Johns Hopkins hospitals reduce sepsis mortality by 20% with AI early warning system detecting cases 6 hours sooner

Sepsis claims over 250,000 lives annually in the US and is notoriously difficult to detect early because its symptoms—fever, confusion—overlap with many other conditions. Traditional electronic detection tools caught fewer than half of sepsis cases and were accurate only 2–5% of the time, leading to frequent false alarms that eroded clinician trust and delayed intervention.

The Targeted Real-Time Early Warning System (TREWS), developed at Johns Hopkins and commercialized by spin-off Bayesian Health, continuously scours EHR data, clinical notes, lab results, and patient history to flag sepsis risk and recommend treatment protocols such as antibiotic initiation. The system tracks patients from admission through discharge across department transitions, and was integrated with Epic and Cerner EHR platforms for broad deployability.

Across a two-year study of 590,000 patients at five hospitals with more than 4,000 clinicians, the AI detected 82% of sepsis cases with ~40% accuracy—roughly 8× better than prior tools. In the most severe cases, the system identified sepsis an average of nearly six hours earlier than traditional methods. Patients treated using the system were 20% less likely to die from sepsis.

• Site-adaptive ML that accounts for local patient population diversity and care delivery patterns is critical to achieving real-world accuracy and clinician adoption.
• Integrating directly with major EHR platforms (Epic, Cerner) at deployment time removes a key adoption barrier for scaling to new hospitals.
• Publishing outcomes in peer-reviewed journals (Nature Medicine, Nature Digital Medicine) with 590,000-patient validation establishes the credibility needed for clinical trust.