Agentic Construction of EPAR‑Structured Toxicity Dossiers: Split Evaluation and a Rapamycin Case Study

Keywords: EPAR, SmPC, pharmacovigilance, MedDRA, sirolimus, rapamycin, safety dossier, LLM agents, evaluation, citation accuracy, side effects, adverse event, toxicology, contraindications

TL;DR: An agent writes EPAR‑style safety dossiers and outperforms a web‑enabled baseline on sirolimus while preserving citation accuracy.

Abstract: We build an AI-driven system that compiles medicine safety dossiers in the structure of the European Public Assessment Report (EPAR)—the format regulators use to document benefits and risks. We focus on the “Undesirable effects” section (side effects) and use sirolimus (rapamycin), a transplant medicine, as a case study. To judge quality, we compare the system’s side-effect list to the official patient leaflet by matching each side-effect term one-to-one and checking whether the attached study citation is correct. On sirolimus, our approach covers more of the leaflet’s terms than a strong, web-enabled baseline while keeping good citation accuracy. We release data tables, figures, and a simple recipe so others can repeat the evaluation on their own agents

Submission Number: 241