Johns Hopkins Health System

Sepsis comprises nearly 27% of in-hospital deaths globally, yet early recognition remains difficult. Prior AI deployments had failed to produce real-world impact due to poor provider adoption, excessive false alerts, and one-size-fits-all models that ignored the diversity of patient populations and care delivery patterns across different health systems.

Bayesian Health's adaptive AI platform (TREWS — Targeted Real-time Early Warning System) was deployed across five hospitals integrating directly with existing EMR systems. The platform adapts to each hospital's unique patient population and provider workflows, delivering early sepsis detection flags with prescriptive clinical workflows, explanations, and escalation pathways (paging, phone) embedded in the clinician's existing tools.

Across 764,707 patient encounters (17,538 with sepsis) with 2,000+ providers, the platform achieved 82% sensitivity with a 5.7-hour lead time over standard detection. Provider adoption reached 89%, and timely use of the AI was associated with an 18.2% relative reduction in sepsis mortality, as validated in three prospective, peer-reviewed studies published in Nature Medicine and npj Digital Medicine.

• Adaptive AI that personalizes to individual hospital characteristics achieves high clinician adoption (89%) where traditional models have failed
• Embedding AI alerts directly into EMR workflows with prescriptive, explainable guidance is key to driving actual behavior change at point of care
• Rigorous prospective multisite studies — not retrospective analyses — are required to credibly demonstrate mortality impact from clinical AI