Nondestructive Direct Determination of Heroin in ... - ACS Publications

Sep 9, 2008 - standard deviation of the population in the validation data set (SD) and the RMSEP for the external validation, defined as the residual ...
13 downloads 5 Views 1MB Size
Anal. Chem. 2008, 80, 7257–7265

Nondestructive Direct Determination of Heroin in Seized Illicit Street Drugs by Diffuse Reflectance near-Infrared Spectroscopy Javier Moros,† Nieves Galipienso,‡ Rocı´o Vilches,‡ Salvador Garrigues,*,† and Miguel de la Guardia† Department of Analytical Chemistry, Universitat de Valencia, Edifici Jeroni Mun˜oz, 50th Dr. Moliner, 46100, Burjassot, ´ rea de Sanidad de Valencia, Valencia, Spain, and Unidad de Inspeccio´n de Farmacia y Control de Drogas del A Muelle de la aduana s/n, 46024, Valencia, Spain A new method has been developed for the fast and nondestructive direct determination of heroin in seized street illicit drugs using partial least-squares regression analysis of diffuse reflectance near-infrared spectra. Data were obtained from untreated samples placed in standard glass chromatography vials. A heterogeneous population of 31 samples, previously analyzed by a reference method, was employed to build the calibration model and to have a separated validation set. Based on the use of zero-order data for a calibration set of 21 samples, after standard normal variate and quadratic linear removed baseline correction (detrending), in the wavelength range from 1111 to 1647 nm, 8 PLS factors were enough to obtain a root-mean-square error of prediction of 1.3% w/w, with a quality coefficient of 10% for the estimation of the accuracy error in the prediction of heroin concentration in unknown samples and a residual predictive deviation of 5.4. Heroin, the 3,6-diacetyl derivative of morphine (hence diacetylmorphine), is an opiate drug synthesized from morphine, a naturally occurring substance extracted from the seedpod of certain varieties of poppy plants, by acetylation.1 Heroin is a highly addictive drug often associated with compulsive or addictive patterns of use, for which reason heroin consumption affects both the individual user and the society.2 Pure heroin is a white powder with a bitter taste rarely sold on the streets, and illicit heroin is frequently a powder varying in color from white to dark brown due to impurities left from the manufacturing process or the presence of additives.3 The purity of street heroin can also vary quite widely, as the drug can be mixed with other white powders, such as chalk, flour, talcum powder, paracetamol, and caffeine. The impurity of the drug often * To whom correspondence should be addressed. Tel.: +34 96 354 3158. Fax: +34 96 354. E-mail: [email protected]. † Universitat de Valencia. ‡ Unidad de Inspeccio´n de Farmacia y Control de Drogas del A´rea de Sanidad de Valencia. (1) National Institute on Drug Abuse, Heroin Abuse and Addiction Research Report, May 2005. (2) U. S. Drug Enforcement Administration. http://www.usdoj.gov/dea/ concern/heroin.html (3) Drug Enforcement Administration, Drugs of Abuse, 2005. 10.1021/ac800781c CCC: $40.75  2008 American Chemical Society Published on Web 09/09/2008

makes it difficult to gauge the strength of the dosage, which runs the risk of overdose. Specific laws in many countries establish different penalties as a function of the amount of illicit drugs seized, and because of that, there is a need for qualitative and quantitative methods to determine as soon as possible the concentration of heroin in seized illicit drugs.4,5 Conventional methods applicable to heroin base and heroin hydrochloride are generally based on separative techniques such as high-performance liquid chromatography (HPLC), gas chromatography (GC) with and without derivatization, or capillary electrophoresis, this last being the method used by the Drug Enforcement Administration (DEA) of the United States of America from 2003.6-11 The aforementioned methods are highly sensitive but are destructive and involve a tedious and solvent consuming sample preparation. Additionally, not all of them are suitable for all types of heroin samples. On the contrary, vibrational spectroscopy-based methods, like Raman, mid-infrared (MIR), or near-infrared (NIR) spectroscopy can be applied to determine rapidly the concentration of compounds present in complex samples at percentage levels without any sample preparation. However, after a bibliographic research, only a few references were found for the determination of heroin by vibrational spectroscopy involving, in the main part of cases, just qualitative or classification purposes. In 1987, Ravreby proposed the first Fourier transform infrared (FT-IR) method for quantitative determination of heroin and cocaine using KBr pellets and working with area integration of the carbonyl band. The interferences were corrected through spectral subtraction.11 (4) European Legal Database on Drugs. http://eldd.emcdda.europa.eu/html.cfm/ index5749EN.html. (5) Report from the Toxicological Information Service of the National Institute of Toxicology from Spain 12691, Spain, 2003. (6) Lurie, I. S.; Carr, S. M. J. Liq. Chromatogr. Relat. Technol. 1986, 9, 2485– 2509. (7) Gloger, M.; Neumann, H. Forensic Sci. Int. 1983, 22, 63–74. (8) Barnfield, C.; Burns, S.; Byrom, D. L.; Kemmenoe, A. V. Forensic Sci. Int. 1988, 39, 107–117. (9) Lurie, I. S.; Hays, P. A.; Garcia, A. E.; Panicker, S. J. Chromatogr., A 2004, 1034, 227–235. (10) United Nations office on drugs and crime. Methods for impurity profiling of heroin and cocaine, United Nations publication: Vienna, Austria, 2005. (11) Ravreby, M. J. Forensic Sci. 1987, 32, 20–37.

Analytical Chemistry, Vol. 80, No. 19, October 1, 2008

7257

Figure 1. Diffuse reflectance NIR spectra of illicit street drug samples containing heroin (B ) 21.8% w/w, C ) 13.2% w/w, D ) 17.0% w/w, and E ) 10.7% w/w) from different seizures compared to the spectrum of pure heroin (A) in the wavelength range from 850 to 2750 nm. Instrumental conditions: 4 cm-1 nominal resolution and 32 cumulated scans/spectrum. Note: Spectra have been shifted along the Y axis in order to appreciate their differences. Inset shows the molecular structure of heroin.

FT-IR has been also employed for qualitative identification of heroin seizures and for identification of heroin from different routes by using cluster analysis.12,13 Raman spectroscopy has been used for characterizing pure amphetamine sulfate, cocaine hydrochloride, and heroin in illicit drugs.14,15 However, sensitivity for the detection using this technique depends on the scattering cross section of drug, fluorescence, and complexity of dilutants and matrix of drug, dilutant Raman spectrum, and spectrometer resolution.15 Principal component analysis has been employed for classification of narcotics in solid mixtures with different materials, but no quantitative data were obtained for heroin in illicit drugs.16 NIR spectroscopy is nowadays a well-established technique, suitable to be employed for the quantitative analysis of solid samples. Surprisingly, there is only a single preliminary study on the use of NIR for the determination of heroin, 6-acethylmorphine, and codeine in drugs, published in a Chinese journal, based on (12) Levy, R.; Ravreby, M.; Meirovich, L.; Shapira-Heiman, O. J. Forensic Sci. 1996, 41, 6–11. (13) Cai, X. L.; Wu, G. P. Spectrosc. Spectral Anal. 2007, 27, 2441–2444. (14) Hodges, C. M.; Hendra, P. J.; Willis, H. A.; Farley, T. J. Raman Spectrosc. 1989, 20, 745–749. (15) Ryder, A. G.; O’Connor, G. M.; Glynnn, T. J. J. Forensic Sci. 1999, 44, 1013–1019. (16) Ryder, A. G. J. Forensic Sci. 2002, 47, 275–284.

7258

Analytical Chemistry, Vol. 80, No. 19, October 1, 2008

the use of synthetic samples to built a partial least-squares (PLS) calibration model for the prediction of heroin content in unknown samples.17 So, the main objective of this study was the development of a fast, nondestructive, and well-validated method for the quantitative determination of heroin in seized illicit street drugs using previously analyzed actual samples to build the calibration model in order to model the behavior of heroin and to minimize the possible spectral contribution of cutting agents. EXPERIMENTAL SECTION Apparatus and Reagents. For the direct diffuse reflectance near-infrared (DR-NIR) determination of heroin, a Bruker Gmbh (Bremen, Germany) model multipurpose analyzer Fourier transform near-infrared spectrometer was employed. The system was equipped with a single detector onto which the radiation was focused by means of an integrating sphere, used as measurement accessory. For instrument control and data acquisition, the OPUS program Version 4.2, also from Bruker Gmbh, was employed. Spectra treatment and data manipulation were carried out using Omnic 6.1 and OmnicMacros 6.1 software from Nicolet (Madison, WI). (17) Wu, G. P.; Xiang, B. R. Chin. J. Anal. Chem. 2007, 35, 552–554.

Table 1. Diffuse Reflectance Near-Infrared Absorption Frequencies of Heroin tentative frequencies of heroin bands (nm)

actually measured peak frequencies (nm)

forms of modes of vibration assignment

1160

1157

1195 1360 1395 1420 1570

1190 1200 1357 1391 1425 1570

1685

1684

1705

1709

1725

1727

1765

1767

1780

1780

1920

1914

1950

1936

1990

2000

2070

2074

2090 2140

2100 2135

CdO stretch fourth overtone C-H second overtone C-H second overtone C-H combination C-H combination O-H first overtone N-H stretch first overtone C-H stretch first overtone C-H stretch first overtone C-H stretch first overtone C-H stretch first overtone C-H stretch first overtone CdO stretch second overtone CdO stretch second overtone N-H stretch/N-H bend combination N-H deformation overtone C-H combination C-H stretch/CdO stretch combination or sym C-H deformation C-H stretch/CdO stretch combination or sym C-H deformation asymmetric C-H stretch/C-H deformation combination N-H bend second overtone or C-H stretch/CdO stretch combination, or CdO stretch C-N stretch; N-H in-plane bend. CH stretch/CdO stretch combination C-H stretch/CH2 deformation C-H bend second overtone CH stretch/CH2 deformation combination CH2 bend second overtone C-H stretch/C-C stretch combination C-H combination or Sym C-N-C stretch overtone C-H stretch/C-C stretch combination asymmetric C-N-C stretch first overtone asymmetric C-N-C stretch first overtone

2144 2170

2172

2180

2178

2200

2194

2280

2284

2300

2300

2325

2320

2352

2352

2380

2384

2470

2454

2488

2485

2530

2524

2530

2537

The HPLC system used to obtain the reference data included an HP/Agilent 1100 series, from Hewlett-Packard (Palo Alto, CA), composed by a G1311A quaternary pump, a G1316A column compartment, and a G1315B diode array detector. A C-18 reversed phase (Kromasil) analytical column (250 mm × 4.6 mm i.d. and 5-µm particle diameter) from Scharlau (Barcelona, Spain) was used for heroin determination.

