Back

Establishing a Bidirectional Correspondence Table between the Japanese Standard Tables of Food Composition 2020 (8th Edition) and the USDA FoodData Central Using Large Language Model-Based Matching

Nakagawa, S.; Yamamoto, A.

2026-05-13 nutrition
10.64898/2026.05.10.26352824 medRxiv
Show abstract

BackgroundNo official correspondence table exists between the Japanese Standard Tables of Food Composition 2020 (8th edition; MEXT) and the USDA FoodData Central (FDC), despite their widespread use in nutritional research. This absence has hindered international comparison of food composition data for over six decades. MethodsWe developed a bidirectional matching pipeline using Claude Haiku (Anthropic), a large language model (LLM), combining food category mapping, 17-nutrient Euclidean distance ranking, and LLM-based conceptual judgment. Survey (FNDDS) data were excluded from FDC, yielding 8,158 items (Foundation Foods and SR Legacy). Matching was performed in both directions: MEXT[->]FDC and FDC[->]MEXT. ResultsOf 2,478 MEXT items, 1,927 (77.8%) were matched to FDC items, while 549 (22.2%) had no FDC equivalent (JP-only foods). Of 8,158 FDC items, 5,445 (66.7%) were matched to MEXT items, while 2,698 (33.1%) had no MEXT equivalent (US-only foods). Bidirectional consensus yielded 435 confirmed food pairs across 13 food categories. Notably, FDC items showed systematically higher calcium (+6.0 mg/100g) across 12 of 13 categories, while MEXT items showed systematically higher potassium (-3.7 mg/100g) across 9 of 13 categories and higher vitamin A as RAE (-3.7 g/100g) across 8 of 13 categories. ConclusionsThis study presents the first systematic bidirectional food correspondence table between MEXT and USDA FDC. The 435 confirmed pairs constitute a validated common vocabulary for international food composition research. The systematic cross-national differences in calcium, potassium, and vitamin A represent novel findings with direct implications for international dietary comparison studies. The complete correspondence table (Version 0.1) is openly available at https://github.com/shnkgw-rincom/jbfd-correspondence-table (DOI: 10.5281/zenodo.20103327).

Matching journals

The top 4 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.