Designing AI-Based Work Processes: How the Timing of AI Advice Affects Diagnostic Decision Making
Management Science · 2025 · Jiamin Yin、Kee Yuan Ngiam、Sharon Swee-Lin Tan、Hock Hai Teo
中文摘要
尽管临床人工智能（AI）系统能够通过提供可靠的第二意见来辅助医学诊断决策，但如何将AI有效整合到常规诊断流程中——例如何时向人类医生呈现AI建议——在很大程度上仍未被探讨。因此，本研究采用出声思维（think-aloud）方法，通过实验考察AI建议的提供时点如何影响诊断决策。医生在三种条件下进行医学诊断：事后建议（在初步诊断之后给出AI建议）、事前建议（AI建议与临床信息同时给出）以及对照条件（不提供AI建议）。我们的结果表明，AI建议的提供时点显著影响诊断准确性和校准度，其中事后建议条件表现最佳，对照条件表现最差。随后我们进行了若干分析以厘清其内在机制。我们发现，事后建议条件下更优的诊断质量可归因于更充分的临床信息加工以及对AI推理依据更积极的认知投入。因此，与事前建议条件下的参与者相比，事后建议条件下的参与者更有能力区分正确与错误的AI建议。此外，他们能从与其初始诊断相矛盾的高质量AI建议中获益更多。为获得更多洞见，我们基于医生和临床病例特征估计了异质性处理效应。我们的发现凸显了在常规诊断流程中于适当时点呈现AI建议、以实现借助AI建议成功增强决策的重要性。
Abstract
Although clinical artificial intelligence (AI) systems can augment medical diagnosis decisions by providing competent second opinions, how to effectively integrate AI into routine diagnostic processes, such as when to present AI advice to human physicians, remains largely unexplored. Therefore, our research experimentally examines how the timing of AI advice affects diagnostic decision making using a think-aloud approach. Physicians perform medical diagnoses under three conditions: ex post advice (AI advice given after an initial diagnosis), ex ante advice (AI advice given concurrently with clinical information), and a control condition (no AI advice). Our results indicate that the timing of AI advice significantly affects diagnostic accuracy and calibration, with the ex post advice condition yielding the best performance and the control condition the worst. We then conduct several analyses to disentangle the underlying mechanism. We reveal that the superior diagnostic quality in the ex post advice condition can be attributed to more thorough clinical information processing and more active cognitive engagement with AI’s reasoning rationale. As a result, participants in the ex post advice condition are more capable of differentiating correct from incorrect AI advice than those in the ex ante advice condition. Additionally, they benefit more from high-quality AI advice that contradicts their initial diagnoses. To gain additional insights, we estimate the heterogeneous treatment effects based on physician and clinical case characteristics. Our findings underscore the importance of presenting AI advice at appropriate times during routine diagnostic processes to achieve successful decision augmentation with AI advice.