Welcome to Real Talk on AI and I care with your host, Dr. Scott Morris. And I'm co host, Dr. Rehan Ahmed.
In each episode, we discuss current news, adding a little AI education, and debate innovative AI tools and topics that will change the industry. Hello, everybody. Welcome to Real Talk. This is your host, Scott Morris.
Today, I have a special guest, one of my good friends, Dina Weitzman, who is the senior director of medical affairs at digital diagnostics. And I am so looking forward to, we have discussions that go for hours. Sometimes we talk about the topic we're going to talk about today is, what do we need to worry about? Ethics, bias.
Are we getting better? Are we getting worse? But before we get into that, you guys know we always do the knowledge bite. And I get a lot of feedback from that that people are really appreciative because we kind of warm up subjects and talk about those.
So today we're going to look at a glitch that isn't software. It's in our heads. Let's call it bias. You think of a bias is kind of like a mental shortcut.
I mean, evolutionarily, I guess it kind of helped us make snap judgments to survive when we were out on a savannah somewhere. But in the modern world, those shortcuts off us and lead us into wrong spots, wrong destinations. Whether it's choosing a job candidate or an AI deciding who's going to get alone or what kind of health care they might get. Bias is this invisible hand that kind of tilts the scales.
And unfortunately, many times not in a way we understand. So, but before we blame AI or even machine learning or technology, sometimes we need to look in the mirror. Human decision making is wringled with cognitive biases. There are over 165 different kinds of cognitive biases that we at humans do all the time.
Let's take the big one right confirmation bias. This is our tendency to only look for information that proves that we're right. Because I don't know about you, but I like thinking I'm right, even though a lot of times I'm not. Or there's anchor bias where we say, well, I read this and then we filter everything else through that kind of bias saying, well, that first piece of information is what helps us make our decisions.
And these aren't just bad habits. I mean, they're like hardwired into how we as humans process reality. And then we make decisions under pressure, like a busy clinic day. And then we make a clinic day.
The shortcuts sometimes take over leading to maybe not the outcomes that we were wanting to have. And so, well, what happens then when we take these biases or we teach a computer how to think. Well, you know, AI doesn't have opinions. It's following some sort of logical tree.
And we're going to get into a dean talking about gray box black box. And maybe clear box here in a few minutes. But AI does have to learn on data. And the problem with algorithmic bias happens if a is trained on data that looks at our own human prejudices, then that's going to be a problem like, for example, I mean, if you're a hiring AI looked at, I don't know, let's say the last.
15 20 years of successful executives and sees that mostly men, it learns that being male might be required success when we all know that's crazy and completely biased and completely wrong. But that's the data that it was fed. And the danger is that that feedback loop is we assume computers are objective. And we trust the results more than humans, but then again, that's that anchor biases that the first thing we get and does it actually make sense.
Are we going to dig in that data is and I love dean is going to give us something I call it dirty data. She's going to call something else. Meaning that's a little biased and it and the problem is AI is that's just going to repeat these mistakes. It's going to just scale them up at lightning speed and spread them everywhere.
That's dangerous, you know, so, but is there a way to fix that I'm looking forward to what dean has to say about that. It isn't just better code. It's diversity. It's auditability.
It's human in the loop to go. Does that make sense? Do we need X AI explainable AI so we can kind of peek under the hood of these algorithms and ask the machine, why do you make that choice. So biases, a human problem that we've exported into our technology, identifying that we have these biases, the first step towards fixing it.
Hopefully that helps everybody today. Now we're going to get on to one of my favorite ladies. Dean white'sman, who is this as I said before is the senior director of medical affairs and she is probably understands ethics and bias about machine learning better than anybody I know. And so I'm super looking forward to the show.
Dina, welcome to our show and thanks for joining. No, thanks for having me and I thank you for the kind words and also the excellent. I've been looking forward to having this conversation with you for a while and bringing this out to your listeners too. I am a big fan of the show and it's really fun to be here.
So thank you so much. Of course looking forward to our discussion. I think I'm going to have to go back to the next slide. I think I'm going to have to go back to the next slide.
Like you said, Scott, I've been working for a healthcare AI company. I've actually been here for almost six years now. Before that, I was practicing clinically and multi specialty setting sometime in academics, even a little bit of time, working with data. With the large medical registry.
