BUSI 505 Week 3 Case Study Discussion Board Forum 2 Example

Case Study Discussion Board Forum 2: Clinical Decision Support and Alert Fatigue

BUSI 505 Healthcare Informatics, Case Study Discussion Board Forum 2

Thread by [Student Name]

The hospital, its alert data and the classmates are composites; the textbooks, Scripture and the peer-reviewed sources cited are real.

Our case hospital turned on its electronic record vendor's sepsis prediction alert two years ago. The alert fires when a patient's score crosses a threshold, interrupting the nurse and prompting a sepsis screen. Last quarter it fired more than 600 times a week across 250 beds, and nurses dismissed about 95% of the alerts within seconds. When the quality department reviewed sepsis cases, it found that several patients who developed septic shock had received alerts that were overridden without a screen, and others developed sepsis without ever triggering an alert. Staff described the alert as noise and admitted they clicked through it without reading. Leaders were left asking whether the problem was the clinicians, the model or the way the alert was built into the workflow. The first instinct of some managers was to require nurses to document a reason for every override, which would have added work without asking whether the alerts deserved attention in the first place.

Research suggests the model deserves scrutiny first. Wong et al. (2021) evaluated a widely implemented proprietary sepsis model across more than 38,000 hospitalizations at an academic health system and found discrimination barely better than a coin toss, with an area under the curve near 0.6, well short of what the developer had reported. Nearly one inpatient in five triggered the alert, yet two thirds of those who went on to develop sepsis were never flagged, and among septic patients whose antibiotics came late under usual care the model caught only a small handful. In plain terms, it alerted often, missed often and rarely found patients clinicians had not already recognized. Our hospital had never tested the model on its own patients before turning it on. A model that performs this way will produce alert fatigue no matter how well trained the staff are.

For Nelson and Staggers (2014), decision support works when the person who can act receives the needed information, in a usable format, at the moment of choice, and Braunstein (2014) emphasizes that decision support succeeds only when it fits clinical workflow. Our alert failed several of those tests: it interrupted nurses for patients who were not septic, it gave no specific recommendation and it fired regardless of what the care team was already doing. It also went to the wrong person at times, interrupting a nurse when the decision to order cultures or antibiotics belonged to a physician or advanced practice provider who never saw the alert.

Redesign can help. Fallon et al. (2024) describe a hospital whose interruptive COVID-19 precautions alert fired about 8,200 times a week. A governance team adjusted triggers and replaced the pop-up with a passive, rule-based order panel, cutting weekly alerts by about 80% to roughly 1,450. At the same time, the share of symptomatic patients with appropriate precautions ordered rose from 23% to 61%, and orders were placed faster. Fewer alerts produced better care.

For our case, I would recommend four steps. First, validate the sepsis model on our own patients before relying on it, and adjust or replace it if performance is poor. Second, assign every interruptive alert an owner and a review date. Third, convert low-value interruptive alerts to passive displays where possible. Fourth, measure each alert's firing rate, override rate and effect on the intended action, and retire alerts that do not improve care. An alert that fires constantly teaches clinicians one lesson: that alerts can be ignored.

Isaiah promises that the people will hear a word behind them saying, "This is the way, walk in it," when they turn to the right or to the left (English Standard Version Bible, 2001/2016, Isaiah 30:21). Good decision support aims at that kind of guidance: a clear word at the moment of decision, not a constant noise that drowns out every warning.

Reply to Hannah

Hannah, you asked how a hospital would know whether its clinicians are experiencing alert fatigue rather than simply making reasonable choices to override alerts. That is a sharper question than it first appears, because a high override rate can mean either that clinicians are ignoring important warnings or that the alerts are wrong. Ray et al. (2026) reviewed 22 systematic reviews on clinical decision support and found that only one offered an operational definition of alert fatigue, even though the term appears constantly in hospital discussions. The most common measures were the number of alerts, the override rate and the acceptance rate, but studies defined and reported them inconsistently, which makes comparisons difficult. Newton et al. (2026) interviewed 20 junior doctors working in Australian hospitals and, using a human factors lens, found that alert fatigue appeared at different stages of information processing: some alerts were never noticed, some were handled with mental shortcuts and some demanded so much effort to interpret that doctors gave up. That finding suggests that a single override rate cannot capture the problem. An alert that is never noticed has a low override rate if overrides require a click, yet it is useless; an alert that is noticed and thoughtfully dismissed may have a high override rate and still be working as intended for the few cases that matter. Nelson and Staggers (2014) recommend evaluating decision support from the user's perspective as well as the system's, and Braunstein (2014) notes that usability is part of whether clinical systems deliver value. Given those limits, a practical measurement approach would combine three views: system data on firing and override rates for each alert, chart review of a sample of overrides to judge whether they were appropriate and brief surveys or interviews with clinicians about which alerts they find useful. Each view catches something the others miss, and together they show whether a problem lies in the alert's logic, its timing or its audience. Proverbs warns that when words are many, transgression is not lacking, but the one who restrains his lips is prudent (English Standard Version Bible, 2001/2016, Proverbs 10:19), and an alert system that says too much loses the power to warn when it matters. Would your organization be willing to review a random sample of overridden alerts each quarter? That review is labor intensive, but it is the only way to tell whether overrides reflect fatigue or good judgment, and it often reveals which alerts should be retired first. A small team of a pharmacist, a nurse and a physician could review fifty overrides in an afternoon and bring recommendations to the decision support committee. Over a year, those reviews would build a local evidence base for which alerts earn their interruptions and which do not.

