A novel model selection strategy using total error concept

Abstract

Our previous work had proved that accuracy profile theory could be employed as a means of validating one PLS model in Chinese material medica system. In this paper, accuracy profile theory is proposed as a powerful decision tool to demonstrate the prediction performance of multi-model at each concentration level rather than all concentration levels. Partial least square (PLS), interval partial least square (iPLS), backward interval partial least square (BiPLS) and moving window partial least square (MWPLS) were selected to construct visible and near-infrared (vis/NIR) spectroscopy models. Chemometric indicators, i.e., determination coefficient (R(2)), root-mean-square error of prediction (RMSEP) and ratio of performance to inter-quartile (RPIQ), were used to select the optimum model. However, the results clarified that these commonly used indicators could not clearly demonstrate different PLS models' ability because these indicators depend on all concentration levels to assess the multi-model. Therefore, "total error concept" (accuracy profile theory) was introduced to assess the ability of multi-model at each concentration level. Analytical methodology parameters, i.e., linearity, relative bias, uncertainty, repeatability, intermediate precision, lower limit of quantification (LLOQ) and risk, were calculated by accuracy profile theory. Final results showed that model selection strategy which was based on model assessment at every concentration level was more sensitive than the one based on all concentration levels. The analytical procedures involved in this work ensure that model selection strategy using total error concept is coherent.