A comprehensive review of AI for food recognition and nutrient estimation with an obese children perspective

Image-based dietary assessment (IBDA) promises lower burden than recalls but its real-world performance, especially in children, remains uncertain. Objective: To map the IBDA knowledge landscape and synthesize evidence on accuracy, usability, and translational readiness for pediatric obesity prevention and care. We searched the Web of Science Core Collection (2015–October 2025) for English-language IBDA studies; conducted science mapping (co-citation, thematic/strategic diagrams, temporal keyword evolution); and performed narrative synthesis across recognition, portion/volume estimation, and nutrient inference, emphasizing free-living validation and pediatric subgroups. Publications have grown and coalesced around dietary assessment, nutrition, and validity. Deep-learning and multimodal advances yield strong recognition and feasible portion/energy estimation on curated datasets, yet performance degrades in free-living use due to domain shift, long-tail foods, and error accumulation. For pediatric deployment, clinically useful outputs should prioritize meal-pattern signals and high-leverage categories, communicate uncertainty, and integrate caregiver-compatible workflows, privacy protections, and fairness audits. Pediatric-first validation against appropriate gold standards and benchmarks reflecting children's food ecologies is needed to translate methodological gains into dependable clinical and public-health tools.