How AI is Transforming Cancer Care with Amber Simpson | Approximately Correct Podcast

Speaker A: So there's been great clinical trials, randomized clinical trials, showing that AI, uh, screening tools for breast cancer are better than humans. And that's actually a really nice application because if humans can't actually look at all the images to begin with, we're not even really talking about replacing humans.

Speaker B: Right.

Speaker A: We're talking about making sure that all the images get looked at.

Speaker B: Hey, welcome back to Approximately. Correct. I'm Elana Fish.

Speaker C: And I'm Scott Lilwall.

Speaker B: And this time on the pod, we have Amber Simpson.

Speaker C: Yes.

Speaker B: Yes.

Speaker C: So Amber is a. She's a new hire here at Amy, but she has a lot of experience, uh, working with AI, both the, uh, Canadian and US Healthcare system. She is a fellow here with Amy. She's also Canada CIFAR AI chair. And she works mostly when it comes to medical imaging and AI, although she has some really interesting thoughts on just AI use in healthcare in general.

Speaker B: Yeah. And across the board with all kinds of different, uh, diseases. But mostly cancer. Yes, we talked a lot about cancer.

Speaker C: That's definitely, like, one of her areas of focus. And the reason we wanted to talk to her is she does have a really unique perspective. It's very clinician focused. She works very closely with surgeons and doctors who are seeing patients to see how AI can be used in a way that's most helpful for physicians and for the patients themselves.

Speaker B: Yeah, it's very refreshing because keeping everybody in the loop when it comes to AI, I think is the right way to move forward. Yeah.

Speaker C: So we wanted to get her kind of unique perspective on things. So, uh, we had our conversation with Amber.

Speaker B: Here's Amber.

Speaker C: So, Amber, thank you for joining us today. We've been excited to talk to you for a while. Your work is so interesting, and we're glad to have you on the podcast.

Speaker A: Thank you.

Speaker C: So, uh, I guess we can start talk a little bit about, like, what would you say your area is?

Speaker A: Yeah, so my area broadly is health, artificial intelligence. Uh, I do a lot of work in cancer imaging. Uh, my background is that, uh, my first faculty position was at a big cancer institute in the US So a lot of my experiences around working with cancer data, and then I also work with other types of data, but I'm very. I really like data. I like thinking about data and thinking about the algorithms we can use to bring that data to bear for patients.

Speaker B: And I'm interested in. Because you're talking about data, but you're also talking about health. So which came first for you?

Speaker A: Uh, yeah. That's funny. So my master's is actually in Comp. Sci. Pure theory. I was in a room by myself doing parallel computation, uh, uh, derivations. And I was kind of bored. I'm kind of a social researcher. I need to be with other people, and I need to be with people that, um, know things that I don't know. I love that. That's my favorite thing. And so for me, the computation came first, and then during my PhD, I switched to. To health or to medicine, and I just loved it. I just loved the idea that we could change the world.

Speaker B: Right.

Speaker A: And so, so the health, I think, came first. And then I started doing some data analysis. And this would have been. Oh, my gosh, my Ph.D. i finished in 2010. So this was before the big, you know, deep learning sort of wave came to all of us. Right. So, um, I think the health piece came first. Then the data piece came. And then in 2016, I was a new professor, and then I was at a cancer institute. And we had all of this data that no one else had. And that caused me to really switch directions because, of course, we had. Deep learning came as an onslaught. Right. And I was in a good position to take advantage of that.

Speaker B: Yeah. So when you're thinking about data and, um, health problems, what are the kinds of problems that are solvable right now with machine learning?

Speaker A: Yeah, that's really interesting. So I think we all have experiences now of going to the doctor and. And there's an AI there, right. It's. It's listening and doing transcription for the doctor. Uh, the doctors like this because they can look at you as the patient instead of, uh. For many years they complained about the fact that they had to sit and stare at their computer and enter in information about you. So now they have these scribes. The scribes can listen to you. And so we, that we see that, um, uh, when we go to the doctor.

Speaker C: Actually, I was at the doctor a couple weeks ago, and, uh, I think it was the first time we've made eye contact while I was talking, because he's always. Because, I mean, he's got to be typing stuff, right. So he's. But like, he was definitely using something that. And it did feel a little. A lot more present. Like, I felt better.

Speaker A: They're looking at you, right. And that's very different. And they're actually. This is going to sound very strange, but they're touching you now. So there was a time when they really. They had 10 minutes to give to you. This is the family doctor. Right. Uh, and they would sit and have to fill in the information in Their system, um, and they didn't really have time to speak to you. They didn't really have time to physically examine you, uh, because that 10 minutes runs out really quickly. We see that. That's a big one that we're all seeing now. I think there's some other really interesting things going on where, uh, we have really great trials now showing that in breast cancer screening. Breast cancer screening is when women go and get a mammography, um, and that determines whether or not they should be screened more for breast cancer. So whether they have a potential, ah, mass or something like that, and then it goes on to another screening. This is a hard problem because there's a lot of women. Right. So, um, it's hard for a person to review every one of those scans. And in some countries there aren't enough radiologists to look at all the scans. And we see that in, uh, India and other parts of Asia that they're just, they don't have enough radiologists. So there's been great clinical trials, randomized clinical trials showing that AI, uh, screening tools for breast cancer are better than humans. And that's actually a really nice application because if humans can't actually look at all the images to begin with, we're not even really talking about replacing humans. Right. We're talking about making sure that all the images get looked at.

Speaker C: Yeah. And so, yeah, it's work that's just not being fully done.

