One patient can generate many interruptions
Suppose a model produces the same warning every hour for a patient whose situation has not changed. Counting one correctly identified patient hides the burden of repeated alerts.
The workflow needs a policy for persistence, suppression, acknowledgment, and meaningful change.
Choose a denominator people experience
| Measure | What it reveals |
|---|---|
| Alerts per patient-day | Burden relative to care volume |
| Alerts per clinician shift | Workload at the receiving end |
| Repeated alerts per episode | Redundant interruption |
| Alerts with an available action | Practical usefulness |
These are complementary to sensitivity and predictive value, not replacements for them.
The response should be possible
A warning that identifies a real issue but offers no feasible action may add workload without improving care. An alert routed to the wrong person can fail even when its prediction is correct.
Before broad deployment, examine the full sequence from alert delivery to decision and follow-up. Reducing alert volume may improve usability, but suppressing warnings also has consequences that need evaluation. The aim is an appropriate balance based on the clinical task, rather than maximizing notifications or minimizing them without regard to what is missed.
