Adaptive Arena-based Contestable Argumentative Network-of-Experts for Open-Ended Care Plan Coordination

Modern healthcare requires synthesizing diverse clinical, psychosocial, and functional data to create safe care plans. Monolithic Large Language Models (LLMs) often struggle with this task because they lack transparency, can hallucinate critical information, and fail to represent the multidisciplinary deliberation necessary for high-stakes clinical decisions. The authors ask: can a multi-agent, neuro-symbolic system provide a more reliable, auditable, and human-contestable approach to care plan coordination?

The authors introduce CANOE (Contestable Argumentative Network-of-Experts), a framework that decomposes care planning into five distinct modules: complexity assessment, adaptive team recruitment, role-based argumentative computation, human-in-the-loop contestation, and final plan synthesis.

Instead of a single model generating a plan, CANOE recruits a team of specialized agents (e.g., pharmacists, nurses, social workers) based on the patient's specific needs. These agents generate supporting and attacking arguments for candidate interventions. The system then uses an Arena-based Quantitative Bipolar Argumentation Framework (A-QBAF) to resolve conflicts between these arguments. By calculating acceptability scores through a formal graph, the system provides a structured, mathematically grounded rationale for every recommendation. Crucially, the framework allows human clinicians to intervene—accepting, rejecting, or editing arguments—and automatically recomputes the final care plan based on these changes.

Evaluations on the Discharge Me! and MedicalRAG benchmarks demonstrate that CANOE effectively bridges the gap between automated generation and clinical accountability. While medically fine-tuned models provide the strongest clinical correctness, CANOE’s argumentative structure offers a significant advantage in transparency. It allows clinicians to trace the specific evidence and professional perspectives that drove a decision, fulfilling the need for contestable AI in clinical settings. The system’s ability to deterministically recompute plans after human input ensures that the final output remains grounded in both expert consensus and human oversight.