Abstract:
Virtual patients (VPs) have long been used to teach and assess clinical reasoning. VPs can be programmed to simulate authentic patient-clinician interactions and to reflect a variety of contextual permutations. However, their use has historically been limited by the high cost and logistical challenges of large-scale implementation. We describe a novel globally-accessible approach to develop low-cost VPs at scale using artificial intelligence (AI) large language models (LLMs). We leveraged OpenAI Generative Pretrained Transformer (GPT) to create and implement two interactive VPs, and created permutations that differed in contextual features. We used systematic prompt engineering to refine a prompt instructing ChatGPT to emulate the patient for a given case scenario, and then provide feedback on clinician performance. We implemented the prompts using GPT-3.5-turbo and GPT-4.0, and created a simple text-only interface using the OpenAI API. GPT-4.0 was far superior. We also conducted limited testing using another LLM (Anthropic Claude), with promising results. We provide the final prompt, case scenarios, and Python code. LLM-VPs represent a \'disruptive innovation\' - an innovation that is unmistakably inferior to existing products but substantially more accessible (due to low cost, global reach, or ease of implementation) and thereby able to reach a previously underserved market. LLM-VPs will lay the foundation for global democratization via low-cost-low-risk scalable development of educational and clinical simulations. These powerful tools could revolutionize the teaching, assessment, and research of management reasoning, shared decision-making, and AI evaluation (e.g. \'software as a medical device\' evaluations).
摘要:
虚拟患者(VP)长期以来一直用于教授和评估临床推理。VP可以被编程为模拟真实的患者-临床医生交互并反映各种上下文排列。然而,它们的使用历来受到大规模实施的高成本和后勤挑战的限制。我们描述了一种新颖的全球可访问方法,该方法使用人工智能(AI)大型语言模型(LLM)大规模开发低成本VPs。我们利用OpenAI生成预训练变压器(GPT)来创建和实现两个交互式VP,并创建了上下文特征不同的排列。我们使用系统的提示工程来改进提示,指导ChatGPT在给定的情况下模仿患者,然后提供关于临床医生表现的反馈。我们使用GPT-3.5-turbo和GPT-4.0实现了提示,并使用OpenAIAPI创建了一个简单的纯文本界面。GPT-4.0远远优于此。我们还使用另一个LLM(AnthropicClaude)进行了有限的测试,有希望的结果。我们提供最后的提示,案例场景,Python代码LLM-VPs代表了一种“破坏性创新”——一种明显逊色于现有产品但更容易获得的创新(由于低成本,全球范围,或易于实施),从而能够达到以前服务不足的市场。LLM-VPs将通过教育和临床模拟的低成本低风险可扩展开发为全球民主化奠定基础。这些强大的工具可以彻底改变教学,评估,以及管理推理的研究,共同决策,和人工智能评估(例如,作为医疗器械的软件评估)。