Ensuring reliable feature importance in food chemistry AI.

Food chemistry's rapid AI adoption (2060 AI articles; 415 in 2025) spans machine learning, logistic regression, random forests, and XGBoost. Yet a skills gap in supervised learning fuels misinterpretation: Models optimize prediction, not true associations, and feature importances lack ground-truth validation. High accuracy does not ensure reliable attributions; importances are model- and data-biased. Using a microplastic-cancer case, we show parametric logistic regression on nonlinear data distorts inference. We propose a standards-based pipeline: Unsupervised structure discovery (e.g., feature agglomeration, highly variable feature selection), nonparametric association tests (spearman with p-values), and explicit stability audits of rankings. This multifaceted approach mitigates label-driven bias, improves robustness, and aligns AI insights with mechanistic understanding, supporting credible risk assessment and safer application of AI in food chemistry.