(580a) Wavenumber Selection for NIR Calibration Modeling: Application to Water and Drug Content Estimation

Wavenumber Selection for NIR Calibration Modeling: Application to Water and Drug Content Estimation

Taku Uchimaru1, Takuya Miyano2, Koichi Fujiwara1, Manabu Kano1, Hideaki Tanabe2, Hiroshi Nakagawa2, Tomoyuki Watanabe2, Naoki Wakiyama3
1 Department of Systems Science, Kyoto University, Kyoto, Japan
2 Pharmaceutical Technology Division, Daiichi Sankyo Co., Ltd., Hiratsuka, Japan
3 Daiichi Sankyo Propharma Co., Ltd., Hiratsuka, Japan
Abstract:
To improve understanding of manufacturing processes, process analytical technology
(PAT) has played an important role. Although a large number of successful applications of PAT have been reported, it is still needed to answer the question: how to develop an accurate and robust calibration model based on near infrared (NIR) spectroscopy. In general, one has to carefully select data for modeling and validation, preprocessing methods, input variables, modeling methods, and so on to build an accurate and robust model. In calibration modeling, partial least squares (PLS) are usually adopted to cope with collineartiry among input variables, i.e., absorbance or reflectance at several hundred or thousand wavenumbers. Not all the input variables are necessary to estimate an output variable. In fact, to achieve highly accurate estimation, input variable selection is crucial.
In the present work, new efficient wavenumber selection methods are proposed to
achieve higher estimation accuracy than conventional methods and also to significantly reduce the computational load required by conventional wavenumber selection methods such as interval PLS. In the first method, referred to as R+VIP, two conventional indexes for input variable selection, i.e., the correlation coefficients R and variable influence on projection (VIP) are combined into one index. As a result, R+VIP has advantages of both indexes and functions effectively. The second method is named as spectral fluctuation dividing (SFD). In this method, a whole spectrum is divided into multiple spectral intervals at local minimum points of the spectral fluctuation profile, which consists of the standard deviation of absorbance at each wavenumber in a calibration set, and the spectral intervals of significant relationships with a target response are selected.
The usefulness of the proposed methods in terms of both improving the estimation accuracy and reducing the computational load is demonstrated through its application to the problems of estimating water and drug contents in granules.
Granules containing a drug substance (Daiichi-Sankyo, Japan) were used as analyte.
In granulation process, real-time NIR measurement was performed using a Fourier-transform NIR spectrometer MPA (Bruker GmbH, Germany) and equivalent Matrix-F (Bruker GmbH, Germany) through a fiber-optic probe mounted in the fluid bed granulator. The NIR spectra were obtained at every one minute during the granulation process. For blended granules, NIR measurement was performed using a Fourier-transform NIR spectrometer MPA (Bruker GmbH, Germany). In addition, granules were sampled from the fluid bed granulator at every ten minutes during granulation and at the end of both the spraying and the drying. Water content of the sampled granules was measured by the loss on drying (LOD) method, and drug content in the blended granules was measured by the high performance liquid chromatography (HPLC)
method.
In this work, the following wavenumber selection indexes/methods were compared.
1) Correlation coefficients (R)
2) PLS-beta
3) Variable influence on projection (VIP)
4) R+VIP
5) Interval PLS (iPLS)
6) Nearest Correlation Spectral Clustering (NCSC)
7) Spectral fluctuation dividing (SFD)
8) Moving averaged SFD (MASFD)
The results demonstrated the usefulness of R+VIP and SFD. In fact, the calibration models developed by using R+VIP were more accurate and robust than the models developed by using the conventional indexes. In addition, SFD improved estimation accuracy with less computational load compared with the conventional methods. Furthermore, MASFD, which combines SFD and the moving average method, reduced the number of spectral intervals without deteriorating the estimation accuracy. Hence, the proposed methods are superior to the conventional methods, and will be widely applicable to calibration modeling in PAT.