150: AI in Pathology – Regulatory Aspects of AI – 7-Part Livestream 5/7

The Most Overlooked Risk in AI for Pathology? It’s Not What You Think…

Welcome, my trailblazing digital pathologists! In this episode, I dive headfirst into the regulatory maze of Artificial Intelligence (AI) in pathology, covering global frameworks, safety risks, ethics, and the future of software as a medical device. While regulation might not be the flashiest part of AI, ignoring it could cost us innovation—or worse, patient safety.

We’re on Part 5 of our 7-part AI in Pathology series, and this one’s vital for anyone developing, using, or simply curious about AI and machine learning tools in healthcare.

If you thought regulation was boring, think again—it’s what separates a helpful algorithm from a dangerous black box.

🎧 Listen to the full episode and reference the latest study discussed here: Modern Pathology Journal Article

🔍 Highlights & Timestamps

[00:00:00] Welcome & Why Regulation Matters
Pixelation aside, I introduce today's critical topic—regulatory frameworks that define how AI tools are used, approved, and reimbursed in pathology.

[00:03:00] The Risks AI Brings to Healthcare
We’re not just talking about patient data—think security, ethical biases, economic consequences, and even environmental impact from heavy computation.

[00:05:00] HIPAA, GDPR, and the Common Rule Explained
What protects patient privacy globally? Dive into U.S. and European legislation like HIPAA and GDPR, and how IRBs and the Common Rule ensure ethical compliance in clinical research.

[00:08:00] FDA & Global Agencies Breakdown
Get to know the role of Health Canada, UK’s MHRA, Japan’s PMDA, and others in approving AI tools. Discover how the U.S. FDA sets the gold standard—and why CE Mark devices often hit Europe first.

[00:17:00] What Makes Software a Medical Device (SaMD)?
Four critical questions to ask to determine if your AI tool is considered a regulated device. If the software diagnoses, directs, or lacks transparency—chances are, it’s a device.

[00:22:00] FDA Pathways: Clearance vs. Approval
I break down Class I, II, and III device categories, and what 510(k) clearance means versus pre-market approval (PMA). Yes, we even cover why some tools like Paige AI needed full PMA.

[00:40:00] Why Reimbursement Is the Elephant in the Room
No billing codes, no incentives. I share a personal story about my husband’s five-year reimbursement battle and the challenges of proving economic and clinical value.

[00:45:00] The LDT (Lab Developed Test) Controversy
In 2024, the FDA formally categorized LDTs as medical devices, igniting debate about oversight, innovation, and compliance.

[00:47:00] Generative AI: A New Beast to Regulate
ChatGPT and similar tools pose fresh challenges: reproducibility, explainability, and dynamic outputs. Current frameworks simply can’t keep up—but regulation must.

[00:51:00] USCAP Event Invite & Digital Pathology Collaboration
I’m thrilled to invite you to Muse Microscopy’s USCAP presentation—join live or virtually! Plus, learn how we’re reshaping the digital pathology workflow with direct-to-digital imaging.

🧠 Resource From This Episode

Referenced Study:
📄 Artificial Intelligence in Pathology: Regulatory Challenges & Opportunities
👉 Read the full article

Transcript

AI IN PATHOLOGY_ REGULATORY ASPECTS OF AI

[00:00:00]

Welcome and Introduction

Aleks: Welcome. My digital pathology trailblazers. I'm just checking if I see anybody and if you see me, but if you do, let me know in the chat that you are here. Yes, I see you. Okay. There is always a little bit of a lag showing me if anybody's joining or not, give me a thumbs up or something in the chat that you hear me well, that you see me well, and I know that, I don't know, since we changed the streaming software I've been pretty pixelated, but I think the audio is okay.

So let me know where you're tuning in from. It's six. All one in Pennsylvania, Fairfield, Pennsylvania. Welcome, trailblazers, and today's topic is regulatory aspects of AI and machine learning. We are on part five of our seven part AI series. It means that we only have two left and that means [00:01:00] that once the series is over, everything is gonna be uploaded to a course.

Announcements and Course Information

Aleks: I'm gonna give you a few announcements before everybody gathers because I still see people joining. So the course and the course is gonna be. Here I'm gonna be showing you some QR codes in the corners and the course. There is a course, there is a course on AI in my shop called Path AI Makeover.

Pathology? No, it's called Pathology AI Makeover and. All the live streams that you are part of, they will be on YouTube as well. They will stay there. That's okay. But they will also be edited without my audience interactions and, without, all the engagement that you usually have in the livestream.

So if you just wanna have the distilled version of this, it's gonna be in this pathology ai makeover. And it's going to be also in the updated, oh wait, I have [00:02:00] to. Hide this one. And let me know in the chat where you're tuning in from. Just give me a thumbs up and give me a shout where you're tuning in from and what time it is for you.

