Sepsis Prediction AI: Three Health Systems, Three Measured Outcomes

Will Drewes

Founder, Fern Strategy · 2 min read

The numbers hospitals are actually posting

Sepsis kills roughly 270,000 Americans a year and accounts for one in three hospital deaths. It also moves fast: every hour of delayed treatment increases mortality risk. That combination, stable clinical rules applied to messy, high-volume patient data streams, is exactly where prediction models have started earning their keep.

Three health systems have published results specific enough to be useful.

UC San Diego Health deployed an AI surveillance model across its inpatient units and published outcomes in early 2024. The results: a 1.9 percentage-point absolute reduction in in-hospital sepsis mortality (a 17% relative decrease), a 5-point increase in sepsis bundle compliance, and a 4% reduction in 72-hour SOFA score change, the last metric being a direct measure of organ function deterioration. SOFA improvements are hard to move. A 4% reduction means the model caught deteriorating patients early enough to change their trajectory, not just flag them.

Johns Hopkins Medicine investigators studied their AI early-warning system across a large hospital cohort and found it identified 82% of sepsis cases before they became critical, with an associated 20% reduction in sepsis mortality. The 82% sensitivity figure matters because the baseline problem with sepsis detection is false negatives: patients who look stable until they aren't. Getting to 82% early identification, at a scale where the alert volume stays manageable for clinicians, is the hard part.

Duke Health integrated an AI model into its clinical workflow in 2018 and tracked outcomes over time. The result was a 27% reduction in sepsis deaths. Duke's case is notable because it predates most of the current wave of clinical AI interest by several years, which means the outcome data is more mature and less likely to reflect novelty effects or Hawthorne bias.

What this means for health system leaders

These three cases share a structural pattern worth noting. None of the outcomes came from dropping a model into a workflow and waiting. Each required integration with existing EHR data pipelines, defined alert thresholds, and clinical protocols that told nurses and physicians what to do when the model fired. The model is the easy part to procure. The integration layer and the governance around it (who gets alerted, at what threshold, with what escalation path, and how you audit false positives over time) is where most implementations stall.

For health systems evaluating build-versus-buy on early-warning tools, the UCSD and Johns Hopkins cases are particularly instructive. Both involved models developed or heavily customized internally, which gave the teams control over threshold tuning and audit trails. That control has real value in a regulatory environment where CMS and accreditation bodies are starting to ask harder questions about how clinical AI decisions are documented.

If your current sepsis alerting is a rules-based score with no feedback loop and no audit log, you're not comparing yourself to where these systems started. You're comparing yourself to where they are now.