Multivariate Analysis Applied to the Study of Spatial Distributions

May 30, 2008 - Strategies for quantifying the Raman response include calculation of drug specific Raman ... microscopy (CRM) as an analytical tool to ...
0 downloads 0 Views 584KB Size
Download PDF
Anal. Chem. 2008, 80, 4853–4859

Multivariate Analysis Applied to the Study of Spatial Distributions Found in Drug-Eluting Stent Coatings by Confocal Raman Microscopy Karin M. Balss,*,† Frederick H. Long,‡ Vladimir Veselov,† Argjenta Orana,† Eugena Akerman-Revis,† George Papandreou,† and Cynthia A. Maryanoff† Cordis Corporation, a Johnson & Johnson Company, Welsh and McKean Roads, Spring House, Pennsylvania 19477, and Spectroscopic Solutions, 665 Millbrook Avenue, Randolph, New Jersey 07869 Multivariate data analysis was applied to confocal Raman measurements on stents coated with the polymers and drug used in the CYPHER Sirolimus-eluting Coronary Stents. Partial least-squares (PLS) regression was used to establish three independent calibration curves for the coating constituents: sirolimus, poly(n-butyl methacrylate) [PBMA], and poly(ethylene-co-vinyl acetate) [PEVA]. The PLS calibrations were based on average spectra generated from each spatial location profiled. The PLS models were tested on six unknown stent samples to assess accuracy and precision. The wt % difference between PLS predictions and laboratory assay values for sirolimus was less than 1 wt % for the composite of the six unknowns, while the polymer models were estimated to be less than 0.5 wt % difference for the combined samples. The linearity and specificity of the three PLS models were also demonstrated with the three PLS models. In contrast to earlier univariate models, the PLS models achieved mass balance with better accuracy. This analysis was extended to evaluate the spatial distribution of the three constituents. Quantitative bitmap images of drug-eluting stent coatings are presented for the first time to assess the local distribution of components. Drug-eluting stents (DES) are slotted tubes, usually made of metal, that contain a combination of drug and polymer. The stent acts as scaffolding to keep the artery open while the drug prevents restenosis, i.e., scarlike tissue growth through the stent. Two commercial products that have been implanted in millions of patients with coronary artery disease worldwide are the CYPHER Sirolimus-eluting Coronary Stents (Cordis Corporation) and the TAXUS Express2 Coronary Stents (Boston Scientific). These products achieve the controlled release of drugs that interfere with the cell cycle in the diseased vessels, and in clinical trials they have demonstrated good long-term safety, while significantly reducing restenosis.1–3 The clinical effectiveness of DES is dependent on the amount of drug contained (drug dose) and how * To whom correspondence should be addressed. E-mail: kbalss@ crdus.jnj.com. † Cordis Corporation. ‡ Spectroscopic Solutions. (1) Klugherz, B. D.; Llanos, G.; Lieuallen, W.; Kopia, G. A.; Papandreou, G.; Narayan, P.; Sasseen, B.; Adelman, S. J.; Falotico, R.; Wilensky, R. L. Coron. Artery Dis. 2002, 13, 183–188. 10.1021/ac7025767 CCC: $40.75  2008 American Chemical Society Published on Web 05/30/2008

