Harnessing Generative Artificial Intelligence and a Persona Prompt to Develop Competence in Addressing Social Determinants of Health

A generative AI persona prompt served as the virtual patient, giving students a structured setting to practice SDOH screening and response when clinical training opportunities are inconsistent. Compared with standard simulated patients or conventional simulation, a 12-sentence prompt defines the roles (patient and FNP), the setting, and patient information, and specifies that at least three SDOH significantly affect the patient's health; students worked in groups of three or four, shared a screen, entered the prompt into Claude Sonnet, and used point-of-care tools such as Neighborhood Navigator and GoodRX. A descriptive educational-design report with 98 participating students; the prompt was revised iteratively until the LLM yielded “consistent and effective” output; the activity was categorized as quality improvement after IRB review.

Students received the activity very positively and reported high comfort after it. It translates acceptability and self-rated comfort into concrete figures for this teaching activity, which similar courses can compare against. Anonymous, optional polling with about 75 responses per poll on average; the most common one-word reaction was “helpful” (n = 83), 95% said the activity facilitated their SDOH learning (n = 82), and 78% (n = 90) selected the two highest comfort levels, average 4.1/5.

SDOH knowledge quiz performance was strong overall, while recognizing specific strategies to address SDOH stood out as the relatively weaker point. It separates “knowing about SDOH” from “knowing how to act on SDOH,” offering a pointer for where teaching emphasis may go next. Quiz average of 94% across 98 students; just over half (54%) answered the item on specific strategies to address SDOH correctly; among 55 respondents, 47% cited learning about SDOH resources as their primary takeaway.

At the operational level, all SDOH domains appeared in student findings, and AI output varied in details across groups. It documents how the same prompt behaved when run in parallel groups and how faculty handled that variation. Among the 98 responses, economic stability, housing, neighborhood safety, and transportation were most frequently cited; faculty judged cross-group differences (for example, the virtual patient's zip code) not to affect the experience and to enrich the debriefing discussion instead.

The result is intended for family nurse practitioner students, delivered online via a video-conferencing platform in groups of three or four, and it depends on the university's custom generative AI platform, Claude Sonnet, and resources such as Neighborhood Navigator and GoodRX. The most directly transferable elements are design choices: building and iterating the 12-sentence persona prompt, a 30-minute prebriefing, about 75 minutes of scenario and action-plan work, a 45-minute debriefing framed by a modified PEARLS approach, and having groups share a screen to mitigate technology issues. The text also suggests that more debriefing time may be needed for complex topics, possibly equal to scenario time, and that other courses or topics would require re-iterating the prompt and deciding whether to use multiple cases.

A careful reader may still want to know several things: in the loaded markdown, some parenthetical citation content is truncated (for example, where the SDOH screening miss rate appears, and where literature on generative AI platform security and data leaks is cited), so those sources and their specific figures cannot be checked here; the rubric items and the scored results of the action plans, as well as the quiz item composition, are not detailed in the text; polling was voluntary and anonymous, so the views of non-respondents are absent; privacy and ethics are raised in the text without a described approach to handling them; and persistence of learning, such as performance in later real clinical practice, is not reported in the text.