University of California San Diego Health

Sepsis accounts for 1 in 3 hospital deaths in the U.S., yet it is largely recognized only after its effects are already visible. Existing rule-based detection alerts are slow and produce high false-positive rates, contributing to alarm fatigue in busy emergency departments where true alerts risk being ignored.

UCSD researchers deployed COMPOSER, a deep-learning AI model, across two hospital emergency departments. The model continuously monitors more than 150 patient variables—vital signs, lab results, medications, and medical history—using multiple layers of artificial neural networks. High-risk detections trigger EHR-integrated alerts to nursing staff, while uncertain cases are flagged as 'unknown' rather than generating unreliable recommendations.

A prospective pre-post study of more than 6,000 patient admissions found a 1.9% absolute decrease in mortality, corresponding to a 17% relative decrease across the two EDs. This translated to approximately 60 lives saved per year at those facilities. The system went live in December 2022 and was reported as the first study to demonstrate improved patient outcomes using a deep-learning sepsis prediction model in clinical practice.

• Models that output 'I don't know' for out-of-distribution cases reduce false alarms and preserve clinician trust, addressing a key barrier to AI adoption in high-alert environments.
• Continuous, silent surveillance of 150+ variables outperforms periodic rule-based checks for time-sensitive conditions like sepsis.
• Even a ~2% absolute mortality reduction at two hospitals extrapolates to thousands of lives nationally, underscoring the population-level stakes of deploying validated AI at scale.