Answers are grounded in this paper and cite their sources.
This exploratory pilot study assessed nutrient estimation from pre-meal and experimentally simulated plate-waste photographs of 15 standardized hospital meals. Thirteen artificial intelligence (AI) models and dietary assessment applications were compared with an aggregate benchmark formed by averaging estimates from 10 experienced registered dietitians (RDs). Weighed nutrient values for energy, protein, lipid, and carbohydrate served as references; simulated consumed intake was calculated by subtracting residual values from pre-meal values. Agreement and error were evaluated using Lin’s concordance correlation coefficient (CCC), Pearson’s correlation coefficient, mean bias, mean absolute error, root mean square error, and Bland–Altman analysis. For the simulated consumed condition, the aggregate RD benchmark showed greater concordance than individual AI methods across all nutrients (CCC range, 0.75–0.94), compared with maximum CCCs of 0.29–0.54 across nutrients for AI methods. All AI methods had negative mean biases for consumed energy, protein, and carbohydrate, whereas lipid bias varied. AI relative mean biases for consumed energy ranged from −43.6% to −121.1%, compared with −6.9% for the aggregate RD benchmark. All AI methods overestimated plate-waste nutrient values on average. These controlled findings do not establish validity for patient-generated plate waste or clinical deployment; larger, independent, real-world validations are required.
No takes yet. Share an insight, caveat, or question.
Lim et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: