Enhancing communication and clinical reasoning in medical education: Building virtual patients with generative AI

Generative AI describes any form of artificial intelligence (AI) that can produce new text, images, video or other content. Large language models (LLMs) are a form of generative AI which can generate human-like text. They are trained on large amounts of text data and have been made popular by the launch of publicly available chatbots, including ChatGPT and Google Bard.1 Virtual patients simulate real-life clinical scenarios and can be used to teach knowledge, communication skills and clinical reasoning.2 Given the challenges learners can face in accessing opportunities to practice communication skills in person, particularly in busy healthcare environments, we aimed to develop a collection of realistic virtual patients using LLM technology. The virtual patients enable learners to practise their communication and clinical reasoning skills using a chatbot interface hosted on the Geeky Medics platform. A database of patient scripts was created, reflecting common patient presentations (e.g. chest pain, breathlessness). Each script consisted of a condensed summary of the patient's presentation, including relevant past medical and psychosocial history. Utilising OpenAI's GPT-3.5 and GPT-4 large language models, we developed an interactive chat interface that allows learners to engage with virtual patients generated from the patient scripts. Users can interact with the virtual patient via text or voice. This interaction mimics a real clinical encounter, facilitating natural conversation with the patient. After the consultation, learners can request an immediate AI-powered review of their transcript, which assesses the breadth and depth of their questioning to provide actionable feedback. Examples of feedback include commending areas covered well and highlighting important missed sections of the patient's history. Since their introduction on the Geeky Medics platform, learners have conducted over 45,000 consultations with virtual patients. Implementation challenges included user feedback about occasional 'hallucinations' and off-topic responses from the virtual patients. User feedback has generally been favourable, highlighting the independent aspect of the interactions and the opportunity for deliberate, repeated practice of complex communication skills, such as taking a sexual history. Our findings demonstrate the viability of using LLM technology, specifically GPT-3.5 and GPT-4, to create realistic virtual patients. Despite challenges like 'hallucinations' and off-topic responses, the high volume of consultations indicates strong engagement and potential for this tool to enhance communication and clinical reasoning skills. This approach offers a scalable, accessible way to supplement traditional clinical training, especially in environments with limited direct patient contact.