A Question-and-Answer System to Extract Data From Free-Text Oncological Pathology Reports (CancerBERT Network):...

Area of Science:

• Oncology
• Computational Linguistics
• Bioinformatics

Background:

• Pathology reports are crucial for cancer care, but extracting information can be challenging.
• Existing natural language processing (NLP) systems often require extensive tuning or are limited in scope.
• Deep learning approaches, particularly BERT, offer powerful new capabilities for information extraction.

Purpose of the Study:

• To develop a BERT-based system for automatic extraction of tumor site and histology from oncological pathology reports.
• To accurately extract tumor site and histology descriptions, accommodating diverse terminology.
• To provide standardized International Classification of Diseases for Oncology, Third Edition (ICD-O-3) codes for downstream applications.

Main Methods:

• Trained a base BERT language model on over 275,000 pathology reports for unsupervised learning.
• Developed a question-and-answer (Q&A) model using supervised learning on 8,197 reports to answer specific pathology questions.
• Fine-tuned additional BERT models to predict ICD-O-3 site and histology codes, creating a network named caBERTnet.

Main Results:

• caBERTnet achieved high accuracies in predicting group-level site (93.53%) and histology (97.6%) codes.
• Top 5 accuracies for fine-grained ICD-O-3 code prediction were 92.95% for site and 96.01% for histology.
• The system demonstrated strong performance across a wide range of tumor sites and histologies.

Conclusions:

• Developed an NLP system (caBERTnet) that surpasses existing algorithms in predicting ICD-O-3 codes.
• The system accurately extracts critical tumor information, standardizing it for clinical use.
• This advancement has the potential to reduce treatment delays and improve patient outcomes.