Development of a Rapid and Affordable Machine Learning-Based Screening Technique for Honey Adulteration Using UV–Visible Absorption Spectra

Honey, a high-value product, is a frequent target of food fraud through adulteration with low-cost sweeteners. Up to 87% of local honey products sold in the Philippines are adulterated, with some containing as much as 100% substitution. Beyond health risks to consumers, adulterated honey undermines consumer trust, compromises nutritional and therapeutic qualities, and distorts market competition, putting authentic producers at a disadvantage. This study assessed the feasibility of ultraviolet–visible (UV-Vis) spectrometry at 220-450 nm, coupled with machine learning (ML), as a rapid and affordable screening technique for honey adulteration. Fifty-nine authentic Philippine honey samples from three bee species were analyzed in their unadulterated and adulterated forms with C3 and C4 syrups at a 10% concentration. C3 and C4 sugars are derived from plants using the C3 (Calvin cycle) or C4 (Hatch–Slack pathway) photosynthetic processes, respectively. In honey, C3 sugars originate from the nectar of fruit tree flowers and wildflowers; C4 sugars used as adulterants are usually sourced from sugarcane or maize. However, C3 sugars are now commercially available and could be used as adulterants, challenging the reliability of established methods based on the isotope ratio. In this study, data analysis indicated possible discrimination of samples by bee species rather than adulteration status. A supervised ML model trained to discriminate adulterated from unadulterated samples initially showed 98% accuracy with 100% sensitivity, but once a grouping function was applied to address temporal data leakage, model performance dropped to 34% accuracy and sensitivity. This finding highlights the need to carefully examine ML implementation issues in honey authentication research, particularly regarding sampling balance, data splitting, and model validation, since reported accuracies in published studies frequently exceed 90%. While the present model does not yet achieve reliable discrimination between pure and adulterated samples, protocol refinements and enhanced chemometric approaches can improve robustness for honey authenticity screening.