Prediction of coffee traits by artificial neural networks and laser-assisted rapid evaporative ionization mass spectrometry

Coffee is an important commodity in the worldwide economy and smart technologies are important for accurate quality control and consumer-oriented product development. Sensory perception is probably the most important information since it is directly related to product acceptance. However, sensory analysis is imprecise and present large deviation related to subjectivity and relying exclusively on the sensory panel. Thus, practical technologies may be developed to assist in making accurate decisions.

This study presents a new method applying laser-assisted rapid evaporative ionization mass spectrometry (LA-REIMS) coupled with high-resolution mass spectrometry to fingerprint coffee samples. Predictive models have estimated sensory properties with accuracies between 87 and 96 % for test samples. The complex relationship between the MS profiles and modelled properties, artificial neural networks (ANN) outperformed partial least square-discriminant analysis (PLS-DA) on estimation of coffee properties. Tentatively identified compounds such as sugars, chlorogenic, and fatty acids were the ones that most affected coffee sensory properties according to a novel approach to evaluate ANN weights.

The proposed method could analyse coffee samples with minimal sample preparation using an automated device. Predictive models can be applied to assist sensory panel on making decision due to accuracies up to 96 % Additionally, a novel algorithm for evaluate m/z importance in ANN models were presented, paving the way for a higher-level of interpretation by using this algorithm.