P202: Early Recognition of Potential Adverse Drug Reactions through LLM Assessment of the Scientific Conference literature
Poster Presenter
Joseph Laudano
Medical Director
IQVIA United States
Objectives
To identify potential safety signals from the scientific conference literature not yet included in the United States Prescribing Information (USPI), PubMed or the FDA Adverse Event Reporting System (FAERS) of newly approved medications in the U.S.
Method
A proprietary LLM-based model processed text to extract adverse events (AEs) and normalized terms to the Medical Dictionary for Regulatory Activities (MedDRA). The text used was from the IQVIA Insightmeme (IM) database of scientific conference abstracts, PubMed abstracts, FAERS, and USPIs
Results
We tested the hypothesis that there is additional and potentially useful safety evidence on newly approved medications in the medical conference literature. The subjects of inquiry for this study were newly approved drugs in the United States for the year 2024 with a 12-month post approval analysis window.
We analyzed 50 drugs across different therapeutic classes. From 1,437 PubMed abstracts, 2,315 conference abstracts, and approximately 6,405 FAERS (2024-2025) reports1, the LLM extracted 3,781 unique mentions of AEs. After MedDRA normalization, a total of 2,402 unique AEs were identified, 648 of which were classified as an Important Medical Event (IME)2. From this retrieval, the following was observed:
• Conferences: 201 AEs were identified from the conference data that were not listed in the USPI, PubMed or FAERS. Of these, 60 were identified as IMEs.
• Conference and FAERS Gap: 332 AEs were identified from the conference data that were not listed in FAERS. Of these, 101 were identified as IMEs.
• Conferences and FAERS: 81 AEs were identified from both the conference and FAERS data, but not in PubMed or the USPIs. Of these, 18 were identified as IMEs.
Conclusion
Early and accurate detection of potential adverse events in a drug’s lifecycle can improve patient safety, guide clinical decision making and reduce economic burden for both patients and the healthcare system. While post-marketing spontaneous reports remain a cornerstone of pharmacovigilance, the scientific literature, including conference abstracts, offers a rich, often untapped source of emerging safety data.
Our findings demonstrate that an AI-driven analysis of medical conference abstracts can uncover potential safety signals up to 12 months before they appear in FAERs, PubMed, or the U.S. Prescribing Information. This approach may not only accelerate signal detection but also addresses a critical gap: only 37.3% (95% CI, 35.3% to 39.3%) of abstracts presented at scientific conferences ever reach peer-reviewed publication, suggesting that valuable safety insights may be underreported or delayed3.
By integrating conference abstract mining into routine pharmacovigilance workflows, pharmaceutical companies may proactively identify and investigate potential safety concerns earlier in a drug’s lifecycle, thus enhancing early signal detection.
1- U.S. Food and Drug Administration. FDA Adverse Event Reporting System (FAERS). Updated quarterly. Last accessed October 18, 2025. https://fis.fda.gov/extensions/FPD-QDE-FAERS/FPD-QDE-FAERS.html
2- European Medicines Agency. (2025). Important Medical Event (IME) Terms list: Inclusion and exclusion criteria (EMA/126913/2021 Rev.). European Medicines Agency. https://www.ema.europa.eu/
3- Randomized-controlled trials show greater estimated likelihood of reaching peer-reviewed publication (68.7%) compared to other study designs (44.9%) (Scherer, et al. Cochrane Database of Systematic Reviews 2018, Issue 11). MR000005_abstract.pdf
