Comparison of Chemometric Models for Quantification of Total Curcuminoids in Powdered Turmeric Using Visible and Near Infrared Spectra

Abstract

Introduction The combination of a light source emitting a desired wavelength range and a detector capturing that signal can help understand the photonic interaction of matter in different regions of the electromagnetic spectrum. The reflectance spectra of the same analyte will vary based on which region of the electromagnetic spectrum it is captured in. In conjunction with machine learning applications, several parameters of the analyte can be simultaneously analyzed and reported. In this study, we have attempted to compare chemometric models for predicting the quantity of total curcuminoids (sum of bisdemethoxycurcumin, demethoxycurcumin and curcumin) in powdered turmeric samples using diffused reflectance spectroscopy, in two regions of the electromagnetic spectrum: (1) The visible range (360-750 nm), and (2) A small region of the near-infrared (NIR) range (1550-1950 nm). Materials and Methods The diffused reflectance spectra of turmeric powders (n = 66) are captured using two different spectrometers — JAZ from Ocean optics for the visible range, and Spectral Engines for the NIR range. The quantity of total curcuminoids in this data set ranges between 1-4%. The spectra from each instrument are subjected to a series of pre-processing steps to condition the data. The pre-processing steps were optimized to obtain the best-possible accuracy in the subsequent regression models. Principal Component Analysis (PCA) is used to qualitatively analyze the data and visualize the clustering patterns, if any. The processed spectra are then evaluated with Partial Least Squared Regression algorithm (PLSR), using High Performance Liquid Chromatography (HPLC) as the reference data, to predict the quantity of curcuminoids in powdered turmeric samples. The model performance is reported with coefficient of determination (R2 ), root-mean-squared error of cross validation (RMSECV) and root-mean-squared error of prediction (RMSEP). Results and Discussions The spectrum of turmeric has a single broad peak in the visible region (360-750 nm), owing to the electronic transitions in the curcumin molecule. There are two peaks observed in the NIR region (1550-1950 nm), corresponding to the first overtone of the C-H stretching vibrations and second overtone of the C=O stretching vibrations in turmeric. On subjection to PCA, there is a subtle clustering when the spectra are divided into 1% curcumin slabs in the visible range, as seen in Fig. 2(a). However, there is no such separation seen for spectra collected in the NIR region. On building regression models based on this data, the model with the visible spectra of turmeric has a better R2 of 0.921 as compared to 0.83 for the NIR spectra. Similarly RMSECV is found to be 0.219 (visible), 0.273 (NIR) and RMSEP as 0.302 (visible), 0.352 (NIR) (Fig. 3). The model for the visible spectra performs better both qualitatively and quantitatively. Both the models are also validated on a test data set of 7 samples (not a part of the training data), and the predicted results are shown in Fig. 4. The predicted curcuminoid values from 7 samples have a mean deviation of 7.7% and 11% from the model based on visible and NIR spectra of turmeric, respectively, when compared with the actual curcuminoid values from HPLC. Conclusion The chemometric model built with the spectra of turmeric in the visible range performs better than its NIR counterpart with the same set of turmeric samples. This observation implies that the cost of a chemometric sensing system to predict total curcuminoids in turmeric can be reduced by using the visible wavelength range due to the relatively lower cost of detectors and light sources. However, this is a preliminary study with limited number of samples to conclusively say which region of the electromagnetic spectrum is better for analyzing and predicting total curcuminoids in turmeric. Upon further addition of samples and exploring different machine learning approaches, a stronger conclusion can be drawn. References H. Suresh, A. R. Behera, and R. Pratap, ECS Meet. Abstr., MA2020-01, 1858–1858 (2020) https://iopscience.iop.org/article/10.1149/MA2020-01261858mtgabs Figure 1