A new method named VERGE was developed to extract six red-flag symptoms and family-history risk status from free-text clinical notes related to early-onset colorectal cancer.
VERGE employs a retrieval-augmented generation approach to propose initial labels and evidence, followed by a bounded verification-refinement cycle that checks textual grounding and clinical validity. This process corrects and rechecks claims until resolved or a limit is reached, escalating unresolved claims for human review.
Evaluation on over 4,000 clinician-labeled note pairs showed that VERGE reduced false positives, improving precision from 0.764 to 0.849 and MCC from 0.681 to 0.730 compared to a single-agent baseline. Most flagged errors were resolved autonomously, with only 1.5% requiring human review.
These results demonstrate that a bounded, verification-based workflow can improve the reliability of clinical language-processing tools, supporting better colorectal cancer risk assessment in younger patients.
Source: https://arxiv.org/abs/2609.04366