Qualitative and quantitative analysis of pharmaceutical content by terahertz spectroscopy and machine learning...

Process analytical technology (PAT) plays an essential role in pharmaceutical manufacturing by enabling effective quality control and improving production efficiency. Terahertz (THz) spectroscopy has attracted considerable attention due to its advantages of high biosafety, distinctive spectral fingerprints, and excellent quantitative performance. However, achieving rapid classification and accurate quantification of pharmaceutical content remains challenging. Therefore, the combination of THz spectroscopy and machine learning algorithms is introduced to address this challenge. Three active pharmaceutical ingredients (APIs), namely isoniazid, 4-aminobenzoic acid, and diflunisal, were selected, and principal component analysis (PCA) was first employed for dimensionality reduction of the THz spectral data. Subsequently, the prediction accuracy and computational efficiency were optimized using four machine learning algorithms, including support vector regression (SVR), multilayer perceptron (MLP), convolutional neural networks (CNNs) and backpropagation artificial neural network (BP-ANN). Among all models, the PCA-SVR model achieved the best prediction performance, with a coefficient of determination of 0.9990 and a root mean square error of 1.2%. In addition, the prediction time is reduced by up to 90%, which greatly improves the efficiency of data processing. The robustness of the proposed framework was further validated using a multi-component pharmaceutical system to simulate practical tablet formulations. This work demonstrates the potential of THz spectroscopy for component identification and content prediction of active pharmaceutical ingredients and sheds light on the nondestructive application for monitoring pharmaceutical quality and improving production efficiency in an online manner.