【摘要】 目的 评价人工智能(AI)大语言模型(LLM)生成的与神经眼科相关典型临床问题的答案，并利用客观评价及专家评估的方式多维度探究神经眼科相关问题在LLM上的表现。方法 多中心、随机、横断面试验研究。从神经眼科疾病定义、病因、临床表现及体征检查和治疗及预后4个角度选取30个神经眼科领域相关典型问题，分别使用Deepseek、文心一言4.0、豆包及Kimi 1.5四种国内开源LLM输出答案文本，采取客观评估法定量分析；同时采取专家评估法，由三位眼科专家分别对120个答案文本进行量化评分。根据问题回答的完整性、准确性和专业性及相关性和实用性分别制定3级、5级及4级李克特量表。选取其中表现最佳的LLM,观察在4类问题中该LLM是否存在表现差异，由另外三位专家评估各LLM是否可以替代真实世界医患沟通。结果 在客观的汉语文本阅读难度分析中，4种LLM的总字数组间差异存在统计学意义(均为P<0.001)。在4种LLM中，Kimi 1.5表现最出色，其完整性最高分(3分)、准确性和专业性最高分(5分)、相关性和实用性最高分(4分)的频率分别为61%、29%、41%。Kimi 1.5在神经眼科疾病定义、病因、临床表现及体征、治疗及预后四个方面的问题中表现较为一致，组间差异均无统计学意义(均为P>0.05)。结论 中文LLM在神经眼科临床应用中具有较大的潜力，Kimi 1.5在完整性、准确性和专业性及相关性和实用性方面表现较其他LLM出色，但仍无法代替真实世界医患沟通，未来需要探索AI+医师的新型诊疗模式。更多还原

【Abstract】 Objective To evaluate answers to typical clinical questions related to neuro-ophthalmology generated by Artificial Intelligence(AI) Large Language Models(LLM) and to explore the performance of neuro-ophthalmology-related questions on LLM in a multidimensional manner using objective and expert assessment. Methods Multicenter, randomized, cross-sectional pilot study. Thirty typical questions related to neuro-ophthalmology were selected based on four perspectives: definition, etiology, clinical manifestations and signs, and treatment and prognosis, and were analyzed quantitatively using Deepseek, Wenxin Yiyin 4.0, Doubao, and Kimi 1.5, which are four open-source LLMs in China, and quantitatively analyzed with objective assessment; and quantitatively rated by three ophthalmologists using expert assessment for 120 answer texts. Three ophthalmology experts quantitatively scored the 120 answer texts. Three ophthalmologists quantitatively scored the 120 answer texts. Level 3, 5, and 4 Likert scales were developed according to the completeness, accuracy, professionalism, relevance, and criticality of the question texts, respectively. The best-performing LLM was selected, and its performance was observed across the four types of questions. Additionally, three other experts assessed whether the best-performing one could be evaluated as a substitute for real-world doctor-patient communication.Results In the objective Chinese text reading difficulty analysis, the differences in total word count among the four LLMs were statistically significant(all P<0.001). Of the four LLMs, Kimi 1.5 performed the best, with frequencies of 61%, 29%, and 41% for the highest scores in completeness(3), accuracy and professionalism(5), and relevance and usefulness(4), respectively. Kimi 1.5 performed more consistently on the questions on the four areas of neuro-ophthalmologic disorders: definition, etiology, clinical manifestations and signs, treatment, and prognosis, with no between-group differences(P>0.05). Conclusion Chinese language LLMs have great potential in the clinical application of neuro-ophthalmology. Kimi 1.5 outperforms other LLMs in terms of completeness, accuracy, professionalism, relevance, and usefulness, but it still cannot replace real-world doctor-patient communication. There is a need to explore new diagnostic and therapeutic model of AI+physician in the future.更多还原