Reply to Darnell

Darnell, you raised the question of whether nurses will trust alerts generated by artificial intelligence, especially after experiences with tools like our sepsis alert. Trust is central, because a decision support tool helps only when clinicians act on it appropriately, neither ignoring it nor following it blindly. Both failure modes are dangerous: ignoring a correct alert delays care, while following a wrong one can lead to unnecessary tests, antibiotics or anxiety for patients. Clej et al. (2026) surveyed 239 registered nurses at a large hospital about AI-supported alerts and found three groups: about 27% tended to over-trust AI alerts, 41% showed balanced trust and 32% resisted them. Alert fatigue was the factor most strongly associated with nurses' intention to override AI alerts, partly through moral distress, and the effect was weaker where nurses reported higher team psychological safety. In other words, nurses who felt safe speaking up about a questionable alert were less likely to slide into blanket dismissal, which points to culture as well as technology as a lever leaders control. Wong et al. (2021) show why some resistance is rational, since a widely used sepsis model performed far below expectations when tested independently. Nurses who learned from experience that the sepsis alert was usually wrong were, in a sense, accurately calibrating their trust to a poor tool, and the problem lay with the organization that deployed it without local testing. Nurses' experience with alerts belongs in decision support design and evaluation, consistent with the interprofessional view of informatics in Nelson and Staggers (2014), since so many interruptive pop-ups land on nurses' screens. Braunstein (2014) adds that clinicians must understand the basis of a recommendation to use it well, which is difficult when a proprietary model's logic is hidden. For managers responsible for adopting these tools, the findings suggest three priorities: publish local validation results so clinicians know how well each model performs, involve nurses in deciding which alerts are interruptive and build a unit culture in which questioning an alert is safe and welcomed, with a simple way to report alerts that seem wrong. John urged believers not to believe every spirit but to test the spirits to see whether they are from God (English Standard Version Bible, 2001/2016, 1 John 4:1), and the same discernment applies to machine recommendations: test them, and trust them only as far as the evidence allows. Would your unit consider sharing each AI model's validation results with staff in plain language, perhaps on a one-page summary posted with each new tool? Knowing that a model is right only half the time, or nine times in ten, would help nurses calibrate their trust rather than swinging between blind acceptance and reflexive dismissal. It would also signal that the organization takes their judgment seriously.

References

Braunstein, M. L. (2014). Contemporary health informatics. AHIMA Press.

Clej, E., Fizedean, C., Gherman, A., Ilie, A. C., Bratu, M. L., & Marc, F. (2026). Why nurses intend to override AI alerts: How alert fatigue, moral distress, and team psychological safety shape self-reported trust calibration toward clinical decision support. Healthcare, 14(14), 2063. https://doi.org/10.3390/healthcare14142063

English Standard Version Bible. (2016). Crossway. (Original work published 2001)

Fallon, A., Haralambides, K., Mazzillo, J., & Gleber, C. (2024). Addressing alert fatigue by replacing a burdensome interruptive alert with passive clinical decision support. Applied Clinical Informatics, 15(1), 101-110. https://doi.org/10.1055/a-2226-8144

Nelson, R., & Staggers, N. (2014). Health informatics: An interprofessional approach. Elsevier Mosby.

Newton, N., Bamgboje-Ayodele, A., Forsyth, R., Tariq, A., Huang, J., Yannam, R., Lalor, D. J., Sobey, A. J., & Baysari, M. T. (2026). Experiences of alert fatigue and its contributing factors in hospitals: Qualitative study. Journal of Medical Internet Research, 28, e78676. https://doi.org/10.2196/78676

Ray, C. E., Wilson, G. M., Hughes, A. M., Cunningham Goedken, C., Liu, E. P., Fitzpatrick, M. A., Suda, K. J., Kota, S. M., Nwankpa, C., & Evans, C. T. (2026). Alert fatigue measurement in clinical decision support: A systematic review. Journal of the American Medical Informatics Association, 33(8), 1523-1531. https://doi.org/10.1093/jamia/ocag064

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