Multicenter prospective validation of Epic's retrained sepsis model (version 2)
Epic's retrained sepsis model, gradient-boosted trees fine-tuned on each hospital's own data, discriminated sepsis at encounter-level AUROCs of 0.82 to 0.92 across 227,091 inpatient encounters at four large US health systems, in line with or above the 0.83 to 0.86 Epic reports internally, but with positive predictive values of only 0.13 to 0.26 and a high alert burden.
- Lab
- Epic Systems
- Model
- Epic Sepsis Model v2 (gradient-boosted trees with local fine-tuning)
- Field
- Medicine
- Date
- 2026-02-02
What was found
Version 2, released in 2022, replaced the penalized logistic regression of version 1, addressed the methodological criticisms raised against it, and lets each site fine-tune on its own historical data. Investigators at four health systems collected data for a consecutive period right after each site went live, between August 2023 and March 2025, and scored the model against Sepsis-3 criteria; 7,401 of the 227,091 encounters (3.3%) met them. At the prediction level, with a 12-hour horizon, AUROC ranged from 0.75 to 0.85; measured against clinicians' own recognition of sepsis (antibiotic, lactate and culture orders), encounter-level AUROC was 0.80 to 0.90. According to Epic, as cited in the paper, 95 organisations representing 731 hospitals used version 2 as of August 2025, while more than 100 organisations still ran version 1.
Novelty check
Europe PMC, "Epic Sepsis Model": the authors and this search found no earlier multicenter external validation of version 2. Epic's internal figures of 0.83 to 0.86 are cited in the paper from a meeting with Epic staff, not from a publication.
Caveats and known objections
Discrimination is not benefit: the study measured no patient outcome such as mortality or time to antibiotics. A positive predictive value of 0.13 to 0.26 means roughly four to eight alerts for every true sepsis case. The spread across sites reflects local fine-tuning, so the result describes the model as each site configured it, and a hospital adopting it inherits the need to validate locally. Measuring a live model means clinicians' responses to its alerts can change the outcomes it predicts. The investigators are unaffiliated with Epic, which is why this is graded independent, but several of them also wrote the version 1 validations, so this is one research community's check repeated rather than a fully separate one.
Independent checks
Wong et al., investigators at four US health systems (JAMA Network Open, 2026): Encounter-level AUROC 0.82 to 0.92 across four health systems, at or above Epic's internally reported 0.83 to 0.86, with positive predictive values of 0.13 to 0.26 and high institutional variability. · link ↗
Disagree with these grades?
Bring a citation: a grade moves on evidence, not on argument.
Or on GitHub: submit a check challenge the grade send a correction or send a pull request
Flag this for triage
Signals order the review queue and nothing else. They are never published, and they never move a grade: that takes a citation.
Entry history (1 event)
- AddedEntered the registry graded Independently checked and AI-assisted.
Entries are never deleted. A grade that does not hold up is downgraded on the record, with the reason beside it.
Graded independently checked for verification and ai-assisted for autonomy. What these mean.
Cite this entry
whataifound.org. (2026). Multicenter prospective validation of Epic's retrained sepsis model (version 2). whataifound.org: A Registry of AI Scientific and Mathematical Discoveries. https://whataifound.org/finding/2026-02-02-epic-sepsis-model-v2BibTeX
@misc{whataifound-epicsystems-2026-v2,
title = {Multicenter prospective validation of Epic's retrained sepsis model (version 2)},
author = {{whataifound.org}},
year = {2026},
howpublished = {whataifound.org: A Registry of AI Scientific and Mathematical Discoveries},
note = {Result by Epic Systems. Verification: Independently checked. Autonomy: AI-assisted.},
url = {https://whataifound.org/finding/2026-02-02-epic-sepsis-model-v2}
}