Fig. 1: Dual-track role-play codesign process for developing, deploying and evaluating P&P Care.

a, Contextual understanding: Community stakeholders and researchers collaboratively mapped infrastructural, cultural and literacy-related barriers to existing AI health consultation tools in resource-limited settings. b, Cocreation: Prototype development of the GPT-4-powered P&P Care (OpenAI; GPT-4o mini) chatbot prioritized low-literacy interfaces, integrated with e-learning modules to address AI health literacy deficits. c, Testing and refinement: The codesign team iteratively tested and refined the prototype models via codesigned evaluation metrics and questionnaires, and the virtual patient experiment validated the chatbot performance. The radar plot illustrates the mean scores for health awareness and communication quality metrics (attention, listenability, conciseness, integrity and empathy) between e-learning plus and consultation-only groups from the virtual patient experiments, with error bars representing the s.d. d, Implementation and evolution: Two on-site pilot studies (urban/rural) refined real-time speech-to-text integration and enhanced chatbot functionality in suboptimal network connectivity. The codesign process shaped the pragmatic RCT’s design, including context-aligned trial arms, patient-centred evaluation rubrics and culturally appropriate recruitment. Base map from Ministry of Civil Affairs of the People’s Republic of China (map approval no. GS (2022)1873). The original base map data are unaltered; colours have been modified for illustrative purposes. This map is schematic only.