279145 Assessing the Impact of Experimental Designs On Pharmaceutical Calibration Models Based On near Infrared Spectroscopy

Wednesday, October 31, 2012
Hall B (Convention Center )
Robert Bondi Jr., Benoit Igne, Carl A. Anderson and James K. Drennen III, Mylan School of Pharmacy, Duquesne University, Pittsburgh, PA

The objective of this work was to determine how the predictive performance of partial least squares (PLS) models is affected by the calibration samples selected for model development.  Experimental designs are widely used to produce a set of calibration samples for generating empirical models based on vibrational spectroscopy (e.g., near-infrared).  There are numerous designs that can be utilized, but the choice often depends on the available resources and the intended use of the model.  However, the predictive performance of the model may be directly influenced as concentration covariance is directly related to regression vector estimation on a theoretical level.  It was hypothesized that the theoretical dependence of regression vector estimation on experimental designs is extended to practical applications.  The following designs were evaluated for a model pharmaceutical composite system: 5-level full factorial, 3-level full factorial, central composite, D-Optimal, and I-Optimal. The factors for all designs were acetaminophen content and ratio of microcrystalline cellulose to lactose monohydrate.  Other constituents included croscarmelose-sodium and magnesium stearate (content remained constant).  Five 13-mm compacts were produced at each level (5 levels for a total of 150 compacts).  Models were generated and independently optimized using data from individual design of experiments. The predictive performance of each model was evaluated based on an independent validation set.  The effect of experimental design was tested by determining the statistical significance of the difference in bias and standard deviation for its model prediction performance.

Extended Abstract: File Not Uploaded