What a Decade of Recalls Taught Me About Evidence

Every device recall I've investigated has had the same shape at its center: a sensor that reported a number with more confidence than the underlying physics supported. Not a broken sensor — a working one, reporting correctly, that had simply never been asked to say 'I'm not sure.'

Engineers are trained to eliminate ambiguity, and in most disciplines that's the right instinct. In implantable devices, it's occasionally the thing that gets someone hurt. A glucose sensor drifting near the edge of its calibration window doesn't fail loudly — it just keeps reporting a plausible-looking number, and plausible is exactly what makes it dangerous.

The fix isn't more precise hardware, though we should keep chasing that too. It's designing the reporting layer to carry uncertainty as data, not as a footnote. A range instead of a point estimate. A confidence interval that a clinician — or another piece of software — can actually act on.

I understand the resistance. A device that says 'I don't know' looks, on a spec sheet, like a worse device than one that always has an answer. But I've never once seen a recall caused by a sensor that correctly reported its own uncertainty. I've seen several caused by ones that didn't.

The regulatory evidence we submit should reflect that. Not just 'this device is accurate under these conditions,' but 'this device knows when those conditions no longer hold, and says so.' That second sentence is harder to prove and far more valuable to a patient wearing the thing.