SOPHIE trial finds AI patient practice improves serious-illness communication scores

A newly published randomized study tested a virtual patient and feedback against reading the same communication framework.

A randomized study published in JMIR Medical Education on 20 October reports that practising with SOPHIE, an AI-powered virtual patient, improved serious-illness communication scores compared with reading material. The research involved 51 clinicians and trainees, with assessments before and after a 30-minute intervention.

SOPHIE presents a patient conversation and then gives feedback organised around three skills: empathising, speaking explicitly and empowering the patient. The comparison group spent the same amount of time reading about the same framework. Blinded standardized-patient actors and independent raters assessed performance.

The study, conducted in 2024, found greater improvement in all three skills and the overall communication score for the SOPHIE group. The October publication is the new development; it does not mean the experiment took place this week.

The learning loop is the product

The design combines conversation with a chance to understand and change behaviour. A learner first has to respond to a patient rather than select an ideal answer from a list. Feedback can then connect an utterance to a specific communication goal, making the next attempt more deliberate.

That distinction matters in communication teaching. Knowing the name of a technique is different from finding useful words during an emotionally difficult exchange. A simulation gives the learner a place to practise pacing, directness and acknowledgement, while an instructor can discuss how those choices affect the conversation.

A useful training session could therefore follow the same broad sequence as a coaching exercise: establish one goal, attempt a conversation, examine a specific moment and try again. The technology becomes valuable when it makes that cycle easier to repeat, rather than simply extending the duration of a chat.

A hybrid system behind the virtual patient

The paper describes a combination of structured dialogue and a language model. The structured component maintains the training scenario and communication goals; the model helps produce more natural responses. Speech recognition, speech synthesis and an animated avatar complete the interaction.

For educators, this creates several independently adjustable parts. The case defines the situation, the dialogue design defines the available responses, and the feedback rubric defines what the learner is practising. A curriculum can make those choices explicit before deciding which model or avatar to use.

It also creates a useful distinction between a teaching tool’s appearance and its instructional behaviour. A visually convincing patient is only one part of the experience. The conversation must produce the situations needed for practice, and the feedback must help the learner act differently in a later encounter.

From one session to curriculum design

The trial supplies a measured comparison for a specific educational intervention. It did not compare SOPHIE with every form of standardized-patient teaching, and its assessed communication outcomes should not be described as demonstrated improvements in patient health.

Its practical contribution is a clearer testable unit: one learner, a defined conversation framework, a short practice period and independently scored performance. Medical schools evaluating similar tools can build on that unit and examine how practice fits alongside faculty teaching, human simulation and clinical experience.

The paper gives educators a concrete starting point for planning those comparisons. It turns the question from whether an AI patient can hold a conversation to whether a defined practice-and-feedback session helps someone communicate more effectively.

Sources

- [Effects of a Generative AI Virtual Patient on Serious Illness Communication Skills: Randomized Controlled Trial](https://mededu.jmir.org/2026/1/e93034) — JMIR Medical Education · 2026-10-20