But this entire time, I've always been a clinician first, right? I don't have that PhD am I am behind my name. I just said, I wonder woman to that. Please keep going.
I love it. No, thank you. But I remember when I was considering taking my current role. This was 2020 and admittedly, I hadn't had much exposure to actually AI specifically for healthcare.
I think it was pretty skeptical. But I realized like it or not, it was here and I can sense this was just the tip of the iceberg. So decided to jump in and learn more. And I was I was skeptical if this AI could have a positive impact.
Like what are we doing as a profession. I think that my colleagues are amazing doctors. I think their support staff are fantastic. Like what are we really actually doing here?
We are really, really good for you Scott. I want to flip it. I already want to flip something right right at you first because I have a question. I was thinking through I have a question for you because I really have never asked you this.
But you know, have you when before because you are such a proponent of AI and healthcare. Was there a point prior that you were skeptical? Like was there an aha moment for you where you were like all it. I think I'm skeptical every day.
You know, I mean, I've been building an AI model for five years now, which hopefully audiences will start seeing you shortly. But I'm skeptical every day. And I think the story is that I also, you know, like you we've we've crossed paths for two decades, you know, on the speaking on speaking circuit or the meeting or in some consulting thing. I'm not only have I been skeptical about what technology can do versus what it's purported to do.
But I'm also equally a skeptic as are we as smart as we think we are. Right. So I think we have to balance that. And the reason I start developing the AI tools that I'm developing is that.
You know, we need we as human beings. I mean, and I've said this quote a couple different times over the last couple years. You know, education of what we need to know is now doubling every 73 days. And I think that by the end of 2026, it could be doubling every 73 hours.
You know, and so how how is our gray matter. How's our gray box going to keep up with that. And the rest we can't. So we're counting on what we learned a long time ago.
And that makes me skeptic like I've spent thousands of hours on the podium and I always wonder. How much of this that I spend hours and hours and hours and hours of best of my time learning to go teach. Actually gets used once people leave the clinical and the answer is my mentor is kind of laral examiner and he goes Scott when I was young. He said they think and remember three things when they leave the lecture hall.
They're going to remember two things by the end of the weekend. They're going to apply one thing next week. And if they remember it six months from now, you are really successful. And I'm like, wow, that's really disappointing.
And you know, it is and it makes me skeptical about can I as a provider, remember and learn and do everything that's out there. And the answer is no because I mean, I remember not as the editor for O M for years. I mean, I got paid to read hours every single day. And I couldn't keep up.
How if you're in clinic eight hours a day and you're busy with this that and the other are you keeping up and the answers were not. And this is where I think tools like, you know, like AI embedded medical intelligence is going to change the way we're going to have the access to those. And so am I skeptical that we're going to get the right information from those tools. Yes.
And my skeptical that I have the right information now. Yes. You know, so, so it's a balance of that skepticism. You know, you did a great like when we last time you and I really got talking about this was at a meeting gosh, it was probably two and a half years ago, I vision expo you were on the panel with me.
And we were talking you put up a picture of the balance and I was like, that's it. That's it right there. That's why I'm doing this. So yeah, I'm always a skeptic.
And I think the moment we start trusting anything computer black box gray box doesn't matter. As soon as we interested in publicity. Implicitly, we have that anchor bias and that is a problem. So we should always be skeptical.
We should always be questioning always. Totally agree. And I think, you know, to your point, I think where I even started and I'm going to be Frank here is I didn't even see a pathway forward with it like I didn't see the balance yet I didn't understand how there was a way to do this. And this was back in, you know, when I first started and kind of got first entered into this AI healthcare field.
I didn't understand how this could be done appropriately. And it was literally thought I think it could have been honestly like my first week on the job that there was an early publication from the collaborative community on up found make innovation. They have a different name now. And it's titled foundational considerations for AI using op found make images.
And what struck me about this was this paper presented a bio ethical foundation. It laid it out. And for me, all these years later, even through this rapid evolution of this technology, those core principles are still absolutely applicable. Right.
It's like the non-military sense first do no harm patient benefit improved clinical outcomes. We do it hard, you know, but I think about that though, Dean, are we are we not doing harm now if we are having this debate. Another innovator from S on from Toku, we were having this debate on innovators, which will come out probably the same week as this is coming out. We were having that exact debate and I said, you know, are we coming up to a point where is providers if we don't use the tools of today.
