GPT-doctor: Customizing Large Language Models for Medical Consultation

The advent of Large Language Models (LLMs) has ushered in a new era for design science in Information Systems, demanding a paradigm shift in tailoring LLMs design for business contexts. This paper proposes a novel framework to customize LLMs for general business contexts that aims to achieve three fundamental objectives simultaneously: (1) aligning conversational patterns, (2) integrating in-depth domain knowledge, and (3) embodying the soft skills and core principles. We design methodologies to combine domain-specific theory with Supervised Fine Tuning (SFT) in LLMs. We instantiate our proposed framework in the context of medical consultation, creating a GPT-doctor model. Specifically, we construct a comprehensive dataset for SFT by collecting large volume of real doctors consultation records from a leading online medical consultation platform and medical knowledge from professional databases. Additionally, drawing on medical theory, we identify three soft skills and core principles of human doctors including professionalism, explainability, and emotional support, and design approaches to integrate these skills into LLMs. We demonstrate the feasibility and performance of our proposed framework using online experiments with real patients as well as evaluation by domain experts and real consumers. Results demonstrate that fine-tuned GPT-doctor performs on par with human doctors across multiple metrics including medical expertise and consumer preference. Finally, we unravel the black box and examine the sources of model performance improvement from the perspectives of horizontal conversation pattern alignment and vertical medical knowledge evolution. Our proposed framework offers step-by-step principles and guidance for customizing LLMs for real-world business problems.