NutriMatch: harmonizing food composition databases with large language models for enhanced nutritional prediction

Missing and inconsistent nutrient values in food-composition databases hinder comparative nutrition research. We present NutriMatch, a scalable harmonization method that embeds food descriptions with a large-language model, aligns nutritionally equivalent items, and imputes missing nutrients to enrich FCDB coverage. Applied to the Israeli Human Phenotype Project, NutriMatch expanded logged nutrient profiles from 21 to 151 nutrients; in the Australian PREDICT cohort, coverage rose from 43 to the same 151 nutrients using a validated external FCDB. Enriched nutrient sets improved 2-year obesity prediction (AUC 0.63 → 0.67) and increased median R² for body-fat percentage, waist circumference, continuous glucose monitoring, and several blood biomarkers. Models trained in HPP transferred to PREDICT without retraining, achieving Pearson r = 0.32 for visceral fat mass versus 0.24 with baseline nutrients. NutriMatch thus delivers rapid, reproducible harmonization of FCDBs, phenotype-informative nutrient enrichment, and enables robust cross-cohort studies.