COMPOSER sepsis alerts are linked to lower mortality

COMPOSER with a nurse-facing alert in Epic. [1] [3]

COMPOSER, a feedforward neural network with conformal prediction. [1] [3]

The deployment report does not give a release or checkpoint identifier. [1] [3]

University of California San Diego. [1] [3]

COMPOSER estimated sepsis risk over the next four hours and reported the measurements that contributed to each alert. [1] [3]

The system assessed risk over the next four hours and issued nurse alerts in Epic. Repeat alerts were suppressed for eight or twelve hours after a response. [1]

COMPOSER analyzed clinical measurements and their timing to estimate sepsis risk and identify the main contributing factors. [1] [3]

Nurses acknowledged the alert and could notify a physician. Clinicians decided on tests and treatment. [1] [3]

Continuous analysis of electronic patient data and alerts delivered to nursing staff. Nurse review, physician assessment, and treatment at UC San Diego Medical Center and Jacobs Medical Center.

Nurses reviewed an alert and chose whether to notify a physician, note that care was already underway, or record that infection was unlikely. Physicians decided on treatment. [1]

COMPOSER analyzed patient data and alerted nurses to possible sepsis. Nurses reviewed patients with physicians, who selected treatment. The outcome concerns AI-assisted clinical care.

COMPOSER analyzed patient data and alerted nurses to possible sepsis. Nurses reviewed patients with physicians, who selected treatment. The outcome concerns AI-assisted clinical care.

AIBD-0579-outcome-01 [1]

Sepsis mortality after deployment was lower than the study’s modeled estimate for care without COMPOSER. [1]

Improved Physical Health

Observed outcomes compared with a modeled counterfactual. [1]

People.

Adults with sepsis in the Hillcrest and La Jolla emergency departments. [1]

The study analyzed 6,217 sepsis encounters. Its estimate of 22 additional survivors comes from the counterfactual model. [1]

The mortality difference was statistically significant at Hillcrest but not at La Jolla. The two hospitals served different patient populations. [1]

The post-deployment evaluation ran from 7 December 2022 through 30 April 2023. Mortality was measured during hospitalization. [1]

Yes

Nurses received about 235 alerts per month across both sites. Unnecessary treatment among patients without sepsis, and its costs or harms, were not evaluated. [1]

UCSD IRB 805726 [1]

A prospective, nonrandomized before-and-after study. A Bayesian time-series model estimated what would have happened without deployment. [1]

Researchers used electronic records to identify Sepsis-3 cases, inpatient deaths, organ dysfunction, and completion of the sepsis care bundle. [1]

The analysis included 6,217 eligible sepsis encounters: 5,065 before deployment and 1,152 afterward. [1]

In-hospital mortality, completion of the sepsis care bundle, and change in organ dysfunction over 72 hours. [1]

A modeled post-deployment counterfactual based on the historical period, with adjustment for measured differences. [1]

Mortality was 9.49%, versus 11.39% in the modeled comparison: a difference of 1.9 percentage points (95% CI, 0.3–3.5). [1]

The estimate concerns COMPOSER, nurse alerts, preparation, and subsequent care together. It does not isolate the prediction model from those changes. [1]

There was no randomized or concurrent untreated group. Unmeasured changes in care could affect the comparison. Results covered five months at one health system. [1]

The historical period ran from 1 January 2021 to 6 December 2022. Deployment began on 7 December 2022. [1]

The supplement describes implementation and statistical methods. De-identified data and analysis code are available only through the authors and institutional approval. [1]

The reviewed sources do not report a monetary valuation or financial return. [1]

The reviewed sources do not report a monetary valuation or financial return. [1]

The reviewed sources do not report a monetary valuation or financial return. [1]

The reviewed sources do not report a monetary valuation or financial return. [1]

The reviewed sources do not report a monetary valuation or financial return. [1]

The reviewed sources do not report a monetary valuation or financial return. [1]

The reviewed sources do not report a monetary valuation or financial return. [1]

The paper reports about six months of preparation, staff training, cloud operation, and monitoring, but no monetary cost or financial return. [1]

The reviewed sources do not report a monetary valuation or financial return. [1]

The reviewed sources do not report a monetary valuation or financial return. [1]

The reviewed sources do not report a monetary valuation or financial return. [1]

UC San Diego researchers developed COMPOSER. Nurses and physicians at UC San Diego Health acted on its alerts. [1]

UC San Diego’s Hillcrest and La Jolla hospitals, California, United States. [1]

The reviewed clinical sources identify the organizations involved, but do not establish their complete ownership or control arrangements. [1]

COMPOSER ran on AWS with electronic-record feeds and an Epic nurse alert. The deployment also required clinical staffing and monitoring of data delivery. [1]

The reviewed sources do not document international transfers of patient data or quantify cross-border financial flows. [1]

Not assessed

The deployment required reliable electronic records, nurse training, physician response, and continuing checks of model performance and data delivery. [1]

The reviewed sources do not report an independent repetition of this clinical result. [1]

The June 2024 correction restored Table 3 and its citation. It did not announce a new trial or change the reported mortality estimate. [2]

The paper describes clinical use at UC San Diego and a startup formed by several authors. It does not establish general public access to the deployed system. [1]

UC San Diego Health introduced a nurse alert for possible sepsis. [1]

The project received IRB approval 805726 and ACQUIRE approval 609, with a waiver of informed consent. [1]

The team trained staff, deployed the alerts, and monitored the system during clinical use. [1]

The evaluation examined mortality, sepsis treatment, and organ dysfunction rather than a single formal pass-or-fail threshold. [1]

3 retained sources.

Original citations and the passages that support each finding. [1] [2] [3]

English. No translation was used. [1]

Developer [1] [2]

The original studies and their supporting materials are linked to the findings that they support. [1] [2] [3]

Access and version details are listed for each reviewed source. [1] [2] [3]

COMPOSER deployment was associated with lower sepsis mortality and more frequent completion of the sepsis care bundle. [1]

The deployment paper supplies clinical outcomes. The 2021 paper describes the model, and the correction restores an omitted table. [1] [2] [3]

An original, peer-reviewed study by the developers and deploying health system, using a historical counterfactual. [1]

The study did not measure harms from false-positive alerts in patients without sepsis, financial returns, or an independently replicated mortality effect. [1]

The benefit described here, for the beneficiary groups shown.

Depth 5 for the reported survival benefit.

Depth 5 for the reported survival benefit.

The depth describes the reported result. Evidence strength is assessed separately.

AIBD’s five-level depth rubric, reviewed on 25 October 2026.

AIBD, 25 October 2026.

The grade describes the reported benefit; it does not convert the historical comparison into a causal trial.

A peer-reviewed study by the developers and deploying health system. The outline is retained.

A prospective comparison before and after clinical deployment. [1]

Not assessed

Healthcare - sepsis. Electronic-health-record early warning system.

Identify emerging sepsis so that clinical teams can assess and treat patients sooner.

Deployment began on 7 December 2022. Follow-up ended on 30 April 2023. [1]

Routine use at two UC San Diego emergency departments from 7 December 2022. [1]

SHA-256: d1042baa617a…

UCSD IRB 805726 and ACQUIRE project 609. [1]

Not assessed

Not assessed

3 revisions. Latest: 25 October 2026. [1] [2]