2026年第51期(总第1195期)
演讲主题：When Explainability Helps or Hurts: Explanation Representation and Cognitive Fit in Medical Human-AI Collaboration
主讲人：周伟华 浙江大学教授
主持人：关旭 教授、供应链管理与系统工程系主任
活动时间：2026年06月26日（周五）14:00-15:30
活动地址：管院大楼107教室
主讲人简介：
周伟华博士，浙江大学求是特聘教授、博士生导师。现任浙江省数智管理与决策技术重点实验室主任、浙江大学数据分析与管理国际研究中心主任，担任浙江大学“大数据+分析与管理”创新团队首席专家，浙江省高水平创新团队“数据分析与管理”负责人。主要研究领域包括数据智能、供应链管理与决策。在国内外顶级期刊物《Management Science》《Operations Research》《Manufacturing & Service Operations Management》《管理科学学报》等发表多篇论文。主持国家自然科学基金重大项目课题、国家自然科学基金重点项目和科技部重点研发计划等多项国家级课题。
活动简介：
Explainability can improve medical human–AI collaboration, but it can also make AI advice harder to use. We argue that the same AI advice can impose different cognitive work depending on how its rationale is displayed. In medical AI, many mainstream XAI methods represent model reasoning statistically, through coefficients, feature weights, or local attributions such as SHAP values. Drawing on bounded rationality and cognitive fit, we theorize statistical XAI as a double-edged cognitive intervention: it can increase physicians’ perceived functional understanding of AI advice, but it can also make explainability hurt by requiring physicians to translate feature-statistical signals into clinically meaningful reasons under limited attention. Case-based XAI, implemented as explanation-by-example, should reduce this burden by representing AI support as a comparable patient case and validated outcome. We test these arguments in a randomized experiment with 418 physicians making 4,180 high-stakes EHR-based clinical decisions. Physicians first made independent judgments and then revised them after receiving identical binary AI advice accompanied by one of four information formats: AI-only, logistic-regression explanation, SHAP explanation, or case-based XAI explanation. Statistical explanations improved final accuracy by 3.14–4.26 correct decisions per 100 evaluations, but increased cognitive load and decision time; their benefits were concentrated among physicians with higher AI literacy. Case-based XAI delivered a comparable gain of 4.22 correct decisions per 100 evaluations, increased perceived functional understanding, did not significantly increase cognitive load or total decision time, and showed weaker AI-literacy dependence. These findings show when explainability helps or hurts: it helps when explanation representation fits physicians’ cognitive operations, and hurts when transparency shifts translation work onto clinicians.
