Contextual Bandit Preference Learning for Ventilator Decision Support

A newly developed multi-agent framework for ventilator decision support employs contextual bandit preference learning to align with the unique tuning styles of individual clinicians. The Ventilator Decision Support System (VDSS) integrates modular decision components via contract-driven structured interfaces, ensuring traceable evidence for review. It adapts preferences in real-time by modifying clinician-specific choices based on the final accepted decision during each adjustment cycle. Targeted replanning is initiated by structured rejection feedback, which aims to minimize unproductive iterations and enhance interaction stability. This system overcomes the limitations of rule-based methods that seldom achieve personalization, as well as the challenges of controlling and auditing end-to-end reinforcement learning or extensive language model systems. Details of the framework can be found in a paper on arXiv (2605.23320).