Research and implementation of medical dialogue mechanism based on LLM and multiagent collaboration

Abstract

To address the dual challenges of information overload and misdiagnosis risks in medical decision-making, this paper proposes a three-layer dynamic decision framework (independent layer, interaction layer, and collaboration layer), which is a structured method that integrates large language models (LLMs), multi-agent collaboration, and knowledge graphs for medical decision-making. The mechanism achieves evidence-based and traceable auxiliary diagnosis through the architecture: the independent decision layer (BERT-ESIM hybrid model parses user intent and medical environment information), the interactive decision layer (horizontal structure agent groups for parallel reasoning and conflict resolution), and the collaborative decision layer (integrating knowledge graphs and external real-time data). Tested on the CMB medical dataset, the system achieves an accuracy of 61.67% in medical exam tasks. Although this accuracy is not particularly high, it represents a significant improvement over single-model baselines, demonstrating that multi-agent reasoning enhances diagnostic reliability. The emphasis on traceable reasoning paths and visualized decision processes addresses the ”black box” problem in AI-assisted diagnosis, enhancing clinical credibility. This design provides a reusable intelligent paradigm for precise decision support in complex medical scenarios.

Ähnliche Arbeiten