VERGE: Verification-Enhanced Refinement for Grounded Extraction of Early-Onset Colorectal Cancer Symptoms in Clinical Notes VERGE：通过验证增强细化实现临床笔记中早发性结直肠癌症状的接地提取

TL;DR

• VERGE is an agentic workflow for extracting six red-flag symptoms and family-history risk status from free-text clinical notes related to early-onset colorectal cancer
• It uses retrieval-augmented generation for initial label and evidence proposal, followed by a bounded verification-refinement cycle that checks textual grounding and clinical validity
• VERGE improved precision from 0.764 to 0.849 and MCC from 0.681 to 0.730 compared to a single-agent baseline on 4,033 clinician-labeled note-finding pairs
• Only 1.5% of claims required human review, with the system autonomously resolving most flagged errors
• The verification-based approach reduces false positives without sacrificing true positive detection, offering a path toward more trustworthy clinical NLP tools

Why It Matters

This research addresses a critical gap in early-onset colorectal cancer detection, where structured encounter data fails to capture essential symptom details like duration, context, and family history. For AI practitioners working in clinical NLP, VERGE demonstrates how bounded verification-refinement cycles can significantly improve precision while maintaining recall—a key challenge in healthcare applications where false positives can lead to unnecessary procedures and patient anxiety.

Technical Details

• VERGE employs an agentic workflow combining retrieval-augmented generation (RAG) for initial extraction with a bounded verification-refinement loop that iteratively checks textual grounding and clinical validity
• The system extracts six red-flag symptoms and family-history risk status from free-text clinical notes, targeting early-onset colorectal cancer screening in younger adults
• Evaluation was conducted on 4,033 clinician-labeled note-finding pairs, comparing VERGE against a single-agent baseline, a rule-based clinical language-processing baseline, and an alternative underlying language model
• Performance metrics showed precision improvement from 0.764 to 0.849 and Matthews Correlation Coefficient (MCC) increase from 0.681 to 0.730, with only 1.5% escalation rate to human reviewers
• The bounded verification cycle corrects and rechecks claims until resolved or a limit is reached, balancing autonomous correction with human oversight for unresolved cases

Industry Insight

• Verification-enhanced agentic workflows represent a scalable pattern for improving reliability in clinical NLP systems, particularly where false positives carry significant downstream costs
• The 1.5% human review rate demonstrates that bounded verification can dramatically reduce human-in-the-loop overhead while maintaining clinical-grade accuracy, making deployment more feasible at scale
• As early-onset cancer rates rise globally, this approach could be adapted to other conditions where structured data is insufficient and free-text clinical notes contain critical diagnostic signals

TL;DR

• 提出VERGE框架，用于从自由文本临床笔记中自动提取早发性结直肠癌的6种红旗症状和家族史风险状态
• 采用检索增强生成（RAG）生成初始标签和证据，再通过有界验证-精炼循环进行文本依据和临床有效性检查
• 在4,033对临床医生标注的笔记-发现对中评估，相比单代理基线，精确率从0.764提升至0.849，MCC从0.681提升至0.730
• 仅需1.5%的声明需要人工审核，其余错误可自主解决
• 验证驱动的工作流可在不牺牲检测能力的情况下减少不必要的阳性发现，为临床NLP工具提供更可靠、可信的解决方案

为什么值得看

早发性结直肠癌在年轻人群中呈上升趋势，但缺乏针对该年龄组的循证随访指南，且结构化就诊数据无法捕捉症状持续时间、背景和家族史等关键信息。VERGE通过引入验证-精炼机制，显著提升了临床文本信息提取的精确度和可信度，为AI辅助癌症早期筛查提供了可复用的技术范式。

技术解析

VERGE采用多阶段代理工作流：首先通过检索增强生成（RAG）提出初始标签和证据，然后进入有界的验证-精炼循环，该循环同时检查文本依据和临床有效性，对声明进行修正并重新验证，直到问题解决或达到上限，最后将未解决的声明升级给人工审核。评估基于4,033对临床医生标注的笔记-发现对，与单代理基线、基于规则的临床语言处理基线以及替代基础语言模型进行对比。相比单代理基线，VERGE将精确率从0.764提升至0.849，MCC从0.681提升至0.730，实现了精确率-召回率权衡的平衡改善，且仅需1.5%的声明需要人工审核。

行业启示

验证驱动的工作流设计为临床NLP系统提供了可复用的架构范式，通过有界的验证-精炼循环在保持检测能力的同时显著降低假阳性率。这种"生成-验证-修正"的代理工作流模式可推广至其他需要高可信度的医疗信息提取场景。对于医疗AI产品，引入验证机制和人工审核升级路径是平衡自动化效率与临床安全性的关键策略。