Development of concentration prediction models for personalized tablet manufacturing using near-infrared spectroscopy

Abstract

In this work, concentration prediction models applicable to a very wide range of concentrations for personalized tablet manufacturing were developed using near-infrared (NIR) spectroscopy. Tablet manufacturing experiments were conducted, and NIR spectral data of the tablets were obtained. To search for an appropriate development method for the prediction models applicable to a very wide range of concentrations, three typical active pharmaceutical ingredients (APIs) were selected and two modeling methods were applied, namely, partial least squares (PLS) regression and extreme gradient boosting (XGBoost). The prediction models developed using the two methods were evaluated according to widely used criteria. The results showed that the model accuracy obtained using PLS regression was higher than that obtained using XGBoost. The modeling method using PLS regression was also applied to other typical APIs and inorganic compounds. These results indicate that the model accuracy was affected by the heat resistance of the compound. In addition, this modeling method could be useful in developing prediction models for ion-binding inorganic compounds and in suggesting the presence of metallic foreign matter in tablets.