生成式人工智能与医师基于入院病历的鉴别诊断与诊疗方案比较
结果显示西医GenAI亚组诊断遗漏评分[4.00(4.00,4.50) vs. 4.00(3.75,4.25)]和诊断适宜性评分[4.25(4.00,4.50) vs. 4.00(4.00,4.25)]均高于西医医师亚组(均P<0.05);中医GenAI亚组诊断遗漏评分[5.00(4.00,5.00) vs. 3.00(2.00,3.00)]和方案完整性评分[5.00(4.00,5.00) vs. 3.00(2.00,3.00)]均高于中医医师亚组(均P<0.05)。诊断冗余、诊断优先级及方案精确性在相应亚组间均无统计学显著差异。结论提示GenAI可作为专科医师临床决策的辅助工具。
要点
- 西医GenAI亚组诊断遗漏评分4.00(4.00,4.50)高于西医医师亚组4.00(3.75,4.25),P<0.05,提示GenAI可能减少西医诊断遗漏。
- 中医GenAI亚组诊断遗漏评分5.00(4.00,5.00)显著高于中医医师亚组3.00(2.00,3.00),P<0.05,提示GenAI在中医诊断遗漏方面可能优势明显。
- 西医GenAI亚组诊断适宜性评分4.25(4.00,4.50)高于西医医师亚组4.00(4.00,4.25),P<0.05,但中医亚组间无显著差异。
- 西医GenAI亚组方案完整性评分4.25(4.00,4.50)高于西医医师亚组4.00(3.75,4.25),中医GenAI亚组5.00(4.00,5.00)高于中医医师亚组3.00(2.00,3.00),均P<0.05;方案精确性均无显著差异。
- 诊断冗余和诊断优先级在西医及中医的GenAI与医师亚组间均无统计学显著差异。
适用人群与场景
- 适用场景为住院患者入院后、准备首次病程记录阶段,由专科医师完成标准化病史采集、体格检查及相关辅助检查并形成文本病历数据后。
诊断标准与关键阈值
- 鉴别诊断评价指标包括诊断冗余、诊断遗漏、诊断优先级和诊断适宜性;诊疗方案评价指标包括方案完整性和方案精确性。
- 由西医或中医专家分别对医师组和GenAI组生成的鉴别诊断及诊疗方案采用5分制评分。
- 原文未提及诊断标准名称、条件、评分界值或启动/停止阈值。
处置流程与药物剂量
- 第一步:由专科医师完成标准化病史采集、体格检查及相关辅助检查,并整理为文本病历数据。
- 第二步:医师组由西医或中医专科医师独立制定鉴别诊断与诊疗方案;GenAI组由研究者使用本地部署的定制DeepSeek模型服务生成鉴别诊断与诊疗建议。
- 第三步:由西医或中医专家分别对两组生成的鉴别诊断和诊疗方案以5分制评分,比较诊断冗余、诊断遗漏、诊断优先级、诊断适宜性、方案完整性和方案精确性。
- 原文未提及具体药物名称、剂量、给药途径、频次、疗程或目标值。
红旗征与不能漏的情况
- 原文未提及
研究设计与关键结果
- 研究类型为双路径对照比较研究,使用同一病历数据;纳入96份病历,来自单中心(天津大学海河医院呼吸与危重症医学科),含肺部感染28例和罕见病9例。
- 西医GenAI亚组诊断遗漏评分4.00(4.00,4.50)高于西医医师亚组4.00(3.75,4.25),P<0.05;中医GenAI亚组5.00(4.00,5.00)高于中医医师亚组3.00(2.00,3.00),P<0.05。
- 西医GenAI亚组诊断适宜性评分4.25(4.00,4.50)高于西医医师亚组4.00(4.00,4.25),P<0.05;中医GenAI与中医医师亚组间无统计学显著差异。
- 西医GenAI亚组方案完整性评分4.25(4.00,4.50)高于西医医师亚组4.00(3.75,4.25),中医GenAI亚组5.00(4.00,5.00)高于中医医师亚组3.00(2.00,3.00),均P<0.05。
- 诊断冗余、诊断优先级及方案精确性在相应GenAI与医师亚组间均无统计学显著差异。
证据类型、等级与局限
- 证据类型为单中心双路径对照比较研究,非随机对照试验。
- 局限包括:单中心、样本量96份病历、仅在呼吸与危重症医学科住院患者首次病程记录阶段进行、GenAI使用本地部署定制DeepSeek模型,结果外推需谨慎。
对床旁实践的改变
- 在住院患者首次病程记录阶段,GenAI可作为专科医师临床决策的辅助工具,可能有助于减少诊断遗漏(西医和中医)及提高西医诊断适宜性和中西医方案完整性。
- GenAI在诊断冗余、诊断优先级和方案精确性方面未显示优于医师,不能替代医师独立决策。
- 适用人群为呼吸与危重症医学科住院患者,尤其是肺部感染和罕见病患者;原文未提及不适用人群。
- 常见陷阱:GenAI生成的建议需由专科医师审核,原文未提供药物剂量或具体治疗方案细节,不能直接作为处方依据。
指南或共识出处
- 原文未提及
以上提炼依据论文摘要(全文未开放获取,或许可不允许翻译转载);全文请点「阅读原文」。
英文原文(摘要)· 点击展开
OBJECTIVE: To compare the quality of differential diagnoses and diagnostic and treatment plans formulated by generative artificial intelligence (GenAI) and by specialist physicians practicing western medicine or traditional Chinese medicine at the stage of preparing the initial progress note for hospitalized patients.
