Characterization of blue cheese volatiles using fingerprinting, self-organizing maps, and entropy-based feature selection

Abstract

Understanding which volatile compounds discriminate between products can be useful for quality, innovation or product authenticity purposes. As dataset size and dimensionality increase, linear chemometric techniques like partial least squares discriminant analysis and variable identification (PLS-DA-VID) may not identify the most discriminant compounds. This research compared the performance of self-organizing maps and entropy-based feature selection (SOM-EFS) and PLS-DA-VID to identify discriminant compounds in 17 blue cheese varieties. A total of 172 volatiles were detected using headspace solid phase microextraction, gas chromatography and mass spectrometry, including 1-nonene and 2,6-dimethylpyridine, which were newly identified in blue cheese. Despite SOM-EFS selecting only 14 volatiles compared to 78 for PLS-DA-VID, SOM-EFS proved more effectively discriminant and improved the median five-fold cross-validated prediction accuracy of the model to 0.94 compared to 0.82 for PLS-DA-VID. These findings introduce SOM-EFS as a powerful non-linear exploratory data analysis approach in the field of volatile analytical chemistry.

Keywords: Artificial neural network; Blue cheese; Chemometrics; Flavour; Information gain; Machine learning.