Artificial Intelligence Software for Medical Devices

[kreid]

Hello,

Does anyone have any experience with artificial intelligent software medical devices?

I am wondering if the training/learning of the algorithm continues when the software is on the marketing. Or is the training/learning of the algorithm part of the design and development process which is then 'frozen' and validated prior to release of the device?

Thanks

 

[japayson]

Apparently there is not even a real definition of what "artificial intelligence" means.

I just read a discussion of the ethics of artificial intelligence in healthcare applications. Try a search for "Bioethics of AI in the Healthcare Industry".

I think that continuous learning is the point of AI. I think this is why when it is spoken of commercially the discussion is about so called "cloud computing" and availability of "big" data. It seems that to be "AI" the software is never "locked down". This is probably not compatible with a variety of standards.

Perhaps there is a point where data from multiple device sources could be analyzed through AI methods and the knowledge gained could lead to distributed software (firmware?) revision. How this is compatible with any standards, I don't know either.

 

[yodon]

No experience but the question is, to me, totally interesting so I wanted to join the discussion. This ties in nicely with the recent revelation that IBM Watson failed pretty dramatically on recommending treatment for cancers. Apparently, the recommended treatments were, in some cases, considered dangerous.

My initial thought is that it will depend on the claims. If you claim the software is making a diagnosis, you have a higher burden of responsibility than if the software is "advising."

The whole idea behind AI is continual learning so I don't see such an application having a stake in the ground.

I think the regulatory bodies will likely struggle with these same questions. FDA put out a paper addressing the topic: Transforming FDA's Approach to Digital Health (the AI section is a little more than 1/2 way down).

There have been a few applications cleared in the US:
* FDA permits marketing of artificial intelligence algorithm for aiding providers in detecting wrist fractures
* FDA permits marketing of artificial intelligence-based device to detect certain diabetes-related eye problems
* FDA permits marketing of clinical decision support software for alerting providers of a potential stroke in patients

I believe all these went through the De Novo process - which makes sense. All submitted, as evidence of efficacy, a large block of cases analyzed. What was not clear, though, is how the continual learning affects the situation. Maybe it's possible to continually or periodically benchmark?

 

[TomaszPuk]

Hello,

Does anyone have any experience with artificial intelligent software medical devices?

I am wondering if the training/learning of the algorithm continues when the software is on the marketing. Or is the training/learning of the algorithm part of the design and development process which is then 'frozen' and validated prior to release of the device?

Thanks

In our organization we have taken part in various AI projects including MD ones. I would say AI training / configuration differs from one project to another but surely I would call it software development. You configure the network and provide/train its configuration data (models). Thus you can easily meet identification and configuration management requirements for AI software component. The verification is a bit more tricky as you do not have classical algorithm specifications (detailed design). Instead, I would recommend using input data / vs. expected outcomes to create test level specification - at the end that's what AI is about . Once you have the test level specification can run verification and reuse the same test level for validation on Staging / Production environments.

After GoLive training of AI may continue, and IMHO should continue. Still, I would train it in parallel to the already validated AI deployment / configuration. This way you use your production data to train new version of AI without impacting your currently validated software system. When you are satisfied with the new AI version you can think about new release with updated AI component/software unit.