AI Sepsis Early Warning Systems: What Worked, What Failed and Why

AI sepsis early warning systems are the best case study in clinical AI, because they include both a widely publicised failure and some of the strongest outcome evidence any clinical AI has produced. The difference between the two is rarely the algorithm alone. It is validation, alert design and the clinical workflow behind the alert.

Sepsis is a good target for AI: it is common, deadly, time-sensitive and often recognised late. But a sepsis alert that fires on a fifth of all inpatients and still misses most cases does harm, by burning out nurses and teaching everyone to ignore alerts.

This guide covers what happened with the major systems, what new 2026 data shows, and how a hospital in India or elsewhere should approach a sepsis or deterioration model.

The cautionary tale: Epic Sepsis Model version 1

Epic's original sepsis model was deployed at hundreds of US hospitals. In 2021, Michigan Medicine published an external validation in JAMA Internal Medicine. As reported at the time:

• The hospitalisation-level AUC was 0.63, well below the 0.76 to 0.83 the vendor had reported.
• At the alert threshold used, it missed 67% of sepsis patients.
• It generated alerts for 18% of all hospitalised patients.
• It identified only 7% of the sepsis cases clinicians had missed.

Epic disputed the methods, including the threshold and how sepsis onset was defined. But the lesson stuck: a proprietary model deployed widely without independent validation can perform very differently in your hospital than in the vendor's slides.

What changed: Epic Sepsis Model version 2

Epic rebuilt the model, and version 2 is trained on each health system's own data. In February 2026 a multicentre prospective validation in JAMA Network Open evaluated it across 227,091 adult inpatient encounters at four US health systems between August 2023 and March 2025.

Version 2 beat version 1 on every metric at every site. But positive predictive value stayed low: at the thresholds used, most alerts were still false positives, and performance varied considerably between hospitals. The authors recommended local validation before deployment and workflows to manage false positives.

The success stories: TREWS and COMPOSER

Two systems have published outcome improvements, not just accuracy.

TREWS (Johns Hopkins / Bayesian Health). A prospective multi-site study in Nature Medicine in July 2022 covered 590,736 encounters at five hospitals. Among 6,877 sepsis patients flagged before antibiotics, when providers confirmed the alert within three hours, median time to first antibiotic order was 1.85 hours shorter, and adjusted in-hospital mortality was 3.3 percentage points lower (an 18.7% relative reduction). The design was observational, not randomised.

COMPOSER (UC San Diego). A before-and-after study in two emergency departments, published in npj Digital Medicine in 2024, covered 6,217 septic patients. Deployment was associated with a 1.9 percentage point absolute (17% relative) reduction in in-hospital sepsis mortality and a 5 percentage point increase in sepsis bundle compliance.

Both share features worth copying:

• The alert went to a clinician who had to evaluate and confirm or dismiss it.
• The alert sat inside the clinical workflow, not in a separate dashboard.
• The teams measured process outcomes (time to antibiotics, bundle compliance), not just AUC.

There is also a regulated product: in April 2024 the FDA authorised Prenosis's Sepsis ImmunoScore through the De Novo route, an AI tool that uses 22 parameters to estimate sepsis risk within 24 hours and outputs a risk score and category.

Rule-based scores versus machine learning

Before buying a model, it helps to compare it with what you can build yourself.

For most Indian hospitals the sensible order is manual screening to rule-based score to ML. Each step should beat the previous one on your own data before you switch, and the rule-based score gives you the comparison baseline.

Why sepsis AI is hard

• No clean definition. Sepsis-3, SIRS-based definitions and billing codes label different patients. A model trained on one may look poor on another.
• Label leakage. Models can learn from clinicians' actions (ordering a lactate or blood culture) and appear to "predict" sepsis that clinicians already suspected.
• Low prevalence. In the 2026 study, sepsis occurred in 3.3% of encounters, so even good models produce many false positives.
• Drift. Changes in lab panels, documentation templates or patient mix shift performance.

Lessons for Indian hospitals

Indian hospitals face extra constraints: vitals often documented on paper or entered late, lab results not always timestamped accurately, and nurse ratios that cannot absorb a flood of alerts. Before any model:

1. Get vitals into the system on time. A model fed vitals entered at end of shift cannot warn early. Our guide to nursing management systems covers electronic vitals and charting.
2. Start with a rules-based early warning score. NEWS2 or a modified version, calculated automatically from electronic vitals, is transparent and well understood. It gives you the workflow and baseline before adding ML.
3. Define escalation. Who receives the alert, how quickly, and what they must do. A rapid response team or senior resident review is common.
4. Then consider ML. Validate locally on your own historical data, compare against your rule-based score, and only switch if it is clearly better.

How to deploy a sepsis or deterioration model

Choose the threshold based on alert volume your nurses can handle, not on the best statistical cut-off. Then decide how repeat alerts are suppressed and how dismissals are documented.

Regulatory and governance notes

Whether a sepsis model is regulated depends on intended use and presentation. In the US, the FDA's January 2026 CDS guidance changed the treatment of some risk scores; in India, CDSCO's medical device software guidance classifies standalone software by condition severity and how much its output drives decisions. Sepsis is severe and time-critical, so assume scrutiny. See our AI medical device regulation guide and our hospital AI governance framework.

Building the data pipeline

The engineering work is usually larger than the modelling. You need near-real-time vitals, labs and medication data, a scoring service, an alert channel that reaches the right person (EMR banner, nurse station screen, mobile notification), and an audit log. That is classic health data analysis and machine learning engineering, and it sits on top of a reliable hospital management system.

Designing the alert itself

How the alert looks and behaves decides whether anyone acts on it. Good designs share a few traits.

• Show the why. List the inputs that drove the score (rising heart rate, falling blood pressure, new lactate) so the nurse or doctor can judge it quickly.
• Require a response. Confirm and start the sepsis pathway, or dismiss with a reason. TREWS outcomes were strongest when providers confirmed alerts within three hours.
• Suppress sensibly. Do not re-fire every hour for a patient already on the sepsis pathway; do re-fire if the picture worsens.
• Route to a person. An alert that goes to a shared screen with nobody responsible is the same as no alert.
• Link to the order set. One click from alert to cultures, lactate and antibiotics shortens time to treatment.

Questions to ask a sepsis AI vendor

1. Has the model been validated outside the vendor's own sites, and are the results published?
2. Can it be trained or calibrated on our data, and how much history is needed?
3. What sepsis definition was used for training and validation?
4. What is the expected number of alerts per 100 patient-days at the proposed threshold?
5. Can we run it silently for two to three months before showing alerts?
6. What is its regulatory status in India and elsewhere?
7. How will updates be handled, and how will we be told?

Conclusion

AI sepsis early warning systems show both what goes wrong and what goes right with clinical AI. The failures came from deploying unvalidated models with high alert burden. The successes came from validated models wired into clear clinical workflows with measured outcomes. New 2026 evidence shows better models still produce many false alerts and vary by hospital.

Start with timely electronic vitals and a transparent early warning score, define the response, validate any ML model on your own data, and monitor it like any other clinical intervention.