Experiences and challenges in deploying a respiratory failure prediction model in the ICU

This work was submitted (not yet accepted) as a talk proposal to AI+HEALTH 2026, Stanford University’s sixth annual online conference on artificial intelligence in health and medicine. It describes technical, methodological, and clinical challenges encountered when preparing a respiratory failure prediction model for deployment and prospective silent validation in the ICU. Drawing on the VentAI project, it offers practical lessons for defining clinically meaningful prediction tasks, designing fair clinician comparisons, and evaluating the entire prediction workflow. This research was supported by INNOSUISSE under project 114.383 IP-LS, “VentAI: Ventilatory decision support system for acute respiratory failure.”