Speaking of those live streams, it's gonna be also in the book when I update the book. So if you don't have the book, there is a cure code on the screen right now and you can get it for free. This is fantastic because it's free resource. I'm looking at the chat. And once I see a few more of you joining I'll give you a few more announcements.

But for now, let's start with our paper. Am I pixelated? Guys? Let me know in the chat if you see very pixelated version of me.

Starting the Discussion on Regulatory Aspects of AI

Aleks: And in the meantime, let's start regulatory [00:03:00] aspects of AI and machine learning. This is a very polarizing, maybe it's not that polarizing but topic, but it is very discussion sparking topic.

And today it's a little bit text heavy. We don't have so many images because it's about regulatory. So you can make it as engaging as it can be made, but it's regulatory. So some people are passionate about it and some are less passionate. Nevertheless, it is important because.

Safety, Security, and Ethical Concerns in AI

Aleks: The adoption of AI in healthcare raises several concerns.

And I even have a exclamation mark here because it includes safety. I always have problems with let me know in the chat where you're tuning in from and if I'm pixelated and if you hear me well. So the concerns are safety, security, ethical biases, and we're gonna dive deeper into this next week.

Sorry, [00:04:00] this can go accountability accountability, trust, economic impact, and. Environmental effects. It's a lot like everything, right? The thing with AI and with a heavy computation is that that is new. Is that maybe not because medical devices, you have to like somehow. Get rid of them as well even if it's not software.

But the environmental effects here are not negligible of the computation that is required for different AI tools.

Global and Regional Regulatory Aspects

Aleks: So that's why we wanna discuss regional and global regulatory aspects of ai, including data privacy, software as a medical device agency approval and clearance pathways. This is gonna be cool because, then you're gonna be like in the know on what it means that something is cleared, what it means that something is approved. And then we're gonna talk about reimbursement and [00:05:00] laboratory developed tests. Also, something that was a hot topic last year when FDA said, oh, we're now gonna take care of the lab developed tests and we're gonna treat it as medical device.

I see more people joining. Let me know in the chat where you're tuning in from. Where are you? I hope the chat is working. I know I cannot post messages to some channels, but I should be able to see everybody's messages. So feel free to message questions where you're tuning in from. Am I pixelated?

And whatever you wanna let me know during the livestream. In general it is recommended that AI regulations are formulated by public bodies rather than profit making private companies. Makes sense. But also often the public bodies, maybe not public buddies, but like associations have benefactors, so there's always a tie between the industry and the public buddies.

There's always. Influences influences [00:06:00] from everyone to everyone, but generally public bodies that are impartial. And an important fact is also that gen ai, generative AI that produces stuff, requires different regulations than non generative AI and non generative ai. The agencies or the regulatory bodies were already, they have some framework frame frameworks in place but not too many yet.

I see less people today. Are you not interested in the regulatory? But those who are here are interested. So thank you for joining. And so Gen AI will require different regulations and the primary goal of these regulations is protecting. Patients. And there are different different guidelines different regulations that address that.

Patient Privacy and Ethical Guidelines

Aleks: So one of them is to protect patient privacy. And in the US this [00:07:00] is the health insurance portability and like how. Accountability Act, the hipaa. Oh, is this HIPAA compliant? Is this software HIPAA compliant? Yes, it is. Then it means it protects patient privacy and we also have GDPR that's in Europe, general Data Protection Regulation.

This is the regulation that is in play in Europe. And also there are rules regard regarding ethics. One of them is common rule. We're gonna be talking about what that is. Common rule is to protect patients and subjects in clinical trials. And then these common rules is then being these common role.

Is then being implemented or proposed, recognized by recognized ethics committees such as institutional review boards, IRBs, and of course we have the agencies. We have the FDA, the Food and Drug, US Food and Drug Administrations. So there, these are [00:08:00] the three, three. Ways of protecting like those regulation, privacy regulation.

Then the institutional review boards and agencies, for example, FDA. And we're gonna be mentioning agencies from different places. And looking forward to seeing your chat messages about where you're tuning in from and what time it is for you.

HIPAA and GDPR. They protect personal health information and the common rule, and IRBs protect human subjects in research and development. And FDA protects public health. So here in this paragraph, that's why I have the exclamation mark is basically who does what in terms of regulating how research and, biomedical research is done when humans are involved. And there was an executive order by President Biden when he was the president to [00:09:00] establish a a contemporary healthcare specifica programs and policies and what these policies were supposed to do. They were supposed to they, they would require agencies to increase remic transparency, include human oversight.

