Partial least squares (PLS) regression was used to develop the prediction model. PLS uses principal component analysis and
is a variation of principle component regression (PCR). The correct number of principal components or factors determination
is aided by the Vision software supplied with the instrument by calculating where the predicted residual error sum of squares
(PRESS) reaches a minimum (3). Although the PRESS does not reach a minimum until 13 factors, 6 were chosen as optimal allowing
more than 10 calibration samples for each factor and considering possible degrees of freedom.
Figure 6 is a plot of the principal component loadings around the 1138-nm absorption band for CPM. The loadings appear spectra-like
and are not noisy, indicating good modeling attributes for the factors chosen. The loading indicate regions of high correlated
variance throughout the spectrum for each factor. Figure 7 is a plot of the PRESS. The model chosen used only six of these
factors, trading decreased error for robustness (4). The PRESS for six factors was 0.0012. The resulting model had a multiple
correlation coefficient (R
2) value of 0.9987 and a standard error of calibration (SEC) of 0.0037 mg. The one-left-out cross validation demonstrates good
predictability with a standard error of cross validation of 0.004 mg (very close to the SEC).
Figure 8 shows the NIR predicted CPM concentrations versus the HPLC results for each tablet in the calibration set using the
synthetic calibration equation CPMsyncal2. The other five tablets scanned (tablet 6a through tablet 10a) on the MasterLab
instrument with nominal levels of 1.0 mg CPM, that were not in the calibration set, were predicted for validation. Figure
9 shows the NIR predictions of the validation set versus the Lab (HPLC) CPM value. Table II shows the predicted values for
the validation set. The standard error of prediction (SEP) of the validation set compared with the HPLC data was 0.0104 mg
and the bias was 0.0139 mg. The SEP is significantly larger then the SEC because the synthetic calibration samples have reduced
error except for the "seed" samples. The SEP is low at only about 1% of batch label claim.
Table II: Predicted values for the validation set with nominal value of 1.0 mg.
The authors demonstrated a method for generating a prediction model from placebo and "seed" batch samples from the production
line with synthetic samples generated from average placebo and batch samples to increase the sample range. A prediction model
with an R
2 value of 0.9987 and a standard error of calibration (SEC) value of 0.0037 mg was developed. The standard error of prediction
(SEP) on the validation set was only 1% of label claim at 0.0104 mg.
Thanks to Om Anand, Maria Gerald Rajan, Namrata R. Trivedi, Wen Qu, Yingxu Peng, and Yichun Sun who were at the University
of Tennessee, Department of Pharmaceutical Sciences, at the time the CPM tablets were made and analyzed with HPLC.
Robert Mattes* is an applications scientist and Denise Root is a product manager, both at FOSS NIRSystems, rmattes@FOSS-NIRSystems.com
*To whom all correspondence should be addressed.
Submitted: July 14, 2008. Accepted: Aug. 12, 2008.
What would you do differently? Submit your comments about this paper in the space below.
1. FDA, Guidance for Industry: PAT-A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance (September, 2004, Rockville, MD).
2. R.A. Mattes et al., "Near-Infrared Assay and Content Uniformity of Tablets," Pharm. Technol.
31 (4), 170–182 2007.
3. R. Kramer, Chemometric Techniques for Quantitative Analysis (Marcel Dekker, Inc. 1998).
4. K.R. Beebe, R.J. Pell, and M.B. Seasholtz, Chemometrics: A Practical Guide (John Wiley & Sons, 1998).