Purpose
Developing effective communication skills in challenging patient interactions is essential for medical students. Simulation labs provide excellent training but lack universal accessibility and scalability. AI-based patient encounter simulations offer an innovative alternative, allowing students to practice in a risk-free environment with formative feedback. We present an integrated instructional module combining educational videos with an interactive AI simulation app to train medical students in managing difficult patient conversations.

Methods
The difficult conversations module includes a didactic component on best practices for delivering bad news using the SPIKES (setting up, perception, invitation, knowledge, emotions, strategy/summary) model, followed by interactive training with an AI app that simulates challenging patient conversations.

The didactic module is a free online video summarizing the SPIKES model that students can apply to simulated clinical encounters. The Difficult Conversations Patient Actor (DCPA) app is an openly available web-based app, developed in Python and powered by a large language model (LLM). DCPA hosts multiple cases, each of which simulates a distinct clinical scenario. Cases include diverse patient personalities based on the Big Five personality traits. The use of both LLM and a multidimensional prompt architecture enables the AI to not only accurately portray various scenarios, but also generate specific patient responses and emotional cues. After each interaction, a second AI agent within DCPA provides real-time formative feedback based on a rubric aligned with SPIKES model criteria.

DCPA cases were generated using a separate LLM-powered Case Creation (CC) app. To design the CC prompts, a pilot case was iterated within DCPA to simulate realistic conversation. Multiple cases were subsequently designed by CC using standardized prompts. CC requires limited input on desired clinical scenario and patient personality to generate a case that can be uploaded to DCPA, enabling educators at other institutions to quickly develop cases that can be tailored to the specific needs of their trainees.

Results
The AI patient actors accurately embodied assigned personalities and emotional responses in various clinical scenarios. In addition to appropriate verbal responses, the app conveyed body language and textual emotive cues, such as ""fidgets nervously"" or ""takes a deep breath, trying to calm down.” The app successfully assessed student performance using SPIKES criteria and provided constructive feedback. The DCPA app is openly available at https://patient-actor-dc.streamlit.app/.

Conclusion
This open-source educational module, integrating instructional videos with an AI simulation app, offers an innovative approach to practicing difficult conversations in medicine. The module will be incorporated into the curriculum at Dartmouth's Geisel School of Medicine, with validation through cohort testing to follow.