Speaker A: That's right.

Speaker C: So would these be like screening, uh, tools that are using the images that already been collected, or is it a completely different sort of.

Speaker A: Yeah, these are using the images that are standardly collected, which is typically, typically mammography for breast cancer. If you have high risk features, they send you for, uh, an MRI or a magnetic resonance image, uh, to get more information. And that's sort of the pipeline you go through in these, in these screening steps. Uh, and so because it's really a population level AI tool that's being used, so. Meaning it's looking at all women. You're looking for a needle in a haystack in that. Right. You're not talking about 50% of those women having a cancer. It's a very low number. And you're really interested in just moving them on to the next point of screening. Is this maybe something we should look at again? Okay, then we'll look at it again. And I think that's really interesting.

Speaker B: Yeah. So it's like a high recall application.

Speaker A: Yeah, exactly. Exactly.

Speaker C: All right, so that was kind of. We talked a little bit about machine Learning is good at in the healthcare. Are there still problems that it really struggles with? We haven't really figured out how to tackle those.

Speaker A: Yeah, sure. So a lot of this, uh, comes down to statistics and statistical power still.

Speaker B: Right.

Speaker A: That we generate a ton of data. We generate. So if we think about cancer patients, and I think a lot about cancer patients and cancer from a machine learning perspective is really interesting because the patients go to the doctor a lot because they're undergoing treatment. They're getting imaged every four to six weeks. So we're getting a lot of data on them. Um, they're having a tumor removed, and then we're sending that tumor for sequencing. Uh, we're sending that tumor to be sliced up and put on a pathology slide and looked at. That's another image. We're generating a lot of data in those patients. How we put that data together is still a big question that I think we need to solve. Some of it is, again, related to statistical power. If you're extracting a lot of variables from a lot of images from a lot of clinical data points, and you don't have that many patients, you still are running into an overfitting problem. You're still doing a stage, this torturing of data. And I joke with my clinical collaborators that I can predict anything from anything if you give me a small enough number of patients and enough data points. So those are still problems that we have. Uh, AI tools don't magically get you past those problems of basic statistics. I think that's still a very open area.

Speaker B: And you're talking. One of the things you just mentioned was putting data together. Is that a problem? Is that difficult?

Speaker A: Yeah, I think that is difficult. So if we think about. If you just think about what a patient, uh, goes through, think about an individual patient. So they're going to the family doctor, right? They're going to. Let's talk about a cancer patient. They go to the family doctor, they get screened in this breast cancer screening, for example, they test positive, and then they go through another pipeline of tests and treatments and things like that. It's still the case that in those patients, we treat them with a known drug that works in a lot of people. And then when that fails, we treat them with another drug that we think works in a lot of people. And then we treat them when they fail, that we treat them with another drug. And that's not really a very targeted approach. Right. And I think there's still an opportunity to say, what is the best first treatment that this patient should get? And often we don't go to these precision treatments until many lines of treatment have failed. And so then we start looking at, okay, let's look at the, the genetics of this tumor, and then we'll come up with something that's more targeted. But we, we treat everyone to help a few. And I think there's a role for machine learning to play in that. That could be a, that could be done better for individual patients. The other thing to think really hard about is there are all these data points collected on these patients, right? So they image, like I said, um, at multiple time points, they're getting a pathology sample or a blood sample or, um, a urine sample, or, uh, it could be a plasma sample taken at multiple time points. Thinking about that data in three dimensions and even four dimensions, because there's this time access, I think, is really interesting. And I don't think a lot of our methods take into account time and what changes over time. I think there's a huge room for improvement in that.

Speaker B: It sounds like what you're describing is precision medicine. I've heard that before. But is that actually something that's happening in the real world right now?

Speaker A: Uh, yes, for sure. There's some really simple methods of precision medicine. We call those statistical nomograms. These are really fascinating devices, uh, that researchers, uh, statisticians, uh, clinicians come up with for making very simple decisions based on multiple variables. If you look at these things, they're sheets of paper that have rulers on them. And you go through each of the steps. The first ruler might be, is this person male or female? They're male. Uh, let's say this is prostate cancer. Then it might come to another variable. Uh, uh, and then we say, okay, is it over this amount or under this amount? And then we move over to this part of the nomogram. As you get down to the bottom of this map, there's a decision at the bottom. If you're over to this side, then you're positive for pancreas cancer. If you're over to the other side, then maybe you're negative or you're low risk of pancreatic of prostate cancer. Um, that's existed for a really long time. That's honestly a version of machine learning. It's a multivariable model. Uh, and this is just a very simple way to produce those models for clinicians to understand and honestly for patients to understand too. Then you can get more complicated. We certainly have models looking at genetic data. We take your tumor, for example, in the cancer case. Again, we uh, we take a piece of that tumor and we subject it to tests. Um, and then we know that, ah, if there's positive for some particular gene, then we know that you're going to have this good response to some other therapy that we could give you. We still do a lot of that in a univariable way, though. We really think about that as, oh, you have this high. And so we're going to address that. But we tend to not think about it in a multivariable way, which I think would be a place where machine learning could do a lot of work. Um, yes, precision medicine exists. I think these models that are coming online that really take a lot more data, um, are not as used, but I think it's just a matter of time before those get pulled in.

Speaker C: We've touched on a little bit, but I want to bring back, from a patient perspective, if a person is dealing with cancer, what would that move to precision medicine mean for like, uh, a specific person who's having to deal with that?

