Validation Methods of Machine learning and Artificial intelligence

Wondering if anyone of you can guide me in the right direction.

I am looking to study AI and Machine leaning validation techniques in pharma. i can find general information on AI/ML validation but not specific to pharma.

Also, if there a noted regulation on AI/ML validation in pharma?

I am familiar with AI and medical devices in the General Hospital category.

The FDA is evolving on this issue however, I suggest releasing a locked model for customers and not learning. If you try a learning adaptive model the FDA will probably water it down so much as to kill it. You also won't require PCCP or ACP documentation in the submission.

Make sure you keep your data sets separate (Training vs validation) and the data has demographics to demonstrate no bias.

We are working with an excellent consultant on this so I am learning every day from her. PM me for her details.

I strongly suggest obtaining the advice of a been there/done that type person for this as AI is so new to the FDA and they are very anxious about how to regulate it.

FDA does seem to be coming to grips with ML / AI. I continue to see new products cleared by FDA.

Obviously, you want it continually learning but I think what @Ed Panek is suggesting with the locked down approach is to have a planned release cycle with the model and not lock it down forever.

FDA has is providing some help:

• Good machine learning practices
• Action Plan - with links to some other resources

Be sure to post later on regarding any experiences you have. There are several of us on here that would benefit from more "real world data" (if you will).

patilrahuld said:

Also, if there a noted regulation on AI/ML validation in pharma?

Click to expand...

I'm not sure what you mean by "in pharma". I think that in most cases, the AI tool would be considered SaMD (so, a medical device), or (FDA-)unregulated SW. Are you referring to an AI tool that is intended to facilitate pharmaceutical decision-making or the likes? I think that would still qualify as a device. Anyway, I'm not aware of any FDA regulation or guidance targeting "pharma AI" (as they do for medical devices). However, pharma isn't my specialty and AI is a fairly new field.

Ed Panek said:

I strongly suggest obtaining the advice of a been there/done that type person for this

Click to expand...

I doubt that there are more than very few people who can rightfully claim that they've "been there" where "there" is "pharma AI".

yodon said:

have a planned release cycle with the model and not lock it down forever.

Click to expand...

I think the main issue FDA is having with AI is the performance "unpredictability" of unlocked AI "running wild" in the field (how would we validate that the learning itself would never go rogue). No one is expecting the AI to be "locked forever" (actually, if the model can be improved through e.g. additional training, it'd be welcome I think); so when we talk about "locked AI" we mean that the learning (+ necessary validation of the updated model) only occurs at the Manufacturer's, while the model stays static in the field.

Ronen E said:

I'm not sure what you mean by "in pharma". I think that in most cases, the AI tool would be considered SaMD (so, a medical device), or (FDA-)unregulated SW. Are you referring to an AI tool that is intended to facilitate pharmaceutical decision-making or the likes? I think that would still qualify as a device. Anyway, I'm not aware of any FDA regulation or guidance targeting "pharma AI" (as they do for medical devices). However, pharma isn't my specialty and AI is a fairly new field.

I doubt that there are more than very few people who can rightfully claim that they've "been there" where "there" is "pharma AI".

Click to expand...

good catch. Depending on the consultant, if you ask, "do you have experience in XYZ?" they will generally say "yes" regardless of their actual experience especially with something brand new like AI. No one is an AI expert yet. Its like hiking a trail in the woods. It takes years of hiking that trail in all weather conditions to state you are an expert. Make sure you measure that experience in practical terms.

patilrahuld said:

I am looking to study AI and Machine leaning validation techniques in pharma. i can find general information on AI/ML validation but not specific to pharma.

Click to expand...

If the OP is asking more along the lines of "trying to study (in a specific field)" and less "specifically trying to release/leverage something in a regulated field" Then I suggest the following:

• Do the studying of a wide variety of Machine Learning circumstances/applications.
• Develop a nuanced appreciation of the particular area of interest, to see how the Machine Learning approach could help tackle which parts of that area of interest.

I can only speculate on what role ML might be playing in pharma; my own limited imagination has trouble guessing which possible uses would require "validation" in the classical sense. For example: any new treatments developed with the help of AI wouldn't be accepted just because an AI was validated, and my gut instinct is that it is unlikely that AI would be used to play a role in a manufacturing process (e.g. mixing).

