Clinical decision support override rates above ninety percent are common and widely reported. They are usually discussed as a behavioural issue. They are better understood as a signal about the alerts.
The arithmetic of ignoring
If an alert is correct one time in twenty, the rational response for a clinician under time pressure is to dismiss it reflexively and rely on other safeguards. That is not negligence; it is a sensible allocation of limited attention. The system taught the behaviour.
Every alert that fires without changing a decision is training clinicians to ignore the next one.
Instrumenting the alert estate
- Fire rate per thousand encounters, per rule.
- Override rate, with free-text reasons sampled and read by a human.
- Action rate: how often the alert changed the order.
- Harm prevented, estimated honestly — most rules cannot demonstrate any.
Any rule that cannot show an action rate above a defensible floor should be retired or narrowed. Retiring alerts is politically harder than adding them, which is why estates only ever grow.
Tiering by consequence
Hard stops are a scarce resource
Reserve interruptive alerts for situations where proceeding would cause serious harm. Everything else belongs in passive display, where it informs without demanding a click.
