Enhancing Clinical Trial Summarization: Leveraging Large Language Models and Knowledge Graphs for Entity Preservation

ClinicalTrials.gov is an accessible online medical resource for researchers, healthcare professionals, and policy designers seeking detailed information on clinical trials. Summarizing these long clinical records can significantly reduce the time needed for the database users as the process transforms comprehensive information into concise synopses, preserving the essential meaning and facilitating understanding. In this paper, we employ the Bidirectional and Auto-Regressive Transformers model to generate the trials' brief summaries. Our contributions provide new preprocessing techniques for model training, which leads to a robust summarization model. The fine-tuned model significantly enhanced ROUGE-1, ROUGE-2, and ROUGEL F1-scores by 14%, 23%, and 20%, respectively, compared to previous studies. Additionally, we present an innovative knowledge graph based on entity classes to assess the generated summaries. This graph not only quantifies the essential entities transformed from the original text to the summaries but also provides insights into their specific order and arrangement in sentences.