Improving Faithfulness in Medical Text Summarization: An LLM-Based Approach

Abstract

Medical text summarization, particularly for consumer health queries, is a critical area of research due to its potential to enhance healthcare delivery. In the context of online consumer health questions, efficient communication is paramount. The abil ity to condense lengthy and complex medical queries while retaining essential details is crucial for facilitating timely and accurate responses from healthcare profession als. This not only improves patient outcomes but also optimizes the workflow within healthcare services, making it an indispensable tool in modern medical practice. However, a persistent concern in the realm of consumer health questions is the faith fulness of the summarized queries. Preserving the accuracy of information, especially when dealing with complex medical terminology, is of utmost importance. While the primary focus of text summarization has traditionally been on accuracy, the aspect of faithfulness—ensuring that the summary accurately represents the source mate rial—has often been overlooked. Our research aims to address this gap by enhancing the faithfulness of consumer health queries in addition to improving their accuracy. To address these challenges, we are leveraging large language models (LLMs), which have recently shown significant promise in text summarization tasks. Ultimately, our objective is to improve both the faithfulness and accuracy of LLMs in summarizing medical texts. To achieve this, we propose a novel framework that fine-tunes LLMs using domain-specific medical knowledge. This approach aims to balance concise summarization with the precise representation of medical information, ensuring that the essential details are faithfully conveyed.