Sepsis and deterioration alerts
Software that uses EHR data to flag hospital patients at risk of sepsis or clinical deterioration.
Sepsis and deterioration alerts use data in the electronic health record to flag hospital patients at rising risk for clinician review. Among systems not reviewed here, a cluster randomized trial of the CONCERN early warning system in two health systems found a 35.6% lower instantaneous risk of death in intervention encounters (adjusted hazard ratio 0.64). Kaiser Permanente Northern California's Advance Alert Monitor, which sends alerts to a remote nurse team, was associated with lower mortality (9.8% versus 14.4%) in a staggered rollout over 19 hospitals.
A 2026 meta-analysis of external validations pooled an area under the curve of 0.79 for the Epic Deterioration Index and 0.65 for the Epic Sepsis Model. FDA guidance on clinical decision support, revised Jan. 29, 2026, cites software that detects sepsis and alerts a clinician as an example of a device software function. Alert burden is an open question: in a 2017 study of drug alerts and reminders in primary care, clinicians became less likely to accept alerts as they received more of them.
Bayesian Health TREWS
Bayesian Health says its real-time clinical intelligence platform, of which early sepsis detection is a core module, continuously monitors every patient and surfaces those who need attention. FDA cleared its sepsis flagging device to aid early detection or risk prediction of sepsis developing within 24 hours, used with clinical assessments and other laboratory data.
Why this grade. In a prospective study at five hospitals, sepsis patients whose TREWS alert a provider confirmed within three hours had lower in-hospital mortality than those whose alert was not confirmed in that time (adjusted absolute reduction 3.3 percentage points); there is no randomized trial.
The evidence
The five-hospital study analyzed 6,877 sepsis patients the system flagged before antibiotics; the adjusted relative mortality reduction with prompt confirmation was 18.7%, and the authors described the result as an association. A companion analysis of 9,805 sepsis cases at the same hospitals found the system identified 82%, physicians or advanced practice providers evaluated 89% of alerts, and providers confirmed 38% of evaluated alerts; confirmation within three hours was associated with a 1.85-hour shorter median time to the first antibiotic order.
In the retrospective validation submitted for FDA clearance, covering 7,732 encounters at three clinical sites, encounter-level sensitivity was 79.4%, encounter-level negative percent agreement was 89.5% and flag-level positive predictive value was 11.7%.
- Used by
- The five-hospital outcome study of the Targeted Real-time Early Warning System (TREWS) monitored 590,736 patients. Bayesian's May 2026 press release quotes clinicians at Cleveland Clinic and University of Rochester Medicine; it does not state the scope of those deployments.
- FDA
- FDA cleared the Bayesian Health Sepsis Flagging Device through the 510(k) pathway as K250680 on April 30, 2026, with Prenosis' Sepsis ImmunoScore as predicate. In the clearance validation, encounter-level sensitivity was 79.4%, below the prespecified acceptance criterion; the summary says the totality of device performance, including flag-level performance, was used to establish substantial equivalence.
- Limits
- The mortality result compares patients whose alerts were confirmed promptly with those whose alerts were not, an association rather than a randomized comparison. Under a license agreement, Johns Hopkins and two of the authors are entitled to revenue from Bayesian Health, and all three studies come from the developers or the company.
eCART
AgileMD says eCART guides care teams to the highest-risk patients using FDA-cleared AI combined with embedded decision support for all-cause clinical deterioration. FDA cleared it for automated risk stratification and early warning of impending deterioration, defined as death or ICU transfer.
Why this grade. In a pragmatic before-and-after study in a four-hospital community-academic health system, hospital mortality in the study's main cohort was 8.8% in the eCART intervention period versus 13.9% in the baseline period (adjusted odds ratio 0.60); there is no randomized trial.
The evidence
In the four-hospital study, clinicians were blinded to scores in the baseline year, and in the intervention year high-risk scores prompted physician assessment for the ICU; the mortality decrease was seen in high- and intermediate-risk patients, with more and earlier ICU transfers.
In validation submitted for FDA clearance, eCART's area under the receiver operating characteristic curve (AUROC) was 0.835 retrospectively and 0.828 in a prospective set of 205,946 encounters, where a score of 93 or above had 48.8% sensitivity and 93.3% specificity. In a retrospective comparison at seven Yale New Haven Health hospitals, eCART had the highest AUROC of six scores (0.895), and a public, non-AI early warning score outperformed the Epic Deterioration Index.
- Used by
- According to AgileMD's homepage, the company's software is live at more than 200 U.S. hospitals; a counter on the same page reads "400+ Hospitals."
- FDA
- FDA cleared the eCARTv5 Clinical Deterioration Suite through the 510(k) pathway as K233253 on June 21, 2024. Its 510(k) summary reports retrospective and prospective validation in adult ward patients from three health systems.
