Johns Hopkins AI System Detects Severe Sepsis Nearly 6 Hours Earlier Than Traditional Methods

Approximately 1.7 million adults develop sepsis annually in the United States, and more than 250,000 die. Early detection is critical — in the most severe cases, an hour delay can mean the difference between life and death. Traditional methods and prior electronic tools caught fewer than half of cases because sepsis symptoms such as fever and confusion overlap with many other conditions.

Johns Hopkins researchers developed the Targeted Real-Time Early Warning System (TREWS), a machine-learning model that combines a patient's medical history, current symptoms, and lab results to alert clinicians when someone is at risk for sepsis and recommend treatment protocols such as antibiotic initiation. Bayesian Health, a Johns Hopkins spinoff, led deployment across five hospitals. The system was integrated with Epic and Cerner EHR platforms to enable broader adoption.

More than 4,000 clinicians used TREWS to treat 590,000 patients across five hospitals. In 82% of sepsis cases the AI flagged accurately, it identified patients nearly 40% more often than previous electronic tools. In the most severe sepsis cases, TREWS detected the condition an average of nearly six hours earlier than traditional methods. Results were published in Nature Medicine and Nature Digital Medicine.

• Bedside AI deployment at scale is achievable and life-saving when validated rigorously across thousands of providers and hundreds of thousands of patients.
• Integration with dominant EHR platforms (Epic, Cerner) is essential to enable rapid adoption beyond the originating institution.
• Combining longitudinal patient history with real-time vitals and labs substantially outperforms single-signal or threshold-based sepsis alerts.