I'm a fan of this one. Of both. That is basically common sense. So thank you for making this an executive order. And when necessary, offer long-term safety and real world performance monitoring so that we can ensure our, the. Who does that executive order? Those who enact this executive order, I guess the companies and the technologies and whoever creates these ensures that these tools are fair, appropriate, valid, effective, and safe.

We want that is common sense logic, and I'm glad it is now a mandate. So let's talk about data [00:10:00] privacy and regulatory environment for AI and machine learning in healthcare. Oh, sorry. I think we jumped a little bit. Sorry, I'm jumping now. Let's talk about where we are. We are here five. We still have six and seven to go.

And my friends, my trailblazers, you've lasted. Five live streams. This is amazing. Thank you so much. We are at the tail end of our journey, and I know today is a topic with less images, let's call it that way. Regulatory is a topic with less images. The healthcare organization that are using or co-developing, co-developing those tools, when they have partners, they need to also ensure that those partners are compliant.

So it's not that oh, I'm gonna hire somebody and it's not my problem anymore to develop something for me. If I am then like. CODOing [00:11:00] it, co-marketing, whatever, and I use a third party. It is my responsibility as an organization to make sure that they're compliant, they're HIPAA compliant, or whatever compliant they need to be.

Whatever regulation, whichever country you are and whatever like the bucket of risk or the bucket of regulations you are in you cannot just delegate and not care about it anymore. You have to you're responsible, right? And the as we said, because the technologies are pretty popular, they have high potential and they also raised moral legal and social concerns, right?

Social concerns was like, okay, will those with resources only have access to that? Why? How can we make it more available? In the world there were around 40 jurisdictions. Noted or recorded by the law Library of Congress [00:12:00] in Washington DC in 20 23, 40 jurisdictions that specifically referred to ai.

That's a lot. And we have different organizations like United Nations. North Atlantic Treat your organization, United Nations educational, scientific and cultural organizations, like different organizations that talk about this. And they the regulatory frameworks necessary for approving devices in Europe are less rigorous than.

In the U us So it's easier to get software as a medical device approved in Europe. That's why many companies go to Europe first get their approval CE Mark. Confirm European. I think and we have it in this in this paper later, but basically often the strategy is oh, we go to Europe.

The approval or the clearance approval is easier. We see how it works, we see how it [00:13:00] benefits we use the data and then we go to the US and try to do it in the US as well.

And regarding GDPR. The general data protection regulation in Europe. It imposes restrictions on automated decision making and similar to many country countries, limits the international transfer of personal data. Something specific to GDPR is and this limitation of automated, decision making. So not only for healthcare, right? In healthcare. And we have a question if, is this presentation recorded? Yes. This presentation is recorded. So if it's a very strange time where you are, you. Please stay with me. But you can view the recording and you can share the recording.

And the best way to get the recording is to be on my email list. And the email list you're gonna get on the email list when you download the book. So if you already [00:14:00] downloaded the book, then you are on my email list, or you go to my website, digital pathology place.com and sign up for the newsletter.

And then you get all the recordings of everything. Most of the things. Speaking of that, I do wanna give you one more announcement. Now let's finish this and then there's gonna be a commercial break. And the what was I talking about? The GDPR, right? You cannot do automated decisions about people and not only without human oversight and not only for healthcare, and obviously for healthcare, we think, okay.

The consequences would be, yeah, it is basically patient's ha patient's health, patient's safety but also for things like financial decisions, insurance, right? You cannot just feed the data to an algorithm and get the input, oh, this person is gonna be insured or not, or this is gonna be the insurance rate.

If there's no human oversight. So this is not okay according [00:15:00] to GDPR. And other countries also have regulations china, Canada is mentioned. Australia, obviously these are just examples because there is a lot more countries in the world that probably have regulations. Most of them have some kind of regulations.

So wherever you are using these tools, you need to. Figure out what the regulations are that you need to comply with. So now let's move to common rule and institutional review boards.

Common Rule and Institutional Review Boards

Aleks: This common rule is a federal policy for the protection of human subjects and dysregulation in the.

United Common Rule is the, like a common name for dysregulation, and it establishes requirements for institutional review boards. So dysregulation and institutional review boards [00:16:00] need to comply with that regulation. And this is for protection of human subjects in biomedical research. So that's not.

Specifically for a device. But if there is a trial, a clinical trial involved in a device clearance or in a no in clearance, you don't need a cl a clinical trial and we're gonna get to it. But in the approval process, then you do need to, adhere to this common rule in the process of designing your studies.

FDA and Software as a Medical Device (SaMD)

Aleks: And then we have our FDA, our beloved agency and other regulatory comp counterparts. Manufacturers of medical devices as a software. The, it's a it's a concept. It's a term software as a medical device. This SAMD. Because it's a medical device, but it's also software. And what is the definition?

