MedAction: 能動的マルチターン臨床診断LLMの実現に向けて MedAction: Towards Active Multi-turn Clinical Diagnostic LLMs

MedAction is a research proposal aimed at building clinical large language models that go beyond passive question answering and actively drive multi-turn diagnostic conversations, including patient interviewing, test ordering, and differential reasoning.

Most existing medical LLMs are designed to respond to symptom descriptions provided up front by the user, producing a single-shot answer. Real clinical practice, however, looks very different: physicians iteratively gather history, decide which labs or imaging to order, and refine a differential diagnosis as new evidence arrives. MedAction reportedly tackles this gap by structuring each dialogue turn around the action the model should take, whether that is asking a clarifying question, requesting a specific test, or committing to a diagnosis. The approach appears to combine action-conditioned training data with evaluation protocols that better mirror end-to-end clinical workflows.

The work sits in a rapidly evolving landscape. Google's Med-PaLM line and, more recently, AMIE demonstrated that conversational agents can match or exceed primary care physicians on text-based diagnostic interviews under controlled study conditions. Microsoft has pursued similar directions with medical copilots, while the open-source community has produced models such as Meditron and various Llama-derived clinical variants. MedAction can be read as part of this broader push toward agentic clinical AI, with particular emphasis on active information seeking rather than reactive answering.

Several open challenges remain. Active test ordering implies trade-offs between diagnostic yield, cost, and patient burden, which raises non-trivial questions about reward design and evaluation metrics. Hallucination, rare disease coverage, guideline adherence, and safety guardrails are also persistent concerns that any deployable system would need to address. It is also worth noting that strong performance on simulated patient benchmarks does not always translate to real clinical environments, where data is messier and stakes are higher.

Without direct access to full experimental details, the specific gains MedAction delivers over prior multi-turn medical dialogue systems are difficult to characterize precisely. Still, the framing is timely: as foundation models begin to be evaluated as clinical agents rather than as static knowledge bases, frameworks that explicitly model action selection and longitudinal reasoning are likely to become increasingly important. MedAction appears to be a step in that direction, and its impact will depend on how robustly its training signal and evaluation generalize beyond curated case vignettes.