Project 4: Drug Interaction Checker¶
Mentors: To be announced
Problem: Letting an LLM state whether two drugs interact (rather than only explain a known interaction) risks a fabricated medical fact reaching a patient.
Context: Backed by a structured lookup against the openFDA drug label API (free, no key required), which returns each drug's official drug_interactions label section. Public and requiring no license. No training. (NLM's RxNav Drug-Drug Interaction API was discontinued in January 2024 with no direct replacement, so this project uses openFDA instead.)
Goals: Combine a structured, authoritative lookup with an LLM used only for the "translate this into something a patient can read" step, not for the factual claim itself.
Deliverables: A two-field form (drug A, drug B) that queries openFDA for each drug's label and pulls the raw drug_interactions text, then passes that record (not the drug names alone) into the LLM prompt so the plain-language explanation is grounded in the actual API result. A deliberate contrast with Project 1: here the LLM never originates the medical fact, it only rephrases one.
Showcase: TBD
References:
- Medical Hallucination in Foundation Models and Their Impact on Healthcare (medRxiv, 2025): documents LLMs fabricating drug interactions, dosages, and contraindications when not grounded in a structured source, the exact failure mode this project's design avoids.
- DDInter 2.0 (Nucleic Acids Research): a free, openly downloadable, peer-reviewed drug-drug interaction database (300,000+ records with mechanism and management notes), a good cross-check or alternative data source to openFDA's label text.