Related Experiment Video
Updated: Aug 6, 2026
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
Real-world pharmacovigilance for anti-Aβ therapies using a large language model
Allan Fong1,2, Azade Tabaie1,2,3, Samantha Paylor4
1Center for Biostatistics, Informatics, and Data Science MedStar Health Research Institute, MedStar Health Washington District of Columbia USA.
Introduction:
Anti-amyloid beta (Aβ) therapies for early Alzheimer's disease require enhanced safety monitoring, yet adverse event (AE) documentation is diffuse across heterogeneous electronic health record documents. Large language models (LLMs) may improve scalable pharmacovigilance.
Methods:
We analyzed 20,123 clinical documents from 46 patients who received at least one dose of anti-Aβ therapy (June 24, 2024-July 30, 2025) at a large mid-Atlantic health-care system. We compared standard expert review versus an LLM-augmented workflow applied to the same documents. Expert reviewers annotated therapy-related AEs (e.g., amyloid-related imaging abnormalities with edema or hemorrhage, headache, syncope, hypersensitivity, gastrointestinal symptoms, infusion reactions). Discordant cases were adjudicated to establish a reference label.
Results:
After adjudication, 76% (35/46) patients had an AE. The LLM-augmented workflow achieved 100% sensitivity (positive predictive value [PPV] 89.7%) versus expert review 88.6% sensitivity (PPV 100%).
Discussion:
Findings provide preliminary indications that LLMs may serve as a pharmacovigilance signal detection tool, with a need for further validation and evaluation of clinical integration.
