Leveraging Natural Language Processing to Extract Features of Colorectal Polyps From Pathology Reports for...

Area of Science:

• Medical Informatics
• Gastroenterology
• Computational Pathology

Background:

• Histopathologic features of colorectal polyps are crucial for risk factor research.
• Manual data extraction from unstructured pathology reports is labor-intensive and costly.
• Electronic Health Record (EHR) data offers valuable insights but requires integration with detailed pathology findings.

Purpose of the Study:

• To develop and evaluate a natural language processing (NLP) pipeline for automated extraction of colorectal polyp histopathologic features from pathology reports.
• To specifically focus on accurately extracting individual polyp size.
• To create an analysis-ready dataset by linking NLP-extracted features with structured EHR data.

Main Methods:

• Utilized 24,584 colonoscopy pathology reports from the University of Utah.
• Developed an annotation scheme and reference standard using 350 manually annotated reports.
• Evaluated pipeline performance against the reference standard for features including location, histology, size, shape, and dysplasia.
• Applied the validated pipeline to 24,225 unseen reports and integrated the extracted data with EHR data.

Main Results:

• The NLP pipeline achieved high performance with 98.4% F1-score, 98.9% precision, and 98.0% recall.
• Accurate extraction rates for key features included 95.6% for size, 97.2% for location, and 97.8% for histology.
• The pipeline processed 28,387 polyps, identifying tubular adenomas (55.9%) and advanced adenomas (8.1%), with a mean polyp size of 0.57 cm.

Conclusions:

• The developed NLP pipeline accurately extracts critical histopathologic features, including polyp size, from colonoscopy reports.
• This automated approach significantly enhances the efficiency of creating research-ready datasets.
• The integration of NLP-extracted pathology data with EHR data facilitates robust epidemiologic studies on colorectal polyp risk factors and outcomes.