So this is a software intended to be used for medical purposes, [00:17:00] but it's not part of actual hardware, medical device. So like a scanner has software, right? There is software inside to generate the images and do whatever you need to do. Stitch the tiles and or lines. But it is inside of. Hardware, right?

So that's not gonna be software as me device, even though software is part of this device. Software as a medical device is gonna be an algorithm that you can deploy on on general purpose computing platforms. So smartphone, tablet, personal computer all these algorithms that would help the medical decisions that you can deploy on.

Something else than a dedicated medical device. Something that's gonna be for clinical decision support tools imaging image, so imaging. No, that would be a hardware. But image analysis device. So I think image analysis software, right? That's gonna be a medical device. If you have an algorithm that [00:18:00] quantifies i 67, the proliferation marker this algorithm is gonna be software as a medical device, and it helps usually the image-based specialties.

So most the most image heavy are pathologists and radiologists detect abnormalities in medical images. And yes, we do have an image. So how do you determine if this software functions as a device or not? You ask yourself a few questions, right? And if the answer is yes, it is a medical device, and if the answer is no, it's not a medical device, it can be used for educational purposes.

The question number one is gonna be, does your software acquire process or analyze medical images? Signals or patterns. And if it does, then you go to software function as a software that functions as a medical device. Then question number two that you're gonna ask yourself is, does your software display complex information that require interpretation from a [00:19:00] professional?

And again, yes. Device, no. Then it can be used for educational, something that is not a device. So then you don't have to go through the regulatory stuff that you have to do to get a device cleared or approved. And so that was question number one. Question number two, does your software provide a.

A directive rather than recommendations, options, or information. So if there is recommendation option or information, then you are gonna go to the educational or non device. But if it's a directive, then you're gonna be a device and you need to comply with device regulations. And question number four, does your software provide a recommendation?

Without providing the basis for recommendation. Image analysis, like everything that has the black box or does not provide the recommendation AI based stuff is going to have [00:20:00] yes as an answer and it's gonna go to the device box if the answer is no. So it will give you basis of this recommendation, and I'm seeing a potential for generative ai if it can explain.

Why a recommendation was made, and maybe then it's not gonna have to be a device, but then you have a question. Oh, but is the like, can you. Can you believe this recommendation if it was generated by the same software that gives the recommendation. So that's, another discussion. And as you will see, there are not really any clear guidances for generative ai.

And these are just my, loose thoughts when I think about it. But basically. Long story short, if a lot is at stake when you use this software, then it is a device. If it's just to teach somebody and you can explain and everything, then it's not gonna be a device. And something I was, I was [00:21:00] smiling when I read it.

Was the FDA encourages good machine learning practices? I didn't know there was something like good machine learning practices. I know good laboratory practice, good manufacturing practice the G Ps GMPs and basically G xps. So now in the GX ps it's not the same, but there is good machine learning practice.

So they they, recommended to ensure that AI systems are transparent, explainable, and reliable. And how do you determine. How does those devices need to what's the level of stringency when you develop advice? How much do you need to check it?

Risk Stratification and Good Machine Learning Practices

Aleks: FDA categorizes medical devices including software as medical advice based on level of risk to patients.

And this a concept of level of risk to patients or in general level of risk when you use something. Is gonna be showing up in [00:22:00] different discussions, different AI discussions. In Europe there is a guidance for use of ai and it's also based on risk stratification, basically based and basically, anyway, the risk stratification concept is gonna be applied to many different validation frameworks like GOP validation relies on risk stratification. The categorization of medical devices relies on risk stratification, and. Like how is it stratified for medical devices? It's gonna be class one that is low risk class two, moderate risk and class three high risk.

And if you know which class are the scanners, let me know in the chat which class of medical device are the whole slide scanners. And speaking of whole slide scanners, I do wanna do this interruption right now and show you something else.

Special Announcement: USCAP Event and Muse Microscopy

Aleks: You guys know [00:23:00] that you ASCAP or ascap and I still don't know what the official pronunciation is coming soon and that I'm gonna be joining a sponsor of Digital Pathology Place.

Shout out to them muse Microscopy. So we're gonna be at. Booth five to eight. And there is something I have for you because there's gonna be a presentation and let's let's be honest, most of you are not gonna come to us, US Cup because only a fraction of you is gonna come to a conference and fraction of those people who go to conferences is gonna come to us, uscap.

But there's an option to join virtually for this. This talk that we're gonna have actually, it's cocktail reception and presentation. Who's gonna be staring here? We have a few cool people, and truly yours is gonna be moderating the whole thing, but there's gonna be cool. And very renowned people who are gonna be talking about the new way of tissue imaging.

