Abstract:
Clinical reasoning is core in medical practice that is susceptible to cognitive errors compromising safety and reliability. Despite being promising in clinical activity, such Large Language Models as ChatGPT and Med-PaLM provide unverifiable content and tend to be in interpretable form, as well as do not possess the feature of collaborative decision-making that is inherent to a real clinical team. In this thesis, a Multi-Agent Collaboration Framework with Retrieval-Augmented Generation (RAG) is proposed to reinforce clinical reasoning. The two teams of four specialized agents, Case Analyzer, Evidence Validator, Treatment Planner, and Clinical Reporter work in parallel and have a central Collaborative Orchestrator Agent. The system creates consensus and minimizes reasoning differences by conducting repeated cycles of distributed analysis, cross-team comparison and refinement based on feedback. RAG makes all intermediate and final outputs to be based on verifiable clinical evidence. This piece of work is a blueprint of how to come up with a more transparent, reliable, and trustworthy Clinical Decision Support Systems.
