Artificial neural networks (ANNs) and partial least squares (PLS) regression in the quantitative analysis of cocrystal formulations by Raman and ATR-FTIR spectroscopy.

The present work describes the development of an efficient, fast and accurate method for the quantification of polymer-based cocrystal formulations. Specifically, the content of carbamazepine-nicotinamide (CBZ/NIC) and ibuprofen-nicotinamide (IBU/NIC) cocrystals in Soluplus®-based formulations was independently determined with the aid of either Raman or Attenuated Total Reflectance Fourier-Transform Infrared Spectroscopy (ATR-FTIR) spectroscopy. Spectra peaks from mixtures of IBU/NIC and CBZ/NIC cocrystals with Soluplus at a ratio ranging from 90/10 to 1/99 w/w (cocrystal to SOL) were evaluated and modelled with the aid of feed-forward, back-propagation artificial neural networks (ANNs) and partial least squares (PLS) regression analysis. A 25 full-factorial experimental design was employed in order to evaluate the effect of ANN's structure (number of hidden units) and training (number of iteration cycles) parameters along with the effect of Raman or FTIR spectra region and data preprocessing (direct orthogonal signal correction - DOSC, second derivative, or no preprocessing) on ANN's fitting performance. Results showed that when DOSC preprocessing was employed excellent ANN fitting in both Raman (root mean squared error of prediction (RMSEp) values of 0.43 and 0.34 for IBU/NIC-SOL and CBZ/NIC-SOL, respectively) and FTIR (RMSEp values of 0.04 and 0.03 for IBU/NIC-SOL and CBZ/NIC-SOL, respectively) spectra was obtained. Comparison of ANNs fitting results with PLS regression (RMSEp for IBU/NIC-SOL was 0.94 and 7.36, and for CBZ/NIC-SOL 7.29 and 15.63, using Raman and FTIR analysis, respectively) revealed ANN's fitting superiority, which can be attributed to their inherent non-linear predictive ability.