So [00:24:00] if you know anything, if you've seen any of the content I produced with Muse already, they are direct to digital tissue imager. You basically skip that process of glass light production to image tissue and we're gonna have Dr. Peril. Chandra, Dr. Rao, whom I interviewed Jim Edwards was a guest for an interview when I joined them for CAP conference.

Dr. Derek Welsh, Dr. Richard Levinson is gonna be in this talk as well. Then Dr. Hong, myself, I'm gonna be moderating Ann Matthew Nunez, who is the CEO of Muse. So they are all very cool, who is super cool as well. You guys are super cool. So if you are there, of course, join let me tell you, when is it here?

Maybe let's do this. Here, Monday 24th. It's a little late in the US because it's a cocktail reception, so five 30 to but as I told you, [00:25:00] you are the coolest because you are my. My digital pathology trailblazers can do now or whenever you want. You can scan this code and join virtually.

Let's see if I can show you what it's gonna take you to. It's gonna take you page, replace screen sharing with something else. Give me one second. Okay. Chrome top.

This one? Yes. Okay, I should be sharing. So when you scan this code that I just showed you here, the US Cup event on the left, in the left upper corner you can register for this presentation and it's for free and you just hang out, make yourself a drink because it's going to be a cocktail reception with alcohol all.

Or alcohol free, whatever you prefer. And you can join us 'cause it's gonna be streamed. I think it's super cool. And there's gonna be officially streaming from the conference. And you can join me there. Let's see, I think you might be able [00:26:00] to ask question there then and there. So yeah, if that's of interest to you, then scan the code.

I'm gonna. Leave this code on the screen and for now, let's go back to our live stream. Sorry, I'm covering Matthew here. I. I have to show him Matthew and I just published a podcast with him as well where he talks about this technology. So you're gonna be receiving an email about the new content coming up and you can learn more with this code as well.

But basically if you are there both. Five to eight is where I'm gonna be hanging out.

Podcast Setup and Digital Pathology Discussion

Aleks: We're gonna have a podcast set up there. So if you just wanna be on my podcast and show up to the booth and get interviewed and tell me about your digital pathology experiences, whatever they are, whether you're like absolute beginner and you wanna ask me questions like from the beginning of your [00:27:00] journey or if you're, later in your journey on that digital pathology train you're getting off at the late AI station, we can discuss that. And if you wanna learn more about direct to digital tissue imaging, which I think is gonna be a game changer that's fantastic.

Technical Difficulties and Event Details

Aleks: Why did it disappear? Why did my screen disappear?

Replace, okay. Going back to our.

To our, why is it not showing anymore? Oh. I cannot share screen. Why can it? I stop sharing. Let's do it again. Window.

Okay. Okay that is the 24th. Join me virtually or in real life. It's gonna be so cool. It's gonna be fun.

Global Regulatory Agencies Overview

Aleks: And now back to regulatory. It's fun too for some It is fun. So what are the agencies in the world that regulate this other than FDA? [00:28:00] Because fDA is famous, but there are other agencies.

So there is Health Canada. There is Medicines and Healthcare Product regulatory agency in the uk that's a long name. Pharmaceuticals and Medical Devices Agency in Japan, national Medical Product Administration in China. Central Drug Standard Control organization in India and therapeutic goods administration in Australia.

And obviously these are just examples because as we said, there are more countries in the world than mentioned in this particular paper. And the mark the European Union as I told you, the mark, the CE mark stands for confirm. For medical devices. So if you see a device with a CE mark, it means it was cleared by the European agencies for use as a medical device, the bar, the scrutiny that these.

The, that the devices undergo in Europe is less than FDA. That's a fact. I check this [00:29:00] double and triple check this in several publications, and this is the case. That is also the reason why in Europe, the adoption of digital pathology is is higher. Like higher percentage of healthcare institutions actually adopted digital pathology.

Yeah, and they say it is currently easier to get confirm mark than FDA approval and FDA clearance or approval is often a more lengthy and extensive process. So recently I published a podcast with another digital pathology place sponsor, ham Matsu, and together with, software company, Prosha.

They went and got the clearance for their, for the scanner viewer software, concentric scanner, ham, matsu scanner, and the concentric viewing platform. If you go to YouTube, you're gonna, you're gonna see the podcast or to the podcast. And they said it was a [00:30:00] year of the whole process submitting stuff and.

They had to prepare more than for a year, but definitely something that does not go fast. I wouldn't expect any like regulatory stuff to go fast or any legal stuff. So this is the fun part, like when is something cleared or when is something approved? And any, any answers which class of a device was the scanner.

And I have some interest in collaboration in ocular histopathology. I'll definitely get back to you Arban later, but I don't see anything in the chat about what was the class of the device the scanner that was. I cannot say if it was cleared or approved, because then you'll know what the class of the device was.

