Partial Least Squares-Discriminant Analysis Classification for Patchouli Oil Adulteration Detection by Fourier Transform Infrared Spectroscopy in Combination with Chemometrics.

ABSTRACT

This study aims to test chemometrically partial least squares-discriminant analysis (PLS-DA) classification models to detect the adulteration of patchouli oil (PO) with gurjun balsam oil (GBO) by utilization of Fourier transform infrared spectroscopy. Unsupervised analysis was tested using the pattern recognition method using the principal component analysis model against the original spectrum at wavenumbers 4000-500 cm-1 and at the fingerprint area (1800-600 cm-1). Model testing was also carried out on the spectrum that had been pre-processed using the standard normal variate, second derivative Savitzky-Golay, and normalization approaches. Variable Y samples used were certified reference material (CRM), PO, GBO, and PO forged with GBO (PGBO) with a counterfeiting ratio of 0.5 (v/v) to 10% (v/v) with an interval of 0.5%. The same treatment was carried out on testing of the PLS-DA model. In pattern recognition tests, the best separation of the original spectrum was obtained at wavenumbers 1800-600 cm-1. The model was further tested on PLS-DA by making assumptions or codes for CRM, PO, GBO, and PGBO as +2, +1, 0, and -1, respectively. The results of the model analysis showed that even at the lowest counterfeiting ratio (0.5%), the presence of counterfeiting material was detected by the PLS-DA model. The RMSEC value is close to zero with a value of 0.22, and the R square is close to 1, which is 0.954. This very significant separation is clearly illustrated in the loading plot and bi-plot due to the contribution of chemical compounds in the GBO that undergo vibration at wavenumbers 603, 786, and 1386 cm-1. Validation of the PLS-DA model was carried out strongly using the PLS model, and it showed that the difference between the calibration concentration and the prediction was very low (average 0.45) with an accuracy percent above 99%. The efficacy of the model is further substantiated by the consistent and precise values of sensitivity and selectivity, obtained from both the training set and test set.