A Synthetic Conversational Dataset for Type 2 Diabetes Management

Abstract

Access to real patient-doctor conversations in the medical domain is often restricted due to privacy concerns, making it difficult to build robust conversational AI systems. To address this, we present a novel methodology for generating a high-quality synthetic dataset designed for conversational triple extraction in Type 2 Diabetes management. Using structured prompting with GPT-4, we generated 16 demographically and medically diverse diabetic personas, and 256 multi-turn conversations between these personas and a caretaker agent, simulating realistic and context-rich interactions. The conversations incorporate critical properties such as personalization, empathy, contextual awareness, and medically grounded advice, as validated through both LLM-based and human expert evaluations. These synthetic conversations are further annotated with Subject-Predicate-Object (SPO) labels at the token level, integrating both manual and LLM-automated methods, forming the foundation for downstream tasks like triple extraction. Our work demonstrates the feasibility of using generative AI to simulate healthcare conversations at scale, offering a solution for data-scarce domains.

Resources

Details