Epic Sepsis Model

A learned model that is a feature inside an electronic health record. The first version (2018) was a penalized logistic regression. Version 2, released in 2022 and referred to in the literature as the Early Detection of Sepsis model, is a gradient-boosted tree model that each health system fine-tunes on its own historical adult inpatient data before go-live, scoring every emergency and hospitalized patient every 15 minutes from demographics, vital signs, laboratory results, medication administration and other chart data, as the independent validations by Wong and colleagues and by Dutta and colleagues (both JAMA Network Open, 2026) describe it.

The model is learned and the alert does not exist without it, but it is one function inside Epic's record system, available only to Epic customers, and no organization buys or keeps Epic because of it. A learned component inside a product whose value stands without it. For a buyer the practical consequence is that the incumbent model is already present, while a competing sepsis product has to be procured, given a data feed and justified against it.

The model advises and a clinician decides. Scores run every 15 minutes, a score above the threshold the health system sets raises an alert, and nothing is ordered or started without a person. Thresholds are local: in the four system validation of version 2 (Wong and colleagues, JAMA Network Open, 2026) the sites set them at 14, 29 to 30, 37 and 35 on a 0 to 100 scale to reach 60 percent sensitivity, and on Epic's recommendation scores in the first hour were suppressed for patients without a complete blood count, which removed 9.8 to 30.1 percent of sepsis cases from the analysis.

What is missing is what happens when the alert is wrong. Positive predictive value at those thresholds ran from 13 to 26 percent, alerts clustered in the hours after admission, and it was the investigators, not Epic, who modeled an eight hour silencing rule to make the volume workable; Epic publishes no expected alert burden, lead time or silencing guidance.

In an emergency department, Currey and Tarabichi (American Journal of Emergency Medicine, 2025) found version 2 alerted before any sign of clinician recognition (an antibiotic, culture or lactate order) in only 33.5 percent of cases at a 60 percent sensitivity threshold. The advisory structure and local threshold are described, and the handling of false alerts is left to each hospital.

The approach and version are named, though by researchers rather than by Epic. Version 1 (2018) was a penalized logistic regression. Version 2 (2022) is a gradient-boosted tree ensemble fine-tuned on each site's historical adult inpatient data, drawing on demographics, vital signs, laboratory results, medication administration and other clinical features and producing a 0 to 100 score every 15 minutes (Wong and colleagues, and Dutta and colleagues, JAMA Network Open, 2026; neither paper has Epic authors).

Epic's own internal validation of version 2, an area under the curve of 0.83 to 0.86 at three sites, is public only because the Wong paper cites a meeting with Epic staff. Epic publishes no model card on its public site: the feature list and weights, the base training population before local fine-tuning, and the retraining and update practice are not public, and model documentation sits behind the customer portal. More than 100 organizations were still running version 1 in August 2025, so the version live at any one hospital cannot be assumed.

A structurally short chain. The model is Epic's own and runs inside the health system's Epic environment, fine-tuned on that system's own historical data; no outside model provider is involved. Epic's Data Privacy Framework policy describes Epic as a data processor or service provider to its healthcare customers and says Epic staff access data in customer systems only with the customer's permission. That removes the need for a continuous outbound data feed.

Nothing public says where local fine tuning runs (in the customer's own environment or on Epic infrastructure), whether customer data or model outputs contribute to the base model Epic ships to other customers, or which Epic subsidiaries, Epic Hosting among them, handle data for hosted customers.

The evidence is independent, prospective, multi site and published in full, and it describes a model that discriminates better than its predecessor and still alerts mostly for patients who do not have sepsis. Version 2: Wong and colleagues (JAMA Network Open, February 2026; University of Michigan, Oregon Health & Science University, Emory and MetroHealth, no Epic authors) followed 227,091 adult inpatient encounters after go-live and reported encounter level area under the curve of 0.82 to 0.92 by site, 0.75 to 0.85 at a 12 hour horizon, positive predictive value of 0.13 to 0.26 at 60 percent sensitivity, and median lead time of 1.9 to 10.3 hours.

Dutta and colleagues (JAMA Network Open, 2026) ran version 2 silently across 9 Mass General Brigham hospitals on 198,494 encounters: area under the curve 0.89 against Sepsis-3, with positive predictive value of 11.4 percent and median lead time of 3.4 hours at the Youden threshold, and results that shifted with the sepsis definition used.

Currey and Tarabichi (American Journal of Emergency Medicine, 2025) found 0.90 in an emergency department, falling to 0.85 when only scores before clinicians acted were counted. An observational study at one Minnesota health system (Applied Clinical Informatics, 2026) associated go-live with lower inpatient mortality than a counterfactual model predicted (1.21 against 1.94 percent).

Version 1: Wong and colleagues (JAMA Internal Medicine, 2021) reported 0.63 across 38,455 hospitalizations, with alerts on 18 percent of patients while 1,709 of 2,552 sepsis cases were missed; a study of 145,885 encounters in two county emergency departments (JAMIA Open, 2024) found 14.7 percent sensitivity, 7.6 percent positive predictive value and a median lead time of zero; and a 2026 meta-analysis (Journal of General Internal Medicine) pooled 0.65 and found Epic's reported ranges consistently higher than external results. The study designs are strong, and their findings include low precision and wide variation between hospitals.

The architecture is favorable, and the commitments that would answer the questions artificial intelligence raises are not published. The model runs inside the health system's Epic environment and each site fine-tunes version 2 on its own historical inpatient data, as the independent validations describe the deployment, so no third party model provider receives patient data.

