MDTeamGPT: A Self-Evolving LLM-based Multi-Agent Framework for Multi-Disciplinary Team Medical Consultation (2503.13856v1)

Abstract: LLMs have made significant progress in various fields. However, challenges remain in Multi-Disciplinary Team (MDT) medical consultations. Current research enhances reasoning through role assignment, task decomposition, and accumulation of medical experience. Multi-role collaboration in MDT consultations often results in excessively long dialogue histories. This increases the model's cognitive burden and degrades both efficiency and accuracy. Some methods only store treatment histories. They do not extract effective experience or reflect on errors. This limits knowledge generalization and system evolution. We propose a multi-agent MDT medical consultation framework based on LLMs to address these issues. Our framework uses consensus aggregation and a residual discussion structure for multi-round consultations. It also employs a Correct Answer Knowledge Base (CorrectKB) and a Chain-of-Thought Knowledge Base (ChainKB) to accumulate consultation experience. These mechanisms enable the framework to evolve and continually improve diagnosis rationality and accuracy. Experimental results on the MedQA and PubMedQA datasets demonstrate that our framework achieves accuracies of 90.1% and 83.9%, respectively, and that the constructed knowledge bases generalize effectively across test sets from both datasets.

Overview of MDTeamGPT: A Self-Evolving LLM-based Multi-Agent Framework for Multi-Disciplinary Team Medical Consultation

The paper "MDTeamGPT: A Self-Evolving LLM-based Multi-Agent Framework for Multi-Disciplinary Team Medical Consultation" introduces an advanced framework leveraging LLMs for facilitating Multi-Disciplinary Team (MDT) consultations in the medical domain. MDT consultations are pivotal for handling complex disease diagnostics and treatment strategies by integrating expertise across various medical disciplines. However, these consultations often suffer from inefficiencies due to prolonged dialogue histories and cognitive overload on the part of human practitioners. The proposed framework, MDTeamGPT, addresses these challenges by incorporating a structured multi-agent system enhanced with self-evolving capabilities.

Methodological Framework

MDTeamGPT employs a multi-agent system comprising various specialized physician agents, including roles such as General Internal Medicine Doctor, Neurologist, and Radiologist, among others. Each agent brings domain-specific insights to the consultation process. A notable innovation in the framework is its use of a lead physician role, tasked with consolidating dialogue outputs into categories—Consistency, Conflict, Independence, and Integration—which aids in reaching consensus efficiently.

To enable continuous learning and improvement, MDTeamGPT incorporates two knowledge bases: the Correct Answer Knowledge Base (CorrectKB) and the Chain-of-Thought Knowledge Base (ChainKB). These knowledge bases systematically store correct diagnostic outcomes and reflective insights from erroneous consultations, respectively. This structure allows the system to evolve and improve its reasoning accuracy over time.

Experimental Results

The framework was evaluated using the MedQA and PubMedQA datasets, achieving accuracies of 90.1% and 83.9%, respectively. These results demonstrate a strong capability of MDTeamGPT in medical consultations. Moreover, the effectiveness of the knowledge bases in generalizing across datasets underscores the framework's potential robustness in varying contexts.

Theoretical and Practical Implications

The proposed architecture not only advances the technical application of LLMs in medical diagnostics but also highlights important theoretical insights into multi-agent collaboration and learning. By efficiently managing cognitive loads and reducing information redundancy, MDTeamGPT stands as a promising tool for enhancing decision support systems in healthcare. Practically, it holds the potential to streamline MDT processes, leading to more accurate and timely patient care outcomes.

Future Directions

While the framework shows promising results, future work could delve into more complex integration of external medical databases and real-time clinical data to further enhance decision-making precision. Additionally, expanding the framework’s applicability to include a broader range of medical conditions and scenarios could prove beneficial.

In conclusion, MDTeamGPT presents a well-structured, multi-agent approach effectively leveraging LLMs for MDT consultations. Its ability to evolve with accumulated experiences is a significant step towards more intelligent and autonomous decision support systems in the medical field.