Enhancing Medical Record Comprehensibility: Using Large Language Models to Produce Simplified Narratives of Image Reports in Electronic Medical Data

Abstract

This thesis investigates the use of language models in clinical applications where input documents are long and questions can be complex and require advanced reasoning. The research aims and objectives in this thesis contain two parts: one is to find and evaluate the appropriate solution that enhances the model performance under clinical settings; the other one is to find the solution to modify the models for better performance under this circumstance. Two approaches are developed to address the problem. In the first study, the RAPTOR framework extends a language model's ability to make sense of local and global information from long documents with its unique hierarchical tree structure datastore. The approach may be beneficial where cloud-based large language models (e.g. GPT-4o) cannot be used due to data privacy or reproducibility issues. Specifically, RAPTOR can be tailored to address clinical tasks including extracting critical patient information and summarizing clinical notes from long documents. In the second study, I developed and tested an optimized language model that uses a continual pre-training process to incorporate domain knowledge with a Llama-3.1-8B language model, with a novelly collected, organized and preprocessed Clinical Trial registration dataset called CiTi. The dataset contains 358870 preprocessed clinical trial registration reports and 1401401 related publication abstracts. This study aimed to develop a language model that adapts clinical trial registration data as its specialty domain and has more understanding of this domain than general large language models. The two solutions evaluated in this thesis show that with the updated configuration, it is possible to achieve state-of-the-art performance using locally implemented language models. Future research should consider how specific configurations or auto-configurations better suit simple and complex questions.