- Limits
- The outcome study compared two consecutive years in one health system, and mortality also fell in the average-risk cohort not subject to the intervention (adjusted odds ratio 0.53).
Epic Sepsis Model
Published studies describe a proprietary algorithm in Epic's EHR that calculates a score correlating with the likelihood of an International Classification of Diseases, Ninth Revision code for sepsis. Each hospital sets the score that triggers an alert, within a range of 5 to 8 suggested by Epic.
Why this grade. Two single-center nonrandomized studies of the model in use conflict: at a 746-bed academic trauma center, using the score as a screening test was associated with 44% lower odds of sepsis-related mortality, while at an emergency department time to first antimicrobial did not differ from sister hospitals using a different sepsis alert; no randomized trial has tested it.
Caution. In a study of 38,455 hospitalizations at Michigan Medicine, the area under the receiver operating characteristic curve (AUROC) was 0.63, against the 0.76 to 0.83 Epic reported in internal documentation, and the model missed 67% of sepsis cases while alerting on 18% of hospitalized patients. When predictions made after clinicians recognized sepsis were excluded, AUROC fell to 0.47 in a second Michigan study; a 2026 meta-analysis pooled 0.65.
The evidence
In the trauma-center study, a before-and-after comparison, mortality among patients with scores of 5 or more who had not yet received antibiotics fell from 24.3% to 15.9% (adjusted odds ratio 0.56). In the emergency department study, time to first antimicrobial was 3.33 hours, versus 3.22 hours at sister hospitals using an alert based on systemic inflammatory response syndrome criteria (P=0.437).
The Michigan validation also found the model identified 183 of 2,552 patients with sepsis (7%) who had not received timely antibiotics. Across nine BJC HealthCare hospitals, the C-statistic ranged from 0.55 to 0.73 and was worse where sepsis incidence, comorbidity burden and cancer prevalence were higher. In two county emergency departments in Houston, the alert had sensitivity of 14.7% and positive predictive value of 7.6%, and the authors concluded it fails to achieve meaningful sensitivity. The 2026 meta-analysis pooled three studies of 922,754 patients for the sepsis model, and Epic's reported confidence intervals did not overlap the pooled estimate.
- Used by
- Wong et al. reported in 2021 that the model is implemented at hundreds of U.S. hospitals.
- FDA
- No FDA clearance or authorization for the model was found. FDA's clinical decision support guidance, revised Jan. 29, 2026, lists software that detects sepsis and alerts a clinician as an example of a device software function. Wong et al. attributed hospitals' adoption of such algorithms to ease of integration within the EHR and loose federal regulations.
- Limits
- All the validations are retrospective, and each hospital sets its own alert threshold. Epic developed the model on 405,000 patient encounters at three health systems from 2013 to 2015, and its reported accuracy comes from internal documentation and a conference proceeding. The two implementation studies are single-center and nonrandomized.
Sepsis ImmunoScore
Prenosis says the Sepsis ImmunoScore integrates biomarkers and clinical data to present real-time information and predict patient outcomes. FDA's decision summary describes AI software that identifies patients at risk for having or developing sepsis from 22 parameters: demographics, vital signs, and hematology, chemistry and sepsis biomarker results.
Why this grade. The evidence is accuracy only: in the FDA De Novo validation, a retrospective analysis of prospectively collected biobank data from three U.S. sites, the area under the receiver operating characteristic curve (AUROC) was 0.81 to 0.84 for sepsis within 24 hours.
The evidence
The De Novo validation drew on a research biobank at three sites, including Beth Israel Deaconess Medical Center, with sepsis adjudicated by physicians; observed sepsis rates were 3.02% in the low-risk category and 69.7% in the very-high-risk category. A seven-site observational study of 6,027 encounters of adults with suspected infection reported an AUROC of 0.82 for sepsis, compared with 0.70 for procalcitonin, 0.61 for C-reactive protein and 0.59 to 0.72 for four clinical scores. Eight of its 21 authors list Prenosis as their only affiliation.
- Used by
- Prenosis' website lists 11 hospitals and says the tool is also available through an exclusive partnership with Roche Diagnostics. LifeBridge Health's Sinai Hospital of Baltimore is putting it into clinical use, according to a Fierce Healthcare article reposted by Prenosis.
- FDA
- FDA granted the Sepsis ImmunoScore De Novo authorization as DEN230036 on April 2, 2024. Prenosis calls it the first FDA-authorized AI diagnostic for sepsis.
- Limits
- All accuracy data come from company-led analyses, largely of a research biobank, and no study in this review shows faster treatment or better outcomes after deployment. The score requires blood tests, including sepsis biomarker concentrations, so it can be calculated only after those tests are drawn.
Grades: A, randomized evidence of benefit; B, evidence from clinical use; C, accuracy studies only; D, little or no independent evidence. How the grades work. Reviews of the published evidence, not medical advice or an endorsement of any product.