Machine learning and evolutionary computation on e-nose datasets: A preliminary approach to ergot alkaloid detection in wheat

Abstract

To the best of the authors' knowledge, this is the first time that an approach based on the use of machine learning (ML) algorithms combined with genetic programming (GP) was used to process small-sample-size e-nose data. The approach was proposed to classify the volatile compound information of wheat samples based on the contamination of ergot alkaloids, a class of emerging mycotoxins which pose a severe threat to food safety and consumer health. Unlike previous studies that applied convolutional neural networks to full e-nose response profiles, our approach focused on a small set of features extracted from the steady-state region of each response curve. Despite the low dimensionality, using GP to generate optimal features significantly improved the classification performance of several ML models. Different classifiers, including Decision Tree, Linear Discriminant Analysis, the Mahalanobis Distance Classifier, an artificial neural network-based method and ensemble methods were assessed and applied to a dataset of 21 wheat samples. These samples were classified according to their compliance with the EU maximum limit of 150 μg/kg for ergot alkaloids in wheat. The combined application of GP-based feature transformations, specifically using M3GP, and ML classifiers resulted in significant improvements in accuracy, F1 score, precision and recall compared to models trained on untransformed features. These findings highlight the unexplored potential of GP as a powerful tool for feature construction in sensor-based classification tasks for food safety signal processing.