Development of an intelligent food nutrition recognition and nutrient intake assessment system in hospital settings.

Background: Hospital nutritional diets significantly impact healthcare service standards, therapeutic outcomes, and patient satisfaction. Traditional manual dietary survey methods suffer from substantial limitations and high resource demands. Despite technological advances, few systems have successfully integrated machine vision, electronic weighing, and standardized databases for real-time nutrient assessment in clinical settings. Objective: To develop and validate an intelligent food nutrition recognition and nutrient intake assessment system for hospital therapeutic diets, enabling precise, large-scale monitoring of patient nutrient intake. Methods: We developed an AI-based image recognition system integrating RGB-D imaging, the Segment Anything Model (SAM) for food image segmentation, and the SE_ResNet50_vd model for food classification. The system comprised hardware components (food weight collector, image collector, computing host) and software modules (nutritional analysis platform developed in Java/Spring Boot). A comprehensive database was constructed from 204 standardized therapeutic diet varieties from Peking University People's Hospital, incorporating the Chinese Food Composition Table and cooking loss factors. Post-meal intake was calculated by comparing pre- and post-meal food volumes using depth camera imaging and density-based weight estimation. Food type recognition accuracy was tested on 1,000 samples, and volume estimation accuracy was validated against gravimetric measurements across 100 food items. Results: The food image recognition model achieved 99.2% accuracy on the test set. Volume estimation accuracy exceeded 90% in 39% of cases and 80-90% in 61% of cases, with a minimum threshold of 80% across all tested items. The SAM model demonstrated robust segmentation performance for diverse food types in standardized meal containers. The integrated system successfully monitored over 20 types of therapeutic diets, matching nutrient compositions to personalized requirements with ≥90% accuracy for meal verification. Conclusions: This intelligent system provides standardized, real-time nutrient intake assessment with superior accuracy compared to traditional dietary survey methods. By automating therapeutic meal supervision and enabling precision nutrition monitoring, it represents a significant advancement from empirical to precision-based clinical nutrition practice, with substantial potential for improving treatment accuracy, patient outcomes, and healthcare resource optimization in hospital settings.