Epic's public privacy documents cover its websites and mobile apps, and its Data Privacy Framework policy says Epic acts as processor for its healthcare customers and that staff access customer data only with permission. Nothing public states how long scores and fine-tuning data are kept, whether site data or outcomes feed the base model Epic ships to other customers, or where fine-tuning runs. No guardrail testing or safety event process specific to the model is published beyond Epic's general vulnerability reporting route. General assurances only.

A privacy document reaches the service, with no business associate statement. Epic's Data Privacy Framework policy (effective 13 March 2026) describes Epic as a data processor or service provider to healthcare customers, says Epic staff access personal data in customer systems only with the customer's permission, and names Epic Hosting among the covered subsidiaries; Epic's other public privacy policies cover its websites and mobile apps.

No public statement of business associate status or agreement terms exists; those sit in each customer's Epic contract. The model adds no new vendor or subprocessor, so the existing Epic agreement is what governs it. A substantive privacy document that reaches the service, without a stated business associate position.

Controls are asserted with nothing independent published. Epic's public site offers a route for reporting security vulnerabilities and says a strong information security management system guides its policies, but publishes no attestation, certification, trust center or penetration testing cadence; compliance documents reach customers through the customer portal.

For self hosted customers much of the security burden sits with the health system, since the model runs in its environment; for customers on Epic Hosting, Epic operates the environment. Before contact, nothing beyond Epic's own statements can be checked.

No clearance, and no regulatory position published. Epic has not said publicly how it classifies the model. When FDA's 2022 clinical decision support guidance gave software that analyzes patient data to detect sepsis and alert a clinician as an example of a device function, Epic declined to comment (MedTech Dive, October 2022), and FDA's January 2026 revision still withholds its new enforcement discretion from software that predicts the risk of a time-critical event.

The model is not described on Epic's public site, and its basis is a learned score: nothing public says what a clinician can see about why a patient scored high. Researchers have framed its evaluation in device terms (the Minnesota outcome study describes its results as supporting safety and effectiveness under software as a medical device principles), but that is their framing, not a status, and the tension between the two stays unresolved.

An independent fairness audit of the current version is published with its figures by group. Wong and colleagues (JAMA Network Open, 2026; investigators from the University of Michigan, Oregon Health & Science University, Emory and MetroHealth, with no Epic authors) reported version 2's area under the curve by age, sex, race and ethnicity at each of the four health systems, in the paper's supplement.

The consistent weak spot is age: patients 65 and over, who carry most of the sepsis, had the lowest discrimination at every site (0.74 at Michigan, 0.77 at Oregon Health & Science University, 0.85 at Emory, 0.87 at MetroHealth), against 0.87 to 0.91 for patients aged 18 to 29. By race and ethnicity the paper found no pattern beyond what baseline sepsis rates explain: Black patients ranged from 0.84 to 0.93 across sites and White patients from 0.80 to 0.91.

Men scored lower than women at Michigan (0.79 against 0.84). Differences between institutions were large, which is why the authors recommend local validation. Epic's public pages do not cite the audit and Epic publishes no subgroup results of its own; what it does publish is seismometer, an open source toolkit for measuring a model's performance and fairness on a health system's own data. An independent audit with results compared across groups.

Nothing stands behind the output, and there are ways to check it. Epic publishes no error rate, warranty or remediation commitment for the model; its internal validation figures for version 2 (area under the curve 0.83 to 0.86 at three sites) became public only through an independent paper citing a meeting with Epic staff.

What exists is a clinician between the score and any action, since the alert advises and a person decides, and a route to test the model: Epic publishes seismometer, an open source toolkit health systems can use to measure a model's performance and fairness on their own patients. The error rates themselves come from independent studies, and Epic has not adopted or published them. That leaves review before action and a local audit route, with no commitment from Epic.

Native to the record and in production at named systems. The model reads vital signs, laboratory results, medication administration and other chart data directly from Epic every 15 minutes and returns its score and alerts into the clinician's workflow, with no interface or data feed to build.

Published production deployments include the University of Michigan, Oregon Health & Science University, Emory, MetroHealth and Mass General Brigham, and Epic reported 95 organizations (731 hospitals) on version 2 in August 2025, as the Wong paper records. A, with the reverse side stated plainly: the model runs only inside Epic and cannot be bought, moved or evaluated separately.

The model follows the customer's own Epic installation. Scores are computed inside the health system's Epic environment and each site fine-tunes version 2 on its own historical data before go-live, so inference runs wherever the customer's Epic runs: on its own infrastructure or with Epic Hosting, a subsidiary named in Epic's Data Privacy Framework policy. No external model call is involved.

The deployment options and where data rests follow the customer's installation, and the fine tuning path is left open, since nothing public says whether local training runs in the customer's environment or on Epic infrastructure.

No price, and the buying route can be established. The model is available only to Epic customers, inside Epic, and is switched on by the health system; the Wong paper records Epic's figure of 95 organizations on version 2 and more than 100 still on version 1 in August 2025. Whether it carries a charge beyond the existing Epic license is not published.

The known costs fall on the health system: fine-tuning before go-live, threshold and silencing configuration, and the local validation every independent study recommends. How it is bought is discoverable, the price is not.

Where the model has been validated is named and spans settings. Version 2 has been evaluated prospectively in adult inpatients at four academic and safety net systems (Michigan, Oregon Health & Science University, Emory, MetroHealth), silently across emergency departments, inpatient units, intensive care units and perioperative areas at 9 Mass General Brigham hospitals, in one emergency department, and in an outcome study at one Minnesota system; version 1 has published results from an academic center and two county emergency departments. Coverage is adults only, with no pediatric model. Performance varies widely by site, so the validated settings show where the model has been tested, not how it will perform at a new hospital.