Depth resolved near-infrared spectroscopy and applications of artificial neural networks in pharmaceutical analysis

This study explored the potential of Artificial Neural Networks as a calibration tool in NIRS. A backpropagation feedforward network was used for nondestructive tablet analysis and for noninvasive depth-resolved measurements in vitro.; Theophylline content was predicted in five different lots of directly compressed tablets using both spectral and principal component inputs. Likewise, spectral and principal component inputs were used to classify tablets. Performance of the ANN prediction model was compared to principal component regression and found to offer no significant advantage in prediction error for this simple linear regression while requiring significantly more development time. Spectral inputs provided ANN prediction results that were superior to those obtained with principal component inputs.; Depth-resolved NIR spectral measurements were accomplished by strategically controlling the amount of reflected light reaching the detectors using a series of apertures with different diameters. Depth resolution was found to be approximately 31 {dollar}mu{dollar}m using a system of polymer films. In a more practical demonstration of the method, concentrations of salicylic acid were predicted during diffusion through a hydrogel matrix. Because of the nonlinear relationship between concentration, time and distance, traditional principal component regression was ineffective for concentration prediction whereas an ANN prediction model allowed prediction of drug concentration at any depth and any time in the experimental system that was studied.