CANOE: A Multi-Agent Framework for Contestable Care Plan Coordination
A new preprint on arXiv (2608.05391) introduces CANOE, which stands for Contestable Argumentative Network-of-Experts. This innovative multi-agent framework aims to improve how care plans are coordinated, focusing on transparency and safety. It addresses the limitations of traditional large language models by effectively integrating various clinical, functional, and psychosocial data. CANOE is made up of five key parts: assessing complexity, recruiting adaptable teams, using an Arena-based Quantitative Bipolar Argumentation Framework (A-QBAF) for role-specific arguments, allowing human involvement in disputes, and synthesizing care plans. Agents generate and debate arguments for interventions, resolving disagreements through a structured process. Care planners can then modify or accept these arguments, with the framework adjusting the final plan accordingly. The study emphasizes CANOE's potential to create more accountable AI systems in healthcare, ensuring that decisions can be explained and contested by human experts.
Key facts
- CANOE (Contestable Argumentative Network-of-Experts) is a multi-agent neuro-symbolic framework.
- It addresses limitations of monolithic LLM pipelines in care plan coordination.
- The framework consists of five modules: complexity assessment, adaptive team recruitment, role-based argumentative computation, human-in-the-loop contestation, and care-plan synthesis.
- It uses an Arena-based Quantitative Bipolar Argumentation Framework (A-QBAF) for argumentation.
- Role-specialized agents generate supporting and attacking arguments for candidate interventions.
- Conflicts are resolved through arena-based clash resolution.
- Acceptability scores propagate across the argumentation graph.
- Care planners can accept, reject, edit, or add arguments, and the framework deterministically recomputes the final plan.
- The paper is available on arXiv with identifier 2608.05391.
- The announcement type is 'new'.
Entities
Institutions
- arXiv