Machine learning modeling practices to support the principles of AI and ethics in nutrition research

The paper addresses the growing use of AI/ML in nutrition research driven by complex, large, and multimodal datasets (e.g., PREDICT and Nutrition for Precision Health). Despite widespread availability of AI/ML tools in commercial software, their complexity and lack of standardized procedures can lead to ethical issues, biased predictions, and methodological flaws. The authors use AI/ML broadly to include machine learning and artificial intelligence due to overlapping concerns. Citing notable failures and reviews showing pervasive methodological flaws in medical imaging ML studies, the authors aim to guide nutrition researchers—especially those with statistical backgrounds—on best practices for developing, evaluating, and implementing AI/ML models ethically and effectively. They propose a tutorial and checklist emphasizing extensions of statistical rigor to AI/ML, careful consideration of sample size, balanced datasets, explainable models, and enhanced data literacy to ensure transparency, reproducibility, and minimized bias.

The authors situate their guidance within existing frameworks and best practice literature: FAIR (Findable, Accessible, Interoperable, Reusable) data principles for stewardship and reuse; best practices for AI/ML in related fields (chemistry, life sciences); general machine learning reviews; and domain-specific checklists (CLAIM for medical imaging, NeurIPS and CVPR reproducibility checklists). They highlight cautionary examples from The Alignment Problem (Christian) and empirical work (e.g., Buolamwini and Gebru’s Gender Shades) demonstrating harm from biased data and opaque models. A review of 62 COVID-19 imaging ML studies found methodological flaws in all, underscoring the need for transparency, validation, and thoughtful modeling decisions. This body of literature motivates a nutrition-focused synthesis and checklist that bridge statistical best practices with AI/ML-specific considerations.

This is a narrative methods/tutorial paper synthesizing the authors’ experiences as AI/ML modelers, reviewers, and journal editors in nutrition and obesity. The methodology involves: (1) identifying common omissions and pitfalls from statistical modeling that persist in AI/ML applications; (2) extending statistical best practices to AI/ML contexts; (3) addressing AI/ML-specific issues not handled by point-and-click software; and (4) providing a structured, reproducible checklist covering study design through deployment. Scope and structure: The tutorial covers measurement error, selection bias, sample size calculation considerations, missing data, data imbalance, explainable AI (XAI), and data literacy. It embeds practical recommendations through a detailed checklist (Table 2) spanning: study design and goals; data provenance; measurement error mitigation; suitability of AI/ML; evaluation criteria; data preprocessing (handling missing data, outliers, class balance, feature selection, dataset size/augmentation); algorithm construction (selection, explainability, reproducibility including hardware/software/hyperparameters, and sharing code/data); evaluation (baseline comparison, internal validation with cross-validation and multiple runs for stochastic models, model selection, external validation); deployment and use-case considerations (limitations, retraining cadence, shelf life, and future research). Key procedural recommendations: iterative justification of sample size tailored to model complexity; transparent reporting of dataset class distributions and balancing methods; thoughtful handling of missing data (favoring kNN, multiple imputation; using missingness indicators; modeling MAR/MNAR when possible); mitigating selection bias via recruitment aligned to the target population and cautious up/down-sampling; and pairing non-explainable with explainable models to ensure interpretability and error propagation understanding. The paper does not conduct empirical experiments; rather, it consolidates best practices into actionable guidance.

- The paper provides a comprehensive, stepwise checklist for ethical and effective AI/ML modeling in nutrition, emphasizing transparency, reproducibility, and bias mitigation from study design to deployment. - Measurement error: Use controlled, precise measurements; report error and provide warning labels; avoid training models on biased self-report as ground truth without correction; consider triangulation and multilevel models to correct known biases and assess error propagation, prioritizing explainable models to understand errors. - Selection bias: Ensure dataset representativeness; document demographic and other characteristics; when necessary, cautiously apply up-sampling/down-sampling and clearly disclose limitations of training data in publications and tools. - Sample size: No universal formula; sample size must reflect model complexity and task. Rules of thumb include 50–1000× the number of classes for classifiers, with iterative evaluation and explicit justification of chosen sample sizes. Regularization (e.g., LASSO/Elastic Net) can help when p ≫ n but introduces bias, suitable in exploratory contexts. - Missing data: Classify missingness (MCAR/MAR/MNAR); avoid mean imputation and last observation carried forward; prefer kNN and multiple imputation for MAR; consider modeling MNAR; use missingness indicators or architectures (e.g., RNNs) that incorporate missingness patterns to improve predictions. - Data imbalance: Report class distributions and balancing approaches; recognize that naive accuracy can be misleading for rare outcomes (e.g., GDM ~4–10%); use resampling cautiously and prioritize collecting balanced data. - Explainable AI: Combine black-box models (e.g., deep nets) with explainable methods (e.g., logistic regression, variable importance in random forests, saliency maps) to detect spurious correlations (e.g., ruler-in-image artifact) and improve trust, safety, and error analysis. - Data literacy: Promote stakeholder education to frame appropriate data-driven questions (descriptive, diagnostic, predictive), understand model assumptions, and uphold FAIR-aligned practices. - Overall, the tutorial and checklist operationalize ethical principles and reduce risk of biased or harmful predictions in nutrition AI/ML.

By translating statistical best practices into AI/ML contexts and adding AI/ML-specific safeguards, the paper directly addresses the need for ethical, reliable modeling in nutrition. The checklist and guidance provide a framework for designing studies with clear goals, transparent data provenance, appropriate preprocessing, justified sample sizes, and rigorous validation. Emphasizing explainability and careful handling of measurement error, selection bias, missingness, and class imbalance helps prevent misleading conclusions and supports equitable performance across subgroups. This framework advances reproducibility, facilitates peer review and editorial assessment, and aligns with FAIR data stewardship and emerging discipline-specific checklists. For nutrition researchers, the guidance bridges the gap between traditional statistics and modern AI/ML, promoting safer deployment in clinical and public health contexts.

The authors conclude that high-quality AI/ML modeling in nutrition requires iterative, tailored processes to mitigate ethical risks and bias. They contribute a practical tutorial and comprehensive checklist that span study design, data handling, model construction, evaluation, and deployment with an emphasis on explainability and reproducibility. Future work should: (1) quantify and correct known biases in self-report data; (2) develop and validate methods for MNAR missingness in nutrition contexts; (3) advance interpretable modeling techniques and error-propagation analyses for complex models; (4) improve strategies for dataset balancing beyond resampling; and (5) foster broader data literacy among nutrition stakeholders to ensure responsible AI/ML use.

This is a narrative tutorial and checklist derived from author experience, not an empirical evaluation; it does not cover every possible scenario or provide universal sample size calculations. Many illustrative examples are from domains outside nutrition due to the emerging state of AI/ML in nutrition. Some recommendations (e.g., handling MNAR, error propagation in deep models) remain challenging and context-dependent. External validation strategies may be constrained by data availability, and not all tools or datasets can be openly shared due to practical or regulatory limits.