国内顶级三甲医院胸外科近期公布了一项关于医疗大模型MedGPT在复杂临床决策中辅助作用的对照实验结果。该研究于2025年12月获批,基于过去两年积累的1.6万多份门诊病例筛选出40例肺癌“灰区病例”,吸引了来自全国98家三甲医院的250名医生参与,其中主任医师和副主任医师占比超过63% 1。
数据显示,在有MedGPT辅助的条件下,参与医生的平均得分较无辅助条件提升了24.1%,具体得分从53.4分上升至66.27分,若仅提供检索参考则得分为61.05分 1。在750次方案比较中,MedGPT独立生成的诊疗方案获得660次更高评分,相较于医生无辅助方案胜率达88% 1。此外,MedGPT的综合得分为96.22,显著高于GPT-5.4 Direct(71.38)、Claude Opus 4.7 Direct(61.81)及Gemini 3.1 Pro Direct(54.62) 1。
这项由医联及未来医生创始人王仕锐推动的研究,旨在通过AI“蒸馏”专家经验来解决高质量医疗资源稀缺问题,其愿景被概括为“继绝学、开太平” 1。MedGPT的核心运作机制包含四个步骤:首先判断会诊需求,其次理解病例,接着生成标准(E-path)和前沿(F-path)治疗路径,最后通过检验闭环(V-path)完成方案输出 1。
A controlled study conducted at a top-tier Chinese tertiary hospital’s thoracic surgery department revealed that medical professionals assisted by the large language model MedGPT achieved an average performance score increase of 24.1% compared to those working without assistance 1. The experiment, approved in December 2025, focused on 40 "gray zone" lung cancer cases selected from over 16,000 outpatient records spanning the previous two years 1. These cases represent complex diagnostic and treatment scenarios where clear-cut answers are often difficult to determine 1.
The research involved 250 doctors from 98 different tertiary hospitals across China, with chief and associate chief physicians comprising more than 63% of the participants 1. The study evaluated physician performance under three distinct conditions: no assistance, access to retrieval references, and support from MedGPT 1. The resulting average scores were 53.4 for unassisted doctors, 61.05 for those using retrieval tools, and 66.27 for those utilizing the MedGPT-assisted workflow 1. Furthermore, when comparing standalone AI-generated plans against unassisted human plans in 750 head-to-head evaluations, MedGPT secured higher ratings in 660 instances, yielding an 88:12 win-loss ratio 1.
In terms of competitive benchmarking against other advanced models, MedGPT achieved a comprehensive score of 96.22, surpassing GPT-5.4 Direct (71.38), Claude Opus 4.7 Direct (61.81), and Gemini 3.1 Pro Direct (54.62) 1. The model operates through a structured four-step process designed to mimic expert clinical reasoning 1. It begins with a "Step Zero" assessment to determine if a consultation is necessary, followed by case understanding, the generation of both standard (E-path) and frontier (F-path) treatment strategies, and finally a validation loop (V-path) to ensure logical consistency 1. Wang Shirui, founder of Yilian and Future Doctor, emphasized that this approach aims to "distill" expert experience to address the scarcity of high-quality medical resources, articulating a vision of preserving established medical wisdom while exploring novel therapeutic pathways 1.
评论
还没有评论,欢迎留下第一条。