Are we going to be negligent? The answer is I think yes, I do believe that if you're not willing to use the tools that give you the best data possible to make the best clinical decisions, are you negligent as a healthcare provider. Yes. I will say that from the stage I'm saying on the podcast, I say it everywhere, I'm like, you know, the world changing really fast and we no longer have the excuse of I didn't remember or I don't know.
And you know, I remember reading that same paper you read and I come out and I was just this was when I was a little frustrated with the industry as a whole and I read that and I go, you know, there's hope we need data. You know, we need data and the data is going to be clean data, not dirty data or what you'll probably get into, you know, it's got to be clean data, but we don't have any clean data. I mean, diagnostic devices is a big piece where you can get somewhat and then it's never perfect, but somewhat clean data. But when you start messing with bio data, right, my brain, your brain, my patients brains and what we're communicating all this stuff, now the data gets really freaking dirty.
Because we don't know whether we're asking the right questions, whether we're hearing the right answers, whether the answer they're communicating is actually way it is or they tell us what we want to hear. You know, I start thinking about man, all of those things that were all those the bullet points you just mentioned in that paper came out, I'm like, but we need to go past, I think even past diagnostic test data, you know, hardwired bit data and get into hardwired people data, right, and go, how do we get that information out? And I think it's got to be a better way, right, than what we're doing because what we did worked, we thought for the last 200 years, I don't think it works anymore. No, I absolutely agree with you and Scott, you've already alluded to it twice.
So I think I have to bring it up now. What does she mean? Like what is he talking about with data that Dina says and it's literally it's so simple, it's that simple philosophy of like garbage in garbage out bias data and bias data out or Scott loves it when I say it, the chocolate chip emoji. And the chocolate chip emoji out.
And every lecture, I show that picture every lecture, Dina, it's my favorite and the whole audience burst out laughing so if the audience, if you can't visualize a Hershey's chocolate chip and what that represents. I probably can't help you on that one. That's one of the three things they're walking away with. Did you make it sparkle, hopefully.
That's exactly. Well, let's talk about that. So I mean, how do we, how do we fix. I always think about this is input processing output everything we do from the time we talk every you and I are talking I'm listening that's input I process some words come out of my mouth.
That's output. That's that's just processing whether it's done through cables or whether it's doing through, you know, biometric or it's done through neurological synapses. It's all just processing. And so how do we fix.
The bias dirty crap data going in, you know, I, I struggle with that every day when I'm building algorithms like how do I get the cleanest data in because if I don't, the data coming out is going to be bad too. No, you're, you're absolutely right and it's something I mean, you just alluded to it even in your, your sound bite at the beginning, your knowledge bite, we were inherently biased. So, you know, it's not too eliminate biases and anything that we're developing with these AI algorithms. It's all about minimizing it.
And there's a really great paper about this called like it's, it talks about the total product life cycle. And it's talking about the needs to minimize bias and to look for this at every stage in the product development. I think that the product knowledge by the way. Absolutely.
But here's where it kind of twists things a little bit for me is I think that a lot of people think that it's the onus is on the developer of the AI to minimize bias. And it just sits in this development circle like you have to develop it. You have to validate it. And this is where we minimize bias.
Think about the data and think about what's happening. And I argue it's actually has to be thought through that total patient life cycle. Every step. You know, I did I was just doing this for the lecture of preparing at Seaco.
You know, from the time a patient perceives that they have a problem. There are over 360 different decision points. Until they actually go I'm fine. 360 different decision points in every single of those every and I'm betting as I sit there and work this out on bettons going to be double that.
But every one of those is influenced by some form of bias. You know, and so. I love what you said is we're going to try to minimize it, but it's not just in the responsibility of the producers of the developers. And so, you know, the more data we get, the more bias we can find the better we can fix it and the cleaner it gets.
You know, I mean algorithms are much like human algorithms. Biostat or not are continually learning and improving. We hope ideally. You know, but are we going to take, you know, synthetic.
If we look at biologic synapses versus synthetic synapses. I always think about this. I don't understand the difference between this perspective, Dean. You know, when I get up or you get up and we go lecture about a subject pick a subject.
