TREWS for sepsis — a structured, explained summary
Paper: Adams et al., Nature Medicine (2022)
TL;DR
This paper studies a deployed sepsis alert system called TREWS and asks the clinically relevant question: once such a system is live inside hospital workflow, is timely use of it associated with better patient outcomes? In a prospective multi-site study, TREWS monitored 590,736 encounters across five hospitals, and the main analysis focused on 6,877 sepsis patients who were alerted before antibiotics started. In that target group, patients whose alert was evaluated and confirmed by a provider within 3 hours had lower adjusted in-hospital mortality, better SOFA progression and shorter length of stay than patients whose alert was not confirmed within 3 hours. The paper matters because it moves beyond retrospective sepsis prediction and looks at what happens after deployment.
Why this paper exists
Sepsis AI has had no shortage of retrospective prediction papers. Many studies had already shown that machine learning models could identify sepsis early in historical EHR data. But that is not the same as showing clinical value after deployment. This paper starts from that gap. The authors say directly that few sepsis early warning systems had undergone real clinical evaluation and that more evidence was needed on patient outcomes, especially for decentralized bedside systems rather than workflows that depend on dedicated response staff.
That gap also mattered more because sepsis AI had already been reminded, very publicly, that good-looking model performance does not automatically translate into useful real-world deployment. The best-known example was the Epic Sepsis Model, which was widely implemented, but an independent external validation at Michigan Medicine reported an AUC of 0.63, poor calibration and a high alert burden, and concluded that the model’s widespread use despite poor performance raised serious concerns. Against that background, this TREWS paper is interesting less because it introduces another sepsis score and more because it asks what happened after implementation.
The clinical background is simple enough. Sepsis is common, dangerous and heterogeneous. Early recognition matters, and early antibiotics are linked to better outcomes, but many patients are still recognized late because sepsis does not present in one clean, uniform way. That is why early warning systems are appealing in the first place.
What this paper is actually studying
The important thing to get right is that this is not mainly a model development paper. The underlying TREWS model, its predictive performance and the deployment details are described in a companion paper. This paper is the outcomes paper. Its main question is whether provider interaction with the deployed TREWS alert is associated with better patient outcomes.
TREWS itself continuously monitors patients using routinely collected EHR data, including vitals, labs, clinician notes, medication history excluding antibiotic orders, procedure history and clinical history. When an alert occurs, providers can open a dedicated TREWS page in the EHR, review the patient and either confirm suspected sepsis or dismiss the alert. In this study, that provider interaction becomes central. The study arm is patients whose alert was evaluated and confirmed within 3 hours, and the comparison arm is patients whose alert was not confirmed within 3 hours.
That design choice matters. The paper is not comparing a hospital with AI to a hospital without AI. It is comparing different realized patterns of use of the same deployed system inside the same hospitals. That makes the study more realistic, but it also tells you immediately where the main caution will come from later.
How the study was set up
The study covered two academic and three community hospitals. Across the study period, TREWS monitored 590,736 adult encounters. Of those, 42,089 triggered an alert and 13,680 met retrospective sepsis criteria. The primary analysis then narrowed to 6,877 encounters in the intended target population: sepsis patients who triggered an alert before receiving antibiotics, received antibiotics within 24 hours of the alert and were not admitted directly to the ICU.
Within that cohort, 4,220 patients, or 61%, had their alert evaluated and confirmed within 3 hours, while 2,657 did not. The primary outcome was all-cause in-hospital mortality. Secondary outcomes were SOFA progression 72 hours after the alert and hospital length of stay among survivors. The analysis adjusted for patient demographics, comorbidity burden, labs, vitals, severity scores and admitting hospital.
Get Alex Yorov’s stories in your inbox
Join Medium for free to get updates from this writer.
The paper also defined a high-risk subgroup. This is one of the more useful ideas in the study. The authors built a secondary model to estimate which patients were likely to be most sensitive to delayed antibiotics, then used that model for analysis only, not as part of the frontline TREWS interface. The point was practical: if some patients benefit much more from very early treatment, can they be identified near alert time and prioritized?
