Rapid Evaporative Ionisation Mass Spectrometry Based prediction of fish species – a comparison of established chemometric and novel machine learning approaches.

Rapid evaporative ionisation mass spectrometry (REIMS) has quickly established itself as a powerful and flexible technique for the rapid and
often in-situ analysis of a wide variety of different sample types, ranging from biofluids and cellular materials to highly processed foodstuffs
and even polymers. Metabolomic experiments have become a major use of REIMS systems, a result of the flexible sampling interface options
for the system and the speed at which samples can be analysed. These factors permit large numbers of samples to be very rapidly analysed
and by reducing or eliminating sample preparation, new types of analysis can be developed.The rapid and flexible nature of sample acquisition
means REIMS systems can generate extremely large quantities of data, for which existing processing techniques have previously been utilised
– particularly chemometric techniques making use of LDA and OPLS-DA discriminant modelling. These techniques may not make the most
effective use of REIMS data and can have difficulty processing the largest of data sets, as a result machine learning techniques are increasingly
being adopted to better process and exploit REIMS data. This lecture will use the prediction of fish species as a case study; fish is a common
source of food-fraud, with substitution and mislabelling frauds being common; additionally freshness is important to resale value, so rapid
confirmatory testing is essential. A total of 2000 samples comprising 14 different commercially valuable species, including cod, haddock, whiting,
herring, mackerel, rainbow trout, wild and farmed salmon were analysed using iKnife-REIMS followed by both established chemometric and
novel machine learning strategies. LDA, PLS-DA and OPLS-DA models, which cross-validated between 95-98% are compared to Random Forest
and Support Vector Machine modelling which showed accuracies of 0.895 and 0.980 respectively. Additionally, the machine learning strategies
eliminated the need to pre-process data, substantially cutting data processing times.