Johns Hopkins Hospital

Sepsis affects over 1.7 million adults in the U.S. annually and contributes to 1 in 3 hospital deaths. Symptoms are subtle and can be mistaken for less serious conditions, and traditional early warning systems relied on static thresholds that failed to provide sufficient lead time for intervention. Clinicians needed a dynamic, real-time approach to detect sepsis before visible symptoms fully manifested.

Johns Hopkins developed and deployed TREWS (Targeted Real-Time Early Warning System), a machine learning platform that continuously analyzes vital signs, lab results, EHR data, medication history, and time-series monitor data. The system assigns a risk score and sends contextual, explainable alerts to attending clinicians—often hours before symptoms fully manifest—embedded directly into existing EHR workflows.

Sepsis mortality rates decreased by 20-30% in departments where TREWS was deployed. Clinicians initiated early interventions up to 3 hours sooner than before, false positive alerts dropped significantly, and patient ICU length of stay decreased—contributing to better resource utilization and lower treatment costs.

• Explainable AI that shows clinicians why an alert was triggered is critical for adoption and trust.
• Embedding alerts into existing EHR workflows—rather than creating separate systems—dramatically improves clinical uptake.
• Continuous model retraining is required to keep predictive systems accurate as protocols, medications, and pathogens evolve.