CAT: computer aided triage improving upon the Bayes risk through ε-refusal triage rules

Area of Science:

• Computational pathology
• Health informatics
• Machine learning applications

Background:

• Manual data extraction for the Surveillance, Epidemiology, and End Result (SEER) database is labor-intensive.
• Automated methods using machine learning (ML) and natural language processing (NLP) are needed to improve SEER data timeliness and efficiency.
• Current ML/NLP algorithms struggle to match human expert precision and face challenges with varied registry formats.

Purpose of the Study:

• To develop triage rules for partially automating registry workflows.
• To enhance the precision of automatically extracted information from electronic pathology reports.
• To improve the overall efficiency and scalability of cancer data collection.

Main Methods:

• Developed a mathematical framework to improve classifier precision beyond the Bayes classifier.
• Introduced a triage rule that selectively classifies items with high confidence.
• Characterized the optimal triage rule for classification tasks.

Main Results:

• Demonstrated the usefulness of the triage rule in classifying cancer sites from electronic pathology reports.
• Achieved a desired level of precision in automated information extraction.
• Showcased significant improvements in classification accuracy through the proposed triage rule.

Conclusions:

• Proposed a heuristic triage rule based on post-processing machine learning model outputs (soft-max).
• The developed triage rule significantly enhances classification accuracy in test cases.
• The approach offers a pathway to more precise and efficient automated data extraction for cancer registries.