Maybe I have 100 people may have 200, maybe I have three people in the room. I'm going to change those 300 people, but that's still maybe if I do a really good job and they remember what that's one of the three points they remember by the end of next week. But the reality is what about the other 41,700 doctors that are out there that didn't get that lecture and didn't get that data. Those biases that are existing out there still exists, but maybe they're going to be changed because I have a bias when I present it.
Whereas I look at the future of machine learning is we look at thousands of patients and we changed the algorithm. And we though we need human in the loop. And I'm sure we're going to get into that here in a minute. You know, though we need human in the loop.
Are we able to change that and. So much more bias data. 30 data when we can start changing an algorithm versus changing 41,000 gray box algorithms. Yeah, no, absolutely.
And and I think kind of like piggybacking off of some of this is there's ways that we can do it when we develop and deploy. And we look at some of the high product for health care that can consider and look for this. I think one of the big pieces, like we talked about that total patient life cycle. it's even looking for ways to minimize bias by making sure That the product that is developed is affordable.
It's easy to use. It's accessible. But also it's. And I think this is often overlooked.
I can consider it the overlooked ethical obligation. the need for post market surveillance. I think does not stop at product development, right? We need to continuously monitor this.
For regulated product, clinical trial looks very different than real world setting. And I have, I work with an incredibly talented team. And one of these is my colleague, her name's Catherine Fairchild. She's our senior director of AIML.
And she emphasizes that outcomes and real world performance, they have to drive everything that she does. And I think she's absolutely right. It goes back to this other phrase. I love this one.
It was coined by our founder, Michael Abramoff. He's a retina specialist, has a PhD in machine learning, is that glamour versus impact AI. Just because an AI is quote-unquote pretty, right? If it doesn't perform well, if it doesn't, you know, actually like close those care gaps, you know, in our case, maybe improve patient workflow, improve the ability to efficiently and effectively see patients.
Or, or medication, education, compliance, adherence, I mean, the clearing house of data, clearing house of financial information, I mean, the list is huge. But I love that, you know, you can, you can make the pig pretty, but it's still a pig. It doesn't work. It's a little chaffler chip emoji.
There's chaff chip emoji, that's true. Yeah. Yeah, but what's the point then? What are we really doing?
And I know that I feel like on almost every real talk, the theme is like, what our health system is doing right now is not working. And I have not found a person that disagrees with this. It doesn't matter your background. This is just sort of agreed upon fact.
And for us, you don't want to develop something that overburdens an already overburdened healthcare system. You can't make it, we talk about this all the time. We can't add another tool. People have so many tools, right, that are going unused.
And this has been always my complaint with, you know, we can't add more work. We have to take something away to add something. And ideally, we're going to add something, and take 10 things away, because we'd all like to be, as you just said, you know, more effective. And I have to say thank you.
I wrote this down, Dina. I have been calling it the five E's for the last couple of years. It has to be something that's, something has to be more effective. It has to be more efficacious.
It has to be easy. It has to be educational. It has to be economic. It has to be experiential.
That's six. You just added another one I hadn't thought of. E is, it's got to be ethical. And I'm going to steal it.
So now I got seven E's, because thank you, Dina. But, you know, I agree. That's a great point is that if we just are making tools that don't solve real world problems, then it's going to be a tool that sits on your tool bench and know whatever you're using to collect dust, right? So how do we make for all of you listening?
You know, and AI is developing so fast that every one of you listening can be a tool developer, right? In some way, shape or form that makes life, you can't just count on the innovators out there, the people on our podcast, the folks like Dina, to go and change the world. Some of you are going to have to jump in and use AI to change your world, right? But I agree if we go through that list, Dina, is if we can't make our life more efficient, it's going to be an unused tool.
And how do we do that? Like, I mean, I wonder that like, do we even understand what we're trying to fix? You know, I thank you for pointing out, Dina, it's funny Rayhan I were talking about the other night is like, I feel like all we do is keep talking about how the system's broke instead of how to fix the system. So we're going to try to pivot starting today, you know, on maybe we need to talk about, hey, here's the tools that fix this system instead of trying to just blast the bad system.