Heroin standard (98%, w/w) was supplied by Riedel-de-Hae¨n (Seelze, Germany), and acetonitrile (analytical reagent grade) was purchased from Scharlau (Barcelona, Spain). Buffer solution for chromatographic analysis was prepared in ultrapure water, with a resistivity of 18.2 MΩ cm obtained from a Millipore Milli-Q system (Bedford, MA). Samples Description. The 31 samples employed in this study were provided by the drug control laboratory of the Inspeccio´n de Sanidad de Valencia (Spain), covering an important range of street drugs, seized from 2005 to 2007, containing heroin from 6 to 34% w/w. Data of heroin content were determined experimentally in our laboratory by using a chromatography reference procedure based on HPLC. Furthermore, in order to have additional information on the samples under study, mainly related to the identification of cutting agents, illicit street drugs were analyzed by a headspace gas chromatography-mass spectrometry (HS-GC/MS) procedure. Reference Procedure. The reference chromatography procedure was adapted from the method established by Lurie and Carr with slight modifications.6 An amount between 10 and 20 mg of sample was accurately weighed inside a 25-mL volumetric flask and diluted to the volume with acetonitrile. Twenty liters of the resultant filtered (through nylon 22-µm filters) solutions were directly injected in the HPLC system using a mobile phase containing acetonitrile (A), as organic modifier, and acetate buffer (B) of 4.83 pH, as the aqueous component of the mobile phase. The chromatographic separation of heroin was made in the gradient mode from an initial proportion of 32% v/v A and 68% v/v B, which was modified at 5 min to 45% v/v A and 55% v/v B and maintained from 6.5 to 10 min with pure A. All of the chromatographic measurements were done by absorbance measurements at 254 nm, with column thermostated at 40 °C with a mobile-phase flow rate of 1.0 mL/min. Area values of the chromatogram peaks obtained at 3.32 min for samples were interpolated in an external calibration line established from eight standard solutions of heroin ranging from 36 to 340 mg/L, treated in the same way as the samples. A typical calibration line obtained by this procedure was mAU ) (37 ± 36) + (20.21 ± 0.18) [heroin] (mg/L) with a coefficient of determination (R2) of 0.9991. The method provided a limit of detection of 2.5 mg/L, a limit of quantification of 8.3 mg/L, and a relative standard deviation of 1.2%. DR-NIR Procedure. Solid samples, previously ground in an agate mortar to obtain a fine homogenized powder, were placed in 2-mL standard glass chromatographic vials (12 mm × 32 mm) of 9.5-mm internal diameter used as measurement cells. The spectra were collected in Kubelka-Munk mode with a nominal resolution of 4 cm-1, accumulating 32 scans/spectrum (28-s data acquisition). The closed integrating sphere was used to collect the corresponding background at the same instrumental conditions as the samples. Zero-order and first-order derivative spectra of samples were analyzed using PLS chemometric treatments. Calibration and validation data sets were established from previously analyzed samples selected as a function of the dendrographic distribution of their NIR spectra. The appropriate PLS models were built using the best preprocessing method, waveAnalytical Chemistry, Vol. 80, No. 19, October 1, 2008

7259

Figure 2. Diffuse reflectance NIR spectra of some pure substances commonly employed as cutting agents compared to spectra of an illicit street drug sample (21.8% w/w heroin) and pure heroin in the wavelength range from 850 to 2750 nm. Instrumental conditions: 4-cm-1 nominal resolution and 32 cumulated scans/spectrum. Note: Spectra have been shifted along the Y axis for best appreciation of their bands. Shady region means wavelength range used for building the best PLS model.

length range, and number of factors for heroin quantization in illicit drugs from a calibration set of drug samples and applied for the prediction of heroin content in a validation set of samples different from those used for calibration. Chemometric Data Treatment. PLS models were built, using the Turbo Quant Analyst 6.0 software developed by Thermo Nicolet Corp., from data exported in JCAMP-DX format from Opus. The appropriate spectral range and the optimum number of PLS factors were selected to provide the minimum value for the predicted residual error sum of squares (PRESS). Through this study, several analytical features, related to the quality of the model fitting to the original data, as well as the prediction capabilities have been employed. Among them the rootmean-square error of calibration (RMSEC), the correlation coefficient (r), and the root-mean-square error of prediction (RMSEP), were used as criteria to evaluate the coherence and the performance of multivariate calibration models. 7260

Analytical Chemistry, Vol. 80, No. 19, October 1, 2008

Additionally to the aforementioned classical parameters, the mean difference (dx-y) between the predicted and reference values, joint to other quality indicators; such as the deviation between triplicates (strip) and the pooled standard error of prediction for validation samples (sreg) were also considered to evaluate the predictive capabilities of the PLS-NIR methodology.18 Prediction accuracy was described through the use of the quality coefficient (QC) parameter, which gives an indication of the percentage error to be expected for the estimated concentration in future predictions. Moreover, evaluation of the predictive ability of the PLS models was made from the ratio between the standard deviation of the population in the validation data set (SD) and the RMSEP for the external validation, defined as the residual predictive deviation (RPD). If the error of prediction for a parameter (RMSEP) is high as compared with the spread in (18) Moros, J.; Garrigues, S.; de la Guardia, M. Anal. Chim. Acta 2007, 582, 174–180.