Anyway, doesn't matter. You can still put it in the chat.

FDA Clearance vs. Approval

Aleks: But when you wanna check if somebody knows what they're [00:31:00] talking about in this clearance approval, FDA acceptance pathway, you're gonna ask them, oh. When was the scanner approved? When was, let's say, let's take Philips Intelli site, the first one.

When was it approved? If they know what they're talking about, they're gonna say no. No, it wasn't approved, it was cleared. It was a clearance. So that tells you that it was a class to device.

Medical Device Classification

Aleks: Because class two device gets a clearance and there are different pathways to, so let's start with class one.

And we said that all this is risk based. So this is low, this is mid and this is high, right? So one low number is low risk. That's good. And. What happened with the scanner? I think it actually was classified as a class three device and then it was reclassified as class two and that's why it went this class two clearance pathway, right?

When something is very [00:32:00] new and it does not have. A predicate device or is high risk, then it's gonna go through this pathway. Class three, pre-market approval and then approval. And then we have also alternate pathways and the humanitarian device exemption. If, there's a humanitarian crisis and then you wanna use a device to alleviate the crisis, or there is something called the Breakthrough Device Program where you can go faster, you get this breakthrough designation and you can go faster.

Yeah, class one, that's gonna be like, I don't know, class one. I don't know if gloves or medical device, but this type of a syringe or something that you use like that. And class two, for example, in the digital pathology space, that's gonna be a scanner. What's gonna be class three? I. I don't know.

I'll have to, I'll have to check that. What in our space would be a Class three device? Yeah. Maybe [00:33:00] the algorithms would be class three. Like for example, I. I think page for their prostate thing, they went through an approval process, so that was class three. I will Google that for you right now so that I don't spread misinformation.

And did page AI get an FDA approval or clearance? I think approval.

It's just AI generated a clearance and approval. Class three so class three. So that must have been an approval. That's because and also this class three definition is is a decision being made about a patient's. Treatment or about a patient in general based on the output of this device.

And in the scanner they said it's actually based on what the pathologist says about the app [00:34:00] output. And here in an algorithm whenever you're quantifying something or this was also computer aided diagnostics. We're gonna talk about it. But in general, like the distinction is, okay, is there a decision being made?

About a patient diagnostic decision based on what's coming out of the device. Sorry. Yeah. But based off what's coming out of the device, is this giving information for the diagnosis and for the scanner, they said actually interpretation comes first. So this is just generating an image.

So yeah, for we have here. A little bit of clarification of that this what FDA does when they approve or when they clear stuff, right? Manufacturer and here we are talking about, what are we talking about exactly approach to software as a medical device. Yeah, so we have pathways for medical device approval by the FDA.

It's gonna be [00:35:00] pre-market approval, pre-market notification, five 10 K five 10 K exempt, the novel classification and human monitoring device exemption. Manufacturers may request the novel classification if their device is novel. So there is no predicate, no similar device. When there is a similar device, you can show that your device is non-inferior to the one that is already cleared.

And then you go the five 10 K path. And. But if there are no device and that what's happened when Philips in Philips was doing the clearance of their pathology solution, which was the scanner, not only the scanner, like the whole workflow, this was there, there wasn't any before.

So they went through everything and then the next one I think that was Leica, said, Hey, Phillips already did that. So we're very non-inferior to Phillips. So let's go end of five 10 K. And in the United States, the humanitarian [00:36:00] device exemption is employed when a medical device is used for patients with rare condition.

So less than. 8,000 people have a condition. Then you can have the humanitarian device exemption that we had here right here on the image. And we also have this breakthrough device program can be used to expedite what you're doing. And I think Paige got the breakthrough designation as well when, aI product warrants providing patients with timely access to their sorry. When it's when FDA says, oh no we have to have it fast. So go fast and do this breakthrough pathway because we need it for patients. When we look at the stringency, pre-market approval is the most stringent type of device.

And this goes for class three devices. And that involves nonclinical and clinical studies. And then FDA. Approves the device. So [00:37:00] if something is approved, you know it's class three and it went through this pre-market approval. Then we have the clearance. So the words are approval and CRE clearance.

I, if you wanna have a bucket word, I would say FDA accepted, but that's my invention. I just the word is not my invention, but. That's like a bucket word that I would use if I didn't know if something was cleared or approved. I would say accepted by the FDA and five 10 K clearance is a pre-market submission made to the FDA to demonstrate that device, including software as a medical device, is at least as safe and effective as another legally marketed device.

Similar device, right? So one scanner is showing, oh, I'm. As good as the other scanner. And the word used here is non-inferior. And the regulatory speech, it's gonna be non-inferior. And what they need to show is that they actually are like the other [00:38:00] device and they have to show. Transparency about the data used for training and validation.

