Data Fusion for the Food Industry: Leveraging Machine Learning and Deep Learning Algorithms

Food quality perception is significantly influenced by off-flavours and taints, which encompass various sensory attributes, including grassy, bitter, and astringent notes. Traditional quality control relies on human inspectors, whose evaluations are often subjective and inconsistent. Molecular analysis techniques provide a fast and objective method for quality assessment. However, while modern analytical tools can quantify food quality markers with high precision, they generate large, complex datasets that are difficult to interpret. Machine Learning (ML) and Deep Learning (DL) algorithms have been proven effective in extracting meaningful insights from these datasets, enabling fast and objective quality assessments. The accuracy of exploratory and predictive models can be further improved by integrating analytical signals from different instruments, creating a synergistic effect. This approach, known as Data Fusion (DF), enhances analytical performance and reliability. This study analysed off-flavour cocoa liquor samples identified by an internal company sensory panel and compared them with ‘industrially accepted’ samples from two Ecuadorian cultivars: CCN51 and Arriba. The samples were examined using Liquid Chromatography and Gas Chromatography coupled with Mass Spectrometry. Initially, data from both instruments were analysed separately, and the results were compared. Different ML and DL models were applied, including K-Nearest Neighbors, Linear Discriminant Analysis, Logistic Regression, Partial Least Squares Discriminant Analysis, Random Forest, Support Vector Machines, and Multi-Layer Perceptron. Finally, DF strategies, Low-Level (LLDF), Mid-Level (MLDF), and High-Level (HLDF), was explored. Among these, HLDF based on Bayesian Consensus proved most effective, achieving 93% accuracy in identifying off-flavor cocoa samples.