Application of machine learning for optimizing biomarker combinations and guiding decisions on meat authentication

Abstract

This paper tested the relevance of two machine learning approaches (decision trees, DTs; and random forest models, RFs) applied to meat authentication. DT allow to select and rank potential biomarkers according to their respective discriminatory power, optimize their combinations, and guide decisions on classification of samples according to their production systems, all of which has so far been under-researched. RFs were also developed as they are particularly robust. We applied both methods on 19 compounds/variables measured on different tissues (perirenal fat (PF), dorsal fat (DF) and longissimus thoracis et lumborum (LTL) muscle) in an experiment using Romane male lambs pasture-finished on lucerne for four durations pre-slaughter (n = 34–36 lambs per group). Several DTs/RFs were constructed including measurements that are relatively easy to carry out in the abattoir/point of sale, or measurements requiring laboratory analyses. The DTs/RFs distinguished carcasses of lambs pasture-finished from stall-fed lambs with an accuracy of up to 95.1–95.7 %, and showed that PF skatole and PF carotenoid pigment content (out of 19 variables) played a prominent role in classification. The DT/RF designed for use at the point of sale, which was based on DF spectrocolorimetric characteristics and LTL muscle colour coordinates, achieved 84.3–85.4 % accuracy. This is the first research to use DTs for meat authentication, and threshold values for classification decisions will probably need to be validated further on larger databases. These findings nevertheless raise prospects for broad application of decision trees for authentication.