Skip to main navigation Skip to search Skip to main content

An AI-powered patient simulator (SimPatient) for communication skills education – safety-informed design and randomised controlled trial

Research output: Contribution to conference › Abstract › peer-review

Abstract

Background
Universities are under unprecedented pressure to train more doctors, increasing the demand for clinical placement capacity. Opportunities for communication skills practice are becoming increasingly scarce. While Generative AI offers a scalable solution to this deficit, 'off-the-shelf' models present significant safety risks regarding demographic bias, factual instability, and limited fidelity.

Summary of Work
We developed SimPatient, a multimodal generative AI tool featuring voice-interactive video avatars, designed to scale communication skills education while addressing AI safety challenges. To mitigate demographic bias, we engineered a ‘Diversity Engine’ that aligns avatar characteristics with census data to ensure representative populations. Clinical accuracy of the underlying models was established by validating the ability of Large Language Models to interpret and generate medical information within educational constraints. Fidelity was enhanced by fine-tuning the model on hundreds of hours of recorded consultations, capturing authentic conversational rhythms.

We evaluated SimPatient in a year-long randomised controlled trial (n=50) comparing it against our standard communication-skills curriculum between an AI group (n=22) and the control group (n=28). Outcomes were analysed and interpreted using a series of analyses (e.g., difference-in-differences, Latent change score modelling, hierarchical regression analysis and sentiment analysis of 6,120 conversation logs).

Summary of Results
The AI group demonstrated greater mean score improvement in OSCE communication stations (+13.84) compared to the control group (+6.11), although the interaction effect was not statistically significant (p=0.63). Modelling revealed that intrinsic motivation (p=0.03) and behavioural learning (p<0.01) were significantly associated with skill enhancement in both groups. Positive interactions between students and SimPatient were significantly associated with improvements in communication skills. Crucially, increased competence in the AI cohort was associated with decreased confidence (p=0.04), indicating a correction of over-confidence and enhanced critical self-appraisal.

Discussion and Conclusion
AI simulation can safely augment educational capacity when grounded in empirical safety research. By mitigating the demographic biases and hallucinations inherent in generic models, SimPatient provides equitable, high-fidelity practice. The inverse relationship observed between confidence and competence suggests the tool may combat the 'Dunning-Kruger' effect, fostering 'conscious incompetence', which is often missed in traditional training.

Take Home Message
Safe, scalable AI simulation requires rigorous design to mitigate bias; when properly implemented, it enhances motivation and aligns learner confidence with clinical competence.
Original languageEnglish
Publication statusPublished - Aug 2026
EventAMEE - Vienna, Austria
Duration: 22 Aug 2026 → 26 Aug 2026
https://amee.org/events/amee-2026/

Conference

ConferenceAMEE
Abbreviated titleAMEE
Country/TerritoryAustria
CityVienna
Period22/08/26 → 26/08/26
Internet address

Fingerprint

Dive into the research topics of 'An AI-powered patient simulator (SimPatient) for communication skills education – safety-informed design and randomised controlled trial'. Together they form a unique fingerprint.

Cite this