TOPLINE:

A locally trained model predicted sepsis early among hospitalized patients, with moderate accuracy, high false-positive rates, and modest lead times that varied with the definition of sepsis used, limiting its potential clinical utility.

METHODOLOGY:

• Researchers conducted a diagnostic study across nine acute care hospitals over 6 months in 2024, including 198,494 adult encounters from 2023 (median age, 55 years; 54.8% women; 69.6% White individuals) across emergency departments, inpatient units, ICUs, or perioperative settings.
• The model was locally trained using the Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3); it incorporated patient demographics, vitals, laboratory results, and medications, and generated predictions every 15 minutes.
• The primary outcome was sepsis model performance, evaluated against three electronically computable sepsis definitions: Sepsis-3, the Centers for Medicare & Medicaid Services Severe Sepsis and Septic Shock: Management Bundle (SEP-1), and the Centers for Disease Control and Prevention Adult Sepsis Event (ASE). Performance metrics included the area under the receiver operating characteristic curve (AUROC), the area under the precision-recall curve (AUPRC), and lead-time analysis.
• Data on patient race and ethnicity were self-reported and extracted from electronic health records for stratified analysis.

TAKEAWAY:

• The incidence of sepsis outcomes in the overall study population was 2.9%, 1.2%, and 2.0% using the Sepsis-3, SEP-1, and ASE definitions, respectively. The model showed the highest discrimination for SEP-1, followed by Sepsis-3 and ASE. Precision was highest under the Sepsis-3 definition, followed by SEP-1 and ASE. AUPRC varied with incidence, whereas AUROC was independent of incidence.
• For Sepsis-3, the model yielded a positive predictive value (PPV) of 11.4%, recall of 82.4%, a false-positive rate of 19.3%, and a median lead time of 3.4 hours, whereas SEP-1 had a lower PPV of 6.8% but a longer median lead time of 4.5 hours, indicating earlier detection with reduced precision.
• Model discrimination decreased with increasing patient acuity, with Sepsis-3 outcomes at 8 hours showing AUROCs of 0.90 in the ED, 0.82 in inpatient wards, and 0.76 in the ICU.
• No significant degradation in model performance occurred for any sepsis definition when evaluated using only predictions made prior to clinical recognition of sepsis.

IN PRACTICE:

"These findings suggest that the locally trained early detection of sepsis model may provide moderate predictive accuracy and early warning for sepsis, but high false-positive rates and variability across definitions underscore the need for careful calibration and tailored implementation," the authors wrote.

SOURCE:

The study was led by Sayon Dutta, MD, MPH, Massachusetts General Hospital in Boston. It was published online on April 7, 2026, in JAMA Network Open.

LIMITATIONS:

The study was conducted within a single health care system in the US, limiting generalizability to settings with different patient populations or disease burdens. It evaluated only one sepsis prediction model without comparison with alternative models and used the existing ASE outcome definition rather than newer versions currently under consideration. Additionally, the short duration of the study may not have captured long-term performance drift or seasonal variation in sepsis incidence.

DISCLOSURES:

The study was partly funded by the Agency for Healthcare Research and Quality. One author reported receiving personal fees from UpToDate and DynaMed as well as grants from the Centers for Disease Control and Prevention outside the submitted work.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.