Speaker A: Yeah, that's a great question. So from my perspective, again, we have a patient, we give them the most common treatment, but that they might not actually fit that most common treatment, but it's most common and it works in lots of people. So we give it to them. Maybe it would be better for them to receive this other treatment that they would, would get downstream much earlier. Maybe that would prevent their cancer from spreading and, uh, from getting worse. And so I think there's an opportunity to rethink this order that things in, um, which we're not currently doing.

Speaker C: And I mean, one thing, I mean with cancer treatment, but I'm sure a lot of other things, I mean, the treatments itself is not always pleasant ordeal, so you can kind of skip, uh, ineffective ones.

Speaker A: Yeah, that's true. And chemotherapy's come a long way. It's now usually a pill, and it's pretty easy to take and to deal with. But something like surgery, you know, if you don't need a complex abdominal surgery, then it would be great to avoid that. Right. And so I think those are the types of things that we think about. Historically, though, surgery is the way to cure cancer. The way that we cure cancer is still that we cut it out. Um, and so rethinking that paradigm does mean a lot of people have to rethink it. But there is that kind of work going on now that people are rethinking how we approach this and where the patient goes first.

Speaker B: I think I would be scared if my doctor was like, you have cancer but we're not going to take it out. I think you're right. It would be. But is that what you're talking about, that that's actually an option now?

Speaker A: Yeah, that's an option now because. So there's cancer that you see, but there's cancer that you don't see. Right. And just because we can see it, you may have additional cancer that we're not seeing yet. We may. You may have tumor cells that are going through your body. And so the idea is if we give you a medication, a chemotherapy or an immunotherapy up front, then maybe we would, uh, destroy all of those cancer cells in your body, and then you wouldn't continue to, uh, uh, have a recurrence of that cancer or that cancer coming back. And so that's a really new thought. It was always, we always thought for decades that a tumor was a tumor and that was the thing you wanted to get rid of. But now we understand that you can have cancer that we're not quite seeing yet. You can have tumor cells that are going all through your body, and that those are the ones that we also want to kill.

Speaker B: That's super interesting. And is that, like, what you're most excited about right now, those sorts of changes, or is there anything else in.

Speaker A: Yeah, I think that kind of thing is really interesting. So I do a lot of work in liver, uh, and pancreas cancer. And I do that because they're really aggressive cancers. And everybody knows about pancreas cancer and how terrible it is. So for example, we know now, and we've done a lot of this in our work, that, um, you can have tumors in your liver that you can't see.

Speaker B: You mean with radiology we cannot see it?

Speaker A: We can't see. But if we do analysis of your liver using advanced imaging AI tools, we can see changes that are precipitating cancer. If you knew that you were going to get a tumor in your liver, what would we do differently to treat you? That's an opportunity for intervention. Uh, this is particularly true in the liver because often your first, first cancer that you get, your primary cancer is in some other place, but then it metastasizes or a, uh, secondary cancer goes to your liver. So if we can, can look at patients, livers and say you're at a high risk of having metastatic cancer, then we can intervene with, uh, a chemotherapy or an immunotherapy to stop that from happening.

Speaker B: So it's kind of like the, the liver is like a

Speaker A: kind of a thermometer Right.

Speaker B: I was trying to think. What's the word? Yeah, it's like a barometer. It's like telling you what's happening. A canary in the coal mine.

Speaker A: Yes, that's right. Yeah, that's right.

Speaker B: And so if you keep your eye on the liver, you can kind of get a sense of what might be

Speaker A: happening, what else is going on. And the spleen too? The spleen. We think the spleen doesn't do anything, but it actually is sort of a barometer too, of everything else that's going on in your body.

Speaker C: Uh, so this is just to be. So I make sure I'm clear. Like, this isn't the case you're talking about with like, breast cancer images, where it's looking for images that if somebody looked at, like, uh, an oncologist looked at it, they'd be able to key. Like these are finding things that even if a trained person was looking at these images, they wouldn't be able to spot that.

Speaker A: That's right. So in our work, we're looking at changes that are really very minute, uh, that we can see now. That said, there are huge advancements in how we take images now. And this is the field of, uh, actually developing imaging. Uh, and so there's a really imaging type that's coming out now. Um, that's computed tomography based, uh, and it's called photon counting ct. And this is, um, we're super excited about this. We think that this is a way to see some of those smaller structures that we couldn't see in other ways. And maybe some of these changes that we saw with some of our analysis, maybe we would be able to now see them with photon counting ct. Wow. And photon counting CT itself is an AI technology. So it was, uh, it's literally counting photons, which is something that couldn't be done before. But now you can get a computer to count the photons. And so you're not doing that by. You're not doing that by.

Speaker B: There's a lot of photons.

Speaker A: There's a lot of photons. And so.

Speaker B: But that's what cameras do, they count photons.

Speaker C: Yeah, exactly.

Speaker A: Exactly. Uh, and so the AI. AI in this case, or machine learning, is really coming at this problem from many different angles. Right. It's helping us produce better images. Also, those images now can be taken with less, uh, radiation dose. So this is another thing we think a lot about is radiation and your exposure to radiation. Now, for cancer patients, we don't worry about this, because your biggest problem as a cancer patient is not that you're going to have a bit more radiation, uh, you have other problems that we're trying to solve. But for the general population, when we have screening tools, having those screening tools be lower dose is really important.

Speaker C: Wow. I had no idea. I mean, I knew there were some, um, advances, but I didn't know just like the breadth of what was being made possible.

Speaker A: It's, I think the most, it was, uh, said by the FDA recently to be the biggest, um, and most promising invention, uh, of the last 10 years in imaging.

Speaker C: Oh my God.