the drug is released temporally in vivo (elution profile). Factors influencing these performance parameters include the stent platform design, drug and polymer formulation, and process.4–10 The CYPHER Stent product contains a thin conformal coating around the stent struts. The coating consists of an immiscible blend of two polymers and an active pharmaceutical ingredient, sirolimus. Qualitative chemical mapping is useful to identify the presence of species within defined spatial locations. However, quantifying the spatial distribution of each component is key to the fundamental understanding of the bulk performance properties observed (drug content and elution profiles). The requirements for an ideal tool to quantify spatial distribution of components are that it must be chemically specific for positive identification of each component, possess the spatial resolution to resolve features within regions of interest, and map the desired spatial profile (surface, depth, or bulk). It is also desirable but not required that the tool be nondestructive so that additional complementary experiments or additional nonrelated measurements are possible. The confocal Raman microscope meets these criteria and is an ideal tool to study drug-eluting stents. Qualitative mapping has been demonstrated by our group on nonerodable drug-polymer systems11 while others have reported on degradable drug-polymer systems.12,13 (2) Suzuki, T.; Kopia, G.; Hayashi, S.-I.; Bailey, L. R.; Llanos, G.; Wilensky, R.; Klugherz, B. D.; Papandreou, G.; Narayan, P.; Leon, M. B.; Yeung, A. C.; Tio, F.; Tsao, P. S.; Falotico, R.; Carter, A. J. Circulation 2001, 104, 1188– 1193. (3) Maeng, M.; Okkels Jensen, L.; Rasmussen, K.; Flensted Lassen, J.; Romer Krusell, L.; Thayssen, P.; Thuesen, L. Heart 2007, 93, 694–697. (4) Suzuki, Y.; Ikeno, F.; Yeung, A. C. J. Invasive Cardiol. 2006, 18, 111–114. (5) Sipos, L.; Som, A.; Faust, R.; Richard, R.; Schwarz, M.; Ranade, S.; Boden, M.; Chan, K. Biomacromolecules 2005, 6, 2570–2582. (6) Ranade, S. V.; Miller, K. M.; Richard, R. E.; Chan, A. K.; Allen, M. J.; Helmus, M. N. J. Biomed. Mater. Res., Part A 2004, 71A, 625–634. (7) Chen, M.-C.; Liang, H.-F.; Chiu, Y.-L.; Chang, Y.; Wei, H.-J.; Sung, H.-W. J. Controlled Release 2005, 108, 178–189. (8) Dong, Y.; Feng, S.-S. J. Biomed. Mater. Res., Part A 2006, 78A, 12–19. (9) Jewell, C. M.; Zhang, J.; Fredin, N. J.; Wolff, M. R.; Hacker, T. A.; Lynn, D. M. Biomacromolecules 2006, 7, 2483–2491. (10) Finkelstein, A.; McClean, D.; Kar, S.; Takizawa, K.; Varghese, K.; Baek, N.; Park, K.; Fishbein, M. C.; Makkar, R.; Litvack, F.; Eigler, N. L. Circulation 2003, 107, 777–784. (11) Balss, K. M.; Llanos, G.; Papandreou, G.; Maryanoff, C. A. J. Biomed. Mater. Res., Part A 2008, 85A, 258–270. (12) Belu, A.; Mahoney, C.; Wormuth, K. J. Controlled Release 2008, 126, 111– 121. (13) Hsu, S. L. Am. Pharm. Rev. 2006, 58–64.

