Online Poster Portal

Context
AI is reshaping medical education through the deployment of virtual simulated patients (VSPs) that help learners practice critical conversations. In oncology, effective communication is vital; physicians must deliver abnormal results in a digestible manner to optimize diagnosis and treatment. While standardized patient (SP) programs are essential, they are limited by competing clinical demands and the significant resources needed to train SP

Objective
The objective of this work is threefold: (1) to develop and validate a realistic, usable web-based interface and backend supporting voice-to-voice clinical simulation; (2) to provide a scalable platform for learners to practice complex conversations with VSPs and to receive LLM-generated feedback; and (3) to evaluate the feasibility, acceptability, and preliminary efficacy of these LLM-driven VSPs through pilot studies with medical students .

Methods
This research builds upon a completed initial pilot study that tested an AI voice-driven simulation for discussing diagnostic mammogram results. The current study utilizes an LLM-driven VSP prototype and a multi-agent framework. Learners engage with a web application via a simulated phone call to discuss abnormal results for a mammogram or lung CT with a VSP Agent. A secondary Evaluator Agent monitors the dialogue in real time to generate a scored assessment with immediate suggestions for improvement. To establish primary outcomes, a crossover study is being conducted with 20 medical students split into two groups. Half of the participants complete the mammogram scenario first, followed by the lung CT scenario approximately one week later, while the other half complete the scenarios in reverse order. A clinician reviews recordings of each simulation and scores performance using a standardized, OSCE-like checklist.

Expected Outcomes and Practical Lessons Learned
The platform is designed to produce high-fidelity verbal encounters and expert-validated performance data. We hypothesize that there will be information and skill transfer between the two simulations; specifically, we expect students to demonstrate improved communication performance and higher OSCE scores during their second simulation encounter, regardless of whether the scenario is a mammogram or lung CT follow-up. Key practical lessons learned include the importance of decoupling the Evaluator Agent from the simulated patient agents, which ensures the VSP remains in character while the Evaluator Agent provides timely, data-driven feedback. Additionally, implementing options such as "push-to-talk" allows users to balance the trade-off between latency and realism. To address risks and potential LLM bias, the team utilized clinician-validated personas and secure data handling, instructing learners not to discuss protected health information during sessions.