One Panel Does Not Fit All: Case-Adaptive Multi-Agent Deliberation for Clinical Prediction
Summary: arXiv:2604.00085v1 Announce Type: new
Abstract
Large language models applied to clinical prediction exhibit case-level heterogeneity: simple cases yield consistent outputs, while complex cases produce divergent predictions under minor prompt changes. Existing single-agent strategies sample from one role-conditioned distribution, and multi-agent frameworks use fixed roles with flat majority voting, discarding the diagnostic signal in disagreement. We propose CAMP (Case-Adaptive Multi-agent Panel), where an attending-physician agent dynamically assembles a specialist panel tailored to each case’s diagnostic uncertainty. Each specialist evaluates candidates via three-valued voting (KEEP/REFUSE/NEUTRAL), enabling principled abstention outside one’s expertise. A hybrid router directs each diagnosis through strong consensus, fallback to the attending physician’s judgment, or evidence-based arbitration that weighs argument quality over vote counts. On diagnostic prediction and brief hospital course generation from MIMIC-IV across four LLM backbones, CAMP consistently outperforms strong baselines while consuming fewer tokens than most competing multi-agent methods, with voting records and arbitration traces offering transparent decision audits.
Introduction
The field of clinical prediction has experienced significant advancements with the introduction of large language models (LLMs). However, these models often struggle with case-level heterogeneity. This inconsistency can lead to varied outcomes, especially when dealing with complex cases that may yield different predictions based on slight variations in prompts. The current methodologies, including single-agent and basic multi-agent frameworks, often fall short in addressing these nuances.
The CAMP Approach
The Case-Adaptive Multi-agent Panel (CAMP) presents an innovative solution to the challenges faced in clinical prediction. Below are the key components of the CAMP framework:
- Dynamic Specialist Panel: An attending-physician agent curates a specialist panel that adapts to the unique diagnostic uncertainty of each case.
- Three-Valued Voting System: Each specialist in the panel can vote to KEEP, REFUSE, or remain NEUTRAL on a proposed diagnosis, allowing for nuanced decision-making.
- Principled Abstention: Specialists can abstain from voting on cases outside their expertise, enhancing the quality of the deliberation.
- Hybrid Router: This component directs the decision-making process, ensuring strong consensus is reached or relying on the attending physician’s judgment when necessary.
Performance and Transparency
In rigorous testing against strong baseline models, CAMP has demonstrated superior performance in both diagnostic prediction and brief hospital course generation, utilizing data from the MIMIC-IV database. Notably, CAMP is also more efficient in terms of token consumption compared to many competing multi-agent methods. This efficiency, combined with detailed voting records and arbitration traces, provides a transparent audit trail for decision-making processes, thereby enhancing the trustworthiness of clinical predictions.
Conclusion
The introduction of CAMP represents a significant step forward in the application of AI to clinical prediction. By addressing the limitations of previous models and offering a more adaptable and transparent framework, CAMP has the potential to improve diagnostic accuracy and ultimately benefit patient care. As the landscape of AI in healthcare continues to evolve, methodologies like CAMP will be crucial in navigating the complexities of clinical decision-making.