Methodological challenges in machine learning and deep learning applied to food analysis: A critical review

Artificial Intelligence (AI) is increasingly applied for food quality control, authenticity assessment, and chemical profiling. However, the reliability and industrial applicability of Machine Learning (ML) and Deep Learning (DL) models critically depend on how the datasets are constructed, validated, and interpreted. Among the different analytical techniques used in food analysis, chromatographic fingerprints and chromatographic hyphenated techniques typically generate high-dimensional datasets characterised by structured peak distributions, retention time variability, and strong multicollinearity among adjacent variables, making model development and validation more challenging. These features, combined with limited sample sizes typical of food studies, pose significant challenges for model generalisation and transferability across analytical batches and industrial settings. This work provides a critical evaluation of methodological practices in ML and DL, supported by a systematic review of 165 peer-reviewed studies published between 2015 and 2025. Particular attention is given to common issues such as class imbalance, the misuse of analytical replicates, and the widespread reliance on internal cross-validation without independent test sets. To experimentally assess these factors, a chromatographic dataset on coffee origins was used as a case study to examine how unbalanced class distributions, varying training set sizes, and replicate handling affect model performance. Our results show that high-fold cross-validation and Leave-One-Out Cross-Validation consistently overestimate model accuracy compared to external validation, while treating analytical replicates as independent samples artificially inflates performance without improving true predictive power. Learning curves reveal that most ML models plateau with moderate training sizes, whereas DL models require substantially larger datasets to avoid overfitting. The findings highlight that rigorous validation strategies and careful management of chromatographic data are essential for developing robust, reproducible, and industrially transferable ML models in food chromatography.