Background. There is a real risk of economically motivated food fraud affecting the food supply chain, particularly in the edible oil market where economic pressure is increasing. Recently, near-infrared spectroscopy in combination with machine learning has been proposed as a method for fast, cost effective and non-destructive food authenticity testing. Whilst many papers report excellent performance of such methods in the laboratory, there are relatively few examples of their successful application by retailers and by official control laboratories. This article is a systematic review of recent work on the use of near-infrared spectroscopy and machine learning for the detection of adulteration of edible oils. It outlines the structural research gaps that need to be addressed in order to translate excellent laboratory performance into effective use in the food supply chain and presents a framework that outlines potential methodological routes to address the identified gaps.
Methods. This systematic review was conducted according to the PRISMA 2020 checklist. A literature search was conducted on four databases, Scopus, Web of Science, PubMed and ScienceDirect. The search terms used were a combination of near-infrared and infrared spectroscopy, hyperspectral imaging, machine learning and chemometric techniques, and edible oil, and oil adulteration. All records identified in the search were screened by title, abstract and full text to determine if they met the criteria for inclusion in the review. The extracted data from the included studies were presented in a table, which detailed the techniques used, the types of edible oils and their respective adulterants, the models used for classification, the corresponding performance of the models and the validation of the studies. The limitations of the studies and potential sources of bias were also assessed qualitatively and presented below. The synthesized evidence indicated that there were high levels of accuracy for the detection of oil adulteration on laboratory-generated data. However, there were ten interrelated gaps in the evidence that were related to a number of issues including the analytical validity, cross-instrument analysis, and field robustness of techniques for the detection of oil adulteration. Additional issues included those related to decision making near thresholds for regulation and the need for greater integration with current regulations. The main weaknesses of the current evidence were that the studies had been validated internally-only, conducted under closed-world conditions, using a single-instrument that lacked the use of certified reference materials.
Conclusion. We do not believe that there are huge accuracy gaps and that all five of these “gaps” can be addressed by a single corresponding enabling innovation that increases the deployment readiness of methods. In essence, we believe that for all five of these “gaps” (cf. Section 2), there exists a set of innovations that will bump the field to a corresponding new tier of readiness. These readiness tiers are ordered from threshold uncertainty of conformal prediction over open-set recognition of as-yet unseen types of adulterants to deep domain adaptation for off-line calibration transfer, self-supervised pretraining, and finally to federated learning with shared certified spectral libraries for on-site oil authentication. Closing these five readiness gaps for on-site oil authentication will most directly addresses the current unsolved problem