CP-BayesNE: Uncertainty-Aware Personalized Nutrient-Density Estimation from Food Images Using Conformal Prediction and Prior-Corrected Dietary Adaptation

Background/Objectives: Reliable longitudinal dietary assessment is a prerequisite for precision nutrition, important for studying nutrition-related changes in health and physical function, but conventional dietary recording can impose substantial user burden. AI-driven, food-image-based assessment can reduce this burden, although uncertainty in food recognition and inter-individual dietary differences may propagate to nutrient estimation. Methods:We developed CP-BayesNE, an uncertainty-aware framework that combines temperature-scaled food classification, conformal candidate sets, training-prior-corrected dietary-history personalization, and posterior-weighted nutrient estimation to support AI-driven personalized nutrition. Results: EfficientNetV2-S achieved 78.01% Top-1 and 93.67% Top-5 accuracy, while temperature scaling reduced expected calibration error from 0.11238 to 0.04345. At α=0.10, regularized adaptive prediction sets reduced the average candidate-set size from 80.78 to 5.77 while maintaining 94.32% empirical category coverage. In FoodLogAthl-218, prior-corrected personalization reduced nutrient density mean absolute error (MAE) by 4.0–6.6% across energy and macronutrients over 51 users. A single-participant 92-meal personal dataset provided preliminary end-to-end feasibility evidence, while also revealing limitations, including slightly worse fat MAE and conservative nutrient intervals. Conclusions: These findings support uncertainty-aware personalized dietary assessment while emphasizing that category-level conformal validity does not directly imply accurate true dietary intake.