Figure 3. Dendrographic classification of illicit street drugs containing heroin using the Euclidean distance after vector normalization spectra and applying the Ward linkage method. For details about cluster group composition, see data in Table 2.

composition of that parameter in the validation sample set, PLSNIR models cannot be considered as robust. In contrast, high values of RPD mean a great power of the model to predict accurately the considered parameter.19,20 In spite of the fact that several spectral windows were tested for building the PLS models and evaluating their prediction capabilities using the validation set, only the most significant information is shown here. (19) Massart, D. L.; Vandeginste, B. G. M.; Buydens, L. M. C.; de Jong, S.; Lewi, P. J.; Smeyers-Verbeke, J. Handbook of Chemometrics and Qualimetrics, (Parts A and B); Elsevier Science B.V.: Amsterdam, 1997.

Cluster Analysis. Hierarchical cluster analysis is an exploratory data tool commonly used for solving classification problems. Cluster analysis encompasses a number of different algorithms and methods to do the partition of a set of objects in mutually exclusive groups and thus can be highly useful to evaluate similarities and differences of objects from their NIR fingerprint. Cluster analysis identifies and classifies objects on the basis of their similarities, so that the degree of association is strong (20) Fearn, T. In Handbook of Vibrational Spectroscopy; Chalmers, J. M.; Griffiths, P. R., Eds.; Wiley: Chichester, 2002; pp 2086-2093.

Analytical Chemistry, Vol. 80, No. 19, October 1, 2008

7261

RESULTS AND DISCUSSION DR-NIR Spectra of Heroin Samples. Figure 1 shows diffuse reflectance NIR spectra of different seized illicit drugs containing heroin (between 10.7 and 21.8%) compared with the spectrum of pure heroin in the NIR range between 850 and 2750 nm. The absorption of NIR radiation is due mainly to overtone and combination bands primarily of OsH, CsH, NsH, and CdO groups, which exhibit their fundamental molecular stretching and bending absorption in the MIR range. These overtones do not behave in a simple way, making NIR spectra complex and not directly interpretable as in other spectral regions.23

In spite of that, it is possible to identify from the pure heroin DR-NIR spectrum in Figure 1 the main wavelengths related to the most common functional groups in heroin. (See bands assignment in Table 1). Moreover, as can be appreciated from Figure 1, there are significant differences between the spectrum of pure heroin and sample spectra. These differences are due to the presence of different components used as cutting agents, which can affect the shape and the intensity of the DR-NIR signals. Some of the volatile cutting components in samples were identified using headspace GC/MS, which showed a wide variability of them, such as caffeine and paracetamol preferentially, also some alkaloids as acetylcodeine (an impurity present in illicitly manufactured heroin), noscapine, and lidocaine, which can naturally ocurring or coming from pharmaceuticals, as well as phenacetin (accompanying caffeine or from acetylation of phenetidin), triacetin, salicylic acid, and even cocaine in different amounts and distribution through all the samples under study. Moreover, the presence of other nonvolatile cutting agents such as chalk, flour, or talcum (powdered substances with high availability) was also taken into consideration. Figure 2 shows DR-NIR spectra of different compounds frequently employed as cutting agents compared with the spectra of pure heroin and a seized illicit street drug sample (containing 21.8% heroin) in the wavelength range between 850 and 2750 nm, and as can be seen, pacetamol seems to be the main cutting agent employed for the considered samples, additionally other than caffeine, acetylsalicylic acid, chalk, or flour. Cluster Classification of Illicit Drug Samples Based on Their DR-NIR Spectra. Taking into account our previous experience, we selected dendrogram classification using Euclidean distance with Ward linkage upon considering the complete DRNIR spectra, recorded between 850 and 2750 nm, after vector normalization treatment.24 It must be noted that vector normalization is made by first calculating the average intensity value and subsequent subtraction of this value from the spectrum. Then the sum of the squared intensities is calculated and the spectrum divided by the square root of this sum. This method is used to account for different samples thickness. Figure 3 shows the dendrographic classification obtained for the 31 samples under study, and as can be seen, from this figure different groups of samples could be identified. The two main groups formed seem to be mainly related to the heroin content: group A related to samples with low content of heroin and group B including samples with heroin high content. However, on considering that dendrographic classification was made based on DR-NIR spectra of seized samples, the presence of cutting agents, which modify the intensity and shape of the DR-NIR bands, has an important contribution to cluster formation. Thus, based on the spectra differences originating from the presence of cutting agents at different levels, 17 clusters could be identified, thus evidencing the availability of a much different kind of samples containing heroin.

