Johns Hopkins TREWS identifies 82% of sepsis cases early, cuts time to antibiotics by 1.85 hours

Sepsis affects nearly 49 million patients annually with ~11 million deaths worldwide. Existing detection protocols including Epic's sepsis algorithm have well-documented shortcomings, often failing to identify patients early enough for timely intervention.

Johns Hopkins investigators developed TREWS (Targeted Real-Time Early Warning System), a machine learning model deployed across five hospitals over two years. The system monitors patients in real time and alerts providers, who can confirm or dismiss the alert within a 3-hour window before antibiotic orders are placed.

Across 9,800+ retrospectively confirmed sepsis cases, TREWS identified 82% of patients early. A prospective multi-site study of 6,800+ patients showed that timely provider confirmation (within 3 hours) reduced median time to first antibiotic order by 1.85 hours and was associated with reduced in-hospital mortality.

• Provider confirmation speed is a critical mediating variable — alert value depends on clinician engagement within a narrow window.
• Both retrospective and prospective validation are needed; retrospective alone may miss confounders.
• Deploying across five hospitals in a real-world setting demonstrates generalizability beyond a single site.