Analytical Chemistry, Vol. 80, No. 13, July 1, 2008

4853

Quantitative Raman spectroscopy for predicting concentrations within mixtures is well established in the literature. Increasingly, Raman spectroscopy and microscopy are used to study active pharmaceutical ingredients in tablets,14–17 transdermal delivery devices,18 and other pharmaceutical formulations.19,20 Raman spectroscopy is also successful at describing polymer mixtures.21–24 Strategies for quantifying the Raman response include calculation of drug specific Raman band areas, ratios of bands, or multivariate analysis. Internal standards are also used routinely. Quantifying chemical entities as a function of depth has also been demonstrated by confocal Raman microscopy in living cancer cells,25 human stratum corneum,26 adhesive-dentin interfaces,27 drugs in solid dispersions,28 and phase separated polymers.29,30 Recently, we demonstrated the feasibility of confocal Raman microscopy (CRM) as an analytical tool to describe both drug and polymer components found in the CYPHER Stent.11 Calibration curves based on ratios of Raman bands specific for each component were generated and tested on independent samples. The individual models performed with reasonable accuracy and precision on unknown test samples. The drug model had the smallest accuracy error when assessed against an HPLC assay value, while the polymer models experienced larger accuracy errors when assessed against the gravimetric spray solution content. The sources of error in the Raman measurements included interference from fluorescence, laser power fluctuations, and signal intensity variations as a function of depth within a coating. Additional sources of error for the univariate models included selection of polymer-specific bands and avoiding interference from other components. Although the previous models were adequate for predicting concentrations, the self-consistency between models needed improvement. We employed baseline correction and peak fitting functions to minimize sources of variation, but this strategy did not produce self-consistency among the three univariate models because they did not achieve mass balance.11 (14) Szostak, R.; Mazurek, S. Analyst (Cambridge, U.K.) 2002, 127, 144–148. (15) Szep, A.; Marosi, G.; Marosfoei, B.; Anna, P.; Mohammed-Ziegler, I.; Viragh, M. Polym. Adv. Technol. 2003, 14, 784–789. (16) Dyrby, M.; Engelsen, S. B.; Norgaard, L.; Bruhn, M.; Lundsberg-Nielsen, L. Appl. Spectrosc. 2002, 56, 579–585. (17) Williams, A. C.; Cooper, V. B.; Thomas, L.; Griffith, L. J.; Petts, C. R.; Booth, S. W. Int. J. Pharm. 2004, 275, 29–39. (18) Armstrong, C. L.; Edwards, H. G. M.; Farwell, D. W.; Williams, A. C. Vib. Spectrosc. 1996, 11, 105–113. (19) Davies, M. C.; Binns, J. S.; Melia, C. D.; Hendra, P. J.; Bourgeois, D.; Church, S. P.; Stephenson, P. J. Int. J. Pharm. 1990, 66, 223–232. (20) Yang, H.; Irudayaraj, J. J. Pharm. Pharmacol. 2002, 54, 1247–1255. (21) Chalmers, J. M.; Everall, N. J. TrAC, Trends Anal. Chem. 1996, 15, 18– 25. (22) Shenton, M. J.; Herman, H.; Stevens, G. C. Polym. Int. 2000, 49, 1007– 1013. (23) Shimoyama, M.; Maeda, H.; Matsukawa, K.; Inoue, H.; Ninomiya, T.; Ozaki, Y. Vib. Spectrosc. 1997, 14, 253–259. (24) Workman, J. J., Jr. Spectrosc. Lett. 1999, 32, 1057–1071. (25) Feofanov, A. V.; Grichine, A. I.; Shitova, L. A.; Karmakova, T. A.; Yakubovskaya, R. I.; Egret-Charlier, M.; Vigny, P. Biophys. J. 2000, 78, 499–512. (26) Caspers, P. J.; Williams, A. C.; Carter, E. A.; Edwards, H. G. M.; Barry, B. W.; Bruining, H. A.; Puppels, G. J. Pharm. Res. 2002, 19, 1577–1580. (27) Wang, Y.; Spencer, P. J. Biomed. Mater. Res. 2002, 59, 46–55. (28) Breitenbach, J.; Schrof, W.; Neumann, J. Pharm. Res. 1999, 16, 1109– 1113. (29) Pudney, P. D. A.; Hancewicz, T. M.; Cunningham, D. G.; Gray, C. Food Hydrocolloids 2003, 17, 345–353. (30) Maeda, Y.; Yamamoto, H.; Ikeda, I. Macromolecules 2003, 36, 5055–5057.

4854

Analytical Chemistry, Vol. 80, No. 13, July 1, 2008

Table 1. DOE Mixture Design for Formulation Standards group

drug wt % (%)

PEVA wt % (%)

PBMA wt % (%)

