(580d) Model Development for NIR-Based Real-Time Monitoring of Ingredient Concentration

Model Development for NIR-Based Real-Time Monitoring of Ingredient Concentration

Hiroshi Nakagawa1, Manabu Kano2, Shinji Hasebe2, Takuya Miyano1, Tomoyuki Watanabe1, Naoki Wakiyama3
1 Pharmaceutical Technology Division, Daiichi Sankyo Co., Ltd., Hiratsuka, Japan
2 Department of Systems Science, Kyoto University, Kyoto, Japan
3 Department of Chemical Engineering, Kyoto University, Kyoto, Japan
4 Daiichi Sankyo Propharma Co., Ltd., Hiratsuka, Japan
Abstract:
There has been considerable research on process analytical technology (PAT) and
real-time monitoring based on NIR, but the model development is still an important issue and persons in charge have difficulty in building good models. In the present work, to verify model development techniques in their applications to real-time monitoring of ingredient concentration with NIR during blending, we evaluated the effect of calibration sets, spectral analysis techniques, modeling parameters and manufacturing conditions on the prediction. Then, we proposed a model development procedure on the basis of the evaluation results. The analytical performance of the developed models under various conditions was evaluated on the basis of the fitting performance to the calibration set, the prediction error for data sets acquired with various blending equipment, and the analytical validation according to the United States pharmacopeia (USP <1119>) and draft guideline on the use of NIR.
Granules including the API (Daiichi-Sankyo, Japan), which was under development, were used in this work. The granules were manufactured with fluid-bed granulators and diffusion mixers with V-blender and bin blenders. The samples with nine different API concentrations (70%, 75%, 85%, 90%, 100%, 110%, 115%, 125%, and 130%) were manufactured with various blenders of scale 1 kg, 50 kg, 200 kg, 400 kg, and 500 kg. Corona (Type: Remote NIR-HR, Carl Zeiss, Germany), which has a diode array type NIR spectrometer with 3 nm resolution, was used for real-time monitoring (dynamic measurement). Blending experiments with the various blenders, scales, and manufacturing conditions were conducted to acquire spectra by dynamic measurement based on the diffuse reflectance method. Static measurement with Corona was also conducted under the same measurement conditions as dynamic measurement.
To build a model that can accurately predict the API concentration, a calibration set, a
spectral analysis technique, and modeling parameters need to be selected appropriately. Here, modeling parameters represent spectral preprocessing techniques and wavelengths.
Conventional partial least squares (PLS) and locally weighted PLS (LW-PLS) were used as spectral analysis techniques. In addition, six spectral preprocessing techniques were evaluated:
1) no preprocessing, 2) standard normal variate (SNV), 3) 1st derivation, 4) 2nd derivation, 5) 1st
derivation and SNV, and 6) 2nd derivation and SNV.
The results demonstrated that it is crucially important to include the dynamic measurement spectra acquired with the target blender in a calibration set. In addition, it was found that the prediction error of LW-PLS models was smaller than or at least similar to that of PLS models, through the model development study and the analytical validation. The present work also clarified that the modeling parameters, i.e., spectral preprocessing and wavelengths, affected the prediction error interactively. In conclusion, we propose to include the dynamic measurement spectra in a calibration set, to apply LW-PLS as spectral analysis technique, and to select modeling parameters by taking into account the interaction in order to construct a model with small prediction error.