探索国内外主流大语言模型（LLMs）在肺癌辅助诊疗中的应用现状和前景。

来自复旦大学附属中山医院的肺癌诊疗多学科团队，结合国内外指南和长期临床实践经验，设计出40个涵盖肺癌基本概念、肺癌筛查、肺癌诊断、肺癌治疗和肺癌病理5个模块的肺癌诊疗相关问题，提问国内外主流LLMs，包括DeepSeek-V3、DeepSeek-R1、豆包、Kimi和GPT-4o，并收集模型的输出结果。随后由两名经验丰富的胸外科医生依据5分类法对回答的准确性和情感支持度进行评分，对比不同模型间的表现差异。

GPT-4o、DeepSeek-V3和DeepSeek-R1表现相似，评分中位数[四分位距（IQR）]为5.00（4.50~5.00），显著优于Kimi[4.25（3.50~4.50）]和豆包[4.50（3.88~4.50）]。亚组分析显示，DeepSeek-R1在基本概念、诊断、治疗和病理多个模块表现出色。DeepSeek-V3整体表现优异，尤其擅长诊断模块。GPT-4o则更擅长筛查模块。情感支持度评估显示，LLMs整体表现显著低于准确性维度，得分中位数集中在3.00附近。其中DeepSeek-R1生成的回答最能让患者感到支持，评分中位数（IQR）为3.50（3.00~4.50）。GPT-4o[2.50（2.50~3.12）]、DeepSeek-V3[3.25（2.50~3.50）]和豆包[3.00（2.50~3.50）]表现相似，优于Kimi[2.50（2.50~3.00）]。亚组分析则显示LLMs在各个模块评分整体偏低，低分占比较高，情感支持不足较为明显。

LLMs在肺癌诊疗领域展现出初步的应用潜力，但在处理复杂临床场景和患者沟通等方面仍存在不足。未来，伴随LLMs不断发展完善，可以预见其在肺癌诊疗领域的广阔应用前景。

To explore the current status and potential of mainstream large language models (LLMs) in lung cancer auxiliary diagnosis and treatment.

A multidisciplinary team from Zhongshan Hospital Affiliated to Fudan University designed 40 questions based on domestic and international guidelines and long-term clinical experience. The questions covered five modules of lung cancer diagnosis and treatment: basic concepts, lung cancer screening, diagnosis, treatment, and pathology. The questions were posed to five mainstream LLMs: DeepSeek-V3, DeepSeek-R1, Doubao, Kimi, and GPT-4o. The models’ outputs were evaluated by two experienced thoracic surgeons using a five-point Likert scale to assess accuracy and emotional support.

GPT-4o, DeepSeek-V3, and DeepSeek-R1 performed similarly, with a median [interquartile range (IQR) ] of 5.00 (4.50–5.00), significantly outperforming Kimi [4.25 (3.50–4.50) ] and Doubao [4.50 (3.88–4.50) ]. Subgroup analysis showed that DeepSeek-R1 excelled in basic concepts, diagnosis, treatment, and pathology modules. DeepSeek-V3 performed excellently overall, particularly in the diagnosis module. GPT-4o was best suited for the screening module. The emotional support assessment revealed that all LLMs scored notably lower in this dimension, around 3.00, compared to their accuracy scores. Among the models, DeepSeek-R1 provided the highest level of emotional support, with a median (IQR) of 3.50 (3.00–4.50). GPT-4o [2.50 (2.50–3.12) ], DeepSeek-V3 [3.25 (2.50–3.50) ], and Doubao [3.00 (2.50–3.50) ] demonstrated comparable performance, while Kimi showed the lowest scores [2.50 (2.50–3.00) ]. Subgroup analysis further indicated that emotional support ratings were consistently lower across all modules, highlighting a critical limitation of current LLMs in patient-centered communication.

LLMs show initial application potential in lung cancer diagnosis and treatment, but shortcomings remain in handling complex clinical scenarios and patient communication. With ongoing development and improvement, LLMs are expected to have broad application prospects in the field of lung cancer diagnosis and treatment. To the best of our knowledge, our study is the first systematic evaluation of domestic LLMs in the context of lung cancer care in China.