Talanta 71 (2007) 806815

Different strategies for the direct determination of amoxicillin in human urine by second-order multivariate analysis of kineticspectrophotometric data
Alejandro Garca-Reiriz, Patricia C. Damiani, Alejandro C. Olivieri
Departamento de Qumica Analtica, Facultad de Ciencias Bioqumicas y Farmac uticas, e Universidad Nacional de Rosario, Suipacha 531, Rosario S2002LRK, Argentina Received 13 March 2006; received in revised form 16 May 2006; accepted 17 May 2006 Available online 30 June 2006

Abstract The kinetic evolution of UVvisible absorption spectra of amoxicillin in the presence of copper(II) ions has been processed by the second-order multivariate methods parallel factor analysis (PARAFAC) and also by a novel approach based on partial least-squares with residual bilinearization (PLS/RBL). The latter one is employed for the rst time to evaluate kineticspectral information. The mechanism of the analyte metal-catalyzed hydrolysis involves a reaction intermediate and a nal reaction product, both with spectra which may allow for the determination of amoxicillin in human urine, even in the presence of unsuspected sample components. This is possible thanks to the second-order advantage exploited by the employed chemometric algorithms, among which PARAFAC and PLS/RBL gave the best results. Amoxicillin was determined in a series of spiked and real urine samples, which allowed to perform, respectively, a recovery study and a comparison with the reference high-performance liquid chromatographic technique. The best gures of merit were obtained with PLS/RBL, namely sensitivity, 0.5 AU L mg1 (AU = absorbance units), analytical sensitivity, 500 L mg1 and limit of detection, 6 mg L1 . Relative advantages and disadvantages of the employed algorithms are discussed. 2006 Elsevier B.V. All rights reserved.
Keywords: Second-order multivariate calibration; Parallel factor analysis; Multivariate curve resolution; Partial least-squares with residual bilinearization; Amoxicillin; Human urine

1. Introduction Amoxicillin, 7-[2-amino-2-(4-hydroxyphenyl)-acetyl]amino-3,3-dimethyl-6-oxo-2- thia - 5 - azabicyclo[3.2.0] heptane-4carboxylic acid, is a -lactam antibiotic used in humans and food-producing animals to treat several diseases. In human urine, amoxicillin is excreted in a ratio of 86 8%, its urine concentration being 1.7 7 g L1 , the rst 2 h after a single oral dose of 500 mg. Its peak plasma concentration is 46 10 mg L1 for intravenous, and 913 mg L1 for oral administration respectively [1,2]. Although serious problems have rarely been reported, the determination in biological uids is very important. This is critical in stomach, kidney or intestinal

Corresponding author. Tel.: +54 341 4372704; fax: +54 341 4372704. E-mail addresses: aolivier@fbioyf.unr.edu.ar, acolivieri@hotmail.com (A.C. Olivieri). 0039-9140/$ see front matter 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.talanta.2006.05.050

diseases, where drug monitoring must be carried out during treatment in order to perform a suitable dosage adjustment. Several analytical methods have been reported for the determination of amoxicillin in biological uids, mostly based on high-performance liquid chromatography (HPLC) [310]. Although HPLC assays are selective and sensitive, they have some disadvantages, requiring expensive equipment and toxic solvents, and often involving complex sample pre-treatments. For these latter reasons, spectroscopic methodologies suitable for routine laboratories are welcomed. However, the effect of the potentially interfering biological background should be taken into account if valid predictions are to be made with methods involving no sample pre-treatment steps. One way this goal can be accomplished is with the aid of multivariate calibration techniques. First-order methods can cope with varying amounts of potential interferences (if they are well represented in the calibration phase). In addition, they enable a sample containing unsuspected components to be diagnosed as an outlier, because

A. Garca-Reiriz et al. / Talanta 71 (2007) 806815

807

Fig. 1. Simplied reaction scheme for the copper-catalyzed hydrolysis of amoxicillin.

its spectrum will t signicantly worse to the model than the calibration spectra. These properties have been pertinently referred to as the rst-order advantage [11]. However, it should be noticed that although outliers can be detected, predictions cannot be corrected for the bias caused by the unsuspected sample components. Only second- (and higher-) order data, coupled to appropriate second-order multivariate calibration methods, are able to model unsuspected sample constituents, a property known as the second-order advantage [11]. In the present work, a kineticspectrophotometric method assisted by chemometrics is developed for the determination of amoxicillin in human urine. Amoxicillin, like other -lactam antibiotics, is hydrolyzed in aqueous solution through the catalysis of metal ions [1216]. This metal-catalyzed hydrolysis seems to have a complex chemistry, and controversial pathways have been proposed, involving the catalyzed effect of the metal ion as well as complex formation. For ampicillin, for example, the overall reaction consists of the opening of the -lactam ring, followed by reorganization to produce a pyrazyne derivative, which then undergoes copper(I) complex formation [12]. We have conjectured that a similar mechanism could be applied to amoxicillin: a fast reaction producing an intermediate with a structure likely to have a red-shifted spectrum with respect to the analyte, followed by a slow reaction to yield a nal product, with a distinct spectrum due to complex formation with copper ions (see Fig. 1). The overall changes taking place during the metal-catalyzed hydrolysis seem to adapt to the above described mechanism, as well as being favorable for a multivariate kineticspectrophotometric analysis, since the reaction products have spectra which: (1) differ from the background urine signal, (2) appear in a more useful region than the analyte absorption and (3) display a higher sensitivity in comparison with unreacted amoxicillin. The complete spectral-time evolution for the coppercatalyzed hydrolysis of amoxicillin was measured for each sample, and then analyzed by two different algorithms. One of them is parallel factor analysis (PARAFAC) [17], the current standard for this type of studies, even when the presence of linear

