Obtaining Knowledge in Pathology Reports Through a Natural Language Processing Approach With Classification,...

Area of Science:

• Computational pathology
• Bioinformatics
• Natural Language Processing

Background:

• Tumor banks require continuous sample curation for research utility.
• Semistructured clinical pathology notes hinder automated data abstraction.
• Accurate specimen identification is crucial for tumor bank management.

Purpose of the Study:

• To develop a novel natural language processing (NLP) method for automated extraction of pathology specimen data.
• To improve the dynamic population of tumor registries by identifying overlooked specimens.
• To create a re-implementable solution for institutional tumor banks.

Main Methods:

• A composite NLP pipeline was developed using supervised machine learning.
• The pipeline included classification, named-entity recognition (NER), and results proofreading.
• The system was trained and validated on gastroesophageal cancer pathology reports.

Main Results:

• The NLP pipeline successfully extracted specimen label, date, and location from 188 pathology reports.
• Up to 24 additional unique samples were identified in external consult notes.
• The classification model achieved 100% accuracy; NER models demonstrated strong performance.

Conclusions:

• A re-implementable, automated NLP and machine learning approach can accurately extract specimen attributes from pathology notes.
• This method dynamically populates tumor registries, enhancing sample curation.
• The developed pipeline addresses a critical need for efficient tumor bank management.