What problem does it solve?
Sepsis and general clinical deterioration are both common and hard to catch early. Johns Hopkins reports that about 1.7 million adults develop sepsis every year in the United States and that more than 250,000 of them die. Symptoms such as fever and confusion overlap with many other conditions, deterioration can develop over hours rather than minutes, and a nurse or physician watching one patient at a time can miss the pattern across dozens of scattered vital sign and lab readings. Traditional early warning scores, computed by hand from a handful of vital signs at intervals, catch some of this, but multivariate models that continuously read the full electronic health record can add signal a periodic score would miss, if they are validated on the hospital's own population first.
- About 1.7 million adults develop sepsis every year in the United States and more than 250,000 of them die; under the current standard of care, sepsis kills 30 percent of the people who develop it.Sepsis detection AI has the potential to prevent thousands of deaths (2022)
How does it work?
- Continuous scanning. The model reads vital signs, lab results, nursing notes and orders from the electronic health record, refreshed hourly or more often, for every admitted patient rather than only those already flagged as high risk.
- Risk scoring. It calculates a probability of impending sepsis or deterioration and compares it against a threshold tuned to the hospital's own patient population.
- Alerting the right person. When the score crosses the threshold, it alerts a bedside nurse directly or, in a centralized model such as Kaiser Permanente's, a remote monitoring team who reviews the case before contacting the local care team.
- Bedside confirmation and action. A clinician assesses the patient at the bedside and decides on tests, antibiotics or escalation; the model suggests but never orders treatment itself.
- Feedback and retuning. Alerts that turn out to be false alarms are reviewed and used to retune the threshold, because too many false alerts cause staff to stop trusting the system.
- Audience
- Employee facing
- Autonomy
- Assist
- Adoption
- Mainstream
- Channels
- Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
No public deployment has disclosed a measurable outcome yet.
Value drivers: Risk and loss reduction, Speed and cycle time.
Who already uses it?
2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Johns Hopkins Medicine
United States · Healthcare · 2022
Johns Hopkins developed the Targeted Real-Time Early Warning System, TREWS, which combines a hospitalized patient's medical history with current symptoms and lab results, scouring medical records and clinical notes to flag sepsis risk. Bayesian Health, a company spun off from Johns Hopkins, led and managed the deployment across five hospitals. Over a two year study of 590,000 patients treated by more than 4,000 clinicians, published in Nature Medicine and Nature Digital Medicine in 2022, the system detected the most severe sepsis cases an average of nearly six hours earlier than the prior standard of care, and researchers found patients were 20 percent less likely to die of sepsis when the system was used.
No outcome disclosed.
Kaiser Permanente Northern California
United States · Healthcare · 2022
Kaiser Permanente Northern California built the Advance Alert Monitor, AAM, a predictive model developed by its own Division of Research that scans almost 100 elements from the electronic health record hourly for patients in medical and surgical units and transitional care units across its 21 Northern California hospitals, giving clinicians a 12 hour lead time before clinical deterioration. A specialized team of Virtual Quality Nurse Consultants monitors the model's output and contacts the patient's local care team when it fires, handling more than 16,000 alerts a year. A physician researcher analysis published in the New England Journal of Medicine found the program prevented an average of 520 deaths a year over a three and a half year study period, and the program has been recognized by The Joint Commission and the National Quality Forum, and honored with the 2021 John M. Eisenberg Award for Local Level Innovation in Patient Safety and Quality.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Streaming or hourly vital signs, laboratory results, orders and notes for every admitted patient
- A clinically confirmed record of who actually developed sepsis or deteriorated, to validate and retune the model on the hospital's own population
- A staffed response pathway, such as a bedside nurse, a rapid response team or a remote monitoring team, that can act on an alert within minutes
Systems to integrate
- Electronic health record, for vitals, labs, notes and orders
- Nurse call, paging or secure messaging system to deliver the alert
- Rapid response or code team workflow for escalation
Complexity: High
This is not a document assistant: it must read structured vitals, labs, orders and notes from the electronic health record in close to real time, integrate with a nurse call or paging workflow, and be validated and retuned locally, because a model built on one hospital's population and staffing pattern does not transfer cleanly to another, as external validations of other sepsis models have shown.
- 1
Validate locally before going live
Test the model's predictions against your own hospital's historical outcomes before relying on it, because performance built on one population and one care model does not transfer automatically to another.
- 2
Route to a named role, not just a chart
Decide exactly who receives the alert, such as a bedside nurse, a centralized monitoring team, or both, and what they must do within a set number of minutes.
- 3
Set and retune the alert threshold
Start conservative, measure the false alert rate in practice, and adjust it; too many false alerts cause staff to ignore even the true ones.