(21) Moros, J.; In ˜o´n, F. A.; Garrigues, S.; de la Guardia, M. Anal. Chim. Acta 2005, 538, 181–193. (22) Moros, J.; In ˜o´n, F. A.; Khanmohammadi, M.; Garrigues, S.; de la Guardia, M. Anal. Bioanal. Chem. 2006, 385, 708–715.

(23) Burns, D. A.; Ciurczak, E. W. Handbook of Near-Infrared Analysis, 2nd ed.; Marcel Dekker: New York, 1992. (24) Moros, J.; In ˜o´n, F. A.; Garrigues, S.; de la Guardia, M. Anal. Chim. Acta 2007, 584, 215–222.

Table 2. Characteristics of Illicit Street Drug Samples Classified into Clusters after Dendrographic Treatment of DR-NIR Dataa

group A

B

cluster index

number of samples

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

1 1 3 2 1 3 1 1 5 3 2 2 1 1 1 2 1

heroin (% w/w) mean 10.7 17 11.9 6.7 8.25 23 17 21.1 24.8 32 18 23.8 34.0 12.1 17.7 22.7 14.2

(s

samples 1 13 29, 26, 30 27, 28 31 2, 11, 9 8 4 3, 22, 25, 23, 24 5, 10, 7 12, 14 20, 21 6 16 15 17, 19 18

1.2 0.4 4 0.7 3 3 1.4

1.3

a Content values are expressed in % w/w ± s refers to the standard deviation (of the mean).

Table 3. Descriptive Statistics of Calibration and Validation Data Sets Used for PLS-DR-NIR Determination of Heroin in Illicit Street Drugsa heroin (% w/w) set

number of samples

mean

(s

calibration validation

21 10

19 21

8 7

a

± s refers to the standard deviation of the mean.

between objects of the same cluster and weak between objects of different clusters. It seeks to minimize within-group variance and maximize between-group variance, being the result of cluster analysis of sample spectra of a number of heterogeneous groups with homogeneous contents. In this study, we have employed cluster analysis in the same way as in previous works to obtain a calibration set representative of all the samples considered.21,22

7262

Analytical Chemistry, Vol. 80, No. 19, October 1, 2008

Table 4. Prediction Capabilities of PLS-DR-NIR Procedure for Heroin Content Estimation in Illicit Street Drug Samplesa models noncorrected spectra

SNV corrected spectra

parameters

zero-order spectra

first-order derivative spectra

zero-order spectra

first-order derivative spectra

spectral region (nm) baseline correction type factors RMSEC (% w/w) r RMSECV (% w/w) RMSEP (% w/w) strip (% w/w) sreg (% w/w) QC (%) dx-y (% w/w) RPD