Speaker B: Interesting. I smell a Nobel. That rhymes.

Speaker C: Make some shirts.

Speaker B: I had also heard, um, you say something in another context about radiomics. Is this radiomics or is that something different?

Speaker A: Yeah, the radiomics is the idea of extracting additional information from regular scans that we acquire for other reasons in patients. Um, we use it interchangeably with the idea of quantitative imaging. It was a term that predated, you know, all of this imaging AI, um, but it's all kind of the same thing. You're taking images and you're mining them for additional information.

Speaker B: So like genomics is, is mining the genome for information. And this is just like all of

Speaker A: radiology, radiological images, radiology images, pathomics is for pathology images, transcriptomics, uh, uh, all of these. Yeah, we love omics.

Speaker C: It's, I mean, it's a good suffix, I guess. Yes, yes. Yeah, I mean, so this sounds like some really amazing stuff, but I assume, like, what are the kind of the challenges that we run into as we sort of develop these new imaging techniques and analysis techniques?

Speaker A: Yeah, that's a great question. So it's still, you know, the data sharing and getting access to really great data is a challenge. So one of the things I've tried to do in my work is release data publicly, uh, so that other really, uh, great teams can work with them.

Speaker C: Is that not the common approach or is that, uh.

Speaker A: It's not common because people are worried about accidentally releasing protected health information. Uh, there are major fines for doing this. There are medical institutions across the world that have been fined for accidentally releasing. So people worry about this. And, and it's a, it's a real concern. Right. So if you accidentally release something and somehow we figured out who that person is, in many parts of the world, that person would be denied insurance.

Speaker B: Right, right.

Speaker A: I forgot about that. So this is a concern. And so, um, we think about this pretty hard. Um, in Canada, the direction that we've moved to is these very high security, uh, platforms for doing this kind of research. Uh, one challenge with that is that you need to have the data next to the very good compute to do the work. And I think that's one thing that we're doing a good job of building out in Alberta. Um, and so the ability to be able to get the data as well as, so not just the images, but also the clinical data, the relevant clinical data that goes along with those images and then have that next to the compute is really key to doing this really well. The other piece, do you mean like,

Speaker C: like physically close or.

Speaker A: Well, sort of, yes. Because images are, uh, take a lot of computational power to, to churn through. Right. And making these big, big deep learning models requires, uh, gpu. And so you need that imaging data in the right computational context to do that work. But the other piece you really need is to talk to the clinicians that actually see those patients. And I see that as a big gap in the field that we have many colleagues that don't do that terribly well. Um, and that's really important because sometimes as AI researchers we come up with these things that we think are brilliant and they're really not, but they could have been, right? They could have been. So one of the things that my group does is takes really great questions, we find the really great data and then we release that data publicly so that we can have grand challenges and different groups can compete, uh, to get the best answer. And so we did that. So, uh, we did this in Ah, 2018. Um, so a really common problem in my field is to figure out in a scan, where's the liver, where is the spleen, where's the pancreas, um, where are the tumors? This is a process called segmentation. Uh, in 2018 we released the largest data set across 10 different types of organs and tumors. Um, this was called the Medical Segmentation Decathlon.

Speaker B: It's such a great name.

Speaker A: It is such a great name. This, this moved the field forward. This moved us into generalizability of image segmentation so that we can train models on one type of data and those models can generalize to other tumors or other organ types or other diseases. That was the first time that happened and that has really made all the difference in our field. Now we can take, let's say you have a CT scan, maybe you've got, ah, maybe you've got some abdominal pain. So we do a CT scan on you and maybe it's fine, but we can actually take that scan now, um, and apply AI tools to it to see if there's other things going on in your scan that A radiologist wouldn't have looked at. So typically when you, when you get your image, so you get imaged and then it goes to the radiologist, they're going to look for the thing that explains the abdominal pain, but maybe you have some degeneration, uh, of one of your vertebra. So that can be something that they wouldn't necessarily look for because that's not what the purpose of the scan was. But we can use AI tools to then mine existing data for other information that could potentially help you.

Speaker B: Okay, so you mentioned this grand challenge, which I've heard you talk about before. Did you run it more than one time?

Speaker A: It continues to run. So it's actually available. You can continue to access the data.

Speaker B: Is it the same data?

Speaker A: Yeah, it's the same. And so you can benchmark any method that you have now and it is often used in papers to benchmark, so it's considered, ah, a very good benchmark. And in fact, when my students join the lab and they want to work on image segmentation, I say, okay, so first go to this challenge and work on it.

Speaker B: Mhm. Cool. And so have you joined the Grand Challenge in subsequent years? And one.

Speaker A: That's a very good question. So the winner of the original challenge, uh, changed the field.

Speaker B: Oh, cool.

Speaker A: And that is considered still a method, uh, that's to beat now. And it still kind of works the best across everything. There are situations where other methods work, uh, but it's really quite amazing.

Speaker B: So it just goes to show that data sets are important. They cause those kind of, of, you know, big advancements.

Speaker A: Yeah, yeah. And data sets that are really considered within the application area of importance. So there was a lot of thought put into that data. The data was annotated by clinicians. I had practicing radiologists and surgeons doing the annotations. So it was really quite good. And really we thought really hard about the data that we used for that.