dependency in the time dimension poses some challenges to this algorithm. We also applied partial least-squares coupled to residual bilinearization (PLS/RBL) [18], a newly introduced method with all the exibility of the latent variable models [19], yet providing the second-order advantage thanks to the useful RBL procedure. The potentiality of PLS/RBL for kineticspectral data remains to be checked, which explains the interest in the present work in comparing its prediction ability with those of more classical approaches. It should be noticed that multivariate curve resolution-alternating least-squares (MCR-ALS) [20], which has been especially developed for spectroscopickinetic data and other systems showing linear dependency, did not provide useful analytical results in the presently discussed case.

Fig. 2. Two basic modes of obtaining the second-order advantage from higherorder information. (A) Combining data from calibration and test sample before computing the regression coefcients. (B) Estimating loadings from calibration data only, and then calculating regression coefcients after the test sample enters the scene.

808

A. Garca-Reiriz et al. / Talanta 71 (2007) 806815 Table 1 Central composite design for response surface optimization pH Cu2+ ( 102 mol L1 ) Absorbance Total A at 325 nm 0.448 1.135 0.614 1.043 0.405 0.963 0.635 1.183 0.918 0.905 0.889 Buffer A at 290 nm 0.118 0.117 0.066 0.151 0.074 0.094 0.144 0.142 0.120 0.121 0.123 Net A at 325 nm 0.330 1.017 0.545 0.891 0.331 0.868 0.491 1.041 0.798 0.785 0.768

The latter calibration methods are capable of achieving the second-order advantage in either of the two modes schematically shown in Fig. 2: in PARAFAC, data for the unknown sample determine (together with calibration data) the regression coefcients leading to prediction (Fig. 2A), while in PLS/RBL calibration is rst performed using only calibration data, with the unknown sample aiding in the obtainment of sample-specic regression coefcients in a subsequent step (Fig. 2B). We discuss the results of recovery studies for spiked urine samples, comparison with a reference HPLC technique, as well as the relative advantages and disadvantages of the applied chemometric models, with particular emphasis in analytical performance and ease of program operation by unskilled analysts. 2. Experimental 2.1. Apparatus Absorbance measurements were done on a Jasco V-530 (Tokyo, Japan) UVvisible spectrophotometer, using 1.00 cm quartz cells. Parameters were: spectral range, 322370 nm, scanning step, 2 nm, scanning rate, 2000 nm/min, number of spectra, 12, interval between spectra, 90 s. The size of the absorbancetime data matrices was 25 12 = 300 data points per sample (25 digitized wavelengths and 12 spectra per sample). The total time to record a spectrum under these conditions was ca. 1 s, i.e., considerably smaller than the total measuring time (16.5 min). The cell has been thermostated at 25 C. Increasing the temperature causes an increase in the reaction rate, but may introduce complications in the data treatment if the reaction progresses during the time each spectrum is acquired. HPLC measurements were done with a Waters (Milford, Massachusetts USA) liquid chromatograph equipped with a 515 Waters high pressure pump operating at 1.0 mL min1 , a Rheodyne injector with a 20 L sample loop and a variable wavelength UV visible detector measuring at 230 nm, a 5 m Zorbax SB C18 column (15 cm 4.6 mm). All chromatograms were run at room temperature using a 95:5 mixture of aqueous KH2 PO4 (pH = 5.0) and acetonitrile. 2.2. Reagents and solutions The following reagents and solutions were used for the spectrophotometric method: sodium acetate (Merck), HCl (Merck), amoxicillin standard solution (LLCM, 1.007 g L1 ), copper(II) sulfate (Merck), sodium acetate/acetic acid 1 mol L1 buffers with different pH values: 3.7, 4.0, 4.7, 5.4 and 5.7. Stock solutions of amoxicillin were prepared by dissolving an appropriate amount of amoxicillin standard in doubly distilled water and sonicating during 15 min. Amoxicillin solutions were prepared every two weeks and stored in a refrigerator at 4 C, while CuSO4 0.210 mol L1 was prepared by dissolving suitable amounts of the reagent in doubly distilled water. 2.3. Optimization of the kinetic parameters In order to select the optimum reaction time for the catalyzed reaction, the absorbance was registered as a function of time at

