction

Propranolol (PRO), amiloride (AMI) and dipyridamole (DIP)

are widely used to treat several diseases. PRO and AMI
1
are
antihypertensive substances, whereas DIP
1
is an antithrombotic
agent. These components are classified as doping substances
and have been included in a list of forbidden substances
2
by the
International Olympic Committee.
The sensitivity of fluorescence spectrometry has enabled it to
be a useful analytical tool for monitoring trace compounds of
biochemical interest. However, when substances whose
individual spectral profiles contain broad bands, which often
overlap, must be determined simultaneously by
spectrofluorometry, it would be difficult to select a pair of
excitation and emission wavelengths which could permit the
determination of one of them without any interference of the
other. For such a reason, it is worth investigating new methods
to solve this problem without resorting to expensive or time-
consuming techniques.
Several approaches, such as synchronous,
3
derivative
4
fluorescence spectrometry as well as variable-angle
fluorescence
5
spectroscopy have been developed to overcome
such problems without having to utilize time-consuming and
expensive procedures. All of these methods are based on
improving the selectivity of the coexisting components.
Chemometric methods, such as factor analysis,
610
have
become even more popular in solving problems that are difficult
to handle using conventional techniques, without having to
resort to expensive or time-consuming procedures. Especially,
three-way resolution and calibration methods
1113
play important
roles in solving the problem of closely overlapping fluorescence
spectra. These methods utilize mathematical separation
procedure to substitute the traditional chemical separation
procedure, which is different from those.
35
In addition, these
approaches can not only determine the concentrations, but can
also provide spectral profiles of the components in the mixtures.
In this work, with both purposes, the present authors resolved
mixtures of the aforementioned three doping substances with
broad and highly overlapping excitation and emission spectra,
using a newly proposed modified PARAFAC algorithm with a
penalty digonalization error (PDE).
14
One purpose is to solve
the spectral overlapping problem and to further extend the
applications of chemometric methods; another is to further
demonstrate the characteristic features of the PDE algorithm.
Theory
Before using the PDE algorithm, a simple introduction to it
should be presented. For details, the readers are recommended
to Ref. 14.
Trilinear model
According to three-dimentional fluorometry, the trilinear
model can be written as
X..k = ckf af bf
t
+ E..k k = 1, 2, 3, , K, (1)
where F is the component number and K is the number of
samples (mixtures); af, bf and cf are the emission, excitation and
concentration profiles of the f th component. In matrix notation,
the trilinear model can also be equivalent to the following three
F

f=1
333 ANALYTICAL SCIENCES MARCH 2002, VOL. 18
2002 The Japan Society for Analytical Chemistry
Spectrofluorometric Resolution of Closely Overlapping Drug
Mixtures Using Chemometrics Methods
Yu-Zhen CAO,*
,
**

Cui-Yun MO,*** Ji-gong LONG,** Hong CHEN,* Hai-Long WU,*** and

Ru-Qin YU***
*Key Laboratory of New Packaging Materials & Technology of China National Packaging Corporation,
Zhuzhou Institute of Technology, Zhuzhou, 412008, China
**Environmental Monitoring Central Station of Guangzhou City, NO. 95, Jixiang Road of Guangzhou City,
Guangzhou, 510030, China
***College of Chemistry and Chemical Engineering, Hunan University, Changsha, 410082, China
A modified parallel factors analysis (PARAFAC) algorithm with a penalty diagonalization error (PDE), newly proposed
by the present authors, was utilized to simultaneously resolve drug mixtures of propranolol (PRO), dipyridamole (DIP)
and amiloride (AMI) without any loss of sensitivity. The analyses were performed in aqueous solution. The
experimental results demonstrated that the profiles of the spectra and the concentrations could be accurately resolved
using the PDE algorithm with a high sensitivity and stable repeatability. That is to say, the closely overlapping problem
of the spectra could be easily solved. Furthermore, simultaneous determinations of three kinds of tablets, which contain
PRO, AMI and DIP, respectively, were successfully performed with satisfactory results.
(Received July 9, 2001; Accepted November 16, 2001)

To whom correspondence should be addressed.

