Multivariate modeling of diffuse reflectance infrared fourier transform (DRIFT) spectra of mixtures with low-content polymorphic impurities with analysis of outliers

Abstract

Diffuse Reflectance Fourier Transform Infrared Spectroscopy (DRIFTS)-based multivariate models were developed to quantify the content of two polymorphic impurities in mixtures with the desired active pharmaceutical ingredient (API) form, with the impurities not exceeding 2% wt/wt. In addition, close attention was paid to the outlier detection criteria: Q residuals; Hotelling T2; and score bi-plot. While reasonably accurate results were obtained for the relatively simple calibration models for both forms of the impurity, the predictions for “blank” samples (separately verified to be impurity-free) were apparently biased. Thus, the model training sets were augmented with spectra from calibration mixtures incorporating some of the API from batches used in the prediction. The performance of the updated models as assessed by cross-validation was somewhat degraded as a result, while predictions against independent batches of API showed a decrease in bias indicating robustness had improved. Nevertheless, the Q residuals criterion disqualified a large number of prediction samples as outliers in contrast to the other two criteria that reported no issues at all. The results here demonstrated the effectiveness of DRIFTS for quantifying low concentration polymorphic impurities, while simultaneously highlighting the variability issues that can be encountered in practice and which need to be understood and managed appropriately to ensure the success of any automated or Good Manufacturing Practice (GMP) compliant application of multivariate modeling.