ratio PBMA/PEVA

A B C D E F G H I J K L

20.0 20.0 20.0 28.4 28.4 28.4 40.0 40.0 40.0 0.0 28.4 28.4

24.2 55.8 31.7 21.6 50.0 28.4 18.1 41.9 23.8 50.0 11.6 60.0

55.8 24.2 48.3 50.0 21.6 43.2 41.9 18.1 36.2 50.0 60.0 11.6

2.31 0.43 1.52 2.31 0.43 1.52 2.31 0.43 1.52 1.00 5.17 0.19

In this work, we seek to enhance the calibration models by applying spectral preprocessing strategies and multivariate data analysis to improve the accuracy and precision of predictions on unknowns. Furthermore, the range of drug and polymer concentrations was narrowed from 0 to 100% to a range more applicable to the commercial drug-polymer formulation. The formulations were part of a design of experiments (DOE) intended to independently vary the three constituents and minimize the necessary number of samples for analysis. We monitored the CRM response as a function of drug-polymer formulation. Partial least-squares (PLS) regression coupled with preprocessing of the spectra generated quantitative models for sirolimus, PEVA, and PBMA. The models were then used to predict the formulation on unknowns. A total of six stents from a single production lot not used for the calibration were analyzed to assess the PLS methods. In addition, the first quantitative spatial distribution maps of individual components within drug-eluting stents are shown. MATERIALS AND METHODS Materials. Sirolimus, poly(ethylene-co-vinyl acetate) [PEVA], and poly(n-butyl methacrylate) [PBMA] were obtained in-house and used as received. Poly(o-chloro-p-xylylene) [parylene-C] was deposited as a discrete layer onto the 6 cell by 33 mm 316L stainless steel stents before the drug-polymer coating. A series of solutions containing sirolimus, PEVA, and PBMA were prepared in tetrahydrafuran (THF) and spray-coated onto the stents. The mass of each individual stent was recorded before and after spraying to obtain the total stent coating mass. An experimental mixture design was used to determine the formulations studied. Table 1 summarizes the composition of the solutions that were spray-coated onto stents. A target weight of 1.2 × 103 µg of each formulation was applied on the 6 cell by 33 mm stents. After spray coating, the stents from each formulation were sterilized with ethylene oxide. Sterile samples were then analyzed by CRM. In addition, samples were extracted and the extracts assayed by highperformance liquid chromatography (HPLC) analysis. The independent test samples used to validate the component quantitative models were prepared on pretreated parylene-C, 7 cell by 33 mm stainless steel stents. A target weight of 1.4 × 103 µg of a formulation containing 28.4, 28.4, and 43.2 wt % sirolimus, PEVA, and PBMA was applied to these stents. DOE Mixture Design. The formulation standards chosen to build the calibration models for sirolimus, PEVA, and PMBA were

based on a full factorial design of experiment, which results in a modified extreme-vertex mixture design. In our previous work, a DOE design contained formulations that spanned a large concentration range with the purpose of gauging the linearity of the Raman response from 0 to 100%. In this study, we were interested in generating standards that would represent concentration ranges that bracket commercial product formulations. We also added a placebo formulation to challenge the sirolimus model. The drug concentration design contained three different concentrations of drug; 20, 28.4, and 40 wt %. The PBMA and PEVA concentrations were varied according to their ratio: 0.43, 1.52, and 2.31. The design focused on concentrations that would represent a small, mid, and large PBMA to PEVA polymer ratio. We added three additional formulations to the model: a placebo (0% drug), an extreme high (5.17), and an extreme low (0.19) PBMA to PEVA ratio. Drug Content Assay. The total amount of sirolimus present in units of micrograms per stent was determined by reverse phase high-performance liquid chromatography with UV detection (RPHPLC-UV). The analysis was performed with an in-house modification of literature-based HPLC methods for sirolimus.31,32 The average drug content of five sterile stents from each formulation was reported and used for the PLS laboratory values. In addition, the average drug content of five sterile stents from the unknown lot was measured for comparison to the PLS predicted value. Confocal Raman Microscopy. Spectral depth profiles of the standard and unknown samples were performed with a CRM200 microscope system from WITec Instruments Corporation (Savoy, IL). The instrument was equipped with a Nd:YAG frequency doubled laser (532 nm excitation), a single monochromator (Acton) employing a 600 groove/mm grating and a thermoelectrically cooled 1024 by 128 pixel array CCD camera (Andor Technology). The microscope was equipped with appropriate collection optics that included a holographic laser bandpass rejection filter (Kaiser Optical Systems Inc.) to minimize Rayleigh scatter into the monochromator. The Raman scattered light was collected with a 50 µm optical fiber. With the use of the “Raman Spectral Imaging” mode of the instrument, spectral images were obtained by scanning the sample in the x, z direction with a piezo driven xyz scan stage and collecting a spectrum at every pixel. Typical integration times were 0.3 s per pixel. The spectral images consisted of 4800 total spectra corresponding to a physical scan dimension of 40 by 20 µm. For presentation of the confocal Raman data, images were generated based on unique properties of the spectra (i.e., integration of a Raman band, band height intensity, or bandwidth). The microscope stage was modified with a custombuilt sample holder that positioned and rotated the stents around their primary axis. The x direction is defined as the direction running parallel to the length of the stent and the z direction refers to the direction penetrating through the coating from the aircoating to the coating-metal interface. Typical laser power was less than 10 mW on the sample stage. All experiments were conducted with a plan achromat objective, 100 × NA ) 0.9 (Nikon). (31) Napoli, K. L.; Kahan, B. D. Clin. Chem. (Washington, DC, U.S.) 1996, 42, 1943–1948. (32) Wang, C. P.; Scatina, J.; Sisenwine, S. F. J. Liq. Chromatogr. 1995, 18, 1801–1808.

