句级交叉参照提升对人工智能生成的患者友好型放射学报告的理解与信心:面向理解的设计
该研究在116名普通读者中采用被试内设计,比较人工智能生成的患者友好型报告的三种呈现格式(词典式、整段式、句级交叉参照式)对客观与主观理解、信心和反应时间的影响。结果显示句级与整段格式在理解、信心和速度上均优于词典式,句级格式信心最高,提示界面设计是可影响患者沟通效果的可调节因素。
要点
- 句级交叉参照格式使普通读者对人工智能生成患者友好型报告的回应信心最高。
- 句级与整段格式的客观理解、主观理解和反应速度均优于词典式格式。
- 报告越长、试验越靠后,读者信心越低;报告越长反应越慢。
核心结论
- 呈现格式是影响普通读者解读人工智能生成患者友好型报告的关键且可调节因素。
- 句级与整段格式较词典式格式表现出更高的理解、信心和更快的反应。
- 新型句级格式引发最高信心,提示人工智能输出与源文本之间的交叉参照可促进信任并支持理解。
研究设计与样本
- 被试内研究,116名参与者同时查看放射学报告与人工智能生成的患者友好型报告版本。
- 比较三种格式:词典式(术语定义)、整段式(全文翻译)和新型句级式(源句与译文配对交叉参照)。
- 测量指标包括客观理解、主观理解、回应信心和反应时间。
诊断标准与关键数值
- 以词典式为参照,客观理解优势比:句级式OR 2.31(95%CI 1.36-3.94),整段式OR 2.23(95%CI 1.30-3.80)。
- 以词典式为参照,主观理解优势比:句级式OR 4.63(95%CI 2.24-9.57),整段式OR 2.18(95%CI 1.13-4.23)。
- 以词典式为参照,回应信心优势比:句级式OR 3.75(95%CI 2.41-5.84),整段式OR 2.21(95%CI 1.45-3.35),句级式高于整段式(校正P < .05)。
- 信心随试验靠后下降(OR 0.82,95%CI 0.69-0.97),随报告变长下降(OR 0.55,95%CI 0.41-0.75)。
- 反应时间较词典式更快:句级式比值0.74(95%CI 0.66-0.83),整段式比值0.71(95%CI 0.64-0.80);报告更长和健康信心更高使反应变慢。
临床应用与阅片要点
- 回应信心更高与客观理解更高相关(OR 2.75,95%CI 1.84-4.13,P < .001)。
- 回应信心更高与主观理解更高相关(b = 0.59,P < .001)。
- 作者认为,人工智能输出与源文本之间的交叉参照可促进信任并支持理解。
局限性
- 原文未报告局限性。
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
OBJECTIVE: Determine whether the format of artificial intelligence (AI)-generated patient-friendly reports (PFRs) affects laypeople's comprehension of medical information.
METHODS: In a within-subjects study, 116 participants viewed radiology reports alongside AI-generated PFR versions in three formats: dictionary (term definitions), block (full translation), and a novel sentence-based format (paired source-translation sentences for cross-referencing). Objective and subjective comprehension, confidence, and response time were measured.
RESULTS: Relative to the dictionary format, objective comprehension odds were higher for sentence (odds ratio [OR] 2.31 [95% confidence interval 1.36-3.94]) and block (OR 2.23 [1.30-3.80]), and subjective comprehension was higher for sentence (OR 4.63 [2.24-9.57]) and block (OR 2.18 [1.13-4.23]). Response confidence was higher for sentence (OR 3.75 [2.41-5.84]) and block (OR 2.21 [1.45-3.35]) than dictionary, with sentence higher than block (adjust P < .05). Confidence decreased on later trials (OR 0.82 [0.69-0.97]) and for longer reports (OR 0.55 [0.41-0.75]). Response times were faster for sentence (ratio 0.74 [0.66-0.83]) and block (ratio 0.71 [0.64-0.80]) than dictionary; longer reports and higher health-confidence slowed responses. Higher response confidence associated with greater objective (OR 2.75 [1.84-4.13], P < .001) and subjective comprehension (b = 0.59, P < .001).
DISCUSSION: Presentation format is a key, modifiable factor influencing how laypeople interpret AI-generated PFRs. Sentence and block formats showed higher comprehension, confidence, and faster responses than a dictionary format. The novel sentence-based format elicited the highest confidence, suggesting that cross-referencing between AI output and source text fosters trust and supports understanding. These findings highlight the critical role of user interface design in effective AI-generated patient communication.
引用
- Armstrong BA, Fang Z, Johnston A, et al. Sentence-Level Cross-Referencing Improves Comprehension and Confidence in AI-Generated Patient-Friendly Radiology Reports: Design for Understanding. J Am Coll Radiol. 2026. PMID: 41812853. doi:10.1016/j.jacr.2026.02.037
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。