Quantification of Total Curcuminoids in Powdered Turmeric Using Ridge Regression and a Coarse-Fine Decision Tree Approach on NIR Spectra

Near-infrared spectroscopy (NIRS), in combination with multivariate statistical analysis, is renowned for its non-invasive, rapid, and environmentally friendly attributes, making it a valuable tool for accessing the quality and quantity of active ingredients in food and agricultural products. This study explores a novel approach for quantifying total curcuminoids in powdered turmeric using NIRS, covering the spectral range of 1550-1950 nm, acquired by a microelectromechanical systems-based handheld spectrometer system assembled in-house, paving the way for point-of-use applications in the field.<br/>It is challenging to have one single model work with reasonable accuracy over a broad range of predictor parameters, which in this case is curcumin content. Although, in some turmeric species curcumin content can be as high as 12% w/w, the prominent range is found to be 1-4%. To ensure uniform prediction accuracy across this range, we propose a coarse-fine decision tree model consisting of two sets of mathematical models- a coarse model for the initial prediction spanning the broad range of 1-4% curcumin, and a second set of five finer models (in the ranges 1-2%, 2-3%, 3-4%, 1.5-2.5%, and 2.5-3.5%) for the final prediction. The decision for the choice of the fine model is reckoned using the error of prediction from the coarse model. We use ridge regression algorithm due to its ability to handle multicollinearity present in NIR spectral data.<br/>The work involves collecting reflectance spectra from 148 turmeric samples, of which 133 were used for training and 15 for blind testing. These samples include both commercially available turmeric powders as well as dry turmeric roots directly from farmers. The roots are procured from different states of India, namely, Andhra Pradesh, Tamil Nadu, Karnataka, Kerala, Maharashtra, and Meghalaya. Initial quantification of curcumin was performed using High-Performance Liquid Chromatography (HPLC), providing a benchmark for the predictive model.<br/>The coarse model demonstrated a coefficient of determination (R2) of 0.91, with Root Mean Square Error of Cross-Validation (RMSECV) and Root Mean Square Error of Prediction (RMSEP) values of 0.06 and 0.22, respectively. For finer models, the error metrics significantly improved, with training and validation errors ranging between 0.01-0.07 and 0.05-0.13, respectively, leading to a 67% improvement in prediction accuracy. The fine models collectively achieved an R2 of 0.98, showcasing the enhanced precision of the proposed method.<br/>The coarse-fine decision tree approach involves an initial coarse prediction, followed by the selection of an appropriate fine model based on the coarse model's error margin. This hierarchical decision-making process ensures the accurate final prediction of curcumin content. The study highlights three distinct scenarios for model selection: ranges fitting within one fine model, ranges spanning two models, and ranges entirely within two models. Each scenario dictates specific strategies for final prediction, thereby optimizing accuracy.<br/>The miniaturized spectrometer, combined with the ridge regression-based coarse-fine decision tree model, provides a robust integrated hardware-software solution for rapid and precise quantification of curcumin in turmeric, holding significant potential for field applications in the food and agricultural industries. It is a reliable alternative to conventional laboratory techniques, which require more time, involve the usage of chemicals and skilled labor. The proposed method simplifies the process, by integrating state-of-the-art hardware with advanced machine-learning techniques for computation, thus, reducing time and resource requirements while maintaining high accuracy.