■通过评估6种不同的人工智能(AI)模型的屈光手术建议(激光原位角膜磨镶术[LASIK]或光屈光性角膜切削术[PRK])来研究其当前决策能力。有瘢痕疙瘩形成史的理论患者。
■克劳德-2(人类,2023),GPT-4(OpenAI,2023),GPT-3.5(OpenAI,2022),双子座1.0(谷歌DeepMind,2023),MicrosoftCopilot(MicrosoftAI,2023),和Google-PaLM(GoogleAI,2022)进行了三个系统的查询,以确定理论患者的最合适的手术计划(LASIK或PRK),双眼屈光度(D)分别为-3.50,-5.00和-7.00屈光度(D),简单的眼部检查,和瘢痕疙瘩形成的历史。然后,他们的任务是提供已发表的科学参考资料来支持他们的回应。将AI模型的建议与一组6名经验丰富的眼科医生的建议进行了比较,作为一个基准。
■眼科医生小组一致推荐LASIK(6/6眼科医生),与AI模型(6/6模型)对PRK的一致初始建议相反。在人工智能模型提供的42个参考文献中,55%是虚构的,45%是真实的。当与有瘢痕疙瘩形成史但近视严重程度增加(-3.50至5.00至7.00D)的同一患者一起就诊时,6种模型中只有1种改变了对LASIK的最初建议。
很明显,当前的AI模型缺乏准确分析和评估临床情景中明显风险因素所需的批判性思维能力。如近视程度较高的PRK术后角膜雾霾的风险,特别是有瘢痕疙瘩形成史的病例。[JRefractSurg.2024;40(8):e533-e538。].
UNASSIGNED: To investigate the current decision-making capabilities of 6 different artificial intelligence (AI) models by assessing their refractive surgery recommendations (laser in-situ keratomileusis [LASIK] or photorefractive keratectomy [PRK]) for a theoretical patient with a history of keloid formation.
UNASSIGNED: Claude-2 (Anthropic, 2023), GPT-4 (OpenAI, 2023), GPT-3.5 (OpenAI, 2022), Gemini 1.0 (Google DeepMind, 2023), Microsoft Copilot (Microsoft AI, 2023), and Google-PaLM (Google AI, 2022) underwent three systematic queries to determine the most appropriate surgical plan (LASIK or PRK) for a theoretical patient with an increasing manifest refraction of -3.50, -5.00, and -7.00 diopters (D) in both eyes, an uncomplicated ocular examination, and history of keloid formation. They were then tasked with providing published scientific references to support their responses. The AI models\' recommendations were compared to those of a group of 6 experienced ophthalmologists, serving as a benchmark.
UNASSIGNED: The group of ophthalmologists unanimously recommended LASIK (6/6 ophthalmologists), in contrast to the unanimous initial recommendation for PRK from the AI models (6/6 models). Of the 42 references provided by the AI models, 55% were fictitious and 45% were authentic. Only 1 of the 6 models altered its initial recommendation to LASIK when presented with the same patient with a history of keloid formation but with increasing severity of myopia (-3.50 to 5.00 to 7.00 D).
UNASSIGNED: It is evident that current AI models lack the critical-thinking abilities required to accurately analyze and assess apparent risk factors in clinical scenarios, such as the risk of corneal haze after PRK at higher levels of myopia, particularly in cases with a history of keloid formation. [J Refract Surg. 2024;40(8):e533-e538.].