As far as using an AI in the development of solutions: I expect the most difficult aspect would be deciding on the metric(s) for evaluating possible "solutions" to whatever the problem is... the "validation" would (IMO) likely be demonstration that the algorithm(s) have been implemented as required. Outputs of the AI would still require external analysis.

A very casual introduction to machine learning for anyone who is interested is Oliver Roeder's Seven Games: A Human History. I found the writing to be uneven, but there are enough pointers to different methods of Machine Learning (at least in what are essentially zero-sum scenarios of "win-or-lose", but there are applications to other areas) that the book still gets a soft recommendation from me. In order to get that information from the book, it is necessary to step between the histories of some personalities and his mostly personal takes on the games themselves. Roeder is a journalist, and it the book reads (to me) like it didn't get a particularly serious pass through an editor. Nevertheless, it offers accessibility on the topic for a broad swath of people.

Ed Panek said:

I am familiar with AI and medical devices in the General Hospital category.

The FDA is evolving on this issue however, I suggest releasing a locked model for customers and not learning. If you try a learning adaptive model the FDA will probably water it down so much as to kill it. You also won't require PCCP or ACP documentation in the submission.

Make sure you keep your data sets separate (Training vs validation) and the data has demographics to demonstrate no bias.

We are working with an excellent consultant on this so I am learning every day from her. PM me for her details.

I strongly suggest obtaining the advice of a been there/done that type person for this as AI is so new to the FDA and they are very anxious about how to regulate it.

Click to expand...

Hi Ed,


Thank you. i will appreciate if you can share any information/documents (non proprietary)? Sorry if i am asking anything that is not ideal but i am stretched in finding any guidance on such.

Rahul

Ed Panek said:

Depending on the consultant, if you ask, "do you have experience in XYZ?" they will generally say "yes" regardless of their actual experience

Click to expand...

Ed Panek said:

No one is an AI expert yet.

Click to expand...

I wouldn't go as far as that. I worked for a company that developed (and is still improving) an AI-based SaMD almost two decades ago, and during that time had the good fortune of meeting a few people who - IMO - can rightfully be dubbed "AI experts". But I agree there aren't many of them yet.

I liked the hiking trail metaphor, having been a hiking guide in my 20s

Tidge said:

I can only speculate on what role ML might be playing in pharma; my own limited imagination has trouble guessing which possible uses would require "validation" in the classical sense. For example: any new treatments developed with the help of AI wouldn't be accepted just because an AI was validated, and my gut instinct is that it is unlikely that AI would be used to play a role in a manufacturing process (e.g. mixing).

Click to expand...

From my acquaintance with AI/ML in/as medical devices, they are mostly about making inferences from a big dataset (e.g. a medical image); too big for a human to analyse methodically and consistently (and non-intuitively, i.e. like a radiologist would observe an X-ray image). In that context, validating the AI does not equate validating the device (or the design). While the latter means building a solid (preferably as objective as possible) case that the device fulfills / will fulfil the user's needs, the former simply means that the AI will "spit out" the right answers ("right" in this case means close enough to what a human expert, e.g. a radiologist in the case of making inferences from a medical image, would "spit out"). Technically speaking, AI validation is similar to AI training, and to use a human parallel, it's like the test at the end of the course, that checks that the training was successful. The main difference is that in the training stage the errors are used for improving (the model), while in validation it's not done - there's only a "pass or fail" result. Of course this description is a little simplistic, but you get the gist I hope.

At this stage I don't think AI is seriously used for "developing new treatments" or the likes. I think it's just not there yet. Maybe it plays some modest role in that arena, but I think it's likely that wherever it's claimed to be more than that, it's mostly marketing inflation.

Tidge said:

the book reads (to me) like it didn't get a particularly serious pass through an editor.

Click to expand...

This seems to be the case with so many nonfiction books today, that would have otherwise been quite good.

My reply was explicitly avoiding diagnostic(direct or assisted), which would be a medical device and not "pharma".

From my recollection, one of the later chapters in the Seven Games book did mention (casually) the potential application of one of the specific AI learning techniques in the development of new therapies, but it was pretty casual. It has been several decades since I did any investigation into machine learning (for problem solving, and not zero-sum attacks) and the aforementioned book didn't really offer much insight (IMO) into machine learning aside from the applications towards "solving" the titular games.