E-mail: yuzhencao@yeah.net
equations:
Xi.. = Bdiag(ai)C
t
+ Ei.. i = 1, 2, , I, (2)
X.j. = Cdiag(bj)A
t
+ E.j. j = 1, 2, , J, (3)
X..k = Adiag(ck)B
t
+ E..k k = 1, 2, , K, (4)
where I and J are the emission wavelength number and the
excitation wavelength number, respectively.
The algorithm
From Eqs. (2) (4), one can obtain the following three
equations for the diagonalization error:
B
+
Xi..(C
+
)
t
diag(ai) = B
+
Ei..(C
+
)
t
, (5)
C
+
X.j.(A
+
)
t
diag(bj) = C
+
E.j.(A
+
)
t
, (6)
A
+
X..k(B
+
)
t
diag(ck) = A
+
E..k(B
+
)
t
. (7)
By combining the diagonalization error with the PARAFAC
error, one can obtain three objective functions, as follows:
F(A) = ||Xi.. Bdiag(ai)C
t
||
F
2
+||B
+
Xi..(C
+
)
t
diag(ai)||
F
2
, (8)
F(B) = ||X.j. Cdiag(bj)A
t
||
F
2
+||C
+
X.j.(A
+
)
t
diag(bj)||
F
2
, (9)
F(C) = ||X..k Adiag(ck)B
t
||
F
2
+||A
+
X..k(B
+
)
t
diag(ck)||
F
2
. (10)
From Eqs. (8) (10), one can acquire the following equations to
express the factor matrices A, B and C:
pi = diag[B
+
Xi..(C
+
)
t
], i = 1, 2, 3, , I;
P = [p1; p2; ;pI]; (11)
A= ( X..kBdiag(ck) + P)(I + diag(ck)B
t
Bdiag(ck))
1
; (12)
qj = diag[C
+
X.j.(A
+
)
t
], j = 1, 2, 3, , J;
Q = [q1; q2; ;qJ]; (13)
B= ( Xi..Cdiag(ai) + Q)(I + diag(ai)C
t
Cdiag(ai))
1
; (14)
rk = diag[A
+
X..k(B
+
)
t
], k = 1, 2, 3, , K;
R = [r1; r2; ;rK]; (15)
C= ( X.j.Adiag(bj) + R)(I + diag(bj)A
t
Adiag(bj))
1
. (16)
The PDE algorithm computes the loading matrix A using Eqs.
(11) and (12) with B and C fixed; then, it fixes C and renews A
to update B utilizing Eqs. (13) and (14); it finally calculates
matrix C with the renewed A and the updated B according to
Eqs. (15) and (16).
Experimental
Apparatus
All of the fluorometric measurements were performed on an
F-4500 fluorescence spectrophotometer (HITACHI).

