A1.2 Leveraging Language Models for Clinical Data Management: A Case Study on Text Cleaning and Mapping

This presentation is part of a full breakout session. To add this presentation to your schedule you must add the full breakout session.

Clinical data management constantly seeks more efficient and accurate processes. This presentation explores an approach using Large Language Models (LLMs), specifically OpenAI’s ChatGPT and embedding models, as alternatives to traditional text cleaning and mapping methods.

We will present a use case on mapping concomitant medication verbatim to standard dictionary terms in the WHO Drug Dictionary. The process begins with ‘prompt engineering’, a technique that refines raw text data by removing non-essential information, such as dosage details and usage instructions. The cleaned data is then processed using embedding and semantic search to identify the most similar term from the WHO Drug Dictionary.

This approach simplifies the task traditionally performed using regular expressions for text cleaning and Fuzzy matching for similarity search. By using ChatGPT for text cleaning and embedding models for term mapping, we offer an alternative that may provide more accurate results and a more efficient process.

Our presentation will share experiences and findings from this project, demonstrating the practical application of LLMs in clinical data management. This discussion will be of interest to Clinical Trial Data Managers (CTDMs) seeking reliable data, as well as a technical audience interested in the application of AI techniques. We aim to foster a dialogue about the potential of technology in clinical data management, making it relevant to both technical and non-technical audiences.