Quantification of palm oil adulteration in Sacha inchi oil: a comparative study between chemometrics and machine learning coupled with FTIR spectroscopy.

Abstract

The growing demand of Sacha inchi (Plukenetia volubilis) oil (SIO) has led to its high vulnerability to fraudulent economic adulteration with lower price alternatives, such as palm oil (PaO). As a consequence, analytical methods capable of authenticity verification are highly urgent. This study evaluated the effectiveness of Fourier transform infrared (FTIR) spectroscopy combined with chemometrics and machine learning algorithms to detect and quantify PaO adulteration in SIO. The chemometrics utilized partial least square (PLS) and principal component regression (PCR) optimized at various wavenumber ranges, whereas six advanced machine learning frameworks were performed. Among the models tested, the machine learning technique of k-nearest neighbors (kNN) algorithm emerged as the most powerful predictive model among others. The optimized kNN model achieved an exceptional goodness of fit, particularly at the wavenumber range of 3100-2800 cm-1 with R2 = 0.999, root mean square error (RMSE) of 0.756, and mean absolute error (MAE) of 0.528. Furthermore, the box-plot, diagnostic scale-location and residual plot analyses confirmed that kNN maintained strict homoscedasticity and eliminated systematic bias across the entire concentration levels of PaO. The ANOVA test demonstrated the highly significant difference in model capabilities (p < 0.001). Moreover, the summary statistics confirmed that the kNN model achieved the highest mean accuracy for all wavenumber ranges used, with R2 of 0.9985 and the lowest mean accuracy of both RMSE (1.1258) and MAE (0.883). These findings highlighted that integrating FTIR spectroscopy with machine learning models offers a precise, reliable, and powerful analytical tool for quality control and ensuring regulatory compliance of high-value edible oils. Further validation using diverse commercial and market samples is required to confirm the robustness, generalizability, and practical applicability of the developed models.