This is important. Also any foreseeable risks associated with ai, machine learning software as a medical device within routine operations. And there were other programs some are discontinued. There was a concept being explored, and that was between 2017 and 2022, where FDA, thought, oh, it's, maybe it's gonna be more streamlined if we don't do this device board by device. But if we go and do it by company, so if a company follows good, what was it? Machine learning practices and all good practices, then maybe they can get like some kind of blanket clearance. It was discontinued.

This was called the pre-certification program, pre-cert. Program that was between 2017 and 2022. And then also there is [00:39:00] something called predetermined change control plan. And this is supposed to be a structured report that outlines how changes to a software product, such as updates and enhancements, because software we know it regularly gets updated, right?

So do you need to go with every update to the FDA and do the same exercise that you already did for the whole thing in the first place? Not always. Sometimes you have to, and FDA has rules for that. But basically this predetermined change control plan includes the types of changes that can be implemented without requirements for additional regulatory.

Review, right? So this is and along with the process for evaluating and documenting these changes. So documentation is a big deal. And also it specifies, okay, when, what's the threshold where you actually have to go to the regulatory agency and do a new pre-market submission, right? So there is a definition of a [00:40:00] threshold.

If so, and so much changes, or if something happens, then you actually have to go again.

Reimbursement Challenges for AI in Healthcare

Aleks: Now let's talk about money, meaning let's talk about reimbursement. And another mention of this QR code that I keep on the screen. Scan it and register for our use cap presentation in cocktail party, which is free.

But what is not free is healthcare in the US and that's why reimbursement is important. And there are yeah, reimbursement. What are the issues reimbursement issues for AI and machine learning in healthcare? I guess they're no different than other reimbursement issues for other stuff.

But this is new, so you have to fight for this reimbursement again. So to secure reimbursement manufacturers in all countries need to demonstrate the clinical and economic value of their technology, and it's difficult. How do I know that it's difficult? Have I done [00:41:00] it? No, who has done it?

My husband, so he is a, he's an MD clinical pathologist, and he used to work for a company that had, I think they had a lab developed test and he was basically for, I don't know, three years or four years, or even five years, they were like fighting, had this major initiative to get it into different healthcare guidelines and get this test reimbursed or at least recommended it so that you can even think about reimbursement.

And then at the end they said, no thank you. It's. No, we're not gonna reimburse it. Yeah, it was difficult. I actually once drove him to some of those what is this? Like those assemblies of those people who reviewed this stuff. We went to Baltimore and he was presenting this, and then in the end, after five years, they said no.

So it's difficult. Let's not underestimate the difficulty of getting reimbursement. [00:42:00] But. If you demonstrate the clinical and economic value, then there's hope. Let's call it hope, maybe even a chance. And of course there's an acronym, A BHS, so algorithm based healthcare services. And it can be different types of algorithms, right?

In terms of a digital pathology, I am thinking about, okay, image analysis, quantification of something, highlighting something, but it can be an algorithm that takes into consideration different type of biomarkers, risks, stratification and things like that. So those algorithm based healthcare services the, they.

If they use AI to answer clinical questions or AI that streamlines operational tasks, and they use AI or machine learning to produce clinical outputs for the diagnosis and or treatment of a patient condition. And it's difficult to get reimbursement for this because the traditional [00:43:00] way of doing this is fee for service fee for service systems.

And there, there are no specific billing codes for the algorithm based of care services. The recent recent years, maybe two years ago, I think two years ago, codes were. I dunno if introduced or invented for digital pathology and the codes didn't go with any reimbursement. People were just asked to use the codes to show how much of this particular service is being used.

And I don't know what the plan to actually like, get codes to reimburse digital pathology, but. I know it was a very cumbersome process and I think there is a DPA Digital Pathology Association webinar on this. But yeah, so it's valuable. Those systems, those A BHS services they, there is a growing recognition of the value of them.

But they are new. They're. They're like, they [00:44:00] don't fit in the existing framework. So you cannot just like plug and get a code. You have to figure out how to do this. And the regulators are also working on it. For example, Medicare Payment Advisory Committee released a 30 page report, and I have a special news for you.

We're gonna go through this report in this live stream. Not at all. Don't worry. That was a maybe not so funny joke, but when I thought about it, I'm like, that's something I am never reading. So the fee for service doesn't work. The the, there would have to be a value-based payment model, like extra payment meant for better patient outcomes.

And I'm like, how like. If there is, that's biology. If there is good biology, outcomes like in terms of like treatment outcomes or whatever. But I read this and I'm like, so those who will have. Will get healthy, they're gonna get their [00:45:00] money back, and those who don't get healthy will not. So a lot of discussions and in short the software reimbursement strategies remain underdeveloped.

