A Chemometric Study of Vis-NIR Reflectance Spectra of Turmeric Powders to Quantify Total Curcuminoids

摘要

This study focuses on the synergistic application of spectroscopy and multivariate analysis for refined quantification of total curcuminoids in powdered turmeric samples. Quantifying the pharmacologically active curcumin is a direct way of analyzing the quality of a particular turmeric sample, and hence paving the way to establish its market value. In this study, the spectra of sieved turmeric powders were collected in the Vis-NIR range (wavelength 360-1030 nm) and subjected to multivariate analysis. The content of total curcuminoids was first accurately quantified by high performance liquid chromatography (HPLC), which served as the reference data. The combination of multiplicative scatter correction, de-resolving, orthogonal scatter correction, and the first derivative of the spectra were used as pre-processing steps to filter the informative features of the spectra by reducing noise. This sample set included curcuminoids in the range of 1-4% as established by HPLC. The correlation between the processed spectra and the curcumin content was examined by the partial least squared regression (PLSR) algorithm through the construction of two kinds of models. A coarse model for broad classification, having 66 samples of 15 scans each (total - 990 spectra), was used as training data, which gave a correlation coefficient (r2) of 0.92 at PLS factor 7. Cross validation was performed on a test data of 10 samples which yielded a root mean squared error of cross validation (RMSECV) of 0.219. Along with this, the ratio of prediction to deviation (RPD) was calculated to estimate the robustness of the model in the long run, and was found to be 3.06. Owing to the diversity in the range of values and the need for poignant accuracy in quantifying the total curcuminoids through its spectrum, two sets of finer models were constructed. The result of the coarse model's prediction along with the error (±Δ) for a test sample determined which of the finer models was best suited to give the final predicted value of curcumin. The first set of fine models covered the range of 1-2%, 2-3% and 3-4%, while the second set comprised models in the range of 1.5-2.5% and 2.5-3.5%, to eliminate ambiguity for predicted values falling on the border. For each of the finer models, r2 was in the range of 0.90-0.96 (Figure 1(a)), RMSECV was between 0.06-0.13 (Figure 1(b)) and RPD was greater than 4, all confirming good accuracy.