For each stent, six locations were selected along the length. The three locations were located within one-third portions of the stents so that the entire length of the stent was represented in the data. The stent was then rotated 180° around the circumference, and an additional three locations were sampled along the length. In each case, the data were collected from the strut portion of the stent. We chose to neglect the link portion for analysis in our models. The strut was chosen because of the simplicity to align an xz profile for reproducible spatial areas of analysis. The Supporting Information contains illustrations of the stent and portions analyzed by CRM. The Raman spectra of each individual component present in the formulation were also collected for comparison and reference. With the use of the instrument software, the average spectra from the spectral image data were calculated by selecting the spectral image pixels that were exclusive to the active drug-polymer layer. Therefore, both the parylene-C polymer layer and air regions were removed prior to data analysis. Spectral Image Processing and Partial Least-Squares (PLS) Regression Model. The PLS calibration regression models were built from the Raman data of the 12 formulation standards shown in Table 1. An additional two stents were analyzed from group F, for a total of 14 stents used in the calibration models. With the use of the WITec software, each spectral image array from individual spatial locations was reduced to an average spectrum. The average spectra were imported in the Unscrambler Software version 9.6 (Camo Software Inc. Woodbridge, NJ). These average spectra were then used to build three separate calibration models for sirolimus, PEVA, and PBMA. The spectral preprocessing, outlier detection, factor selection, and prediction were performed using the Unscrambler software. The spectral preprocessing included a moving average with segment 3 followed by a standard normal variate (SNV) function. We found this to be adequate in reducing spectral variations due to laser power fluctuations, scattering effects, coating thickness and depth profiling, and fluorescence background. A Savitsky Golay first derivative with 11-point smoothing was applied as an additional preprocessing step for the sirolimus model. PLS Model for Quantitative Images. Quantitative images were generated in Isys software version 4 (Malvern Instruments). PLS models were generated with the same spectral data described in the previous section. For all three models SNV preprocessing was used. Intensity thresholding was used to remove the air region and the parylene-C layer below the polymer-drug coating. RESULTS AND DISCUSSION Confocal Raman Spectroscopy. The CYPHER Stent coating is composed of an active drug-polymer layer containing sirolimus, PBMA, and PEVA as well as an inert layer of parylene-C deposited on the metal stent. The Raman spectra of each component found in the CYPHER Stent product are illustrated in Figure 1. Each component has a unique spectral profile with several individual Raman band assignments reported previously.11 Previous work in our group characterized the Raman spectra of each individual component and identified unique spectral signatures for each component. The triene band at 1634 cm-1 is particularly useful for qualitative identification of sirolimus (Figure 1d). Similarly, the band between 1265-1365 cm-1 is assigned to a CH in-plane Analytical Chemistry, Vol. 80, No. 13, July 1, 2008

4855

Figure 1. Raman spectra collected from neat components of (a) parylene-C, (b) PBMA, (c) PEVA, and (d) sirolimus.

