Comparison of Gaussian process regression, partial least squares, random forest and support vector machines for a near infrared calibration of paracetamol samples

Abstract

In this article, we analyze the near-infrared (NIR) spectra of fifty-eight (58) commercial tablets of 500 mg of paracetamol from different origins (that is, with different batch numbers) in the local markets in Bamako. The NIR spectra were recorded in the spectral range 930 nm-1700 nm. The samples are divided into forty-eight (48) samples forming the set of calibration (training set) and ten (10) samples used as the validation or test set. To perform multivariate calibration, we apply-three nonlinear regression techniques (Gaussian processes regression (GPR), Random Forest (RF), Support vector machine (KSVM)), along with the traditional linear partial least-squares regression (PLSR) to several data pretreatments of the 58 samples. The results show that the three nonlinear regression calibrations have better prediction performance than PLS as far as RMSE is concerned. To decide the best regression model, we avoid R2 since this quantity is not a good parameter for this purpose. We will instead consider RMSE when comparing the different multivariate models. Additionally, to assess the impact of data preprocessing, we apply the above regression techniques to the original data, Multi-scattering correction (MSC), standard variate normalization (SNV) correction, smoothing correction, first derivative (FD), and second derivative correction (SD). The overall results reveal that Gaussian Processes Regression (GPR) applied to smooth correction gives the lowest RMSEP = 2.303053e-06 for validation (prediction) and RMSEC = 2.112316e-06 for calibration. In our investigation, one also notices that the developed GPR model is more accurate and exhibits enhanced behavior no matter which data preprocessing is used.