1111 - 1720 linear removed 7 1.8 0.97 4 1.6 0.14 1.7 9 -0.13 4.4

1284 - 1874 linear removed 4 1.4 0.98 3 1.8 0.2 1.8 9 -0.9 3.9

1111 - 1647 quadratic removed 8 1.5 0.98 3.8 1.3 0.4 1.3 10 -0.7 5.4

1111 - 1881 quadratic removed 5 0.8 0.995 2.5 1.6 0.4 1.6 10 -0.8 4.4

a Factors were selected looking for the best prediction capabilities of the model for the validation set. r is the linear correlation coefficient for calibration. RMSEC is the root-mean-square error of calibration; RMSECV is the root-mean-square error of cross-validation; RMSEP is the rootmean-square error of prediction. strip is the standard deviation between triplicates. sreg is the reproducibility of the determination, established from the standard error of prediction. d(x-y) is the mean difference between predicted vs actual heroin content values (% w/w), respectively. QC is the quality coefficient. RPD is the residual predictive deviation defined as the ratio between the standard deviation of the population in the validation data set (SD) and the RMSEP obtained.

Through this study, the hierarchical cluster analysis of DRNIR spectra was used for an appropriate selection of the calibration data set, and it is clear that the cluster classification of new unknown samples could be extremely useful to verify their nature in order to evaluate the possibilities of the NIR technique to appropriately predict their heroin concentration. Table 2 indicates the number of samples included in each cluster as well as the mean and the standard deviation values of heroin content in the 17 clusters, ordered from the top to the bottom of Figure 3. Selection of the Calibration Set. The number and nature of samples integrating calibration and validation subsets employed for PLS modeling are always critical factors in multivariate analysis. Because of that, based on the use of a dendrogram of Figure 3 and following the approaches stated in previous contributions of our team,24 calibration and validation sets were selected using the following criterion: at least one sample from each cluster was considered for calibration and in the case of clusters containing more than one sample, the rounded values of the square root of the total number of samples included in each cluster was used for calibration. The remaining samples were incorporated to the validation set. On the basis of the aforementioned considerations, we built multivariate calibration models using 21 samples, validating the predictive capabilities and analytical features of the models using the remaining 10 samples. Table 3 shows the average data and data dispersion for calibration and validation sets obtained following aforementioned criterion. NIR Partial Least-Squares Modelization for Heroin Determination. For heroin content prediction, PLS calibration models must be optimized in terms of both spectral range and number of factors employed. Moreover, a critical comparison about the use of zero-order and first-order derivative spectra, as

well as the application of standard normal variate (SNV) combined with detrending as correction algorithm for particle size and scattering effects was also tested. Processes and results obtained are detailed in Table 4. Spectral ranges were chosen using the moving window strategy, thus minimizing the RMSEP values. Table 4 also shows the best prediction capabilities achieved by the PLS-DR-NIR technique in each case. Selected NIR interval is indicated in Figure 2, and it can be seen that in this range spectral perturbations of studied cutting agents are reduced as compared with other intervals. It is interesting to note that there are slight modifications in the selected spectral ranges and in the number of PLS factors used for building the appropriate models on considering the use of zero- or first-order derivative spectra with or without the use of the SNV algorithm. From data reported in Table 4. it can be appreciated that the model built on using zero-order SNV corrected provided the best RMSEP and the highest RPD. The reproducibility of the determination, established from the mean standard deviation of each replicate (strip) and the standard error of prediction (sreg), which includes the uncertainty in the model, ranges from 0.14 to 0.4 and from 1.3 to 1.8% w/w, respectively. The QC values obtained showed no significant differences among all models obtained. PLS-DR-NIR models developed for determining heroin in seized illicit street drug samples were compared with an HPLC reference method (Figure 4). A statistical analysis of the four models built for heroin determination was performed by applying linear regression analysis of predicted versus reference data. In the present work, the accuracy of PLS-DR-NIR method was studied using ordinary Analytical Chemistry, Vol. 80, No. 19, October 1, 2008

