Predicting low-level concentrations of spectrally similar amino acids using Variable Strength Coefficient approach with Parallel Factor Analysis and Neural Network on Near-infrared Images of tablets

Predicting low-level concentrations of structurally related components, with similar near-infrared spectra, is one major challenge in the field of near-infrared chemical imaging. In this study we propose a new non-linear regression model, named VSC-PARAFAC-ANN that integrates the newly introduced variable selection and dimensionality reduction technique, known as Variable Strength Coefficient (VSC), together with Parallel Factor Analysis (PARAFAC) decomposition and Artificial Neural Network (ANN). The proposed model overcomes the problem of spectral similarity and high prediction errors for low concentration components through building a robust model that; maximizes the spectral differences between components to improve the model’s selectivity; and capable of predicting, accurately, low concentrations of components that exhibit non-linear behavior. The VSC-PARAFAC-ANN model performance was compared to different regression models as PLS, CLS, PARAFAC-regression with and without applying the VSC optimization. VSC-PARAFAC-ANN showed superior performance over all other techniques with the ability to predict, accurately, low concentrations of structurally related and spectrally similar components with percentage prediction error as low as 0.01%.