Enhancing Hospital Nutrition Assessment Through Artificial Intelligence: A Prospective Tray-Level Pilot Study

Background/Objectives: Disease-related malnutrition affects 30-50% of hospitalized patients and is associated with adverse outcomes and increased healthcare costs. Routine monitoring of dietary intake typically relies on nursing dietary diaries, which are limited by subjectivity and workload constraints. Artificial intelligence (AI) may improve the accuracy and efficiency of nutritional assessment. This exploratory pilot study evaluated the feasibility of an AI-based system for estimating food intake in hospitalized adults by comparing its performance with gold-standard meal weighing and nurse-completed diaries. Methods: A prospective observational study was conducted in the General Medicine Unit of St. Antonio Hospital (Padua, Italy) between June and August 2025. Food intake was assessed using three methods: manual weighing (reference), nursing dietary diaries, and AI-based image analysis. Analyses were performed at the tray level. Results: A total of 362 meals from 67 patients were analyzed. Concordance between weighed intake and nursing diaries was 60.8%, with diaries frequently overestimating consumption. In a real-world subset, the AI system achieved a mean absolute error of approximately 40 g (≈10% of average tray weight). As multiple trays could originate from the same patient, uncertainty estimates may be optimistic and should be interpreted with caution. Overall food waste was 30.7% of food served. Conclusions: This pilot study shows the feasibility of AI-based intake monitoring in a real-world hospital setting. Our findings are exploratory and based on tray-level analyses; a systematic underestimation bias was observed, and superiority of the AI system over routine documentation cannot be established.