French Virtual Patients: LLM-Based System for Clinical Training With 240 OSCE Dialogues
A new system uses AI to generate realistic doctor-patient interactions with automatic feedback.
A team of French researchers from Avignon Université and Sorbonne Université has introduced a new controllable virtual patient system for medical training, built around a dataset of 240 real Objective Structured Clinical Examination (OSCE) dialogues. The dataset captures student-patient interactions in French, addressing the shortage of human standardized patients. The researchers then developed an LLM-based pipeline that generates synthetic dialogues with modular control components: retrieval-based grounding ensures the virtual patient stays on topic, while a reflection loop maintains character consistency and realism over multi-turn conversations.
The system includes a multi-level evaluation framework that uses an LLM-as-a-Judge approach to assess three dimensions: patient simulation quality, student performance, and linguistic quality. Experiments show that the controllability modules significantly improve patient fidelity and student evaluation consistency. The team also implemented an interactive prototype where medical students can practice with the virtual patient and receive automatic, structured feedback, making it a practical tool for scaling clinical communication training. The work was accepted at SIGDIAL 2026.
- 240 real French OSCE dialogues form the training dataset
- LLM pipeline uses retrieval-based grounding and a reflection loop to ensure patient fidelity
- Interactive prototype gives students automatic feedback on performance
Why It Matters
Scalable virtual patients could democratize clinical communication training and reduce reliance on human actors.