Large Language Models advance clinical data extraction and patient safety

Recent research highlights the growing role of Large Language Models (LLMs) in enhancing clinical data processing and patient safety. One study proposes a framework using GPT-4o-mini to generate synthetic data for clinical named entity recognition. By integrating self-verification and semantic mapping with SNOMED-CT, researchers found that a 1:1 ratio of human-annotated to synthetic data optimizes model performance for fine-tuning LLaMA-3-8B.

In the field of neurosurgery, researchers are testing Retrieval-Augmented Generation (RAG) models to predict patient outcomes. While current datasets are insufficient for full implementation, proof-of-concept testing indicates that RAG-LLM models can provide information derived strictly from datasets without the typical hallucinations associated with standard LLMs.

Additionally, new methodologies are being developed to improve patient safety event studies. By combining an extract-transform-load (ETL) pipeline with the OpenFDA API and LLMs, researchers can more accurately analyze free-text narratives in medical device reports from the MAUDE database, enhancing the efficiency of event categorization.