Speaker B: And how long did it take to

Speaker A: do all of that, to do the annotations? So we had done them as part of our research program for other things that we were doing with that data. Um, and then actually the way this challenge came about is that I was at mackay, which is our big medical image computing conference that takes place annually. And I was walking around and looking at all these. Excuse me, I was walking around looking at all of these posters and thinking, wow, they're not benchmarked. No one has access to good data that's annotated really well. And remember, this would have been 2017. Right. So we had to do, uh, different world then. Um, and I thought, wow, I have all this amazing data that we've worked so hard to annotate for really great clinical questions. Uh, and let's see if we can put this to work. And then I collaborated with somebody from ucl, uh, in London to make it happen. Yeah.

Speaker B: So is the data from America, Is it from.

Speaker A: Yeah, it's mostly. So it's actually mostly the test data is from my group, which at the time was at Memorial Sloan Kettering Cancer center in New York.

Speaker C: York.

Speaker B: Uh, which brings up a different question, which is about bias. So if you have a data set from the Western world, how does it generalize to other populations?

Speaker A: Yeah, that's a great question. And it's a bit of a difficult one to unpack. So there's biological reasons why there's differences in data. For example, if we're looking at liver cancer, liver, uh, cancer in Asia is different because of hepatitis C. So hepatitis C is more widespread there. Um, and so it's also a little bit related to geography. I hesitate to say ethnicity. So if you are Asian and in North America, because you don't have that propensity to hepatitis C, this is different. Right. So there's definitely genetic variation, there's biologic variation, um, that comes into play that you have to think about. There's also clearly, in the US not everybody has access to the same care. We know in the US for example, that AI algorithms are optimized on cost savings in hospitals. This is one of the reasons why we need to retrain algorithms and have our own models in Canada, because actually, we don't optimize on cost savings in the health system. There's many models that are made in the US that don't apply to us,

Speaker B: to be clear, because we're optimizing for health outcomes.

Speaker A: Because we optimize for health outcomes, we actually have this crazy idea that people should get better if they're treated. Right.

Speaker B: Right.

Speaker A: And so we have other metrics that

Speaker B: we use, and cost must be like a factor, but.

Speaker A: And cost is certainly a factor, but it's not the only factor.

Speaker B: Exactly. Yeah. Right.

Speaker C: Well, so, I mean, you've mentioned that there isn't always this. This environment of sharing that data. Uh, and it was something that was very important to you. What has to happen to make that happen more? Is it just a willingness or are there structural issues?

Speaker A: Yeah, that's a great question. And maybe I could talk a little bit about Canada and how we organize our health data and how we organize our health systems. Our health systems are provincially based in Canada. And this was one of the reasons I actually moved back to Canada from the U.S. uh, in the U.S. it's really, uh, uh, hospital by hospital. There are some health systems. So there's the University of California Health System, which is many hospitals and tens of millions of patients because it's California. But this is not uniformly true. There's also in the U.S. uh, insurance companies that have a tremendous amount of data. There's pharma companies that make drugs that have an incredible amount of data. But generally speaking, hospitals keep their own data and they don't share. Um, and again, it's because they're worried about, uh, leaking information about patients. They also just don't know how to do it.

Speaker B: Well, they're also businesses.

Speaker A: And they're businesses. It's just a very different environment. So in Canada, each of our provinces runs their own health authority. Um, and so in Alberta, we actually decided years ago to enter into a common medical record system for all patients. So every patient in the province of Alberta is part of, uh, what we. It's called software. Uh, it's software called epic. And so we have this EPIC electronic medical record system that everyone's data is going into. So that, number one, is going to just set us off on a much easier foot to take advantage of the data. We are the only province in Canada that's done that. Um, and in fact, I'm from Ontario. Ontario tried to do this in 2015. I think this was the E health scandal. You can talk, you can look up about this. And it killed the common medical record for Ontario.

Speaker B: Yeah, that's too bad.

Speaker A: And it's too bad. Um, so Alberta, we have five million Albertans in a common medical record system. Um, and so we're building a trusted research environment to work with that information, um, and pairing that data with the compute, like what we were talking about before.

Speaker C: And would that data also be like, we talked about generalization, the stuff from Alberta applicable even to the rest of the country? Like, I know, for instance, like Alberta, I think has an average age, is much lower than most other provinces. Or are there still generalization issues even within the country?

Speaker A: Yeah, I think that's tbd. I don't think we know the answer to that, but I think for, for some very simple problems that would make a big difference in the health care system. I think there's a lot that we could do to generalize it across the country. Um, I'm hoping that Alberta is really demonstrating to the rest of the country the need to get the act together. I think it's actually ethically imperative that we have this data available. It's not just about privacy legislation. That's obviously super important. We should probably talk about privacy legislation. Um, but there is an imperative, an ethical imperative for us to work with this data in collaboration with the people that actually see these patients. Our privacy legislation in Alberta is quite strong. And so there are rules by which we have to follow to use this data. And that's why we have it in secure systems. Those systems can be audited. Uh, you need to do the thing with the data that you said you were going to do. And frankly, I think that's fair. Right. I think that's an expectation. Um, I'm certainly a public servant. Right. So I'm in service to the public. So to me, this is, this is a reasonable use of data and I'm happy to meet that expectation.

Speaker C: Right. Yeah, I guess. Yeah. It's important to point out when we talk about, like, you know, privacy concerns and stuff, like, those are there for a very good reason.

Speaker A: Yes, there's very good reasons for that. Uh, and we need to. We need to protect the population, and we need to protect populations, um, that have been historically taken advantage of with data, uh, including indigenous populations. And so actually, in Alberta, we also have very strong ties to our indigenous communities. So if you make a request of health data in Alberta and you want to use indigenous data points, you need to get the approval of those communities. And that's actually factored into how this works in the province, which I think is really cool. And I think, again, we're the only province that does that.

