夜雨聆风AI尚不具备独立临床诊疗的能力
AI is not yet capable of independent clinical diagnosis and treatment
尽管人工智能（AI）在医疗领域的应用日益增多，但其“像医生一样思考”的能力仍存在明显短板。由美国麻省总医院MESH孵化器团队开展的一项最新研究发现，生成式AI在临床推理关键环节仍显不足，尚不具备独立承担临床诊疗任务的能力。相关成果发表在最新一期《JAMA Network Open》上。
Despite the increasing application of artificial intelligence in the medical field, its ability to "think like a doctor" still has significant shortcomings. A recent study conducted by the MESH Incubator team at Massachusetts General Hospital in the U.S. found that generative AI still falls short in key aspects of clinical reasoning and is not yet capable of independently undertaking clinical diagnosis and treatment tasks. The findings were published in the latest issue of JAMA Network Open.
团队选取包括ChatGPT、DeepSeek、Claude、Gemini和Grok在内的21种大语言模型，在29个已发表的临床病例中进行测试，并通过逐步提供患者信息（从基本症状到实验室和影像结果）来模拟真实诊疗过程。结果显示，当获得完整信息时，所有模型在超过90%的案例中能给出正确的最终诊断。
The team selected 21 large language models, including ChatGPT, DeepSeek, Claude, Gemini, and Grok, and tested them on 29 published clinical cases. They simulated the real-world diagnostic process by providing patient information step by step, from basic symptoms to laboratory and imaging results. The results showed that when given complete information, all models could provide the correct final diagnosis in over 90% of the cases.
然而，在关键的早期诊断阶段，这些模型普遍表现欠佳。研究发现，超过80%的情况下，模型未能提出合理的“鉴别诊断”，即对多种可能疾病进行系统性分析与筛选。这一能力被认为是临床推理的核心，也是医生决策的重要基础。
However, during the crucial early diagnostic stage, these models generally performed poorly. The study found that in more than 80% of instances, the models failed to propose a reasonable "differential diagnosis"—that is, a systematic analysis and selection among multiple possible diseases. This ability is considered the core of clinical reasoning and a crucial foundation for physician decision-making.
为更全面评估模型能力，团队提出了一种名为PrIME-LLM的新指标，从提出潜在诊断、选择检查手段、给出最终诊断到制定治疗方案等多个环节对模型进行综合评价。结果显示，各模型整体评分在64%至78%之间，表现存在明显差异。
To evaluate the models' capabilities more comprehensively, the team proposed a new metric called PrIME-LLM, which assesses the models across multiple stages, from proposing potential diagnoses and selecting tests to delivering a final diagnosis and formulating a treatment plan. The results show that the models' overall scores ranged from 64% to 78%, indicating significant performance variation.
团队指出，大语言模型更擅长在信息完备的情况下“给出答案”，但在信息不充分、需要开放性推理的情境中表现较弱。随着实验室数据和影像资料的加入，模型表现有所提升，且新一代模型整体优于旧版本，表明相关技术正在持续改进。
The team pointed out that LLMs are better at "providing answers" when information is complete but perform weaker in scenarios requiring open-ended reasoning with insufficient information. The performance of LLMs improved with the addition of laboratory data and imaging results. Moreover, newer-generation models generally outperformed older versions, indicating that the underlying technology is continuously improving.
团队表示，当前大语言模型尚不适合在缺乏监督的情况下直接用于临床实践，其价值更在于辅助医生决策，而非取而代之。
The team stated that current LLMs are not yet suitable for direct use in clinical practice without supervision. Their value lies more in assisting physicians with decision-making rather than replacing them.
翻译要点说明：
1. 首先必须明确本文中的一个核心词“临床诊疗”，它的翻译是“clinical diagnosis and treatment”，其次还有一个易出错的地方就是文中出现的“鉴别诊断”，其翻译为“differential diagnosis”，另外就是机构名称的翻译，文章中只出现了一次，但大家也需要了解，麻省总医院的翻译为“Massachusetts General Hospital”。
2. 这篇文章中还出现了几个四字格的短语，英文翻译时要对其进行恰当的转化，例如：
“日益增多” → increasing（形容词）或 on the rise
“明显短板” → significant shortcomings或 noticeable deficiencies
“表现欠佳” → performed poorly或 showed suboptimal performance
“持续改进” → continuously improving（进行时态体现动态过程）
中文的四字格常表达程度、频率或状态，英文可用副词、形容词或特定动词短语对应，不必逐字翻译。
3. 这篇文章中多次出现了数字以及百分比，在这里了解一下搭配：
·百分比与介词搭配：“在超过90%的案例中” → in over 90% of the cases
注意：over=more than，但over更简洁。cases用复数。
·范围的表达：“在64%至78%之间” → ranged from 64% to 78%（标准表达）或：were between 64% and 78%
“超过80%的情况下”：译为：in more than 80% of instances（instances比cases范围更广）或：in over 80% of cases
材料来源：新华网
