TLDR: This research introduces a hierarchical multi-agent AI system designed to transform passive medical pre-consultation into a proactive, structured inquiry process. It addresses challenges like short consultation times and limitations of existing AI by using eight specialized agents coordinated by a central controller. The system effectively manages tasks like triage and medical history collection, achieving high accuracy in department classification, superior task completion rates, and excellent clinical quality scores, as validated by physicians. It’s also model-agnostic and maintains data privacy.
Global healthcare systems are grappling with significant challenges, including increasing patient numbers and severely limited consultation times. In many countries, primary care visits average less than five minutes, a duration often insufficient for comprehensive patient care and accurate diagnosis. While pre-consultation processes, which involve patient triage and structured history-taking, offer a potential solution, existing AI systems in this area have been limited by passive interactions and difficulties in managing context over extended conversations.
A new study introduces a hierarchical multi-agent framework designed to transform these passive medical AI systems into proactive inquiry agents. This innovative approach aims to overcome the limitations of current pre-consultation technologies by autonomously orchestrating tasks and intelligently coordinating specialized agents.
The Multi-Agent Architecture
The researchers developed an eight-agent architecture with a centralized control mechanism for dynamic medical consultation. This framework breaks down the pre-consultation process into four main tasks: Triage (determining the appropriate department), History of Present Illness collection, Past History collection, and Chief Complaint generation. The first three tasks are further divided into 13 domain-specific subtasks to ensure thorough information gathering.
A central ‘Controller’ agent plays a crucial role, coordinating the specialized agents through dynamic completion assessment, adaptive prompt generation, and hierarchical task management. This allows the system to proactively guide the conversation, much like a human physician would, rather than simply reacting to patient input.
Key Innovations
The framework introduces three core innovations:
- Dynamic Subtask Completion Assessment: The system monitors information gathering progress across 13 predefined medical domains using medically-informed evaluation criteria.
- Adaptive Prompt Generation: It synthesizes patient responses with accumulated consultation context to formulate clinically optimal follow-up questions.
- Hierarchical Task Management: This balances the broader diagnostic progression with the detailed collection of symptom information through priority-based orchestration.
Evaluation and Results
The system was rigorously evaluated using 1,372 validated electronic health records from a Chinese medical platform. Its performance was assessed across various foundation models, including GPT-OSS 20B, Qwen3-8B, and Phi4-14B.
The results were highly promising:
- The framework achieved 87.0% accuracy for primary department triage and 80.5% for secondary department classification.
- Task completion rates reached an impressive 98.2% when using agent-driven scheduling, significantly outperforming sequential processing, which achieved 93.1%.
- Clinical quality scores, as assessed by 18 physicians on a 5-point scale, averaged 4.56 for Chief Complaints, 4.48 for History of Present Illness, and 4.69 for Past History.
The system demonstrated robust performance across different language models without requiring task-specific fine-tuning, highlighting its model-agnostic nature. Consultations were completed efficiently, averaging 12.7 rounds for History of Present Illness and 16.9 rounds for Past History collection.
Also Read:
- AI Learns to Ask: Bridging the Human-AI Intent Gap
- VERIMOA: Enhancing Automated Hardware Description Language Generation with a Smart Agent Framework
Impact and Implications
This hierarchical multi-agent framework successfully enables proactive, structured medical inquiry through intelligent coordination, marking a significant improvement over conventional passive approaches. Its model-agnostic architecture maintains high performance across different foundation models while preserving data privacy through local deployment.
These findings suggest that autonomous AI systems have the potential to greatly enhance pre-consultation efficiency and quality in clinical settings, reducing physician workload and ensuring more thorough patient assessments. For more details, you can read the full research paper here.