3.7 5.7 4.7 4.7 4.0 5.4 4.0 5.4 4.7 4.7 4.7

1.58 1.58 0.83 2.31 1.05 1.05 2.10 2.10 1.58 1.58 1.58

a single wavelength. For optimizing the pH value and the Cu2+ concentration, a surface response methodology was employed. Both factors were varied according to a central composite design, with three replicates at the central point, and thus 11 samples were required. The factors were in the following ranges: pH 3.75.7 and Cu 2+ : 8.3 103 to 2.31 102 mol L1 . Table 1 shows the specic values of both factors in all experiments, along with the absorbances recorded for each solution at 325 nm (where product II, furnished by degradation of the analyte in the presence of copper ions, displays a spectral maximum), for a buffer solution not containing the analyte at 290 nm (where the buffer signal is maximum), and the corresponding net analyte absorbance at 325 nm (i.e., the absorbance difference between a solution of the analyte and the buffer, both recorded at 325 nm). 2.4. Kinetic experiments In 2.00 mL volumetric asks 1.00 mL of acetate/acetic acid buffer (pH = 5.7) was placed, adding suitable volumes of amoxicillin stock solution, and nally 0.150 mL of 0.210 mol L1 CuSO4 . The metal-catalyzed hydrolysis was spectrophotometrically monitored measuring change in absorbance during the selected time. 2.5. Calibration samples A calibration set having six equally spaced concentration levels of amoxicillin in the range from 0.00 to 0.190 g L1 was prepared. All calibration solutions were prepared in duplicate. This calibration set is similar to those employed in univariate calibration of a single analyte, since the present report is focused on the achievement of the second-order advantage, which transforms the multivariate problem into a pseudo-univariate calibration case. The concentrations were selected considering the linear calibration range, as well as the usual concentration values of amoxicillin in urine. All calibration samples were prepared by triplicate and measured in random order under the optimal conditions. Kineticspectral data matrices were processed applying the second-order multivariate calibration methods PARAFAC and PLS/RBL.

A. Garca-Reiriz et al. / Talanta 71 (2007) 806815

809

2.6. Test samples Five articial samples were prepared starting from suitable volumes of the stock solution of amoxicillin, by selecting concentrations at random order from the calibration range. Each articial sample was measured by triplicate in random order under the optimal conditions. Kineticspectral data matrices were processed as describe above. 2.7. Urine samples Both spiked urine and real urine samples were analyzed. Samples were only diluted (1:10) taking into account the linear calibration range for the analyte. Spectral measurements were done as described above and analyzed using the chemometric methodologies, in this case in a manner which exploits the second-order advantage. 2.8. Chromatographic procedure For the chromatographic method, working solutions of amoxicillin were prepared by dilution of the standard solution described above, in order to assess the linear range, which was attained from 0.00 to 0.30 g L1 . Real urine samples were diluted taking into account the linear calibration range as well as therapeutic urine levels. They were then ltered through a 0.5 m membrane lter, and 20 L were injected with detection proceeding at 230 nm. The average retention time was 170 10 s. 3. Theory 3.1. Parafac A set of data matrices (those for calibration and for each of the unknown samples) are stacked to yield a three-dimensional array of size (I + 1) J K (I is the number of calibration samples, J the number of absorption wavelengths and K the number of time values). PARAFAC ts these data according to a trilinear model [21], yielding the score or relative concentration [(I + 1) 1], absorption (J 1) and time (K 1) proles for component n in all mixtures (N is the total number of components). Notice that this PARAFAC operation mode implies that it achieves the second-order advantage in the manner depicted in Fig. 2A. The number of components to be modeled can be estimated by different methods [22,23]. In the present case, information concerning the properties of the system is available, and thus for pure analyte standards N was set to two, on account of the presence of two responsive species in the hydrolysis of amoxicillin. For real urine samples N is likely to be larger, since a background arising from the biological matrix is also present, consisting of the superposition of spectrally responsive components, which may also be time-dependent. However, satisfactory results were obtained by including a single unsuspected component, which was considered as an average background effect. In any case, the optimal choice of N was guided by the quality of the nal least-squares t.

PARAFAC requires to be supplied with initial values of the scores and loadings. This is usually done by resorting to a direct trilinear decomposition (DTLD) of the three-way array. However, this may not be feasible for spectral-time data, which are linearly dependent in the time dimension, because reagents and products of chemical reaction are bound by closure relationships among their concentrations. PARAFAC is not ideal for handling data which are not strictly trilinear in the latter sense [24], but the array which would obtained on stacking spectraltime matrices for a single species is indeed trilinear, and thus PARAFAC should give acceptable results provided suitable initial conditions and restrictions are set, and the model is initialized with proper guesses for scores and loadings. The spectral loadings can initially be set as equal to spectra of the analyte at time zero at a time of reaction completion, and, for real samples, including a spectrum of a pool of different urines (all normalized to unit length). The time proles are normalized linear functions for each of the analyte species (one decreasing and one increasing), and a constant for the background urine in real sample applications. Finally, an approximation to the scores is provided in terms of: (1) the known calibration concentrations of the analyte, (2) the mean calibration concentration for the analyte in the unknown, and (3) a single value for the background urine, obtained after normalizing its spectrum. Once the chemical constituent under investigation is identied with the aid of the recovered proles, absolute analyte concentrations are obtained after calibration using the nominal analyte concentrations. In the presently studied case, the analyte of interest has two kinetically related species, and hence two separate pseudo-univariate graphs are expected, one for each of these species. PARAFAC will therefore provide two concentration values, which should ideally be equal, obtained from the score for the unknown sample. For further details on the application of restricted PARAFAC to kineticspectrophotometric data, see Refs. [2529]. 3.2. PLS/BRL In this method, concentration information is rst employed during the calibration step, without including data for the unknown sample, leaving the second-order advantage for a later stage, i.e., in the mode shown in Fig. 2B [30]. The I calibration data matrices are rst vectorized into JK 1 vectors, and then a usual PLS model is built using these data together with the vector of calibration concentrations. The number of PLS factors can be selected by techniques such as leave-one-out cross-validation [31]. Notice that PLS is a latent variable method, and hence no prior information as to the spectral or time evolution of the analyte are in principle required for its successful operation. When unsuspected constituents occur in the unknown data matrices, a separate procedure called residual bilinearization is applied, which has already been described in the literature, and is based on the singular value decomposition (SVD) to model the unsuspected effects [18]. The PLS/RBL has not been applied, to the best of our knowledge, to resolve kineticspectral data systems, except for a small simulation study [19].