- 4
Build the bedside response into the workflow
An alert is only useful if it leads to a defined clinical action, such as a sepsis bundle or a rapid response call; write that pathway down and train staff on it before go live.
- 5
Track outcomes, not only alerts
Measure what happened to alerted patients compared with similar patients who were not alerted, not just how many alerts fired.
Guardrails
- The system only alerts; it never orders a test, a medication or a transfer itself
- A clinician always assesses the patient at the bedside before any treatment decision is made
- Alert thresholds are validated on the hospital's own population before go live and retuned after any major change in patient mix
KPIs to instrument
- Time from alert to bedside assessment
- False alert rate and the share of alerts staff act on
- Sepsis or deterioration related mortality and ICU transfers, before and after, on a comparable population
- Time to antibiotics or the relevant treatment bundle after an alert, versus without one
Human in the loop
A nurse or physician always evaluates the patient before any action is taken. The model's role is to shorten the time between a patient beginning to deteriorate and a clinician noticing, not to replace the clinician's judgment.
Common failure modes
- Alert fatigue
- Too many low value alerts and staff start ignoring all of them, including the true ones. Track the false alert rate, retune the threshold, and keep the number of alerts a clinician gets per shift to a workable level.
- A model that does not transfer
- A model tuned on one hospital's population and staffing pattern can perform far worse at another. An external validation of a different, widely deployed proprietary sepsis model found it caught only a third of true sepsis cases at one academic medical center; revalidate locally before go live and after any major change in patient mix.
- Alerts without an owner
- An alert nobody is accountable for acting on is worse than no alert. Name the role that must respond and the maximum time to respond, and track it.
What are the risks and rules?
EU AI Act
High risk
Article 6(1)(a) and (b) and Annex I: a system is high risk when it is a safety component of, or itself is, a product covered by EU harmonisation legislation listed in Annex I, and that product is required to undergo a third party conformity assessment under that legislation. Software that predicts sepsis or deterioration to guide treatment typically qualifies as a Class IIa or higher medical device under the EU Medical Device Regulation, Rule 11, which brings it into the high risk tier, though MDR Article 5(5) provides an in house exemption that can apply to a tool a health institution builds and uses only within its own organization.
Rules that apply
Controls to put in place
- Local validation of model performance against the hospital's own outcomes before go live and after major changes
- A named clinical role responsible for every alert, with a maximum response time
- Ongoing monitoring of the false alert rate and the clinician response rate
- Change control and revalidation when the patient population or care model changes materially
When it went wrong elsewhere
- External validation at the University of Michigan found a different, widely deployed proprietary sepsis model missed most true cases. A study of 38,455 hospitalizations found the Epic Sepsis Model, a different product from the deployments described on this page, had a sensitivity of only 33% and a positive predictive value of 12%, and that its predictive performance (AUC of 0.63) was well below the range reported internally by its developer (0.76 to 0.83). It illustrates why a sepsis or deterioration model built on one hospital's population may not transfer to another. Both deployments on this page were developed on their own health system's patient data by its own researchers, rather than adopted as an off the shelf product.
Frequently asked questions
- How much earlier does AI detect sepsis than standard care?
- Johns Hopkins reports that its Targeted Real-Time Early Warning System, TREWS, detected the most severe sepsis cases an average of nearly six hours earlier than the prior standard of care, in a study of 590,000 patients treated by more than 4,000 clinicians at five hospitals.
- Do these systems replace clinical judgment?
- No. Every deployment described here only alerts; a nurse or physician still assesses the patient at the bedside and decides on treatment. Kaiser Permanente's Advance Alert Monitor routes alerts to a virtual nursing team that contacts the bedside team, who make the clinical call.
- Can an early warning model be trusted straight out of the box?
- No. An external validation of a different, widely used proprietary sepsis model at the University of Michigan found it caught only a third of true sepsis cases and had a positive predictive value of just 12%. Both deployments described here were developed on their own health system's patient data by its own researchers, not bought and deployed as an off the shelf product.
- What is the evidence that these systems save lives?
- Kaiser Permanente Northern California's physician researchers estimate that its Advance Alert Monitor program prevents an average of 520 deaths a year over a three and a half year study period, published in the New England Journal of Medicine. Johns Hopkins reports that patients were 20% less likely to die of sepsis because of TREWS, in a study of 590,000 patients treated by more than 4,000 clinicians at five hospitals.
How to cite this page
Blits.ai AI Use Case Library, "AI early warning for sepsis and in hospital clinical deterioration", last verified 29 September 2026, https://www.blits.ai/ai-use-cases/sepsis-and-deterioration-early-warning. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 29 September 2026: First published