7263

Figure 4. Prediction plots for PLS-DR-NIR determination of heroin in seized illicit street drugs evaluated using the actual reference data established by HPLC. Round points indicate predicted values obtained using a PLS model built from zero-order (A) and first-order derivative (B) spectra without SNV scattering correction and from zero-order (C) and first-order derivative (D) spectra previously corrected using the SNV algorithm. Solid line belongs to the ideal regression line. Together with them, plots of the joint confidence intervals for the intercept and the slope obtained using ordinary least-squares (OLS) and weighed least-squares (WLS) techniques, in which round points indicate the ellipse centroid while the cross indicates the theoretical point of zero intercept and unity slope, are also presented. The level of significance chosen for the joint confidence interval for the intercept and the slope was R ) 5% in all cases.

least-squares (OLS) (eq 1) and weighted least-squares (WLS) (eq 2) for the calculation of the regression parameters, xpredicted ) a + bxreference

(1)

wxpredicted ) awsqroot + bwxreference

(2)

which were calculated from the linear fit and were conveniently compared with the ideal values 1 and 0 using the elliptic joint confidence region test for the true slope β and intercept R, according to

(∑ x )(a - R)(b - β) + (∑ x ) (b - β) ) 2s F

n(a - R)2 + 2

2

i

2

2 2

i

(3) 7264

Analytical Chemistry, Vol. 80, No. 19, October 1, 2008

In eq 3, n is the number of points, s2 the regression variance, and F the critical value with 2 and n - 2 degrees of freedom at a given confidence level. The joint confidence intervals for the intercept and the slope obtained using both regression techniques, OLS and WLS, are shown in Figure 4. The level of significance chosen for the joint confidence interval was R ) 5% (confidence level of 95%) in all cases. The boundary of the ellipse was determined by the magnitude of experimental errors and by the degrees of confidence chosen. According to Figure 4, regression techniques include the ideal point (1, 0) inside the ellipses, for uncorrected and corrected PLS models built using zero-order spectra and first-order derivative spectra corrected using SNV. So it can be concluded that proportional and constant biases are absent and that the heroin predictions may be considered satisfactory.

Consequently, it can be concluded that the best results were obtained using zero-order SNV corrected PLS-DR-NIR spectra (Figure 4C). An additional confirmation of the robustness of the model selected was based on a segmented cross-validation carried out leaving out one full cluster each time. RMSECV value of 2.7% w/w, highly similar to 3.8% w/w value found for a full cross-validation, was obtained, thus indicating that the developed method has a high robustness considering completely new samples.

ing). Through the employment of this model, 1.3% w/w and 5.4, as RMSEP and RPD values, respectively, were obtained. A QC value of 10% was also obtained. The use of hierarchical cluster analysis, based on the DR-NIR spectra, proposed as a tool for the proper selection of calibration and validation sample sets, could be also useful to verify the nature of new seized samples and to evaluate the capabilities of the developed calibration model to correctly predict the heroin concentration in unknown samples.

CONCLUSIONS Results obtained in this work evidence that PLS-DR-NIR provides a rapid, accurate, and nondestructive analytical method for the determination of heroin in street-seized illicit drug samples. Since neither chemicals nor excessive time-consuming sample preparation processes are necessary, NIR spectroscopy provides an ideal analytical method for direct and instantaneous measurements of seized drugs. In short, the best PLS model obtained was that built using 8 factors from the spectral range between 1111 and 1647 nm for the zero-order spectra, previously corrected using the SNV algorithm joint together a quadratic removed baseline (detrend-

ACKNOWLEDGMENT The authors acknowledge the financial support of Ministerio de Educacio´n y Ciencia (Project CTQ2005-05604, FEDER), Direccio´ General d’Investigacio´ i Transfere`ncia Tecnolo`gica de la Generalitat Valenciana (Project ACOMP/2007/131) and Universitat de Vale`ncia (Convocato`ria d’Accions Especials, Project UVAE-20070213).

Received for review April 20, 2008. Accepted July 24, 2008. AC800781C

Analytical Chemistry, Vol. 80, No. 19, October 1, 2008

7265