810

A. Garca-Reiriz et al. / Talanta 71 (2007) 806815

3.3. Figures of merit Figures of merit are regularly employed for method comparison. They are best understood by resorting to the useful concept of net analyte signal (NAS), rst developed by Lorber [32], although there appears to be conicting reports as how the NAS should be computed for higher order methodologies [33,34]. In the case of PARAFAC, an equation which seems to be appropriate for cases where the second-order advantage operates is [35]: SENn = sn {[(BT Pb,uns Bsus ) (CT Pc,uns Csus )] sus sus
1 1/2 }nn

where P is the (JK A) loading matrix provided by the unfolded PLS model, Pc,uns and Pb,uns have the same meaning as above, and implies the Kronecker product. Finally, the limit of detection can be estimated by Eq. (4), as with PARAFAC. 3.4. Software All routines employed to carry out the calculations described in this paper were written in MATLAB [39]. Those for applying PARAFAC are available in the internet thanks to Bro [40]. PLS/RBL is available from the authors on request, including a useful graphical user interface for data input and parameter setting, of the type already described for rst-order multivariate calibration [41], and which also works for PARAFAC. 4. Results and discussion 4.1. Spectral characteristics The absorption spectra for a standard solution of amoxicillin and a typical urine sample are shown in Fig. 3A in a rather wide spectral range. Not only a severe overlapping can be noticed in the latter wavelength range, but the low-intensity analyte peak lies in a region where high absorbance is recorded for the

(1)

where SENn is the sensitivity for component n, sn is the integrated total signal for component n at unit concentration, Bsus and Csus are the matrices containing the proles for all suspected components (i.e., those present in the training set of samples) in each data dimension, nn implies selecting the (n,n) element corresponding to the nth. analyte of interest, * implies the Hadamard matrix product, and the projection matrices Pb,uns and Pb,uns are given by: Pb,uns = I Buns B+ uns Pc,uns = I Cuns C+ uns (2) (3)

where Buns and Cuns contain the proles for the unsuspected components as columns. Notice that when the second-order advantage is employed, Eq. (1) implies that SENn is sample-specic and cannot be dened for the multivariate method as a whole. In such cases an average value for a set of samples can be estimated and reported [36]. More useful than SENn appears to be the analytical sensitivity n , dened, in analogy with univariate calibration, as the quotient between SENn and the instrumental noise level. Its inverse establishes the minimum concentration difference which can be appreciated across the linear range, and is independent on instrument or scale [37]. The limit of detection (LODn ) should also be considered. An approximation to the LODn can be gathered from the expression: LODn = 3sX /SENn (4)

where sX is the instrumental noise level. Eq. (4) stems from the consideration of a signal to noise ratio equal to 3, but does not account for calibration uncertainties, and thus it generally provides overoptimistic values. Inasmuch as SENn is given as an average value over a test sample set, LODn is also reported as an average gure. In the case of PLS/RBL, the estimation of the sensitivity has been recently discussed [38]. The appropriate expression is:
+ SENn = 1/ (Peff ) v T

(5)

where v is the (A 1) latent vector of regression coefcients for the PLS model, and Peff is a matrix given by: Peff = (Pc,uns Pb,uns )T P (6)

Fig. 3. (A) Absorption spectra of amoxicillin (0.079 g L1 , solid line) and a typical urine blank sample (diluted 1:10, dashed line) in a wide spectral range. (B) Absorption spectra at a reaction time of 16.5 min for an aqueous calibration solution of amoxicillin 0.103 g L1 (solid line), for a typical urine blank (dashed line), and for a urine sample spiked with 0.103 g L1 of the analyte (dotted line). The spectra for the restricted working spectral range are shown in the insert.