What the results actually show
The main result is fairly clear. In the primary cohort, patients whose alert was confirmed within 3 hours had lower adjusted mortality than those whose alert was not confirmed within 3 hours. The adjusted absolute reduction was 3.34%, and the adjusted relative reduction was 18.18%. The same group also had better SOFA progression and about 11.6 fewer hours of median length of stay among survivors.
The high-risk subgroup is where the paper becomes more interesting than a simple “alerts help” story. Among patients identified as high risk, timely confirmation was associated with a larger adjusted absolute mortality reduction of 4.50%, along with better SOFA progression. So the paper is not only saying that earlier recognition may matter. It is saying that the patients most likely to benefit may be identifiable at alert time.
The antibiotic timing analysis supports that story from another angle. Patients who received antibiotics within 3 hours of the alert had lower mortality, better SOFA progression and shorter stay than those treated later. The paper also reports that each hour of delay from alert to antibiotics was associated with higher mortality, with adjusted odds ratios of 1.08 for patients treated within 6 hours and 1.05 for those treated within 12 hours. That matters because it makes the alert function as a clinically meaningful time zero rather than just a software event.
What is technically interesting here
One of the most useful operational points in the paper is that TREWS was deployed as a decentralized bedside system. It did not depend on pop-up interruptions, automated phone calls or a separate team that had to review every alert. Providers could choose whether, when and how to respond. That makes the system harder to study cleanly, but more realistic as a hospital workflow tool.
Another strong point is methodological rather than architectural. The authors explicitly avoid a simple before-versus-after deployment design, which can be distorted by surveillance bias if more patients get coded as septic after an alert system is introduced. Instead, they use post-deployment data only and rely on natural variability in provider response to form the comparison. That does not eliminate confounding, but it is a more careful design than many implementation papers use.
The paper also treats alerting less as a pure prediction problem and more as a treatment-timing problem. That is a better framing. The point is not merely that TREWS can produce a score. The point is whether it can create an earlier actionable moment that shifts antibiotics earlier enough to matter for outcomes.
The important caution
This paper should not be read as proving, in the strongest causal sense, that TREWS itself reduced mortality. It is a prospective observational cohort study, not a randomized trial. The authors adjusted for many measured patient variables, but they also state clearly that residual confounding may remain. Faster alert confirmation may reflect the value of the alert, faster clinician recognition, lower diagnostic ambiguity, stronger teams or some combination of these.
That caution matters because the study arm is partly defined by provider behavior. A clinician who confirms an alert quickly may differ in important ways from one who does not, and not all of those differences are necessarily captured in the EHR. So the right reading is association with improved outcomes, not definitive proof that the model alone caused those improvements.
There are other limits too. This was one hospital system in one region, the study relied on a particular retrospective sepsis phenotype and the authors did not have data on whether antibiotics were always microbiologically appropriate for the infection. So the paper is strong on internal deployment evidence, but it is not the last word on generalizability or causal effect.
What this paper really contributed
The cleanest way to read this paper is not “machine learning predicted sepsis well.” That is too easy and too general. The more useful reading is this: the paper studies whether a deployed sepsis alert can define an earlier treatment-relevant decision point inside routine clinical workflow, and whether timely use of that point is associated with better outcomes.
That is why the paper matters. Its contribution is not a flashy new architecture. It is prospective deployment evidence tied to mortality, organ dysfunction and length of stay. In a field where many sepsis models looked promising in retrospective evaluation, and where at least one very widely deployed system became controversial after independent validation, that shift toward outcome-linked deployment evidence is important.
It also contributes a practical idea about prioritization. The paper suggests that future alert systems may not need to treat every alert as equally urgent. If high-risk patients can be identified near alert time, then systems could reduce alert burden by giving more priority to the cases most likely to benefit from immediate action.
If you keep one idea from this paper
Do not remember this paper as “the one where AI lowered mortality.”
Remember it as a deployment paper showing that a high-precision sepsis alert, integrated into ordinary EHR workflow, can create an earlier treatment-relevant decision point, and that timely clinician confirmation of that alert was associated with better patient outcomes.
