TL;DR
This paper introduces KACR, a novel framework that enhances clinical decision-making by integrating structured knowledge graphs (KGs) into a multi-agent collaboration system. KACR uses reinforcement learning to guide discipline selection and a confidence-aware mechanism for inter-agent discussion, leading to more accurate and robust diagnoses.
What Problem Does This Paper Address?
Large Language Models (LLMs) show promise in clinical decision support, but current multi-agent approaches often rely on prompt engineering for discipline selection and lack robust knowledge integration. This can lead to superficial reasoning, especially in complex cases. Existing methods struggle to precisely guide agents to relevant clinical knowledge and effectively coordinate multidisciplinary expertise, hindering accurate and evidence-based diagnostic conclusions.
Methods
KACR employs a two-stage approach:
- Knowledge-Aware Co-Reasoning (KACR) Framework:
- Reinforcement Learning (RL)-Optimized Discipline Agent: An Actor-Critic architecture trained with Proximal Policy Optimization (PPO) dynamically identifies relevant nodes (symptoms, diseases, disciplines) on a clinical knowledge graph (CKG). The Actor navigates the CKG using a predefined meta-path, while the Critic evaluates the reasoning path. This module uses a frozen LLM backbone augmented with LoRA for efficient training.
- Multidisciplinary Collaboration Module: This module coordinates multiple rounds of discussion among specialized agents. It incorporates:
- Uncertainty Quantification: Dynamically weights agent opinions based on their prediction confidence.
- KG-Enhanced Confidence Reflection: Agents iteratively refine their positions through evidence-based persuasion, with persuasion strength guided by the Critic's value estimates. This ensures consensus aligns with the clinical knowledge structure.
Datasets
The paper evaluates KACR on several datasets:
- Primary Benchmarks: MedQA, MedMCQA, PubMedQA, and MMLU (medical topics).
- Additional Datasets for Generalizability: DDXPlus (EHR data), SymCat (synthetic symptom-condition data), JAMA (clinical cases), and Medbullets (USMLE-style questions).
- CKG Source: Unified Medical Language System (UMLS).
A re-labeled sub-dataset was created for training the discipline reasoning module, with annotations from three medical professionals.
Key Insights and Ideas
- Bridging LLMs and KGs is Crucial: Integrating structured knowledge from KGs significantly enhances LLM performance in complex reasoning tasks like clinical diagnosis, moving beyond simple prompt engineering.
- RL for Dynamic Discipline Selection: Using RL to train an agent to navigate a CKG for discipline identification is more effective than static or prompt-based methods, especially for ambiguous queries.
- Confidence-Aware Collaboration: A multidisciplinary collaboration that explicitly quantifies and leverages agent confidence leads to more robust and reliable consensus, mimicking real-world expert discussions.
- Dual-Phase Reasoning Mimics Clinical Workflow: The KACR framework's two-phase approach (patient guidance and multidisciplinary consultation) effectively mirrors established clinical decision-making processes.
Key Results
- KACR outperforms state-of-the-art clinical LLMs and multi-agent frameworks across eight benchmark datasets, achieving the best performance.
- In online A/B testing, KACR improved accuracy by 27% and user satisfaction by 25% compared to traditional rule-based systems. It also reduced the need for human customer service intervention by 61% and increased conversion rates by 58%.
- A case study demonstrated KACR's ability to correctly identify a complex diagnosis (Meningioma) by integrating insights from Neurology, Oncology, and Genetics, highlighting its holistic assessment capabilities.
Limitations and Future Work
- Prompt Engineering Optimization: While hand-crafted prompts were used, future work could involve developing a systematic framework for automated prompt optimization using techniques like neural architecture search.
- KG Construction vs. Integration: The current work focuses on integrating existing KGs rather than constructing them. Exploring better KG alternatives could further improve performance.
- Theoretical Optimality: The current implementations of prompts and meta-paths may not represent theoretically optimal configurations.
- Future Directions: The authors suggest further investigation into integrating knowledge graph embeddings and neural architecture search for automated prompt optimization. They also note that the framework can be generalized to other domains (e.g., law, finance) by adapting the KG and agent roles.
Summary updated