Alert Fatigue in Automated Clinical Decision Support

emerging Updated: 2026-06-13
decision-support clinical-decision-support alert-fatigue cry-wolf false-positives positive-predictive-value trust-calibration automation failure-modes healthcare-ai

Alert Fatigue in Automated Clinical Decision Support

Status: emerging
Last updated: 2026-06-13
Sources: Piis2213260018300729.Pdf
Tags: [decision-support, clinical-decision-support, alert-fatigue, cry-wolf, false-positives, positive-predictive-value, trust-calibration, automation, failure-modes, healthcare-ai]

Summary

An automated decision-support system is only useful if the people it advises keep trusting and acting on its output. This editorial in The Lancet Respiratory Medicine (2018) describes what happens when they stop: automated sepsis alerts built into UK primary-care record systems were triggered so often, on such low thresholds, that clinicians became desensitised, began ignoring or switching them off, and real cases of sepsis risked being missed. The piece is a concrete, recent case of the cry-wolf / false-alarm problem in a deployed algorithmic decision-support tool, and of the gap between an alerting algorithm's intent and its behaviour in the field. It belongs in this knowledge base not for its clinical subject but as a real-world failure mode of AI-assisted decision support — over-triggering, poor positive predictive value, and the collapse of calibrated trust — directly paralleling the automation critiques in Ironies Of Automation.

Body

Context

The Lancet Respiratory Medicine (2018) is a one-page editorial prompted by a January 2018 admission from NHS England that automated sepsis alerts in primary-care systems were being over-triggered. These alerts are a series of escalating prompts fired when a clinician records potential sepsis symptoms in a patient's electronic health record, designed and rolled out across UK primary-care informatics to support earlier sepsis detection. Within this knowledge base the article is held as a deployment case study of decision-support failure: where Ironies Of Automation argues from control engineering that automation reshapes rather than removes the human's burden, this editorial shows an AI-assisted alerting system whose poor calibration shifts clinicians from over-reliance toward disuse. (The source is a clinical editorial; it is filed here for that transferable lesson, and its fit is flagged as borderline in the routing note — see Alert Fatigue Clinical Decision Support's META.)

Key Points

The alerts were over-triggered on thresholds too low to be useful. Clinicians reported alerts firing on weak or irrelevant signals — for example when a patient's temperature was merely measured, with no fever recorded. Because sepsis is hard to diagnose early and its early signs resemble common infections, while only a very small proportion of patients presenting with infection symptoms actually have sepsis, a sensitive alerting rule produces a large volume of false positives in primary care.

The consequence is alert fatigue, a trust-and-reliance failure. Frequent, low-value alarms desensitise clinicians, who then ignore the alerts or switch them off — so a system meant to catch missed sepsis can instead let real cases slip through. This is the cry-wolf / false-alarm effect operating on an algorithmic decision-support tool: unreliable warnings erode the warning system's credibility and drive its disuse, the same mechanism that the automation-trust literature describes for human–automation reliance.

The cause is ambiguous, and the evidence base is thin. The editorial notes it is unclear whether over-triggering stems from how the algorithm was implemented or from the algorithm itself, with NHS England committing to work with software providers to update the alert algorithms, and NICE stressing that its recommendations are guidelines requiring clinical judgement rather than automatic execution. There is little published evidence on the efficacy of sepsis alerts in primary care; a 2015 systematic review in hospital settings found that automated electronic sepsis alerts might improve care processes but tend to have poor positive predictive value and do not reduce mortality or length of stay, with minimal data on the harms of false positives.

A usable alert must balance false against valid alarms. The editorial closes on the design requirement: alerting systems should combine high specificity and sensitivity, present clear information, support safe and efficient handling of alerts, and avoid needlessly disrupting workflow — striking the correct balance between false and valid alarms so that alerts retain their intended urgency and attention-grabbing power. This is a trust-calibration problem: too many false alarms and the signal is discounted; too few and genuine cases are missed.

Conclusion

The Lancet Respiratory Medicine (2018) concludes that the value of an automated alerting system depends on preserving the correct balance between false and valid alarms; an alert that fires too readily destroys the very urgency and trust that justify it. For this knowledge base the editorial is a compact, real-world demonstration that the central challenges of AI-assisted decision support are not only algorithmic accuracy but calibrated trust and deployment design: an algorithm with poor positive predictive value, however well intentioned, fails operationally by training its users to ignore it. The lesson generalises beyond healthcare to any domain deploying automated alerts or decision aids, and it grounds the abstract automation-disuse and cry-wolf concepts elsewhere in the corpus in a documented case.

References

The Lancet Respiratory Medicine (2018) 'Crying wolf: the growing fatigue around sepsis alerts', The Lancet Respiratory Medicine, 6(3), p. 161. doi: 10.1016/S2213-2600(18)30072-9. lancet2018cryingwolf

Open Questions

  • The editorial cannot say whether over-triggering came from implementation or from the algorithm itself. Distinguishing a miscalibrated deployment from a fundamentally weak model is the core diagnostic question for any decision-support failure, and it is left open here.
  • The reported design target — high specificity and sensitivity with a correct false/valid balance — understates the base-rate problem: with very low sepsis prevalence in primary care, even a highly specific rule yields many false positives. How decision-support alerts should be tuned (or suppressed) under low base rates is unresolved in the corpus.
  • This is a single clinical editorial standing in for a general AI decision-support lesson. Whether the KB should hold it, or instead ingest a primary human-factors / ML source on alert fatigue and trust calibration, is a scope decision for the user.