Challenges and solutions for the development of medical GPTs

To systematically summarize the major challenges in developing medical GPT systems and, with reference to recent international reviews, evaluation frameworks, ethical and regulatory guidance, as well as Prof Chunxue Bai' s BAIMGPT White Paper, to outline practical solutions for translating large language models into clinically usable systems. Recent high-impact systematic reviews, methodological studies, real-world workflow evaluations, and governance guidance were synthesized to examine the main issues in medical GPT development, including factual reliability, knowledge updating, data governance, multimodal integration, workflow adaptation, explainability, bias, fairness, and accountability. Current evidence indicates that the bottlenecks of medical GPT go well beyond imperfect accuracy. Major challenges include hallucinations and factual inconsistency, limited ability to absorb newly updated medical knowledge, heterogeneous clinical data and unstable labels, insufficient support for multimodal decision-making, weak adaptation to real-world workflows, incomplete explainability and accountability, and concerns regarding bias, fairness, and ethics. Current LLMs remain sensitive to information order and quantity and are not ready for autonomous clinical decision-making. The mission of medical GPT development is not simply to improve language generation, but to transform large models into trustworthy medical intelligence systems with reliable knowledge, workflow compatibility, traceability, and governance readiness. At present, medical GPT should be positioned as a tool for cognitive augmentation and workflow support rather than a substitute for clinical judgment.

Key Words: BAIMGPT/medical GPT; large language model; clinical decision support; retrieval-augmented generation; data governance; human-AI collaboration; disease-specific agent; BAIMGPT