The 510(k) Gap: What Hospital Radiology Departments Haven't Resolved Before Their Next AI Model Update
Viz.ai is deployed in 1,100+ hospitals with 20 FDA clearances. Aidoc covers 1,200+ health systems. FDA has cleared 950+ AI/ML-based SaMD through Q1 2025 — 75%+ in imaging, most under the 510(k) predicate pathway that validates a specific algorithm version. When vendors update their radiology AI models, hospitals continue running them under the original clearance. Most radiology departments cannot confirm what version is in production, what the false positive rate is on their patient population, or whether alert threshold configuration matches their clinical protocols. This episode dissects the version staleness problem and the governance posture every health system needs.
The Deployment Debrief · Host: Elise · AI Insight Lab
Key takeaways
- 1
The 510(k) predicate pathway validates a specific algorithm version. When a vendor updates the underlying radiology AI model — even a performance improvement update — the hospital may be running a version not covered by the original clearance without a corresponding regulatory event.
- 2
Most radiology departments cannot answer: what version of the AI model is currently running in their PACS? Whether that version matches their 510(k) clearance documentation? What the false positive rate is on their specific patient population?
- 3
Alert fatigue from high false positive rates is an undocumented risk at most radiology departments — aggregate accuracy metrics from vendor clearance documentation do not reflect performance on sub-populations that differ from the training dataset.
- 4
The FDA's predetermined change control plan (PCCP) framework places documentation obligations on the vendor, but the hospital remains responsible for confirming that the version running in their system is covered by a valid clearance — and most hospitals have no process to do this.
The Deployment Memo
One enterprise AI deployment, dissected every Tuesday.
Every issue covers the same format as this episode: what broke, why it broke, and how to avoid it before it happens to you.
Episode sections
Why 950+ cleared radiology AI devices and the 510(k) predicate pathway that cleared them creates a version staleness problem no hospital has formally solved.
What these platforms do in the diagnostic imaging workflow, how they integrate with PACS systems, and where the alert enters the radiologist's reading queue.
Why the 510(k) predicate pathway validates a specific algorithm version — and what happens to the clearance status when the vendor updates the underlying model.
The FDA's predetermined change control plan (PCCP) framework and why most hospitals cannot confirm whether their vendor has obtained updated clearance for each model version running in their PACS.
Why aggregate accuracy metrics from the vendor clearance documentation don't reflect performance on your patient population — and why most radiology departments don't track their AI tool's false positive rate by sub-population.
Vendor-reliant, version-tracking, performance-monitored, and full SaMD governance program — what each requires and which health system posture your radiology department is implicitly running.
Running a model version not covered by the original 510(k) clearance, FDA audit exposure, malpractice liability from AI-assisted miss, alert fatigue-driven protocol deviation, and PACS integration failure after unannounced model update.
The version and performance questions every radiology department should be able to answer before the next Joint Commission review.