文献速递技术与安全

RadTranslateGPT:一种改进的基于人工智能的放射学报告翻译与简化系统

RadTranslateGPT: An Improved AI-Based System for Translation and Simplification of Structured Radiology Reports.
PubMed 文献 · J Am Coll Radiol · 2026-04-09

本研究从机构数据库选取30份去标识化急诊放射学报告,由3名急诊放射科医生评估GPT-o1简化报告,并由9名医疗口译员比较GPT-o1与Google Translate的翻译质量。GPT-o1简化报告在76.9%的评估中被评为「极其准确」或「非常准确」,其翻译准确度中位数显著高于Google Translate(4.0 vs 3.0,P < .001),所有测试语言中语言语域评分和可理解性均更优。

要点

  • GPT-o1简化报告76.9%被评为极其或非常准确,98.9%对患者理解有用。
  • GPT-o1翻译准确度中位数4.0 vs Google Translate 3.0,P < .001。
  • 作者指出本研究为样本量有限的先导研究,临床整合前需进一步验证。

核心结论

  • GPT-o1能生成简化、患者友好的放射学报告,并能高质量翻译为西班牙语、阿拉伯语和普通话;但作者指出这些发现应结合本先导研究样本量有限的背景解读,且大型语言模型在临床整合前需进一步验证。

研究设计与样本

  • 先导研究,从机构数据库选取30份去标识化急诊放射学报告;第一阶段由3名委员会认证的急诊放射科医生评估GPT-o1简化报告,第二阶段由9名医疗口译员(每种语言3名)评估GPT-o1和Google Translate的输出,两组均采用5点Likert型量表评分。

诊断标准与关键数值

  • 共90次简化报告评估:69份(76.9%)被评为「极其准确」或「非常准确」,41份(45.6%)被评为「非常清晰」,89份(98.9%)被评为对患者理解「有用」。
  • GPT-o1翻译准确度中位数4.0 vs Google Translate 3.0(P < .001)。
  • GPT-o1在所有测试语言中语言语域评分更高,且被评为比Google Translate更易理解。

临床应用与阅片要点

  • 作者认为大型语言模型可作为工具,用于提升有限英语水平人群的健康素养。
  • 作者指出在临床整合前需要对这些大型语言模型进行进一步验证。

局限性

  • 作者指出这些发现应结合本先导研究样本量有限的背景解读。
  • 作者指出大型语言模型在临床整合前需进一步验证。

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

PURPOSE: Radiology reports often contain complex medical jargon that can be difficult for patients to understand, especially those with limited English proficiency. With increased access to electronic health records, there is a need for patient-friendly and multilingual interpretation of radiology reports. The aim of this study was to evaluate the performance of GPT-o1 in simplifying and translating emergency radiology reports into Spanish, Arabic, and Mandarin and to compare it with Google Translate.

METHODS: Thirty deidentified emergency radiology reports were selected from an institutional database. Phase 1 involved the evaluation of the GPT-o1-simplified reports by three board-certified emergency radiologists. Phase 2 involved nine medical interpreters (three per language) assessing both GPT-o1 and Google Translate outputs. Both groups rated outputs on a 5-point, Likert-type scale.

RESULTS: There were a total of 90 evaluations of the simplified radiology reports. Sixty-nine the reports (76.9%) were rated as "extremely accurate" or "very accurate," 41 (45.6%) were rated as "very clear," and 89 (98.9%) were marked as "useful" for patient comprehension. GPT-o1 achieved significantly higher translational accuracy (median, 4.0 vs 3.0; P < .001) and higher language register scores and was rated more comprehensible than Google Translate in all languages tested.

CONCLUSIONS: GPT-o1 generated simplified, patient-friendly radiology reports and also produced high-quality translations into the three languages tested; however, these findings should be interpreted in the context of this pilot study with a limited sample size. These findings suggest that large language models could serve as tools to enhance health literacy for limited English proficiency populations. Further validation of these large language models is needed before clinical integration.

来源:J Am Coll Radiol,查看原文。版权归原作者所有。

引用

  1. Khanna P, Mazumder A, Samuel JKR, et al. RadTranslateGPT: An Improved AI-Based System for Translation and Simplification of Structured Radiology Reports. J Am Coll Radiol. 2026. PMID: 41966443. doi:10.1016/j.jacr.2026.04.006

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。