University of California, Irvine–Pathology Extraction Pipeline: The pathology extraction pipeline for information extraction from pathology reports

Abstract

We describe Pathology Extraction Pipeline (PEP)—a new Open Health Natural Language Processing pipeline that we have developed for information extraction from pathology reports, with the goal of populating the extracted data into a research data warehouse. Specifically, we have built upon Medical Knowledge Analysis Tool pipeline (MedKATp), which is an extraction framework focused on pathology reports. Our particular contributions include additional customization and development on MedKATp to extract data elements and relationships from cancer pathology reports in richer detail than at present, an abstraction layer that provides significantly easier configuration of MedKATp for extraction tasks, and a machine-learning-based approach that makes the extraction more resilient to deviations from the common reporting format in a pathology reports corpus. We present experimental results demonstrating the effectiveness of our pipeline for information extraction in a real-world task, demonstrating performance improvement due to our approach for increasing extractor resilience to format deviation, and finally demonstrating the scalability of the pipeline across pathology reports for different cancer types.