Why is sepsis so hard to catch in the ER?
Sepsis, sometimes called blood poisoning, happens when the body's response to an infection spins out of control. It can damage organs like the kidneys or lungs and can be deadly.
Treatments such as antibiotics, IV fluids, and removing the source of infection work best when they start early. But early on, the signs are vague and differ from one patient to the next. Many sepsis alerts already built into hospital records go off too often for nothing, which wears down staff trust.
What did the San Diego team do?
The UC San Diego Health team added COMPOSER to two emergency departments. COMPOSER is a deep learning model, an AI that learns patterns from large amounts of data. It reads real-time data from the electronic health record, such as vital signs and blood tests. It was built to raise fewer false alarms: when a case does not look enough like the ones it knows, it labels it "indeterminate" instead of firing an alert.
In an unusual choice, the alert goes to the nurse, not the doctor. The nurse decides whether the doctor needs to know right away. The researchers compared 6217 patients with sepsis: 5065 before the tool arrived and 1152 during the 5 months after. A statistical model adjusted for seasons, ER volume, and how sick patients were, to estimate what would have happened without the alert.
Results: fewer deaths and faster care
After launch, 9.5% of patients with sepsis died in the hospital, while the researchers' model expected 11.4% without the alert. That is 1.9 percentage points lower, a 17% relative drop. Compliance with the sepsis care bundle (timely blood cultures, antibiotics, and fluids) rose by 5.0 points, and organ damage over the first 72 hours worsened a little less.
In more than half of the alerts, nurses chose to tell the doctor immediately. In those cases, patients got antibiotics sooner. The authors see this as a likely reason for the drop in deaths.
Key figures
- 6217adult patients with sepsis included in the study
- 17%relative drop in in-hospital sepsis deaths after the alert went live
- 9.5%in-hospital death rate after the alert, vs. 11.4% expected without it according to the researchers' model
- +5.0 pointsin compliance with the sepsis care bundle (recommended tests and treatments)
After sepsis, keeping track with Mima
This kind of alert happens inside the hospital, between the AI and the ER team: there is nothing you need to do. But sepsis often leaves questions about what happened and what comes next.
Mima helps you keep track of your hospital stay, your prescriptions, and how tired you feel day by day, then get your questions ready for your follow-up visit, such as: "Which signs should send me back to the ER?"
What this means for patients
This is one of the first times a deep learning sepsis tool used in real ER care has been linked to better survival. It also shows that a well-designed alert, one that rarely fires for nothing and goes through nurses, can speed up care.
For patients and families, the message stays simple: an infection that gets worse fast (fever or feeling very cold, fast breathing, confusion) needs emergency care. The AI supports the care team; it does not replace the exam or anyone's judgment. The CDC offers a plain-language guide to the warning signs of sepsis.
What are the limits of this study?
The study was not randomized: it compared a period before with a period after. The authors say so themselves, so it cannot prove for sure that the AI caused the drop. The fall in deaths was also only clear at one of the two hospitals.
Everything took place in one large academic center with a strong focus on sepsis, over a short 5-month window. The study did not look at patients flagged by mistake, for example those who may have received antibiotics they did not need. The authors call for randomized trials across many hospitals.
Frequently asked questions
What is sepsis?
Sepsis is the body's extreme, out-of-control response to an infection, which can damage organs. It is a medical emergency: the earlier treatment starts, the better.
What is the COMPOSER sepsis model?
COMPOSER is a deep learning model built at UC San Diego Health. It continuously reads ER patients' electronic health records and alerts nurses when sepsis seems to be developing.
Does AI reduce sepsis deaths?
In this study, in-hospital sepsis deaths fell by 17% (relative) after COMPOSER went live. The study was not randomized, so it shows a strong link, not final proof.
Is COMPOSER FDA approved?
The study does not mention FDA clearance. It describes a tool built and used inside the UC San Diego Health system.
The original study
Boussina A, Shashikumar SP, Malhotra A, et al. “Impact of a deep learning sepsis prediction model on quality of care and survival” npj Digital Medicine, January 23, 2024. DOI : 10.1038/s41746-023-00986-6
Read the full study in npj Digital MedicineThis article summarizes a scientific study and is not medical advice.