94→82%
Real-world sensitivity drop measured in a published device study
AI-enabled medical devices pass validation, then drift in the field — by site, by scanner, by population. VarunaForge AI is the independent layer that detects that drift and turns it into the post-market evidence regulators expect.
Validation is a snapshot. Real-world performance moves with the population, the equipment, and the site. The regulation now treats this as the manufacturer's continuing obligation, not a one-time study.
A published real-world study of an FDA-cleared AI diagnostic, across 17 sites and 100,000+ cases, found sensitivity falling from roughly 94% to 82% — and lower still on the hardest cases. The risk is measured, not hypothetical.
One site can diverge while the rest hold. Without an independent view across sites, a manufacturer cannot tell a device problem from a site problem — opposite responses to opposite causes.
Across the EU and the US, regulators increasingly treat post-market performance monitoring of AI-enabled devices as a lifecycle expectation. Falling short puts market access at risk — and for a maker without the in-house infrastructure, that is existential.
We track sensitivity and specificity against the manufacturer's registered validation — with confidence intervals, not a single accuracy number — and flag only when performance moves outside the expected range and stays there. True drift, not natural variation.
A view across sites is the one thing you cannot assemble from your own deployments alone. When three sites hold and one diverges, that tells you where to act — and only an independent benchmark can give you that answer.
The result drops into your post-market evidence as a deliberately-collected, standalone component — proactive monitoring, audit-ready, and yours to file with your notified body or the FDA.
The post-market performance expectation is already here, in the EU and the US alike. We can show you what independent, cross-site monitoring looks like on a device like yours — starting with your own retrospective validation data, no hospital integration required.