A. Garca-Reiriz et al. / Talanta 71 (2007) 806815

811

time scale of the present experiments. Only the transformation of product I into product II is monitored by the spectral-time evolution shown in Fig. 4A. Fig. 4B shows the corresponding time evolution of the spectra for a typical urine blank sample, i.e., not containing the analyte. As can be appreciated in this latter gure, there is a small spectral drift associated with the time axis in Fig. 4B, which could be ascribed to copper complex formation with urine components. 4.2. Optimization of kinetic variables The UVvisible spectra of amoxicillin in the presence of Cu(II), acetic acid/acetate buffer (pH ca. 5.7) presents an absorption maximum at about 330 nm. The time evolution of the absorbance at this particular wavelength showed an increase with time, while the absorbance of urine under the same conditions remained almost constant. A total time of ca. 20 min seemed to be appropriate for kinetic studies, rendering a reasonable absorbance value which then increased very slowly. Once the optimum measuring time was established, the chemical variables pH and Cu(II) concentration were optimized using a surface response methodology (Table 1 shows the experimental design). When the total absorbance of the solution is optimized, the least-squares t to a quadratic surface response yields the following adjusted expression: (Total A at 325 nm) = b0 + b1 pH + b2 [Cu(II)] + b11 pH2 + b12 pH [Cu(II)] + b22 [Cu(II)]2 , where b0 = 4.3(8), b1 = 1.5(3), b2 = 0.7(3), b11 = 0.12(3), b12 = 0.01(5) and b22 = 0.15(6) (standard deviation in parenthesis), indicating that all coefcients are signicant (p < 0.05) except b12 . The quality of the t can be assessed by the global residual standard deviation st = 0.035 AU (absorbance units). It can also be concluded that pH 5.7 and Cu 2.10 102 mol L1 are the optimum conditions. However, as the Cu concentration increases, the absorbance of a background signal arising from the interaction between Cu and the buffer does also increase. Therefore, the optimum conditions would be those corresponding to a trade-off between maximum absorbance of the reaction product and minimum background absorbance. In order to obtain these optimum conditions, two response surfaces were estimated: one corresponding to the absorbance of the reaction product, and a second one corresponding to the blank (i.e., a buffer solution not containing the analyte). The absorbance values are given in the last two columns of Table 1. Once both response surfaces are obtained (one for each response), the best trade-off can be estimated by resorting to the so-called desirability function [42]. The construction of the latter function rst involves the transformation of each estimated response variable into a desirability value dm for the mth variable, where 0 dm 1. The value of dm increases as the desirability of the response increases: in our case, dm increases with the net analyte absorbance, and with the inverse of the buffer absorbance. The individual desirabilities are then combined using the geometric mean into a global desirability D: D = [(d1 )w1 (d2 )w2 . . . (dm )wm ]
1/M

Fig. 4. (A) Three-dimensional time-spectral surface corresponding to the catalyzed hydrolysis of an aqueous calibration solution of amoxicillin 0.103 g L1 . (B) Same as (A) for a typical urine blank.

background urine. This prevents the application of a conventional UVvisible methods for the determination of amoxicillin in urine, even those involving the use of multivariate calibration methodologies. Moreover, techniques relying on the use of spectral derivatives may not be useful due to the intrinsic variability of urine spectra among individuals. Therefore, the alternative was explored of studying the kinetics of the metal catalyzed hydrolysis, in order to: (1) generate spectralkinetic second-order data which would allow one to achieve the important second-order advantage, and (2) enhance the selectivity in the determination by converting the analyte into reaction products having red-shifted spectra which would show a lower degree of overlap with that of urine. As shown in Fig. 3B, reaction of the analyte provides the desired spectral shift, with an observable absorbance change in going from the urine blank to a spiked urine sample. It is required to restrict the working spectral range as illustrated in this latter gure, in order to avoid a highabsorbance region at wavelengths lower than 322 nm. Fig. 3B shows that the range 322370 is adequate for the analytical purposes pursued in the present work. The complete time-spectral evolution for a sample only containing the analyte is shown in Fig. 4A. Notice that at t = 0, the spectrum of this solution corresponds to that of product I (see Fig. 1), since the rst step leading to this product is fast enough to preclude observing the transformation of the analyte in the

(7)

812

A. Garca-Reiriz et al. / Talanta 71 (2007) 806815 Table 3 Statistical analysis of prediction results for PARAFAC and PLS/RBL in the test and spiked urine sample sets Parametera General features Number of training samples Spectral range (nm) Time range (min) Matrix data size Calibration range (g L1 ) Test set Number of samples Concentration range (g L1 ) Number of components RMSEP (g L1 ) REP% Spiked urine set Number of samples Concentration range (g L1 ) Number of components RMSEP (g L1 ) REP%
a

Table 2 Prediction results for the test (T) and spiked urine (S) samples Sample Nominal concentration (g L1 ) Predicted concentration (g L1 ) PARAFACa 0.036(8) 0.082(2) 0.122(2) 0.144(3) 0.199(9) 0.067(6) 0.105(7) 0.063(3) 0.067(3) 0.105(4) 0.025(3) 0.069(5) 0.027(4) 0.100(8) 0.000(1) 0.008(3) 0.004(2) PLS/RBLa 0.039(2) 0.071(2) 0.116(3) 0.135(3) 0.192(8) 0.059(5) 0.105(5) 0.050(5) 0.059(5) 0.096(5) 0.026(5) 0.076(5) 0.028(1) 0.127(2) 0.000(1) 0.004(1) 0.003(2)

PARAFAC 6 322370 each 2 nm 018 each 1.5 min 25 12 0.0000.190 5 0.0360.185 N=2 0.008 7 12 0.0000.120 N=3 0.009 10

PLS/RBL

Test set T1 T2 T3 T4 T5

0.037 0.074 0.120 0.147 0.185

Spiked urine set S1 0.063 S2 0.094 S3 0.064 S4 0.064 S5 0.093 S6 0.030 S7 0.073 S8 0.030 S9 0.120 S10 0.000 S11 0.000 S12 0.000
a

A = 3, Nuns = 0 0.007 6

A = 3, Nuns = 2 0.009 10

Average of three determinations with the standard deviation in parenthesis.