deformation for identification of parylene-C (Figure 1a).33 Unique bands exist for PBMA and PEVA at 600 and 630 cm-1, respectively (Figure 1b,c). Additional PBMA bands occur at 850 and 940 cm-1. Confocal spectral images can be generated based on the integral of these spectral regions. We exploit the parylene-C signal between 1265-1365 cm-1 to distinguish pixels corresponding to the active drug-polymer layer from the parylene-C layer. These spatial regions are then used to generate average spectra exclusively for the active drug-polymer layer. Spectral depth profiles of drug-polymer coatings on stents are shown in Figure 2. The cross-sectional images represent the drug-polymer coating from the air-coating to coating-metal

interface at three different sirolimus concentrations. Each pixel in the images contains an entire Raman spectrum. The images were constructed by integration of the unique triene band. The signal for the triene band is observed throughout the entire coating with intensity variations (indicated by light and dark regions) observed within the 40 by 20 µm spatial region (Figure 2a-c). The intensity variations become more pronounced at higher concentrations. To properly assess the distribution of drug within a polymer matrix, the spectra within the active drug-polymer coating containing sirolimus are separated from spectra that are exclusive to the parylene-C layer as well as the regions of air. The average spectra calculated from these images are illustrated in Figure 2d. Notice the signal for the triene band at 1634 cm-1 increases with increasing sirolimus concentration. To develop the PLS models, the average spectrum was calculated for each spatial location profiled (Figure 2d) from the calibration set samples (Table 1). A total of 6 spatial locations per stent for a total of 84 spatial locations were included in the models. The average spectrum is representative of pixels corresponding to the active drug-polymer layer. The number and size of the spatial locations collected per stent is assumed to represent the bulk concentration. Our previous work established that additional spatial sampling does not reduce the variance observed.11 However, we recognize that spatial heterogeneity both in spatial location and depth profile exists in our coating samples. Atomic force microscopy supports immiscibility on the micrometer to submicrometer length scale.34 This heterogeneity is the reason we desire to develop a quantitative method; therefore, we assume that by sampling an area larger than the

Figure 2. Qualitative images taken from the integrated area 1600-1700 cm-1 representative of sirolimus distribution at (a) 20 wt %, (b) 28 wt %, and (c) 40 wt % drug concentrations. The average spectra from the drug-polymer coatings are shown in part d. The yellow scale bar is 8 µm. 4856

Analytical Chemistry, Vol. 80, No. 13, July 1, 2008

Figure 3. A comparison of the raw average spectrum (top) and the processed spectrum (bottom) after (a) a moving average, SNV, and Savitsky Golay 1st derivative was applied, (b) regression of the predicted sirolimus concentration versus HPLC assay results, (c) regression of the predicted PBMA concentration versus solution content, and (d) regression of the predicted PEVA concentration versus solution content. Table 2. Calibration Model Correlation Coefficients (r), SEP, and SEC for Each Component

sirolimus PBMA PEVA

r

SEPCV (%)

SEC (%)

0.941 0.955 0.958

3.2 4.1 2.8

3.1 3.8 2.7

heterogeneity, we can achieve average values representative of the bulk. Sirolimus, PBMA, and PEVA Coating Components PLS Regression Models. Figure 3 contains an example of spectral preprocessing as well as the regression curves for each component model. Figure 3a highlights the effectiveness of spectral preprocessing to eliminate background. The regression of the predicted concentration versus laboratory assay is shown in Figure 3b-d. A PLS model with two factors was developed for the sirolimus calibration. In the case of sirolimus, the spectrum from one spatial location of the group A standard (Table 1) was removed as an outlier. A summary of the correlation coefficient and the standard error of calibration and prediction (SEC and SEPCV, respectively) are shown in Table 2. A reasonable fit is obtained with the standard errors of calibration and prediction around 3 wt %. For comparison, the HPLC assay has an acceptable precision of 2% RSD. For example, a formulation of 28.4 wt % translates to a variation of ±1.4 wt %. After examination of the residual variance plots, three-factor models were chosen for the PBMA and PEVA models. After inspection of the score plots, three spectra were removed from the PBMA model and four spectra removed from the PEVA model. The standard error of prediction and calibration

