Artificial neural network and partial least square in predicting blood hemoglobin using near-infrared spectrum

Abstract

Predictive models are crucial in near-infrared (NIR) spectroscopic analysis. Partial least square - artificial neural network (PLS-ANN) is a hybrid method that may improve the performance of prediction in NIR spectroscopic analysis. This study investigates the advantage of PLS-ANN over the well-known linear and non-linear modelling approaches in spectroscopy analysis that are partial least square (PLS) and artificial neural network (ANN). The results show that ANN that coupled with first order SG derivatives achieved the best prediction with root mean square error of prediction (RMSEP) of 0.3517 gd/L and coefficient of determination (R) of 0.9849 followed by PLS-ANN with RMSEP of 0.4368 gd/L and R of 0.9787, and PLS with RMSEP of 0.4669 gd/L and R of 0.9727. This suggests that the spectrum information may unable to be totally represented by the first few latent variables of PLS and a nonlinear model is crucial to model these nonlinear information in NIR spectroscopic analysis.