But so, well, that brings me a question. So, you know, I know you have some political, you got to be some politically savvy on this question, but I also know Dina well enough to the audience to know that she's going to say what exactly what she thinks anyway. That's one of the things I love about her. But, you know, so what do you look at and go, these are the tools we really need.
That's, you know, that's a really great question. And I think some of it is, even if we, some of the times the things that the tools that we need or we think that we need, is it a tool that our patients need, is it a tool? You said something really interesting. You said that if we don't think about developing these impact AIs, right, that they're just going to sit on the shelf.
But I also worry that if we develop these glamoury eyes, they could be adopted and have worse patient outcomes. And so I think a lot of this goes back to your question, it goes back to, but, you know, following those patient outcomes, because we are going to make mistakes. We are going to think that we need this. We are going to have to get right.
Oh, absolutely. Absolutely. And I want to, you know, you alluded to this at the beginning of the talk. I want to kind of pivot that all through back over here.
And you talked about, yes, the ethical design and development adoption, but also the ethical cost of inaction, of not appropriately using those tools, of not ethically adopting them and using them. You know, we're no surprise. We're facing a projected global shortfall of 11 million health workers by 2030, according to the World Health Organization. And obviously, we're feeling that in eye care now, these gaps are widening, patients are being diagnosed too late, preventable harms still happens every day, and then let alone the burnout.
You know, it's, we've already talked about this. It's not working. But when we've seen real world evidence that when deployed thoughtfully, AI can allow clinicians to practice at the top of their license. It can allow them to make those complex decision-making that they need to do.
It can automate and handle scalable pattern recognition tasks. Is it now ethically justified not to use it? I totally agree. You know, I knew you were saying that.
I was thinking, you know, I hear that. I get in that debate all the time with, especially some of the deans of the universities and some of the governing bodies and go, we're such a shortfall, we're such a shortfall, we're such a shortfall. I'm like, but what we're going to do is going to change in the next few years too. And if we stop becoming so much of our actual patient time being data collectors and start being data interpreters and data communicators, I think we all build to see a lot more people.
So maybe we don't have as much of a shortage of human manpower as we think if we would just address what are, maybe we don't have a workforce problem if we would fix our workflow problem. I love that. I love that. Way to flip the script there.
I do. I love that. What a great concept. And I think it also brings to the next thing which we would be remiss to not talk about.
I mean, is there an AI and I care real talk without bringing it up, but acuomics, right? I can't nearly disclose who's having this conversation with because she would kill me. But the reality is that we, I think Dean and I are very curious to see your response to this. We're at the precipice of this entire professions, successor failure because either we as I care, and I'm not, I mean, a top of tree, ophthalmology, opticianary, it doesn't matter.
I mean, any of the three O's, we're either going to wrap our arms around acuomics and we're going to own it or we're going to avoid it and we're going to get owned by it because the rest of medicine is going there, right? I mean, when we think about neurodegenerative disease and cardiovascular disease and endocrine disease and the future of all that, if we don't as this industry, and if I could make a begging plea, and this is one of those, but probably just need a video clip this and put it on LinkedIn every day. If we don't as an industry, get our arms around acuomics, we will get left behind on this subject. It is an inevitable future.
The question is, are we going to be in it? Absolutely. I mean, I completely agree with this. And for me, the opportunity there is just so large and it actually, it makes me so excited.
I feel lucky to be part of this eye care profession because we have the ability to really tap into something that could just help every other healthcare discipline, help these patients complete systemically everything that they do. One of the biggest points I know Rupa Hansara, he had a great real talk a few weeks ago and he brought up the point that many patients see their eye care provider more regularly than their primary care provider. So, absolutely. AI-enabled eye exams can flag the systemic risk in population to otherwise wouldn't be screen, declining to use that capability could disproportionately affect underserved communities.
We have, I feel, an obligation to bring this to our patients who think of the world communities where many patients don't have access to cardinal... For decades. For decades. Absolutely, absolutely.
But they're gonna find a way hopefully to an eye care provider. And it's all about thinking about meeting the patient where they present, right? And so if this routine retinal imaging enhanced by AI can flag elevated cardiovascular risk or early neurodegenerative patterns, absolutely, you know, sign me up. I think this is just another way that we can kind of help equitably close that care gap.