RMSEP, root mean square error of prediction; REP%, relative error of prediction (percentage of RMSEP with respect to mean training concentration).

where wm is a relative weight which can be used to enhance the effect of variable m, and M is the sum of all values of wm . The parameter D provides the overall assessment of the desirability of the combined responses. We applied this methodology to the variables represented by the last two columns of Table 1, with the following results (coefcients of a quadratic surface response): for the buffer absorbance at 290 nm, b0 = 0.23(7), b1 = 0.06(3), b2 = 0.20(3), b11 = 0.004(3), b12 = 0.015(4) and b22 = 0.022(5), with all coefcients signicant (p < 0.05) except b11 , st = 0.003 AU, and for the net absorbance at 325 nm, b0 = 4.1(8), b1 = 1.5(3), b2 = 0.6(3), b11 = 0.12(3), b12 = 0.01(5) and b22 = 0.13(5), with all coefcients signicant (p < 0.05) except b12 , st = 0.034 AU. Then, using w1 = 1 for the inverse of the buffer absorbance and w2 = 6 for the net absorbance in Eq. (7), in order to give more weight to the analyte absorbance over the buffer, it can be concluded that a pH of 5.7 and Cu 1.60 102 mol L1 are the optimum vales. 4.3. Multivariate calibration results 4.3.1. Test samples A set of ve test samples (T1T5 in Table 2), only containing the analyte, was analyzed by the presently proposed methodology, in order to monitor the predictive ability of the different models. Analysis using PARAFAC required the introduction of two spectral components, corresponding to the reaction intermediate I and product II of the analyte hydrolysis, and to constrain scores and loadings to be non-negative. The recovery of component proles was successful, but the results for this simple test set do not convey the importance of those obtained in the presence of unsuspected components (see below).

In the case of PLS/RBL, three latent variables were required to describe the variability present in the calibration set of samples, as dictated by leave-one-out cross-validation. Error analysis suggested that Nuns = 0, as expected from the known composition of the test samples, making unnecessary to resort to RBL and the second-order advantage. The prediction results from PARAFAC and PLS/RBL for this test of samples are shown in Table 2, while the detailed statistical analysis is presented in Table 3. As can be seen, good recoveries are obtained, and the results provided by both methodologies do not signicantly differ. 4.3.2. Spiked urine samples Urine samples display a strong background signal, which varies among individuals. The study of these samples with the PARAFAC model required, therefore, the introduction of additional spectral components, other than those employed for the analysis of the test samples of Section 4.3.1. In all of the studied urine samples N = 3 seemed to be an adequate choice, implying that a major urine component dominates the overall absorption of the biological background. PARAFAC models t under non-negativity restrictions for both scores and loadings, and initialized in the manner described in Section 3.1 led to satisfactory results, including component prole recovery and concentration estimation. As an example, Fig. 5A and B show the proles extracted by PARAFAC when processing a typical spiked urine sample. As can be seen, the spectral features for both species intervening in the analyte hydrolysis are correctly recovered, as well as the background signal from urine. Likewise, the time proles are reasonably reproduced if they are compared with those expected on the basis of the kinetic behaviour of the analyte and urine (Fig. 4A and B). The success in recovering unsuspected proles is essential for the achievement of the second-order

A. Garca-Reiriz et al. / Talanta 71 (2007) 806815

813

Fig. 5. (A) Spectral proles recovered for all components by PARAFAC (solid lines) and for the rst RBL principal component of the unsuspected signal by PLS/RBL (dashed line). A typical urine sample has been processed, spiked with amoxicillin 0.103 g L1 . (B) The corresponding time proles for the same sample. In both cases, component numbering corresponds to the overall contribution to the PARAFAC spectral variance, and are ascribed to: component 1, reaction product II; component 2, urine background; component 3, reaction intermediate I.

Beyond the global RMSEP and REP% values, an accuracy test was applied involving the consideration of the elliptical joint condence region (EJCR) for the slope and intercept of predicted vs. nominal concentration values [43]. Since standard deviations for each predicted concentration are available, a weighted least-squares regression was performed, using the inverse variance at each point as weight. Both for PARAFAC and PLS/RBL, the elliptical regions contained the theoretical point of slope = 1, intercept = 0, indicating that they yield accurate results on the spiked urine sample set. Specic results concerning the regression parameters are: PARAFAC, slope = 1.00(4), intercept = 0.001(1), and PLS/RBL, slope = 1.01(3), intercept = 0.001(1). Again, no signicant difference is found between both algorithms. Intermediate precision was estimated analyzing triplicates of urine spiked samples, prepared and measured on different days and using different stock solutions. Under these conditions, the difference in the prediction of replicates is a good measure of reproducibility of the method, expressed as the ARSD (average relative standard deviation), given by 100 s/c, where s is the standard deviation and c the group average concentration. ARSD values are ca. 3.5%, which are reasonable considering the complexity of the samples. Additional urine samples were studied including other therapeutic drugs which are regularly administered along with amoxicillin, such as diclofenac, ibuprofen, acetilsalicylic acid, naproxen and acetaminophen. However, in all of these cases the spectral features of the added drugs were undistinguishable from that of the background urine, which dominates the sample spectrum within the useful working range. Consequently, similar results to those discussed above for the spiked urine sample set were obtained. 4.3.3. Figures of merit Figures of merit were estimated for the analyte in all samples studied, and for both multivariate methods (PARAFAC and PLS/RBL). The most interesting values correspond to the analysis of urine-containing samples, where the full sensitivity and selectivity would decrease due to signicant spectral overlap. The average analyte sensitivity, as calculated for PARAFAC and PLS/RBL according to Eqs. (1) and (5) respectively, were 0.4 and 0.5 AU L mg1 . This result implies that both of these multivariate methodologies, although employing a very different scheme for achieving the second-order advantage, display similar sensitivities towards the analyte of interest. The analytical sensitivities n , calculated using Eqs. (1) and (5), are 400 and 500 L mg1 for PARAFAC and PLS/RBL respectively, using an instrumental noise level of ca. 0.001 AU. 1 This leads to inverse values (n ) of 0.0025 and 0.002 mg L1 , both given in useful concentration units. These gures of merit are reasonably low in view of the analyte concentration range. The corresponding limits of detection were computed as 7.5 and 6 mg L1 for PARAFAC and PLS/RBL respectively, indicating no serious difference between the employed algorithms, although slightly favouring the PLS method. The obtained gures of merit are similar to those for the HPLC method, which shows a limit of detection of ca.

