文献速递乳腺病理血液淋巴病理细胞病理与分子病理

病理简化:ChatGPT 对病理报告的摘要能力评估

Pathology Made Simple: ChatGPT's Summarization of Pathology Reports.
PubMed 文献 · Arch Pathol Lab Med · 2025-08-21

该研究回顾性分析 2022-2023 年 120 份匿名病理报告,聚焦乳腺病理、黑素细胞病变和淋巴瘤 3 个亚专科各 40 例复杂病例,用标准化简短提示让 ChatGPT-4o 生成一句话摘要,并由两名独立病理学家评估。结果显示摘要可读性高,但关键错误率在不同亚专科间差异明显。

要点

  • ChatGPT-4o 生成的一句话摘要在 120 例中 119 例可读性评分为 10/10。
  • 淋巴瘤摘要关键信息纳入评分最高(9.55),黑素细胞病变最低(8.09)。
  • 无关键错误率:淋巴瘤 95%、乳腺癌 77.5%、黑素细胞病变 62.5%。

核心结论

  • 在标准化提示和专家监督下,ChatGPT-4o 生成简洁可读的病理报告摘要具有前景,但偶发错误及处理复杂或多部分病例的局限性仍存在。

研究设计与样本

  • 回顾性分析 2022-2023 年 120 份匿名病理报告,聚焦乳腺病理、黑素细胞病变和淋巴瘤 3 个亚专科各 40 例复杂病例。
  • 使用标准化简短提示让 ChatGPT-4o 为每份报告生成一句话摘要,由两名独立病理学家评估。

主要发现

  • 关键信息纳入平均分:黑素细胞病变 8.09、乳腺癌 8.15、淋巴瘤 9.55。
  • 无关键错误率:黑素细胞病变 62.5%、乳腺癌 77.5%、淋巴瘤 95%。
  • 排除非必要信息在各亚专科评分 consistently 高,可读性在 120 例中 119 例评为 10/10。

临床怎么用

  • ChatGPT-4o 可在标准化提示和专家监督下辅助生成病理报告摘要,但需警惕关键错误,尤其黑素细胞病变。

局限性

  • 偶发错误及处理复杂或多部分病例的局限性被记录。
  • 需进一步优化和领域特定模型训练以提升可靠性和临床效用。

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)。

原文

CONTEXT.—: Pathology reports are essential for guiding clinical decisions but are often complex and lengthy. Artificial intelligence tools like ChatGPT may offer a way to distill these reports into clear, concise summaries to improve communication and efficiency in clinical settings. OBJECTIVE.—: To evaluate the performance of ChatGPT-4o in summarizing detailed pathology reports into 1-sentence diagnoses that retain critical clinical information and are accessible to medical professionals. DESIGN.—: We retrospectively analyzed 120 anonymized pathology reports from 2022-2023, focusing on 40 complex cases from 3 subspecialties: breast pathology, melanocytic lesions, and lymphomas. Using a standardized brief prompt, ChatGPT-4o generated 1-sentence summaries for each report. Two independent pathologists assessed each summary for inclusion of essential information, exclusion of irrelevant details, presence of critical errors, and overall readability. RESULTS.—: The mean scores for inclusion of essential information were 8.09 (melanocytic lesions), 8.15 (breast cancers), and 9.55 (lymphomas). Critical error-free rates were 62.5%, 77.5%, and 95%, respectively. Exclusion of nonessential information scored consistently high across subspecialties, and readability was rated 10/10 in 119 of 120 cases. CONCLUSIONS.—: ChatGPT-4o, when used with a standardized prompt and expert oversight, shows promising ability to generate concise and readable summaries of pathology reports. While overall performance was strong, occasional errors and limitations in handling complex or multipart cases were noted. Further refinement and domain-specific model training may enhance the reliability and clinical utility of artificial intelligence-assisted reporting.

来源:Arch Pathol Lab Med,查看原文。版权归原作者所有。

引用

  1. Zabarsky Shasha G, Balint-Lahat N, Schiby G, et al. Pathology Made Simple: ChatGPT's Summarization of Pathology Reports. Arch Pathol Lab Med. 2025. PMID: 40956700. doi:10.5858/arpa.2025-0072-OA

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。