Shaping AI in Pelvic Floor Physiotherapy: The Impact of Role-Play Prompting on ChatGPT Response Quality.

INTRODUCTION AND HYPOTHESIS: Pelvic floor physiotherapy (PFP) is a highly specialized field requiring complex and multidisiplinary clinical reasoning. While large language models (LLMs) are increasingly being utilized to support clinical education and decision-making in this domain, the relevance, accuracy, and comprehensiveness of their outputs vary significantly depending on prompt engineering. The aim of this study was to examine the impact of persona-based role-play prompting on the quality of ChatGPT responses to clinical questions in PFP.

METHODS

Twenty-one open-ended clinical questions covering assessment and management of common pelvic floor dysfunctions were presented to ChatGPT (GPT-4.1) under two conditions: (1) a neutral prompt and (2) a role-play prompt instructing the model to respond as an experienced pelvic floor physiotherapist. Two independent physiotherapists rated all responses across four domains-relevance, accuracy, comprehensiveness, and clarity-using a five-point rubric.

RESULTS

Persona-based prompting significantly improved response quality across all domains (p < 0.001). The largest enhancement was observed in comprehensiveness (mean difference; 1.45), followed by relevance, accuracy, and clarity (mean difference; 1.12, 0.88, and 1.48, respectively). Effect sizes were large to very large (Cohen's d; 1.66-2.11). Inter-rater reliability ranged from moderate to excellent (ICC; 0.61-0.81).

CONCLUSIONS

Persona-based role-play prompting markedly enhances the quality of LLM-generated responses in PFP. For clinicians, educators, and students, adopting structured prompts will substantially improve output quality; however, because accuracy remains imperfect, all generated responses still require careful professional oversight.