I love this sentence, so I have it. In red here. Yeah, it's difficult. It's not, there's no framework that you can put your software as a medical device in and yeah, the lack of standardized frameworks and that's a problem. So let's talk about another hot topic.

Laboratory Developed Tests and AI Regulations

Aleks: The LDTs Laboratory developed tests and artificial intelligence and machine learning platforms.

So when was that? They will tell us here in 2024 FDA, published new regulations that explicitly placed LDTs within the medical device regulatory framework. It used to be like a alternate path if you wanted or if you had, something that was a good test, but there was no option or you didn't wanna [00:46:00] go through the medical device route, or it was just in one lab.

You could do the lab develop test, you still can do it, but then FDA is gonna look at it. So that was a hot topic last year, caused a lot of dilemmas. If you see words in red, I like them for one or other reason for some reason dilemmas. So yeah, I love dilemmas in the regulatory space.

Generative AI and Regulatory Challenges

Aleks: And then there's gonna be more dilemmas because we need to think about regulating generative ai, right?

So that's gonna be more dilemmas. 'cause gen AI is at totally different. Beast. So what are these unique challenges? Reproducibility, right? It's not gonna be the same input always gives you the same output. It's prompt. It's talking, like talking to a person. Maybe I'm gonna give you the same answer every time you ask me what my name is.

But actually not even that, because in Poland I'm gonna tell you a different thing than in the us. Or in [00:47:00] Spain because people in Spain would call me ale. Whereas in Poland, everybody calls me Ola. And in the US and Germany and other countries, they call me Alex. For you I'm probably Dr. Alex, so not even that will give you the same answer, even though the question is, what's your name?

So the same problem with reproducibility is for generative ai, and those models are dynamic non-deterministic, and they like. We said software as a medical device doesn't fit into the regulatory framework. This like totally doesn't fit. Gen AI doesn't fit. And even though Chad GPT was released in 2022, we have, we are in 2025 right now.

There is nothing that is guiding this. And it will require it's work in progress. There are already tools out there, but those tools are like. Assistant tools rather than diagnostic tools. There is a lot of work [00:48:00] going on helping pathologists to be more efficient, better doctors for patients, faster.

Doctors for patients to get patients the diagnosis faster but more. Of Okay. Assisted by ai. So there's no determination made by ai. Different companies are doing that and I'm gonna show you some of them. At uscap there's gonna be videos and vlogs and if you've seen any of my conference content, you will see that as well.

Yeah. So it's gonna be cha a challenge. It is a challenge. Software as a medical device is a challenge, but also, the current regulatory frameworks assume a level of predictability and consistency. The generative AI cannot always guarantee. So we need the development of adaptive regulatory standards where these already developed, no.

That's what we need in paragraph of these of our modern pathology papers is always like the outlook into the future. So [00:49:00] that's what's needed. Whenever there is a black box and whenever we have deep learning, there is a black box. Then we face the problem of explainability, obviously accountability, bias mitigation, and a potential misuse.

And next week we're gonna be talking about the ethics. I know that regulatory might not be for a lot of people, but I think. And using it in a way that's that's fair, that's just is gonna of discussions. Or the same issues for Gen AI as any type of ai. Like data privacy, security issues, copyright infringement, intellectual property rights.

And there is no currently no separate regulatory review process dedicated just for AI based tools. So you will see some lists of, oh, so many tools are AI based. I think I was showing a list at some [00:50:00] point with like hundreds of, hundreds of radiology, cardiology and other AI based tool. And there were like three tools for pathology and only one tool was image-based.

So work in progress. Things to remember from today's livestream is gonna be this medical device classification course 1, 2, 3, in which paths they. Go through and also what makes a software a device. Because if it's a device, then it has to go through this through the regulatory pathways.

Conclusion and Event Invitation

Aleks: Other than that, I wanna just invite you again to our.

US cap presentation. So I should move myself to me. No, sorry. I'm covering Matthew. Okay. Anyway, you can join us for free virtually, and if you are there, come to this reception together with Muse microscopy. We're gonna show you a direct [00:51:00] to digital imaging device which is super cool. You can pick it up.

It's very portable. And I think it's gonna disrupt the pathology workflow. And before you go, if you don't have my book. I got the book. There's the book it's gonna be updated the ai. This is the free digital copy and whoever has the digital copy. So if you don't wanna buy the physical book, it's okay because it's gonna be updated soon.

It's on Amazon right now, but this is gonna take you to the virtual copy that is free. If you're not on my list yet, get on my list because then you're gonna get the recording of this particular live stream and. All the other things that are happening in the digital pathology space. And if you are here, you are a true digital pathology.

Trailblazers. I appreciate you. Thank you so much for your attention, and I talk to you in the next episode.