Sentence-Level Five-Class Classification of Food-Disease Relationships in Biomedical Abstracts: A Comparative Evaluation of Large Language Models

Le résumé fourni par la source

Synthesizing evidence on food-disease relationships is critical for advancing nutrition science, yet manual curation cannot keep pace with the rapid expansion of the biomedical literature. This study evaluates large language model (LLM)-based approaches for sentence‑level classification of food-disease relationships in PubMed abstracts. Using 1,500 manually annotated sentences across four major non-communicable diseases, the performance of ChatGPT‑4o, with and without Chain‑of‑Thought (CoT) prompting, was compared with that of supervised classifiers and a rule‑based approach. Sentences from RESULTS and CONCLUSIONS were classified into “Effective”, “Detrimental”, “No Effect”, “Inconclusive”, and “Unrelated”, assuming prior dictionary‑based entity identification. Rule-based classification performed poorly, whereas both the supervised model and ChatGPT-4o achieved high accuracy (>0.8). Prompting strategy and few-shot examples had minimal impact, suggesting that ChatGPT-4o baseline reasoning is sufficient for this task. GhatGPT-4o outperformed the supervised model on complex, multi-association RESULTS sentences and better distinguished between background context, improving “Unrelated” classification. In contrast, the supervised model was more reliable for directional phrasing in CONCLUSIONS sentences and less susceptible to hallucination. These complementary strengths, together with knowledge graphs visualization, demonstrate the potential of LLM‑based frameworks for scalable extraction of food-disease relationship, reducing manual effort. Future work will extend this diseases coverage and integrate metabolite-level and full-text.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.