Clinical Trial Eligibility Criteria Decomposition and Parsing with Large Language Models

Clinical trial eligibility criteria, often presented as complex free text, pose significant challenges for automated processing. This study introduces a Decomposition and Parsing (DP) workflow to address these challenges by systematically breaking down criteria into "study traits"-the smallest meaningful units-and structuring them with components such as entities, modifiers, constraints, and negations. Leveraging advanced large language models (LLMs) like GPT-4o and Llama3.3 with Chain-of-Thought prompting, the workflow successfully processes Alzheimer's disease trial datasets, achieving strong performance in tasks like logical relationship extraction and trait computability determination. However, challenges remain in capturing nuanced elements like modifiers. The study also proposes innovative evaluation metrics that outperform traditional approaches in assessing the quality of automated extractions. This scalable and intuitive framework advances the representation of clinical trial eligibility criteria, paving the way for improved biomedical informatics applications and highlighting the need for domain-specific fine-tuning and broader dataset integration.