Title: VentAgent: When LLMs Learn to Breathe -- Multi-Objective Arbitration for ARDS Ventilation

Abstract:

Managing Acute Respiratory Distress Syndrome (ARDS) through mechanical ventilation demands a delicate equilibrium among conflicting physiological objectives, specifically maintaining acid-base balance, protecting lung tissue, and ensuring adequate oxygenation. Existing data-driven approaches, particularly those that mimic retrospective Electronic Health Records (EHR), frequently fall prey to imitation bias. These models tend to pick up on superficial correlations derived from inconsistent clinical practices; for instance, they might incorrectly link survival outcomes to passive ventilator settings simply because such settings are prevalent among stable patients. Consequently, these methods often lack the ability to generalize effectively to unstable or out-of-distribution patient phenotypes. Furthermore, conventional Reinforcement Learning (RL) techniques face difficulties in navigating the adversarial trade-offs inherent to critical care, frequently generating black-box policies that lack clinical interpretability.

To overcome these challenges, we present VentAgent, a hierarchical framework that employs Large Language Models (LLMs) as transparent arbitrators for mechanical ventilation. We redefine ventilation control as a dynamic Multi-Objective Arbitration process, moving away from single-objective optimization. VentAgent structures decision-making into three distinct, interpretable phases: Perception, Planning, and Orchestration. By utilizing the semantic reasoning strengths of LLMs, the system synthesizes strategies from diverse experts and addresses conflicting clinical priorities via an explicit coordination mechanism. Assessments conducted on a high-fidelity physiological simulator demonstrate that VentAgent surpasses state-of-the-art RL and classical control baselines. Additionally, it translates control decisions into human-readable reasoning chains, establishing a safer, more adaptable, and interpretable paradigm for automation in critical care.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC