Decisions, Not Alerts

An alert says a threshold was crossed. A decision says what should be done, how sure the system is, what happens if it is wrong, and who is allowed to say yes.

Alert fatigue is not a volume problem that can be tuned away. It is a category error. An alert transfers an unprocessed observation to a person and asks them to supply everything the system did not: significance, context, options and consequence. Dismissal is the rational response to most of them, and dismissal rates are what the resulting system optimises for.

A decision card carries the missing work. In the reference implementation each one states the recommended action, a confidence figure with an explicit statement of what remains uncertain, the reasoning in plain language, the relevant trend, links to evidence, the policy in force, the downside if the recommendation is wrong, the alternatives considered and why they were not chosen, the deadline, and the role required to authorize it.

That last element is the quiet one. A stream that shows a care-management lead work only a nurse may authorize is not a work queue, it is a routing failure. Authority belongs on the card.

There is also a category for honest deadlock. When two models disagree inside their confidence bands, the correct output is not a picked side but an unresolved uncertainty with a proposed test.

That their platform also supports prioritized worklists and configurable severity; that confidence figures are model artifacts that clinicians will not read correctly; and that recommending an action shifts liability from the clinician to the software vendor in a way no one has yet tested.

The liability point is the serious one. It is also why the alert persists: an alert transfers risk to the reader. A system that recommends has to accept that transfer running the other way, which is why the evidence record and the authorization model are not optional features of this pattern.

Two measures over the same population. First, the disposition rate: what proportion of items are acted on, modified, or rejected with a reason, against the dismissal rate of the alert system being replaced. Second, calibration: when the system says it is 80 percent confident, does it turn out to be right about 80 percent of the time. A stream with a good disposition rate and poor calibration is a persuasion engine.

Whether confidence should be shown as a number at all, or only as a band with the uncertainty named. What the right maximum length of the stream is before it becomes the worklist it replaced. Whether rejection reasons can be captured richly enough to retrain against without turning every rejection into a form.