AI Automation ·
What should a Korean hospital revalidate after a brain CTA reading aid is updated?
On November 13, Bracco and Avicenna.AI announced US FDA 510(k) clearance for an aneurysm imaging aid [1]. FDA’s public K262062 documents include review of a predetermined change control plan [2].
Problem definition and research question
Korean radiology and device management teams need to understand how a new version changes what clinicians see. The review burden could shift across scanners or acquisition conditions when annotations change. What should a hospital check before allowing a new version into clinical operations?
Evidence and causal analysis
FDA documents explain the assistive use with qualified review of original images and the US change plan [2]. Korea’s MFDS publishes separate review guidance for AI medical device change plans [3]. These sources do not establish Korean clearance or a local performance gain for this product. Errors clustering by scanner or image characteristics are a hypothesis for evaluation.
| Dimension | International evidence | Korean evidence and application |
|---|---|---|
| Verified scope | K262062 includes a predetermined change plan review [2] | MFDS publishes guidance for AI device change plans [3] |
| Local decision | US documentation sets device-specific boundaries | Check Korean authorization and hospital image characteristics separately |
Comparison of alternatives
Option A retains the operating version and tests the update in an environment hidden from routine patient reading. This avoids workflow disruption but costs evaluation effort and time. After separate testing, option B observes small changes over a limited period before expanding. It yields workflow evidence sooner but needs rollback procedures and staff support. Both presuppose confirmed Korean authorization and institutional procedures.
Implementation and evaluation
Radiology and device staff should apply both versions to the same deidentified test images under the applicable institutional review process. Compare annotation differences by scanner, disagreements with specialist reading, review time and ability to trace version history. A proposed gate is to defer expansion and investigate if disagreements increase under a predefined acquisition condition. Qualified clinicians retain diagnostic review.
Conclusion and open questions
The change plan describes the permitted updates, while local validation examines those updates in the actual reading environment. Local scanner-specific outcomes are unavailable. Which device and image conditions belong in a mandatory validation sample?