Overview
The Medical Safety Review and Evaluation team at a leading global pharmaceutical company is responsible for ensuring patient safety by reviewing product safety labels and documenting adverse reactions in a regulatory database. Their work directly impacts compliance, reputation, and patient well-being — all under tight timelines and scrutiny from regulators.
Challenge
Each safety label spans 10,000 to 40,000 words across nearly 100 pages, requiring painstaking manual review. The process demanded tracking and coding adverse events using MedDRA terminology — a task so detailed that even experienced physicians found it time-consuming and error-prone.
Adding to the challenge, the team relied on Excel spreadsheets to document findings — a system highly vulnerable to human error and inconsistent tracking. The complexity and scale of these reviews not only slowed operations but also exposed the company to compliance risks.
A single missed adverse reaction could lead to severe consequences: regulatory penalties, lawsuits, or damage to patient trust. The manual process drained physician time, consumed valuable resources, and limited how many labels could be reviewed per year. The emotional toll was also real — frustration, burnout, and fear of missing something critical.
Solution
To eliminate human bottlenecks and strengthen compliance, the team adopted an AI-driven adverse event labeling agent — powered by Large Language Models (LLMs) and a custom semantic re-ranking engine. The approach focused on automation with assurance, ensuring human experts remained in control while AI handled the heavy lifting.
The AI agent:
- Extracts adverse events directly from safety labels using advanced LLMs
- Maps each event to MedDRA codes through a combination of exact, synonym, and semantic matching — achieving nearly 100% exact match rate across all identified terms
- Generates structured safety reports ready for upload to Veeva Safety after human verification
- Prioritizes semantic accuracy with a custom LLM re-ranker, ensuring top-quality MedDRA matches
The system was validated on 10 safety labels, achieving an average recall of 88%, with top cases reaching 94% and 99% accuracy. Notably, the agent identified adverse events that had been missed during manual review.
Results
- 150–600 physician hours saved annually — with potential to double these savings when rolled out across all product lines
- Reduced human error by eliminating manual data entry and coding
- Enhanced compliance through standardized, auditable AI outputs
- Improved patient safety by increasing the accuracy and completeness of labeling
- Higher consistency and faster quality control across departments
What once took days of meticulous manual work can now be completed in a fraction of the time — with greater precision and regulatory confidence.