Unpacking Unstructured Data: A Pilot Study on Extracting Insights from Neuropathological Reports of Parkinson’s Disease Patients using Large Language Models

Abstract

ABSTRACTObjectiveThe aim of this study was to make unstructured neuropathological data, located in the NeuroBioBank (NBB), follow FAIR principles, and investigate the potential of Large Language Models (LLMs) in wrangling unstructured neuropathological reports. By making the currently inconsistent and disparate data findable, our overarching goal was to enhance research output and speed.Materials and MethodsThe NBB catalog currently includes information from medical records, interview results, and neuropathological reports. These reports contain crucial information necessary for conducting in-depth analysis of NBB data but have multiple formats that vary across sites and change over time. In this study we focused on a subset of donors with Parkinson’s Disease (PD). We developed a data model with combined Brain Region and Pathological Findings data at its core. This approach made it easier to build an extraction pipeline and was flexible enough to convert resulting data to Common Data Elements (CDEs) used by the community.ResultsThis pilot study demonstrated the potential of LLMs in structuring unstructured neuropathological reports of PD patients available in the NBB. The pipeline enabled successful extraction of microscopic and macroscopic findings and staging information from pathology reports, with extraction quality comparable to results of manual curation. To our knowledge, this is the first attempt to automatically standardize neuropathological information at this scale. The collected data has the potential to serve as a valuable resource for PD researchers, bridging the gap between clinical information and genetic data, thereby facilitating a more comprehensive understanding of the disease.