were only slightly higher (4 wt %) compared to the sirolimus model for the PBMA model while the PEVA model was similar (≈3 wt %). Evaluation of PLS Models. Further evaluation of the PLS models can be found in the Supporting Information. The PLS analysis described is highly specific. The specificity of the polymer models is greatly improved compared to our previous univariate modeling.11 The regression coefficient curves from the PLS models describe the relative importance of different spectral regions to the calibration model. Inspection of the regression coefficient curves (Supporting Information) reveals unique trends attributed to each component. For example, the spectral region 700-900 cm-1 shows a positive response for PBMA and negative for PEVA. Peaks in the PBMA regression coefficient are consistent with Raman bands in the pure component spectrum at 850 and 940 cm-1. The methylene stretching region 2600-3300 cm-1 are mirror images of each other, highlighting the specificity of each model to the individual polymer component. The CH region exhibits a unique peak for PEVA at 2850 cm-1. The spectral region between 1000-1700 cm-1 has a unique response for sirolimus compared to PEVA and PBMA. The 1634 cm-1 band for sirolimus is clearly seen in the PLS regression coefficient. The PLS regression models are also linear (see Supporting Information). Sirolimus and PEVA show excellent linearity across the concentration range tested (0-40 wt % for sirolimus and 12-60 wt % for PEVA). The PBMA normality plot shows a slight nonlinearity. Because the PBMA calibration was constructed using (33) Mathur, M. S.; Weir, N. A. J. Mol. Struct. 1973, 15, 459–463. (34) Cordis Corporation, unpublished data, 2007.

Analytical Chemistry, Vol. 80, No. 13, July 1, 2008

4857

Table 3. Summary of Predictions Compared to Laboratory Values

avg sirolimus avg PBMA avg PEVA a

predicted amount (%)

SD (%)

laboratory value (%)

wt % difference (%)

accuracy (%)

28.3 42.8 28.8

3.4 2.7 2.2

27.3a 43.2b 28.4b

-1.0 0.4 -0.4

-3.5 1.0 -1.5

HPLC assay value. b Gravimetric solution content.

only formulation values and not laboratory measurements, this slight nonlinearity is not a critical issue. Independent Test Samples. The judgment of whether a PLS model is performing well can be examined by the various outputs of the regression model, such as the predicted vs laboratory assay, standard errors of prediction and calibration, score plots, residual variance, and normality plots. However, a true measure of the performance of a calibration model is to analyze independent sample sets or “unknowns”. Test samples from the same lot were analyzed by CRM and HPLC. For sirolimus concentration, the drug content is expressed as a wt %; the HPLC assay content (composite of five stents) is divided by the average coating mass of the five stents. The polymer content is estimated from the gravimetric solution content divided by the coating mass. With the use of the PLS calibration curves, predictions for sirolimus, PBMA, and PEVA were performed. Table 3 contains a summary of the predictions for each component. The individual stent predictions can be found in the Supporting Information. In all individual stent samples, the sirolimus predictions were within 4 wt % of the laboratory assay value and the standard deviation within an individual sample was less than the model standard error of prediction (SEPCV) and standard error of calibration (SEC). This translates to accuracy errors less than 15%. Similar agreement between predictions and theoretical concentrations is found with the polymer predictions. The predicted content was within 3.1 wt % of the solution content, and the standard deviation was less than the SEPCV and SEC for the model. For PEVA, the predicted content was within 2.7 wt % of the solution content, and the standard deviation was less than the SEPCV and SEC for the model. Previously, the univariate polymer model predictions were within 7 wt % of the theoretical concentration, resulting in accuracy errors greater than 20% compared to less than 10% error achieved in this work. Furthermore, the three independent models achieved mass balance indicating that the models were self-consistent and could successfully predict each individual component. The sum of all predictions ranged from 99.1 to 100.8 wt %. Spatial Distribution of Components. The power of quantitative CRM was demonstrated by our ability to predict the bulk concentrations of each individual component in unknown samples even though the data is collected within a much smaller portion of the stent (spatial regions of 400 µm2) compared to a bulk measurement. To further extend the quantitative stent coating analysis method, we wish to probe the spatial distribution on the smallest length scale possible. The qualitative images of the formulation blend such as those seen in Figure 2 illustrate variations in sirolimus signal with concentration rich regions that vary in diameter from a few micrometers to several tens of 4858

