Foundation models have yet to demonstrate feasibility, safety, or effectiveness for data analysis or decision support in the ICU

Critical care medicine, a specialty defined by high-stakes, time-sensitive decisions made under uncertainty, would seem to benefit from machine learning (ML)-based clinical decision support (CDS). Indeed, the vast amount of clinical data collected in the ICU has led to the publication of hundreds of ML models intended for CDS in critical care. But only a handful have been employed at the bedside and even fewer evaluated in rigorous clinical trials. We must overcome the implementation and evidence gaps for ML in critical care to realize the promise of this technology and ensure its effectiveness and safety. A foundation model for critical care would be an exciting technical innovation that could contribute to improved performance and generalizability of ML systems. In this viewpoint, we review some of the potential benefits of foundation models and argue that, despite these benefits, these models do not address the primary obstacles that have thus far limited the adoption of ML models for bedside use in critical care.

National Institutes of Health

Alexander T Moffett, Gary E Weissman