.png)
UC San Diego Health
Healthcare · US
17% fewer sepsis deaths
UC San Diego Health saves 22 extra lives in 5 months with real-time AI sepsis detection
Deep-learning model COMPOSER monitors 150+ patient variables continuously and only triggers an alert when sepsis risk crosses a meaningful threshold, reducing alert fatigue while saving lives.
Friction
Sepsis is a leading cause of hospital mortality, but early signals are hard to distinguish from normal variation. Existing systems either missed cases or flooded nurses with false alerts, causing staff to ignore them.
Breakthrough
COMPOSER runs continuously in the Epic EHR, monitoring each patient in real time. When risk is high, it triggers a Best Practice Advisory. Uncertain cases are marked "indeterminate" rather than forcing a binary decision, deliberately reducing alert fatigue.
Impact
1.9 percentage point absolute reduction and 17% relative reduction in sepsis mortality. 22 additional survivors in a 5-month intervention period. Sepsis bundle compliance rose 5 percentage points. Study covered 6,217 adult sepsis patients across two emergency departments.
Unlock the full analysis with breakthrough, impact, what made it smart and its technical approach below!
Problem
Sepsis requires fast treatment, but early symptoms are subtle and easily missed. Manual monitoring across hundreds of patients simultaneously is impossible. Existing rule-based alert systems generated so many false positives that clinical staff began ignoring them.
What made it smart
COMPOSER is not a generic warning tool. It is a real-time deep learning model built specifically to limit unnecessary alerts. Its "indeterminate" output category for borderline cases is a key innovation: the model knows when it does not know, which makes clinicians trust it more.
Technical approach
The model ingests 150+ variables per patient including lab values, vitals, medications, demographics, and medical history, and runs continuously from patient check-in. Alerts are sent to nurses via Epic. Nurses and physicians then determine the next step together. Human-in-the-loop is a hard design requirement, not an afterthought.
Strategic lesson
The most effective AI in high-stakes environments is not the one that flags the most. It is the one that flags so precisely that people actually act on it.
Reflection question
In your organisation, where are warning signals being ignored because there are too many of them? And what would it take to make alerts trustworthy again?