advantage, and for the unbiased prediction of the analyte concentration in real urine samples. For using the PLS/RBL model, the number of latent variables A was identical to that employed for calibration and in the study of the test samples (i.e., A = 3), while Nuns had to be increased to two in order to accommodate for the signal of the unsuspected urine background. The proles recovered for the unsuspected components when Nuns = 2 are given by SVD analysis and therefore they do not necessarily resemble true spectra. However, the rst RBL principal component in each dimension for the unsuspected signal can be compared with those furnished by PARAFAC. The good agreement of the results provided by these two techniques is apparent on inspection of Fig. 5A and B. Specic prediction results from PARAFAC and PLS/RBL are provided by Table 2, with the corresponding statistical information collected in Table 3. The overall prediction results are satisfactory for both algorithms, judging from the reasonably low values of RMSEP (root mean square error of prediction). These encouraging results imply that the second-order advantage is fully exploited the new PLS/RBL algorithm, suggesting that the latter one is a reasonable alternative for the processing of the presently study kineticspectral second-order data.

814

A. Garca-Reiriz et al. / Talanta 71 (2007) 806815

Table 4 Comparison of the predicted amoxicillin concentration in real urine samples by both chemometric algorithms and HPLC Sample HPLC Concentrationa 1 2 3 4
a b

PARAFAC (g L1 ) Concentrationa 1.31(7) 1.21(1) 2.61(9) 2.07(9) (g L1 ) tb 1.51 0.64 1.81 1.29

PLS/RBL Concentrationa (g L1 ) 1.20(3) 1.20(1) 2.64(6) 1.84(8) tb 0.18 0.43 1.34 0.86

1.21(9) 1.18(8) 2.71(3) 1.80(9)

Average of four determinations, with standard deviation in parenthesis. Calculated t coefcient (six degrees of freedom) for the comparison of each algorithm with HPLC. The critical t-value ( = 0.05) is 2.45.

0.2 mg L1 , and a sensitivity estimated as ca. 2.0 AU L mg1 based on UV detection at 230 nm. The times to complete a given urine analysis are however lower for HPLC, i.e. ca. 10 min (after the column is stabilized), in comparison with 20 min for the present method. Nevertheless, the simplicity and easy implementation of the kineticspectrophotometric method are important characteristics which are favourably compared to the signicantly more laborious HPLC technique. 4.3.4. Real urine samples A series or urine samples obtained from patients administered with amoxicillin were analyzed according to the presently described procedure, and the results were compared to those rendered by a reference HPLC technique. Samples were collected within 2 or 3 h after a dose of 500 mg, and processed as described in the Section 2. Both PARAFAC and PLS/RBL were applied in order to process the second-order data and to predict the analyte concentration. The results are shown in Table 4. They are statistically comparable to those obtained using the reference method HPLC (Table 4), although PLS/RBL seems to provide results which are closer to the HPLC ones. 5. Conclusions Given that the above employed second-order multivariate methodologies are all able to model unsuspected sample components by achieving the second-order advantage, the following additional characteristics should be considered in assessing their relative advantages and disadvantages: (1) analytical performance, (2) ease and speed of program operation and (3) model interpretability. PARAFAC provides the analyst with information having a rich physical signicance. However, the price paid is a somewhat complex program operation, requiring the introduction of initial guesses of component spectra and time proles for successful data decomposition, because the latter shows a linear dependency in the time dimension. PLS with second-order advantage is certainly the easiest of both employed multivariate methods, requiring almost no knowledge of the spectral or kinetic properties of the analytes for program operation. However, it provides no discernible physical interpretation as to the intervening chemical phenomena, except perhaps in what concerns the spectral and time proles of a single unsuspected sample component. From a purely analytical

