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

Area of Science:

• Food Science
• Analytical Chemistry
• Data Science

Background:

• Artificial Intelligence (AI), Machine Learning (ML), and Deep Learning (DL) are increasingly used in food analysis for quality control, authenticity, and chemical profiling.
• Chromatographic data, common in food analysis, presents challenges like high dimensionality, retention time variability, and multicollinearity, complicating ML/DL model development.
• Limited sample sizes in food studies hinder model generalization and transferability across different settings.

Purpose of the Study:

• To critically evaluate methodological practices in ML and DL for food chromatography.
• To assess the impact of data construction, validation, and interpretation on model reliability and industrial applicability.
• To provide recommendations for developing robust and transferable ML models in food chromatography.

Main Methods:

• Systematic review of 165 peer-reviewed studies (2015-2025) on ML/DL in food chromatography.
• Case study using a coffee origin chromatographic dataset to examine class imbalance, training set size, and replicate handling.
• Analysis of validation strategies, including internal cross-validation versus external validation.

Main Results:

• High-fold cross-validation and Leave-One-Out Cross-Validation overestimate model accuracy compared to external validation.
• Treating analytical replicates as independent samples inflates performance metrics without enhancing true predictive power.
• Deep Learning (DL) models require larger datasets than Machine Learning (ML) models to prevent overfitting, with ML models often plateauing at moderate training sizes.

Conclusions:

• Rigorous validation strategies, particularly external validation, are essential for accurate assessment of ML/DL model performance in food chromatography.
• Careful management of chromatographic data, including proper handling of analytical replicates and class imbalance, is critical.
• Implementing these practices will lead to more robust, reproducible, and industrially transferable AI models for food analysis.