METHODS: A dual-path controlled comparative study was conducted using the same medical record data. Hospitalized patients admitted to the Department of Respiratory and Critical Care Medicine at Haihe Hospital, Tianjin University, from April to November 2025 were selected. After admission, standardized history taking, physical examination, and relevant ancillary examinations were performed by specialist physicians, and textual medical record data were compiled. According to the method used to formulate the diagnostic and treatment plans, two groups were established: a physician group, comprising a western medicine physician subgroup and a traditional Chinese medicine physician subgroup and a GenAI group, comprising a western medicine GenAI subgroup and a traditional Chinese medicine GenAI subgroup. In the physician group, diagnostic and treatment plans were independently formulated by specialist physicians. In the GenAI group, diagnostic and treatment recommendations were generated by investigators using a locally deployed, customized DeepSeek model service. Evaluation indices for differential diagnosis included diagnostic redundancy, diagnostic omission, diagnostic prioritization, and diagnostic appropriateness. Evaluation indices for diagnostic and treatment plans included plan completeness and plan precision. The differential diagnoses and diagnostic and treatment plans generated by the physician and GenAI groups were separately rated by experts in western medicine or traditional Chinese medicine using a 5-point scale.
RESULTS: A total of 96 medical records were included, including records from 28 patients with pulmonary infections and 9 patients with rare diseases. (1) Quality of differential diagnosis: For diagnostic omission, scores were higher in the western medicine GenAI subgroup than in the western medicine physician subgroup [score: 4.00 (4.00, 4.50) vs. 4.00 (3.75, 4.25), P<0.05], and scores were higher in the traditional Chinese medicine GenAI subgroup than in the traditional Chinese medicine physician subgroup [score: 5.00 (4.00, 5.00) vs. 3.00 (2.00, 3.00), P<0.05]. For diagnostic appropriateness, scores were higher in the western medicine GenAI subgroup than in the western medicine physician subgroup [score: 4.25 (4.00, 4.50) vs. 4.00 (4.00, 4.25), P<0.05], whereas no statistically significant difference was found between the traditional Chinese medicine GenAI and traditional Chinese medicine physician subgroups. For diagnostic redundancy and diagnostic prioritization, no statistically significant differences were found between the corresponding GenAI and physician subgroups in either western medicine or traditional Chinese medicine. (2) Quality of diagnostic and treatment plans: For plan completeness, scores were higher in the western medicine GenAI subgroup scored higher than in the western medicine physician subgroup [score: 4.25 (4.00, 4.50) vs. 4.00 (3.75, 4.25)], and were also higher in the traditional Chinese medicine GenAI subgroup than in the traditional Chinese medicine physician subgroup [score: 5.00 (4.00, 5.00) vs. 3.00 (2.00, 3.00); both P<0.05]. For plan precision, no statistically significant differences were found between the corresponding GenAI and physician subgroups in either western medicine or traditional Chinese medicine.
CONCLUSIONS: At the stage of preparing the initial progress note for hospitalized patients, GenAI may offer potential advantages in reducing diagnostic omissions and traditional Chinese medicine, as well as in improving diagnostic appropriateness in western medicine. GenAI may serve as an adjunctive tool to support clinical decision-making by specialist physicians.
参考来源
- Yang W, Fu H, Meng X, et al. [Comparison of generative artificial intelligence and physicians in differential diagnosis and treatment planning based on initial hospital medical records]. Zhonghua Wei Zhong Bing Ji Jiu Yi Xue. 2026. PMID: 42875245. doi:10.3760/cma.j.cn121430-20260313-00118
供执业医师抢救与诊疗参考,不替代临床判断;剂量请结合体重、肝肾功能和本院药品规格复核,以指南原文为准。