k=1

j=1

i=1
Reagents
All of the experiments were performed with analytical
reagent-grade chemicals, pure solvent and doubly distilled
water. Stock solutions of PRO (The Fourth Pharmaceutical
Factory of Changzhou; 25 mg dissolved in 100 ml of water),
AMI (The Fourth Pharmaceutical Factory of Changzhou; 25 mg
dissolved in 100 ml of water) and DIP (The Fourth
Pharmaceutical Factory of Changzhou; 25 mg dissolved in 100
ml of ethanol) were diluted to prepare working solutions by
suitable dilutions using water. The stock solutions were stored
and protected from light. Three kinds of tablets (Amiloride,
Propranolol hydrochloride tablets and Persantin) containing
AMI, PRO and DIP, respectively, with nominal contents, were
randomly purchased from local pharmacies.
Procedure
The steady state fluorescence of 9 samples of the
aforementioned three species was measured in aqueous solution.
The concentrations of these species are listed in Table 1. The
excitation wavelength was set from 220 nm to 320 nm at an
interval of 4 nm, and the emission wavelength varied from 330
nm to 522 nm with an interval of 4 nm. The scan rate was 1200
nm/min. The effect of Rayleigh scattering was compensated by
subtracting the measurement matrix of a blank from the sample
measurements. Thus, a 49 26 9 data array was collected.
This data array was treated using the PDE algorithm.
In order to analyze three tablets containing PRO, AMI and
DIP respectively, the authors combined the mixture of the three
tablets with nine samples. The same treatment was made to
these samples and the mixture. Thus a 49 26 10 data array
was collected. The data array was resolved using the PDE
algorithm.
Results and Discussion
Linear range
The linear ranges of PRO, AMI and DIP were investigated
before undertaking an experimental design. When PRO, AMI
and DIP were, respectively, in the range of 9 100 ng/ml, 10
150 ng/ml and 20 110 ng/ml, the correlation coefficients were
all above 0.9990.
Resolution of profiles in three modes
Figure 1 shows the resolved profiles using the PDE approach
334 ANALYTICAL SCIENCES MARCH 2002, VOL. 18
Table 1 Resolved concentrations with N = 3 and real
concentrations of 9 samples (g/ml)
a. MAE is defined as the mean absolute error.
Sample
1 0.0000 0.0387 0.0602 0.0005 0.0389 0.0638
2 0.1044 0.0581 0.0602 0.1054 0.0553 0.0683
3 0.0835 0.0387 0.0803 0.0863 0.0397 0.0708
4 0.0626 0.0387 0.1004 0.0656 0.0416 0.0980
5 0.0835 0.0581 0.0602 0.0826 0.0549 0.0707
6 0.0626 0.0581 0.0803 0.0617 0.0524 0.0961
7 0.0835 0.0387 0.0000 0.0796 0.0355 0.0018
8 0.0626 0.0581 0.0000 0.0624 0.0530 0.0016
9 0.0835 0.0000 0.0803 0.0819 0.0015 0.0652
MAE
a
0.0014 0.0025 0.0061
Real Resolved
AMI PRO DIP AMI PRO DIP
with N = 3 and the actual ones of the PRO, AMI and DIP. The
concentration profiles of 9 samples are listed in Table 1. These
results are nearly the same as those of the true ones, which
illustrates that the PDE algorithm will provide an accurate
resolution of the substances that have seriously overlapping
spectral profiles when the correct component number is used in
the computation. The required iteration number for analyzing
this data array was merely 18, and the required time was only 3
s. These phenomena showed that the PDE algorithm was able
to recover the profiles of the data array accurately and rapidly,
which are characteristic features of the PDE algorithm.
14
Figure 2 shows the resolved spectra of the data array using the
PDE algorithm with N = 4, in which an excess component was
used for the unknown background and the true profiles were
plotted for a comparison. The concentration profiles of the
three components of interest are listed in Table 2. These
recovered results further illustrated that the PDE algorithm is
insensitive for estimating the component number. That is to
say, the PDE algorithm can provide accurate solutions provided
that the number of factors used in a calculation is not less than
