Sepsis prediction: why one widely used model missed two-thirds of cases and another cut deaths

Sepsis-3 defines sepsis as life-threatening organ dysfunction caused by a dysregulated host response to infection ^[1]. A global burden study estimated 48.9 million cases and 11 million sepsis-related deaths in 2017, about one in five deaths worldwide ^[2]. Early recognition is hard because early signs overlap with many other conditions.

Wong and colleagues externally validated the proprietary Epic Sepsis Model on 27,697 patients at Michigan Medicine in 2021 ^[3]. It achieved an AUC of 0.63, far below the developer's reported performance, missed 67% of patients with sepsis, and generated alerts for 18% of all hospitalised patients. That combination means both missed cases and alert fatigue.

Adams and colleagues studied the Targeted Real-time Early Warning System (TREWS) across five hospitals, published in Nature Medicine in 2022 ^[4]. Among patients with sepsis whose alerts were confirmed by a clinician within three hours, in-hospital mortality was 3.3 percentage points lower in adjusted analysis, an 18.7% relative reduction. Clinicians evaluated and confirmed a high share of alerts, which the authors credit to the system's design and workflow integration. The study is observational, so it shows an association rather than proof that the alerts caused the reduction.

• Validate on your own patients before go-live; vendor performance figures do not transfer automatically.
• Design the alert for the clinician: show the reasons, ask for a confirm or dismiss, and track response time.
• Measure outcomes such as time to antibiotics and mortality, not just AUC.
• Monitor after launch; case mix and documentation habits change.

Found this useful?

Share the key finding on X.