Using LLMs to Classify FDA Adverse Event Reports

This study evaluates whether expert-curated few-shot examples improve LLM classification of FDA adverse event reports compared to zero-shot prompting. Across 200 FAERS reports classified by severity and organ system, few-shot prompting reduced misclassification in clinically significant edge cases by correcting severity over-estimation, capturing hepatotoxicity signals, and resolving organ-system ambiguity, demonstrating that targeted data curation has outsized impact on the cases that matter most in pharmacovigilance.