Understanding What Patients Really Need: A Multi-Intention Recognition and Planning Framework for Complex Medical Queries

ACL ARR 2025 February Submission8120 Authors

Abstract:

LLM-based multi-agent systems have shown promise in healthcare, enhancing diagnostic accuracy and efficiency. However, most existing systems rely on simplistic and naive doctor-patient dialogues, which fail to capture the complexity of real-world clinical interactions. In practice, patients’ self-descriptions are often verbose and contain hidden intents. Accurately extracting these needs and providing appropriate feedback is crucial for improving medical decision-making.
To address these challenges, we propose MIRPF, a Multi-Intention Recognition and Planning Framework designed to understand patients' complex intentions in healthcare settings. MIRPF first introduces an Intention Recognition module to extract and interpret precise medical intents from verbose queries. Next, a Dynamic Intent Orchestration Agent plans the execution sequence, taking into account the urgency and interdependencies of identified intents. Finally, based on this plan, a Multi-Agent Collaboration System, comprising intention-specific agents and a novel Chain of Thought (CoT)-based Hierarchical Progressive Decision-Making Agent, works collaboratively to complete the diagnostic process.
We evaluate MIRPF on two medical dialogue benchmark datasets. The results, measured using automated metrics and expert doctor evaluations, show that MIRPF outperforms existing methods, significantly improving medical proficiency and strategic reasoning.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: healthcare applications, clinical NLP

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 8120