Leveraging LLMs to Understand Narratives in MAUDE Reports

Area of Science:

• Medical device safety surveillance
• Natural Language Processing (NLP) in healthcare

Background:

• The MAUDE database contains valuable narrative reports on medical device adverse events.
• These narratives are largely unexplored, limiting comprehensive analysis of device-related issues.

Purpose of the Study:

• To explore the utility of large language models (LLMs) in analyzing MAUDE database narrative sections.
• To identify uncoded surgical procedures and extract additional insights from reports on endoscopic clips.

Main Methods:

• Utilized OpenAI's GPT-4-turbo model for natural language processing of MAUDE report narratives.
• Focused analysis on reports involving endoscopic clips to identify specific procedural details.

Main Results:

• Demonstrated LLMs' capability to process and interpret unstructured narrative data from MAUDE reports.
• Identified uncoded surgical procedures and uncovered additional relevant information within the narratives.

Conclusions:

• LLMs offer an efficient and cost-effective method for analyzing MAUDE narrative data.
• This approach enhances the translation of MAUDE reports into actionable clinical knowledge for medical device safety.