A perspective on foundation models in intensive care medicine.

Abstract

Foundation models (FMs) pretrained with self-supervised objectives on heterogeneous ICU data have capabilities which supersede siloed, single-task predictive analytic paradigms. Here we present a framework for applying FMs in intensive care and a roadmap for safe, effective deployment. We argue that FMs applied in the time-critical, data-dense environment of intensive care could support rapid learning cycles and real-time, patient-level specific decisions. We highlight enabling resources, including large public databases, and outline potentially high-impact applications: continuous risk prediction and triage support; workflow-aware monitoring that reduces cognitive load; and personalized, data-informed interventions. We also describe potential generative uses, from noninvasive surrogates of invasively measured signals and short-horizon physiologic forecasting to synthetic data for rare conditions. To ensure trust and clinical utility, we advocate grounded explainability, prospective and post-deployment evaluation, and alignment with emerging regulatory pathways. Finally, we propose a staged path from passive surveillance to co-pilot decision support, bounded closed-loop control, and system-level orchestration under human oversight. Realizing this vision requires advances in multimodal fusion at scale, uncertainty-aware real-time inference, decision-focused learning, and human-centered design and governance; whether such systems improve ICU outcomes remains to be established prospectively.