Speaker B: That I want to know, though. So the reason we're doing that is because we want to keep or build the trust. How important is that?

Speaker A: I think it's really important to build trust. And we saw this in the pandemic. This happened in Ontario. So we were doing Pandemic. Yeah, the pandemic happened in Ontario. We saw in the. During the pandemic, we were doing, um, wastewater surveillance, which we still do in Alberta. Yeah. So samples of wastewater are taken and then various tests can be run on that. That those are population level tests. And one of the things that we were doing in Ontario is taking samples from indigenous communities and running them for STDs and then reporting. Oh, this community has all these STDs. Well, it turns out all of Ontario had STDs. It turns out that the pandemic was an interesting time for many, for some people, and a busy time. And so there were reports on the CBC. Oh, the indigenous populations have all these STDs, when in fact that was not true. It was the entire population of different.

Speaker C: They just weren't doing the same testing on. Well, they weren't reporting it or they weren't reporting it. Okay.

Speaker B: Yeah.

Speaker A: And this is how we, we do bad things with data.

Speaker B: Yeah, yeah. And it's not. Yeah. Racism is not over.

Speaker A: No, no.

Speaker C: Yeah. But it's interesting because I mean, I think when people hear doing things with data, that sounds like a neutral.

Speaker B: Right.

Speaker C: Uh, activity, but like it's, you're still making choices along the way to what to do, what to report. Uh, so on the kind of, the trust thing, because I mean, obviously working with clinicians very directly is a very important part of like what your approach is. So what is the temperature? Like, what are you hearing from clinicians when it comes to bringing in AI to their practices?

Speaker A: I think it's changed a lot in the last five to 10 years. I think, um, initially they were unsure of it and then now a lot of clinicians are trained with AI. Residents are typically exposed to this. Medical students, they're becoming, uh, our clinicians and they have some understanding of AI. Already I think people are getting a lot more savvy. Um, LLMs really changed the game. LLMs, uh, have made it so clinicians can do more research with data, which is I think a double edged sword. Um, so I think in general it's much more accepted now. You know, Jeff Hinton saying that all the radiologists are going to be replaced by AI did not do our field any favors. And I think making those kinds of statements, I understand as academics we like to be controversial, but it wasn't helpful. And um, I think we've moved past that now. And we understand that clinicians that don't know how to use AI are the ones that are going to be disadvantaged, but that it's not going to replace them because there is a special sauce to what humans do clinically.

Speaker B: Yeah.

Speaker C: And I mean like I, when I go to the doctor, I, I don't want a computer to tell me things like I want that human.

Speaker A: Exactly.

Speaker C: Connection. And I mean like the, the care that they like I have a caring doctor. Like that's, that's part of it as well.

Speaker A: That's right. And there's a gestalt that clinicians have so they can they. So okay, a medical record only contains information that's collected by the medical record. So uh, there's a lot of other information about patients that we don't have in medical record records that are. Can be perceived by the human eye. Right. And there's things about your gait. Right. So if you're walking a little funny, you know, we had, uh, when I was in New York, we had something called the get up and go test. And so if you could get up out of bed and go, then you could be eligible for surgery. That's not something we. That's not information that we collect. Right. But it's just information that we know that if a patient is well enough to get up and go, they're well enough that they can undergo an aggressive surgery. Actually, we had a very nice model that we had made that predicted outcomes in pancreas cancer patients from imaging. I had, uh, one of my surgical colleagues said, well, I can predict what's going to happen to patients from their images. Um, he said, actually, I also use their ankles. There's a biological reason for this. If you have swollen ankles, it has to do with diabetes and all these other things, things that I don't fully understand. Um, but we did that. We sat, uh, a series of surgeons and radiologists down at a computer. We gave them basic clinical information as well as CT scans that they could review. And we said, is this patient going to survive past two years? In that study, the two radiologists, uh, with the most experience beat everybody and beat the AI. There is this special thing that humans do with gestalt, uh, that is still not something that a machine can do. And I don't think we fully know how to, uh, how to. How to capture that. Maybe if we have rooms like this that have cameras all over them and we have our patients going in those rooms, maybe that would help, I don't know. But there's, you know, there's. I have so many stories about clinicians perceiving something and understanding that something was about to happen. There's a great story that one of my former colleagues gives. Uh, she's an eye surgeon. So when they do surgery on your eye, uh, this is going to get really gross. They actually peel your face forward, Amber. And so we may have to cut this. So they actually peel your face forward. And she had an endoscope in this patient's eye. Um, and she didn't know why, but she just felt like the patient was waking up.

Speaker B: Oh, my God. Oh, wow.

Speaker A: And she just, in that second, pulled the endoscope out, and the patient woke up. But it was okay because she pulled it out, flipped the face back, and then it was. It was okay. And then they sedated the patient again. But I don't know how you train a robot to do that.

Speaker B: Yeah, yeah.

