D2MDT uses multi-agent deliberation for efficient clinical predictions

A new research paper from multiple institutions introduces D2MDT (Department-aware MultiDisciplinary Team Consultation with Deliberation), a multi-agent system designed to improve clinical prediction from electronic health records (EHRs). Traditional approaches rely either on correlation-driven deep models or single large language models (LLMs), both of which struggle with multidisciplinary clinical reasoning. D2MDT addresses this by simulating a real hospital consultation: it first builds structured EHR evidence and consultation-ready semantic evidence, then assigns patient-specific department perspectives—like cardiology or oncology—to different doctor agents. Each agent retrieves complementary evidence and collaborates through a structured deliberation process.

To avoid the inefficiency of full multi-round discussions, D2MDT introduces residual deliberation, which updates only the unresolved parts of the consensus report rather than replaying the entire conversation history. The final prediction fuses this refined consensus with structured EHR representations. In experiments on mortality prediction, D2MDT outperforms both existing deep learning models and single-LLM approaches, while also reducing consultation rounds. The authors have released code to ease reproducibility. This work advances multi-agent AI for critical healthcare tasks where nuanced, department-aware reasoning is essential.