Analytical Chemistry, Vol. 80, No. 13, July 1, 2008

micrometers. Subtle variations in PEVA and PBMA signal are harder to detect with single band sum images. However, the concentration variations become apparent when multivariate analysis is utilized. PLS models created by a spectral image analysis software package were used to create quantitative bitmap images that describe the concentration of sirolimus, PEVA, and PBMA. The same set of Raman data of formulation standards was used to build the PLS models as described in the previous section. We chose a three-factor model for sirolimus and PBMA and a four-factor model for PEVA based on the plot of the PRESS (predicted residual sum of squares) versus the number of factors. An example is illustrated in Figure 4. In this location, the average concentrations of drug, PEVA, and PBMA were 27.6, 24.3, and 38.2 wt %, respectively. The bulk concentrations of drug, PEVA, and PBMA for the samples were 28.4, 28.4, and 43.2 wt %, respectively. However, as shown in the bitmap image in Figure 4a, the color fluctuations suggest sirolimus-rich regions close to the interface of the coating with air, containing drug with greater than the average bulk concentration. Similarly, PEVA-rich regions were observed close to the parylene-C interface, Figure 4b. This visualization is not possible with single Raman bands for PEVA due to the low signal-to-noise for PEVA specific Raman bands. PBMA-rich regions were also observed that are not easily visualized by PBMA specific Raman bands, Figure 4c. The selfconsistency of the three PLS image models is approximately 90 wt % compared to the 100 wt % achieved with the average prediction models, Figure 4d. The relatively large error in mass balance is expected for image generation compared to average predictions due to the decrease in signal-to-noise for an individual pixel. A challenge for processing our images is elimination of the spatial regions corresponding to the discrete parylene-C polymer layer. The removal of the parylene-C layer pixels was accomplished by setting an intensity thresholding function. In some spectral images, incomplete removal of pixels corresponding to parylene-C was observed. Comparison of the 3-bitmap image scale bars reveals that none of the pixels contains 100% of an individual component. As mentioned previously, it is known from theory and other experiments in our group that the drug and the two polymers are immiscible.34 However, the typical length scale of the polymer domains is much smaller than the length scale of the confocal Raman measurements. The sampling in the lateral direction is approximately 300 nm while the depth direction is 500 nm. We are therefore sampling at approximately twice the depth resolution (1 µm). With this resolution, we do not observe regions containing a single component of either drug or polymer. In addition, we recognize there will be a small contribution of signal variations as a function of depth through the coating. This variation, however, does not account for the concentration rich regions of sirolimus and PBMA observed near the surface of the coating. CONCLUSIONS AND FUTURE DIRECTIONS PLS is an appropriate multivariate strategy to quantitate drug-polymer coatings on stents. Multivariate data analysis of drug-polymer coatings found on DES was successful in describing the average content of sirolimus, PBMA, and PEVA on six unknown test samples from the same manufactured lot. Furthermore, the sum of the predictions achieved mass balance for each

Figure 4. Quantitative single component image maps for (a) sirolimus, (b) PEVA, and (c) PBMA. The scale bar is in units of wt %. The sum image of the three models is shown in part d.

test sample. In contrast, univariate models did not demonstrate mass balance. This work summarized the method accuracy, precision, linearity, and specificity. Our future efforts will focus on the further development and interpretation of quantitative PLS image results. The quantitative images will prove valuable in understanding spatial distribution changes as a function of manufacturing process conditions. ACKNOWLEDGMENT The authors would like to acknowledge Cordis Corporation for funding this work, the Cordis Analytical Technologies Group for performing the HPLC assays, and the Cordis Process Develop-

ment team for preparing the samples used in this study. The authors also acknowledge Fred Koehler of Malvern Instruments for assistance with image analysis in Isys software. SUPPORTING INFORMATION AVAILABLE Additional information describing the PLS models as noted in the text. This material is available free of charge via the Internet at http://pubs.acs.org. Received for review December 19, 2007. Accepted April 22, 2008. AC7025767

Analytical Chemistry, Vol. 80, No. 13, July 1, 2008

4859