point of view, however, its prediction ability is comparable to PARAFAC. Acknowledgments Universidad Nacional de Rosario, CONICET (Consejo Nacional de Investigaciones Cientcas y T cnicas, Project e No. PIP 5303) and ANPCyT (Agencia Nacional de Promoci n o Cientca y Tecnol gica, Project No. PICT05-25825) are grate o fully acknowledged for nancial support. A.G.R. thanks CONICET for a fellowship. References
[1] A. Goodman-Hillman, T. Rall, A. Nier, P. Taylor, The Pharmacological Basis of Therapeutics, McGraw Hill, New York, 1996. [2] J. Reynold (Ed.), Martindale. The Extra Pharmacopeia, The Pharmaceutical Press, London, 1993. [3] J. Wibawa, D. Fowkes, P.N. Shaw, D.A. Barrett, J. Chromatogr. B 774 (2002) 141148. [4] W.H. Luo, C.Y.M. Ang, J. AOAC Int. 83 (2000) 2025. [5] L. Valvo, E. Ciranni, R. Alimenti, S. Alimonnti, R. Draisci, L. Giannetti, L. Lucentini, J. Chromatogr. B 797 (1998) 311316. [6] W.A. Moats, J. AOAC Int. 77 (1994) 4145. [7] S. Chulavatnatol, B.G. Charles, J. Chromatogr. Biomed. Appl. 126 (1985) 9196. [8] H. Huang, J. Wu, M. Chen, J. Chromatogr. Biomed. Appl. 102 (1991) 195203. [9] H. Mascher, C. Kikuta, J. Chromatogr. 506 (1990) 417421. [10] J. Haginaka, J. Wakai, Analyst 110 (1985) 12771281. [11] K.S. Booksh, B.R. Kowalski, Anal. Chem. 66 (1994) 782A791A. [12] A. Fern ndez-Gonz lez, R. Bada, M.E. Daz-Garca, Anal. Biochem. 341 a a (2005) 113121. [13] P. Campomanes, M. Men ndez, R. L pez, T. Sordo, Phys. Chem. Phys. e o Chem. 6 (2005) 344351. [14] P. Gutierrez Navarro, A.E. Bekkouri, E. Rodriguez Reinoso, Analyst 123 (1998) 22632266. [15] S.J. Beard, D.T. Ciccognani, M.N. Hughes, R.K. Poole, FEMS-Microbiol. Leu. 96 (1992) 207212. [16] P. Marquez Garca, P. Gutierrez Navarro, P.J. Martnez de los Parras, Talanta 40 (1998) 101107. [17] R. Bro, Chemometr. Intell. Lab. Syst. 38 (1997) 149171. [18] J. Ohman, P. Geladi, S.J. Wold, J. Chemometr. 4 (1990) 7990. [19] A.C. Olivieri, J. Chemometr. 19 (2005) 253265. [20] R. Tauler, Chemometr. Intell. Lab. Syst. 30 (1995) 133146. [21] A. Smilde, R. Bro, P. Geladi, Multi-way Analysis: Application to the Chemical Sciences, Wiley, 2004. [22] R. Bro, H.A.L. Kiers, J. Chemometr. 17 (2003) 274286. [23] A. Mu oz de la Pe a, A. Espinosa Mansilla, D. Gonz lez G mez, A.C. n n a o Olivieri, H.C. Goicoechea, Anal. Chem. 75 (2003) 26402646. [24] R. Bro, N.D. Sidiropoulos, J. Chemometrics 12 (1998) 223247.

A. Garca-Reiriz et al. / Talanta 71 (2007) 806815 [25] M. Azubel, F.M. Fern ndez, M.B. Tudino, O.E. Troccoli, Anal. Chim. Acta a 398 (1999) 93102. [26] J.C.G. Esteves da Silva, C.J.S. Oliveira, Talanta 49 (1999) 889898. [27] S.R. Crouch, J. Coello, S. Maspoch, M. Porcel, Anal. Chim. Acta 424 (2000) 115126. [28] A. Espinosa-Mansilla, A. Mu oz de la Pe a, H.C. Goicoechea, A.C. n n Olivieri, Appl. Spectrosc. 58 (2004) 8390. [29] M.A.B. Levi, I.S. Scarminio, R.J. Poppi, M.G. Trevisan, Talanta 62 (2004) 299305. [30] S. Wold, P. Geladi, K. Esbensen, J. hman, J. Chemometr. 1 (1987) 4156. [31] D.M. Haaland, E.V. Thomas, Anal. Chem. 60 (1988) 11931202. [32] A. Lorber, Anal. Chem. 58 (1986) 11671172. [33] K. Faber, A. Lorber, B. Kowalski, J. Chemometr. 11 (1997) 419461.

815

[34] A.C. Olivieri, Anal. Chem. 77 (2005) 49364946. [35] A.C. Olivieri, N.M. Faber, J. Chemometr., in press. [36] A.C. Olivieri, N.M. Faber, J. Ferr , R. Boqu , J.H. Kalivas, H. Mark, Pure e e Appl. Chem. 78 (2006) 633661. [37] L. Cuadros Rodrguez, A.M. Garca Campa a, C. Jim nez Linares, M. n e Rom n Ceba, Anal. Lett. 26 (1993) 12431258. a [38] A.C. Olivieri, J. Chemometr. 19 (2005) 253265. [39] MATLAB 6.0, The MathWorks Inc., Natick, MA, USA, 2000. [40] http://www.models.kvl.dk/source/. [41] A.C. Olivieri, H.C. Goicoechea, F.A. I on, Chemometr. Intell. Lab. Syst. n 73 (2004) 189197. [42] E.C. Harrington Jr., Ind. Qual. Control 21 (1965) 494498. [43] A.G. Gonz lez, M.A. Herrador, A.G. Asuero, Talanta 48 (1999) 729736. a