NR 500 Week 6 Sepsis Prediction Model Brief Example

Technology in Practice Brief: The Vendor Sepsis Prediction Model in Our Record, What the Company Shares, What Independent Studies Found, and the Question an Informatics Nurse Asks

Student Name

College of Nursing, Chamberlain University

NR 500: Foundational Concepts and Applications

Instructor Name

Month Day, Year

Technology in Practice Brief: The Vendor Sepsis Prediction Model in Our Record, What the Company Shares, What Independent Studies Found, and the Question an Informatics Nurse Asks

One System, Named

This brief is about one system: the sepsis prediction model supplied by our electronic health record vendor, Epic Systems, which our hospital uses to generate the sepsis alert discussed in Week 4. In ordinary words, the model is a program that runs in the background of the record, looks every 15 minutes at a patient's vital signs, laboratory results, medications and other data, and produces a score that estimates how likely the patient is to be developing sepsis. When the score crosses a threshold chosen by the hospital, an alert appears for the nurse and, depending on configuration, the provider. The nurse is then expected to assess the patient and follow the sepsis screening protocol.

The vendor has produced more than one version of the model. The first, often called the Epic Sepsis Model, was used widely for several years; a second version, trained on each hospital's own data, has been introduced more recently. Which version our hospital uses matters, because the published evidence differs between them.

The Company as a Source

The vendor is a source, and an important one, but its material has two limits. First, the model is proprietary: its detailed design and the vendor's own validation results are provided to customer hospitals under license rather than published, so I could not read them and I have not tried to summarize them from secondhand descriptions. Second, a vendor has an interest in how its product is judged. That does not make its claims wrong, but it means they need independent confirmation. What I can say from our own hospital's materials is how the alert has been configured here: the threshold, who receives the alert and what response is expected. Those local settings are documented by our informatics team and are the part of the vendor's system I can describe accurately.

What Independent Studies Found

Independent evidence on the first version is substantial and mostly unfavorable. Wong et al. (2021), studying 38,455 hospitalizations at the University of Michigan, reported discrimination of only 0.63 on the usual curve-based measure, that it failed to identify 67% of patients who developed sepsis, and that at the commonly used alert threshold it fired for 18% of all hospitalized patients. Lyons et al. (2023) examined the same proprietary model across nine networked hospitals and reported that its performance varied between hospitals, and studied the factors associated with that variation. Together, these studies suggest that the first version's performance was weaker than its wide adoption implied, and that a result at one hospital cannot be assumed at another.

Evidence on the second version is newer and more mixed. Dutta et al. (2026) evaluated the locally trained version 2 in nine hospitals, run silently so that clinicians did not see its alerts, and found discrimination that varied with the definition of sepsis used: an area under the curve of 0.89 against the Sepsis-3 definition, 0.94 against the federal SEP-1 definition and 0.85 against another surveillance definition. At the threshold that best balanced sensitivity and specificity for Sepsis-3, the positive predictive value was 11.4%, meaning roughly nine of every ten alerts would be for patients who did not develop sepsis, and the median warning before onset was 3.4 hours. The authors concluded that the model showed moderate performance and that a high false-positive rate could limit its usefulness without careful threshold selection.

What These Findings Mean at the Bedside

For a nurse on a medical unit, the numbers translate into something practical. If most alerts are for patients who do not have sepsis, nurses will learn, reasonably, to treat alerts as low priority, and the rare true alert may be missed in the noise. If the model misses many patients who do develop sepsis, a nurse who waits for the alert before worrying will be late. The alert can be useful as one more reason to look at a patient, but it cannot replace the nurse's own assessment of vital signs, mental status and the patient's appearance. That is the message our unit's teaching should carry, and at present it does not.

The findings also shape how the alert should be described to new nurses during orientation, which is where most of them first meet it. A short, accurate explanation, that the alert is a prompt to look again and not a diagnosis, is more useful than a claim that it finds sepsis early.

Where Nothing Has Been Published

Three things about this system in our hospital have not been published anywhere I could find. The first is its performance here: our sensitivity, false alarm rate and lead time, in our patients, under our threshold. The second is how nurses on our units respond to the alert: how often they assess the patient within an hour, how often they dismiss it, and whether response differs by shift. The third is whether the alert has changed any patient outcome in our hospital. None of these can be borrowed from the published studies, because those studies show that performance varies from one hospital to another and with the definition of sepsis used.

The Question an Informatics Nurse Asks

The question my specialty brings to this system is not whether the model is good in general. It is whether this configuration, in this hospital, helps nurses find sepsis earlier without burying them in false alarms, and how we would know. Answering that requires three things an informatics nurse is positioned to arrange: a clear local definition of sepsis for evaluation, data from the record on alerts and outcomes, and a look at nursing workflow when the alert fires. It also requires honesty in the teaching materials, replacing the intranet claim that the alert "identifies patients early" with a description of what the alert is for and what it cannot do.

I have asked our informatics manager which version of the model we use and whether a local evaluation has been done. If the answer to the second question is no, a silent evaluation of the kind Dutta et al. (2026) describe, in which the model runs without alerting and its scores are compared with actual outcomes, would be a reasonable first project before any change to the threshold.

References

Dutta, S., McMurry, R., Tasi, M. C., Dunham, L., McEvoy, D. S., Stump, T., Filbin, M., & Rhee, C. (2026). Performance of a sepsis prediction model across different sepsis definitions. JAMA Network Open, 9(4), Article e265599. https://doi.org/10.1001/jamanetworkopen.2026.5599

Lyons, P. G., Hofford, M. R., Yu, S. C., Michelson, A. P., Payne, P. R. O., Hough, C. L., & Singh, K. (2023). Factors associated with variability in the performance of a proprietary sepsis prediction model across 9 networked hospitals in the US. JAMA Internal Medicine, 183(6), 611. https://doi.org/10.1001/jamainternmed.2022.7182

Wong, A., Otles, E., Donnelly, J. P., Krumm, A., McCullough, J., DeTroyer-Cooley, O., Pestrue, J., Phillips, M., Konye, J., Penoza, C., Ghous, M., & Singh, K. (2021). External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Internal Medicine, 181(8), 1065-1070. https://doi.org/10.1001/jamainternmed.2021.2626