that of the actual underlying factors, which is one of the
characteristic features of the PDE approach.
14
Simultaneous determination of contents of three kinds of tablets
Since the PDE algorithm was insensitive to estimating the
component number, the present authors resolved a 49 26 10
data array using the PDE algorithm with N = 4, in which an
unknown background component was taken into consideration.
With the concentrations of the nine samples (mixtures) known,
the contents of three components were obtained using a second-
order calibration. The experiment was performed in parallel
four times; the calculated concentrations are given in Table 3.
These results demonstrate that the PDE algorithm can be used to
simultaneously determine practical samples of PRO, AMI and
DIP with an unknown background.
Conclusion
In this paper, the newly proposed modified PARAFAC
algorithm using a penalty diagonalization error (PDE) was used
to resolve the system of propranolol (PRO), amiloride (AMI)
and dipyridamole (DIP). The resolved results showed that the
PDE approach could rapidly solve the problem of serious
fluorescence spectral overlapping of the three components with
an unknown background, and were insensitive to an estimation
of the component number.
Acknowledgements
This work was financially supported by Natural Science
335 ANALYTICAL SCIENCES MARCH 2002, VOL. 18
Fig. 1 Resolved spectra (solid lines) and true spectra (dotted lines)
using the PDE algorithm with N = 3. (a) Excitation spectra and (b)
emission spectra. Fig. 2 Recovered profiles (solid lines) and actual ones (dotted
lines) in three modes utilizing the PDE algorithm with N = 4. (a)
Excitation spectra and (b) emission spectra.
Foundation of Hunan (Grant No. SSY2066) and the Key
Laboratory of New Packaging Materials and Technology of
China National Packaging Corporation.
References
1. A. C. Moffat, Clarkes Isolation and Identification of
Drugs, 2nd ed., 1986, The Pharmaceutical Press, London,
362, 363, 562, 563, 1037 and 1038.
2. C. Rodrguez Burno, Dopaje, 1st ed., 1992, Mc-Graw
Hill, Madrid.
3. T. Vo-Dinh, Anal. Chem., 1978, 50, 396.
4. P. John and I. Soutar, Anal. Chem., 1976, 48, 520.
5. J. A. Murillo Pulgarn, A. Alann Molina, and P. F. Lpez,
Anal. Chim. Acta, 1998, 370, 9.
6. T. Adzuhata, J. Inotsume, T. Okamura, R. Kikuchi, T.
Ozeki, M. Kajikawa, and N. Ogawa, Anal. Sci., 2001, 17,
71.
7. N. Yoshimura, M. Okazaki, and N. Nakagawa, Anal. Sci.,
2000, 16, 1331.
8. I. S. Scarminio, W. J. Barreto, D. N. Ishikawa, E. L.
Paczkowski, and I. C. Arruda, Anal. Sci., 2000, 16, 1337.
9. E. Forgacs and T. Cserhati, Anal. Sci., 1998, 14, 991.
10. L. F. Capitan-Vallvey, N. Navas, R. Avidad, I. Deorbe, and
J. J. Berzas-Navado, Anal. Sci., 1997, 13, 493.
11. H.-L. Wu, Y.-Q. Feng, M. Shibukawa, and K. Oguma,
Anal. Sci., 1997, 13, 99.
12. H.-L. Wu, M. Shibukawa, and K. Oguma, Anal. Sci., 1997,
13, 53.
13. H.-L. Wu, K. Oguma, and R.-Q. Yu, Anal. Sci., 1994, 10,
875.
14. Y.-Z. Cao, Z.-P. Chen, C.-Y. Mo, H.-L. Wu, and R.-Q. Yu,
Analyst, 2000, 125, 2303.
336 ANALYTICAL SCIENCES MARCH 2002, VOL. 18
Table 2 Recovered concentrations using the PDE algorithm
with N = 4 (g/ml)
a. MAE is defined as the mean absolute error.
Sample
1 0.0000 0.0387 0.0602 0.0010 0.0389 0.0637
2 0.1044 0.0581 0.0602 0.1055 0.0554 0.0682
3 0.0835 0.0387 0.0803 0.0859 0.0397 0.0707
4 0.0626 0.0387 0.1004 0.0656 0.0416 0.0978
5 0.0835 0.0581 0.0602 0.0828 0.0550 0.0708
6 0.0626 0.0581 0.0803 0.0623 0.0524 0.0961
7 0.0835 0.0387 0.0000 0.0798 0.0355 0.0017
8 0.0626 0.0581 0.0000 0.0623 0.0532 0.0016
9 0.0835 0.0000 0.0803 0.0821 0.0015 0.0653
MAE
a
0.0014 0.0025 0.0061
Real Resolved
AMI PRO DIP AMI PRO DIP
Table 3 Calculated contents of three drugs for four parallel
experiments (mg)
Nominal contents 5 10 25
First time 5.15 9.90 25.20
Second time 5.12 9.89 25.34
Third time 5.09 10.18 25.27
Four time 4.97 10.20 25.29
Mean values 5.08 10.04 25.28
Amiloride (AMI) Propranolol (PRO) Persantin (DIP)