Clinical Decision Support

Clinical decision support is older than medical AI by decades, and most of what is known about it is discouraging. Drug interaction alerts have override rates above 90 percent. Order set recommendations get clicked past. The failure has never been that the advice was wrong ... it is that the advice arrived in a volume and at a moment that made attending to it impossible.

AI-based decision support enters that history, not a blank field. Any tool generating a notification is competing for attention that is already saturated, and alert fatigue is the constraint that determines whether it survives deployment.

The regulatory line runs through clinician review

Some decision support is a regulated device and some is not, and the distinguishing question is whether a clinician can independently review the basis for the recommendation. Software that explains its reasoning well enough to be evaluated, and that does not drive a time-critical decision, generally falls outside device regulation.

That exemption assumes the review actually happens. Under load, review degrades toward acceptance, which is the same behavior alert fatigue research documented thirty years ago. The regulatory logic and the deployment reality are not describing the same clinician.

The newer failure mode

Generative decision support adds a problem the older systems did not have. A rule-based alert is either triggered or not. A language model answering a clinical question can produce a fluent, well-formatted, entirely fabricated answer, and nothing in its presentation distinguishes that from a correct one.