Speaker C: I. I actually had. I'm not going to get too many details because it's not my story to tell, but I did have a friend who about a year ago, went into their doctor's office, and the doctor did not. Could not say why, but they're just like, I'd like to get, like, a blood test from you. Because he just. He saw him every six months, I think, and he was like, you're different in a way that I can't. And it did reveal, I mean, what could have been a pretty serious problem. Uh, but luckily, you know, it was able to get treatment fairly easily and fast. But it was just like. Like, the doctor even said, I can't tell you why, but, like, this is

Speaker A: something m. There's something going on. There's some deviation from. From normal that I'm seeing. Yeah, Yeah. I have a friend and former colleague that's a radiologist, and he can see things that other people don't see. And it's because he's looked at so many images for so long. When I was at the cancer center, one of the things that we do is have what's called tumor board meetings. These are for specific cancers. Um, we were in the liver and pancreas. Cancer, tumor board meeting. This is radiologists, oncologists, surgeons, everybody that could see these patients. The radiologist sits at the front and they put up images, and then they discuss the images, and everybody asks questions. So the radiologist puts up the image of the liver, and we're all looking at it, and everyone's getting cranky because there's no tumor in the liver. Why are we looking at this patient? We need to get on with our lives. Um, and he starts playing with it, and he says, no, everyone be quiet. And he just stops listening, and he starts playing with the image and playing with how the contrast is working. And all of a sudden you see this 8 centimeter liver cancer just kind of appear out of the background. Right. And could we have seen that with AI Maybe. Maybe if we trained it right. But that was magical to me, that this was a patient that had been passed around to a number of different institutions with all of these symptoms that nobody could. Could understand and pointed to liver cancer, but nobody could see a liver cancer. And then this guy sees it.

Speaker B: I wanted to point that out that, like, we often talk about, like, oh, AI is. It surpassed human performance on a particular task, and it often cannot Surpass the best human. The best humans are still better.

Speaker A: The best humans are still better. So how do we make sure we keep training those best humans?

Speaker B: Right, right. We can't get rid of all radiologists because AI is average. Like we have to keep.

Speaker A: Yeah, exactly. We can't get rid of all the programmers.

Speaker B: Right.

Speaker A: Because there's really good programmers that have that special magic. And so how do we use it to augment but still train people? I don't. Yeah, I think a lot about that.

Speaker B: That. That's a really good question.

Speaker A: Yeah.

Speaker C: Is it one we have an answer to or is it still.

Speaker A: I don't think we know. I don't think we know.

Speaker B: Yeah.

Speaker C: Uh, so I want to take things in a bit of a different direction still healthcare. But I've heard you talk before about how machine learning might change the way we test drugs, specifically our use of animals in testing. So what would be the connection there?

Speaker A: Yeah, so let me take a step back. So in drug development, we typically test drugs on animals. Animals. Uh, and this doesn't work very well. So something like 80 to 90% of drugs, uh, that are tested in animals and that work in animals do not replicate in humans.

Speaker C: It's that many.

Speaker A: It's that many.

Speaker B: And the inverse must also be true. Then we like. It doesn't work in rats. So we're just not going to continue.

Speaker A: Exactly, exactly. And so that's fascinating to me that. And this is a whole machinery m of animal testing that goes on to try to look for these needles in haystack. So something we have in animal testing is called a mouse knockout.

Speaker B: Nothing to do with boxing.

Speaker A: Nothing to do with boxing. So the idea of a mouse knockout is that we remove one gene from that mouse. And so because we've removed that gene, then we can do different tests on that mouse to look at how a drug works that affects that gene.

Speaker B: Right.

Speaker A: And this is the basis of much animal testing. Something that's happened over the last decade, is that we've invested in the idea of generating a lot of data on patients, whether or not those patients have disease. We have. The UK Biobank is a great example of this, where we've, uh, prospectively enrolled patients, we've consented them, and then we collect all kinds of data on those patients. Longitudinal imaging, so imaging over time, uh, genomic studies, all different types of studies. Um, these patients at the time of enrollment may not have had a disease. But as we study them over 10 years and we keep collecting information on them, they start to generate, they start to develop cancer. They start to develop heart disease. And then we can start looking at different signals within those populations. In the US we have something called all of us, nih, all of us, uh, where they're doing something similar in collecting the state data. We always thought that there wouldn't be enough data, that there wouldn't be enough variation in the data. If you're looking at pancreas cancer, for example, it's still an extremely rare disease. So you would need tens of millions of healthy patients to follow who's going to get a pancreas cancer. So that's what we always thought. Well, this doesn't seem to be the case. It seems to be the case that we actually have enough information in these databases, banks to do some of this analysis. Okay, so we've got animals that are, that, uh, with drugs that are developed based on them that aren't replicating. We have these huge data banks. Um, in the 2020s, this idea, uh, idea of Mendelian randomization came up. And the idea of Mendelian randomization is that within the population of humans, there are genes that are turned off and on naturally just by randomization. This is the idea that we developed of a human knockout. This idea that you could have, if you had enough humans and their data to study, that you could have genes turned on and off and then you could study naturally. Naturally without turning their genes on and off. Because they hate when we do that in humans.

Speaker C: Right. I wouldn't appreciate having some of my genes. I think I'm using most of them.

Speaker A: What we showed is that you can do analysis of UK Biobank data and you can show that this natural, uh, randomization of turning these genes on and off gives, uh, you enough variability that you can then test. In our case, we looked at, uh, um, cardiovascular outcomes. The reason we did that is because cardiovascular outcomes are very common. So that means that within these 500,000 patients, there's enough of this natural variation in the genes and then there's enough outcomes in cardiovascular diseases that we can actually start making these human knockouts. And so we actually showed that you could replicate. I think it was 27 cardiovascular trials of these different genes turned on and off and targets made for these genes in UK Biobank data. We were shocked that it worked. We really didn't think it was going to work, but it worked really well. And now we've done this across many, uh, different targets. Um, these genes that are turned on and off, then the idea would be, as the UK Biobank data gets more mature, these patients start to have other things wrong with them, like cancer, like, uh, other diseases that you could actually study. These human knockouts. And that's, that's our idea. And it's actually working. And, and we're the most surprised that it's working of everyone.

