Classifying Tumor Reportability Status From Unstructured Electronic Pathology Reports Using Language Models in a...

Area of Science:

• Computational oncology
• Natural Language Processing (NLP)
• Machine Learning in Healthcare

Background:

• Population-based cancer registries (PBCRs) are crucial for cancer surveillance, relying on pathology reports.
• Current PBCR data extraction often uses manual or rule-based solutions, which are labor-intensive and sensitive to linguistic variations.

Purpose of the Study:

• To develop and deploy a state-of-the-art NLP pipeline for automated detection of reportable tumors.
• To enhance the efficiency and accuracy of cancer data collection within the British Columbia Cancer Registry (BCCR).

Main Methods:

• Fine-tuning two clinical language models (GatorTron and BlueBERT) using 40,000 electronic pathology reports from BCCR.
• Combining the outputs of both fine-tuned models using an OR approach for final decision-making.
• Evaluating the pipeline on extensive test datasets spanning diagnosis years 2021-2023.

Main Results:

• The NLP pipeline demonstrated boosted reportable accuracy.
• The system successfully maintained a true reportable threshold of 98% in retrospective evaluations.
• The approach offers a significant improvement over traditional rule-based methods.

Conclusions:

• Deep learning-based NLP methods significantly outperform rule-based approaches in cancer surveillance.
• The developed pipeline offers a more robust and accurate solution for distinguishing reportability status in electronic pathology reports.
• This advancement provides substantial advantages for PBCRs in managing and analyzing cancer data.