FoodScribe, an open-source semantic framework, converts free-text meal descriptions into quantitative nutrient profiles by combining large language model (LLM) ingredient parsing with USDA FoodData Central database queries. Benchmarked across three LLM providers on the Nutribench dataset, the tool annotated 3,807 meal descriptions in 2.5 hours — work that would otherwise demand substantial effort from trained nutritionists — achieving macronutrient F1 scores of 0.79–0.89, with stronger performance for protein than fat estimation. Applied to a Mediterranean diet cohort, model-derived estimates captured expected dietary shifts, and metabolomics integration linked higher fiber and vegetable intake to a distinct fecal metabolite cluster.
The bottleneck FoodScribe addresses is real and underappreciated: manual dietary coding is so labor-intensive that most large epidemiological studies rely on blunt instruments like food frequency questionnaires, sacrificing granularity. An automated, free-text pipeline could democratize high-resolution dietary phenotyping for under-resourced research groups globally. The metabolomics linkage — fiber and vegetables correlating with a fecal metabolite cluster — aligns with established gut-microbiome literature but here emerges computationally rather than manually, a proof-of-concept with meaningful implications for diet-disease research at scale.
Limitations are notable: F1 scores around 0.79–0.89 mean non-trivial error rates, fat estimation lags protein, and validation against gold-standard weighed food records is absent. As a preprint posted to medRxiv and not yet peer-reviewed, these accuracy benchmarks and cohort findings require independent replication before FoodScribe is trusted for clinical or regulatory decisions. Still, as an infrastructure advance, this is more than incremental.
