A novel variable selection strategy in MCUVE-PLS-DA combined with mid-infrared spectroscopy for the discrimination of Brazilian aged cachaças

Abstract

This study presents an enhanced modification of the Monte Carlo Uninformative Variable Elimination (MCUVE) method for variable selection in Partial Least Squares Discriminant Analysis (PLS-DA), incorporating regression coefficient stability and classification error rate as a dual-criterion strategy. Unlike the conventional MCUVE approach, which relies solely on minimizing the root mean square error of cross-validation (RMSECV), the proposed method aligns more directly with classification objectives, enhancing robustness, interpretability, and performance. This strategy was applied to discriminate cachaça samples aged in amburana, oak, and freijó barrels using mid-infrared (MIR) spectra acquired from only 10 μL of sample, without any sample preparation. The best model, employing Savitzky-Golay first derivative preprocessing, achieved a predictive accuracy of 97.8 %. Compared to traditional MCUVE and Variable Importance in Projection (VIP) techniques, the proposed method demonstrated superior variable reduction efficiency while maintaining high classification accuracy, offering a robust and environmentally sustainable alternative for spectroscopic data in food analysis.