识别与理解人工智能辅助放射学中认知偏倚的框架
本文提出一个概念框架,将人工智能辅助放射学中的认知偏倚与放射科医生和AI的交互过程相联系。该框架基于双加工理论,在七个决策场景中分析AI输出时机、医-AI一致性、信任校准和读片者专业水平对决策的影响,并识别出多种认知偏倚及AI悖论,提出相应策略。
要点
- AI辅助读片时需警惕自满、确认、自动化、不信任、锚定和通常含义偏倚。
- AI输出时机、医-AI一致性和信任校准影响决策偏倚风险。
- 建议通过教育、工作流程与界面设计、临床监督和性能监测减少偏倚。
核心结论
- AI辅助放射学引入与放射科医生认知过程相关的新诊断错误易感性,需通过校准信任来平衡人类专业知识和AI优势。
研究设计与样本
- 本文为观点文章(Perspective),提出一个概念框架,使用双加工理论作为组织模型,未报告样本量、中心数或设备信息。
临床应用与阅片要点
- 日常阅片应关注AI输出时机和医-AI一致性对判断的影响,避免在AI辅助下不加限定地使用结果。
- 需校准对AI的信任,防止过度依赖或完全不信任,并注意自满、确认、自动化、锚定和通常含义偏倚。
- 建议结合教育、工作流程与界面设计、临床监督和性能监测来管理认知偏倚。
局限性
- 本文为概念框架,未提供实证数据验证;原文未报告样本量、中心数或统计结果。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
Artificial intelligence (AI) creates new susceptibility to diagnostic error in radiology related to the cognitive processes by which radiologists' interpretations incorporate AI outputs. This Perspective presents a conceptual framework for cognitive biases in AI-assisted radiology linking specific biases to radiologist-AI interactions. The framework uses dual-process theory, recognizing fast intuitive and deliberate analytic modes of information processing, as an organizing model for understanding cognitive processes underlying radiologists' decisions and their potential contribution to bias. Across seven decision scenarios spanning image viewing, findings interpretation, and contextual integration, the framework examines how AI output timing, radiologist-AI agreement, trust calibration, and reader expertise influence decision-making. We identify adaptive heuristics that support efficient decisions but, when misapplied or overrelied upon, increase susceptibility to complacency, confirmation, automation, distrust, anchoring, and usual implication bias, and increase behavioral risk of use without qualification. We highlight AI paradoxes emerging from the framework of despecialization, system fragility, and deskilling. Finally, we propose strategies targeting these cognitive biases, including education and experiential learning, workflow and interface design, clinical oversight, and performance monitoring. This framework provides a structured basis for generating testable hypotheses and designing AI-assisted radiology systems that preserve the strengths of human expertise and AI while supporting calibrated reliance.
引用
- Armstrong BA, Cheuy L, Kottler N, et al. A Framework for Identifying and Understanding Cognitive Biases in Artificial Intelligence-Assisted Radiology. AJR Am J Roentgenol. 2026. PMID: 42949357. doi:10.2214/AJR.26.35826
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。