One of the limiting factors in developing any untargeted analysis is sourcing samples for the reference database. The samples labelled as “authentic” in the model developed must be of absolute trustworthiness (fully traceable back to authentic production) and also fully representative of every variable within the “authentic” scope (e.g. different permitted agricultural inputs, harvest seasons, species, variety and geographic origin).
This pilot study (open access) uses a statistical approach to circumvent this need. The authors still attempted to source reference samples that were representative of an “authentic” scope but they made no attempt to verify the samples; all reference samples were purchased from online direct-to-consumer vendors. They then built a results dataset of 38 different fatty acid methyl esters, tocopherols and phytosterols measured by GC-MS and LC-MS. They selected from within this dataset using Monte Carlo sampling to choose different “reference databases” and build a large number of One Class Classification models. They then used statistical analysis to see if each of these models appeared internally consistent (i.e. suggestive that all of the reference samples had been “authentic”) or not (i.e. suggestive that some of the reference samples had been “inauthentic”). They chose the model with best internal consistency.
They piloted this approach on 40 avodado oils, identifying 6 of them a potentially adulterated. Subsequent targeted chemical analysis showed that these 6 adulterated samples had been correctly identified.