Speaker C: Uh, so, like, it would be, uh, less, not only more effective than the, like, the animal trials, but, I mean, obviously you're reducing that ethical concern about, you know, using animals.

Speaker A: Yes. And I will say it's not entirely altruistic. Right. So the National Institutes of Health in the US Wants to see these types of computational models used instead of animals. The FDA in the U.S. uh, Health Canada, as well as, uh, Europe, uh, is also doing this kind of work. And it's a cost savings. It's not to save the animals, to be completely honest, but it's a huge amount of money in this machine of pharma and big pharma to use all of these different animals and then have a very low success rate of the. Those drugs.

Speaker B: So to do this work that you've been talking about across all of these different kinds of cancers with, uh, all of these different kinds of computational techniques, you must have people from different fields in the same room.

Speaker A: Yes.

Speaker B: What's that like?

Speaker A: Yeah, I think the big one. Okay. There's so many fields, Right. So first of all, you need to talk to the people that actually see the patients or the ones that are actually involved in that particular type of research. So, for example, in our UK Biobank study, uh, the initial idea came about that came from a geneticist, of course, because he understands, uh, Mendelian randomization. The last biology course I took was grade 10. We dissected the fetal pig. I can tell you that we did not talk about Mendelian randomization. So this is where you have to collaborate with the people. And he came to me and he said, you know, I have this idea. And I said, oh, I know how to do that. And then we did it with a computer. Like, I know how to solve this problem with a computer. And that's, I think, what needs to happen when you want to do things that are really big ideas that are working at the edge of your knowledge and at the edge of the knowledge of your collaborators and of medicine and, and all of the things, then you need to bring those people together and you need to figure out how to talk to each other a little bit. But, uh, for me, I think epidemiology and biostatistics is still super important. If you're doing AI on some cohort that doesn't make sense. Garbage in is garbage out. You really want to work with the people that understand these populations, bringing in the surgeons that see them, the radiologists that look at the images, the pathologists that are assessing benign, uh, and malignant things like that. I think that's really important to solve the most pressing problems of our time.

Speaker C: So kind of onto those pressing problems, uh, we're going to ask you to look to the future a little bit.

Speaker A: Okay.

Speaker C: Uh, so I mean, let's say like, 10. I don't know if 10 years is too much, but, like, what would you like to see when it comes to AI use in the healthcare area within the next 10 years?

Speaker A: Yeah, I think what we're going to see is an acceptance of AI, at least from a healthcare perspective, because I think there's real ways that we can make improvements for actual people using artificial intelligence tools, probably ones that already exist. We probably honestly don't even need new ones, um, that make a difference for patients. Right. And I think that would be a huge win for AI. It would be a huge win for medicine. It would be a huge win for patients. Um, and so I think we're going to see that. And it's going to be, uh, tools that are simple, it's going to be tools that are complex. And I think that's really exciting. And I think LLMs really brought AI to the clinic for better or for worse. So you're getting experience now with AI in the regular world. And so I think certainly in medicine, we have the opportunity to be the good guys and to be the ones that do it right. And to be the ones that do it while mitigating bias and considering AI safety and all of these very important things. Um, and I think in 10 years, we'll be at a position where we've actually changed outcomes for patients and changed and changed the world, really.

Speaker C: I hope that, too. I like that vision. Well, uh, thanks very much for joining us. Uh, I feel like we've really only really scratched the surface here, and I'd love to keep you here and ask so many more questions. Questions. But we know you have to go on and do the work that we're talking about, so we can't keep you.

Speaker A: We'll do a sequel. We'll have a sequel.

Speaker C: We absolutely will want you on again. 100%.

Speaker A: Yeah.

Speaker B: Thanks for coming.

Speaker A: Thank you so much.

Speaker B: What a cool conversation with Amber Simpson.

Speaker C: Yeah, it was. I could have kept going. Like, there's so much that we only really touched on, and I want to

Speaker B: know more about yeah, I'm sure we'll have her back.

Speaker C: Yes, yes, absolutely. And I mean, yeah, the talk she was talking about, the grand challenge, that was. I mean, you obviously have a lot of thoughts on.

Speaker B: Um, yeah, I think it's. I mean, data, the way that we choose how to test AI has an impact on the AI that is built. And sometimes we kind of. We act as if data is an afterthought, and we really. We should be putting it more front and center. And I think that's one of the things that Amber does.

Speaker C: Absolutely. And I did, like, you know, I definitely got the impression she's a conversant.

Speaker B: She saw.

Speaker C: She saw something that was missing, and she's like, I'll solve that.

Speaker B: Yeah. Yeah, I like that better. I'm glad she's here at Amy.

Speaker C: Yeah, it was a joy to talk to her. Uh, so every month we bring you more stories about the science behind AI. Uh, and if you'd like to stay up to date on those, the best way to do that is to subscribe. So we are on Spotify, we're on Apple podcasts, and for every episode, we do a video on YouTube.

Speaker B: Yeah. So if you're looking at us right now on the youtubes, uh, jump down in the comments and tell us, have you seen AI in the doctor's office? That's one of the things we talked about with Amber. Um, do you think it's helpful? Have you enjoyed seeing your doctor's eyes pointed at you more often?

Speaker C: Making more eye contact? Yeah. Yeah. We love to hear everybody's thoughts, experiences. So just, uh, connect with us there.

Speaker B: Yeah. And with that, I'm Elana Fish.

Speaker C: I'm Scott Litwall.

Speaker B: And this has been approximately correct.