Predetermined Change Control Plan

[QuinnM]

Hi,
The FDA identifies about 700 AI/ML medical devices in the US market. I'm interested to know of the 700, how many have a predetermined change control plan ("PCCP")? If public information what companies have successfully submitted the PCCP?
Quinn

 

[Ed Panek]

Few. The real power in AI is in "self-learning" and updating. Most colleagues I've spoken to have seen FDA nitpick their PCCP down so far they gave up and will just submit a special 510(k) to update the model.

 

[yodon]

FDA nitpick their PCCP down so far...

Hmmm... that's discouraging. I wish we could get specifics on what the issues were.

 

[QuinnM]

For anyone are there details that could be shared, both for companies that have PCCPs and those who tried to have a PCCP.

 

[Bev D]

Is this “Predetermined Change Control Plan” like a “Master Validation Plan”?

 

[yodon]

@Bev D - see this draft guidance.

 

[Ed Panek]

My suspicion is fda isn’t aware fully how ai is different.

For example you have a ten problem benchmark. For a release candidate it must score 70% or higher. Your initial release scores 80%. It only misses problem 9 and 10. A new model scores 90%. It solves 9 and 10 now but misses #5. It should be released but fda can’t wrap their head around patient #5 being worse.

Ai is just trained on data. Performance is aggregated. It’s possible to tweak the model but addressing item #5 isn’t a direct pathway.

If you have a learning model after 5 years its performance may be substantially better but for new companies looking to enter your market their performance metric is not state of the art. Its state of the art 5 years ago and fda may worry approving a similar device with much worse performance.