Well, you know, and we're gonna go back to our main theme which was bias. You know, is how... I think a lot of, a lot of patients don't go see their healthcare providers because like, I don't have anything wrong. Like we had a gentleman yesterday, right?
35 year old healthy guy comes in. He's got a multiple turpees, bilateral, multiple turpees. And we're like, my student doctor's like, hey, do you have any family history of any GI problems? Oh, yeah, my dad has GI as colon cancer.
And I'm just like, oh my gosh, like we all learned, you know, boom, boom, boom, guard or syndrome. This is a huge risk FAP, that kind of stuff. And I'm like, he could have went on, I said, one last thing we read in general doctors, he goes, I don't know, when I was 21, he's now 35. But why did he come see us?
Because he's a minus two and he can't pass his driver's test, right? So, we become the gatekeeper for solving a problem with he won't know he has for a decade by the time he knows it, he's got a serious problem. And I look at and go with some of like technologies, you guys are working on and some of the other like, well, we be screening for stuff that catches that stuff and prevents, you know, I can't even, colon cancer is a horrible disease, right? I mean, can we prevent that kind of stuff?
And I think about, you know, but on the other hand, are we going to, as I care where I was going with this, are we with I care going to say we can overcome some of the biases and challenges of people getting to healthcare providers, like, we're just going to do this screening while you're here. We're going to take quick look at your retina and we're going to know, you know, not just, do you have something like that, which is obviously visual, but are we going to be able to screen them on a microscopic level and go, hey, you're having changes in the oxidative stuff that's, you know, some of the oxidative metabolites that are being released, we should probably go get, you know, what's the word I'm looking for? You know, a C-reactive protein, a high resonance, a C-reactive protein, the C-R-U churn out some type of oxidative issues that are going on and prevent a heart attack from ever happening in the first place. I look at then go, I think back to the situation, you know, the systems biased against healthy care.
And I'm hoping that we, as a profession, as an industry of I care, are going to be able to change that so we can be proactive and preventative instead of reactive, reactive, which is what we are now. Yeah, no, I so well said, Scott, so well said. I think I can feel the passion, right? I can tell, like, that's why you and I do this, right?
This is why we're in here. And I hope we can convey this to your listeners too, because I think sometimes, you know, we speak about AI and there's so much concern around it. And it's like, I want to convey the potential, right? I want to convey what we can do.
But even you mentioned even the example with auculomics, I think sometimes what's overlooked too, is the power to use it for things like not just beyond diagnosing in early screening, but like prescreening, but triage, disease monitoring, personalized medicine. And I see this, you know, this image, this retinal image, perhaps being a very valuable tool for our other healthcare professionals and specialties, think that cardiologist, maybe we can provide them with some more information. And so it really keeps us part of that primary care team. So well said, you know, so well said.
I think that's one of the things that the recurring theme as you said is everybody keeps coming to, hey, we have this great opportunity. We just need to not waste it. And you know, that we go back to the original theme was, there's so many naysayers, oh, well, it's going to make mistakes. Well, so do we, right?
It's going to be biased. So are we. It, you know, it's going to take my job. No, failure to adopt new tools is going to allow someone else human to take your job, right?
AI is, we always go back, we say this in every podcast. AI is not here to replace you. It's here to optimize you. It's here to augment you.
But if you don't, others will, right? So, Deena, thank you so much. You know, we try to keep these a half hour otherwise of our readers or listeners go, you're too long, Scott, you're too long-winded. Deena, this is part one of about ten of these that I like to do with you, because we just scratched the surface of what we want to talk about.
Deena, any, so I always try to ask this question too. So what's the one thing you see happening next three to five years that just truly excites you about AI in the eye care industry? I think that we are just beginning to tap into that tip of the iceberg in terms of diagnostics and personalized medicine. I think that we really are at an exciting point.
I mean, in the next few years, where we're really going to break through and show exactly what I's plus I care providers can do. And it excites me. I'm so glad to be working in this field since jumping into this role. I haven't looked back.
I love, you know, this, I don't get a case of the Sunday as I wake up excited to work because it's changing and evolving so fast. Well said, young lady, well said. For the audience, this is real talk. This is Scott moisture host pay attention every week.
When we talk to ex we talk to you but listening to real talk an AI in eye care, your weekly podcast to keep you informed about AI technologies revolutionizing eye care.