Ensuring reliable feature importance in food chemistry AI

Area of Science:

• Food chemistry
• Artificial intelligence
• Toxicology

Background:

• AI adoption in food chemistry is rapidly increasing, utilizing methods like machine learning, logistic regression, random forests, and XGBoost.
• A significant skills gap exists in supervised learning, leading to misinterpretation of AI model outputs.
• Current AI models optimize for prediction accuracy, not necessarily true associations, and feature importances often lack ground-truth validation, being prone to model and data bias.

Purpose of the Study:

• To highlight the limitations and potential misinterpretations arising from the current application of AI in food chemistry, particularly concerning supervised learning.
• To propose a new standards-based pipeline for AI analysis in food chemistry to mitigate bias and improve the reliability of insights.
• To demonstrate the proposed pipeline using a microplastic-cancer association case study.

Main Methods:

• The study critiques the use of parametric logistic regression on nonlinear data, using a microplastic-cancer case as an example of distorted inference.
• A proposed standards-based pipeline includes unsupervised structure discovery (feature agglomeration, highly variable feature selection).
• Nonparametric association tests (Spearman with p-values) and explicit stability audits of rankings are integral parts of the proposed methodology.

Main Results:

• Parametric logistic regression on nonlinear data was shown to distort inference in the microplastic-cancer case study.
• The proposed pipeline aims to mitigate label-driven bias inherent in supervised learning approaches.
• The multifaceted approach is designed to enhance the robustness of AI-driven insights and align them with mechanistic understanding.

Conclusions:

• Current AI practices in food chemistry, especially supervised learning, can lead to misinterpretations due to prediction optimization and lack of validation.
• The proposed standards-based pipeline offers a more robust and reliable method for AI application in food chemistry.
• Implementing this approach supports credible risk assessment and the safer, more effective application of AI in the field.