WEBVTT
1
00:00:02.000 -->
00:01:16.000Welcome to UVA Data Points. I'm your host, Margaux Jacks. Lung cancer kills more people than almost any other cancer, largely because it's caught too late. In this episode, we explore who should be screened and the tricky balance between early detection and the risks that go with over-screening. Joining the conversation are Jeffrey Bloom, a data science biostatistician and Interim Stevenson Dean at the University of Virginia School of Data Science, and Dr. Melinda Aldrich, a cancer epidemiologist and professor in the Department of Medicine and Division of Genetic Medicine at Vanderbilt University. They unpack why deciding who should be screened is far more complicated than just asking about smoking history. From the rise of lung cancer in people who've never smoked, to environmental exposures like radon, to genetic differences in how bodies process nicotine, let's take a look at the science and the stakes in catching lung cancer sooner.
2
00:01:16.000 -->
00:01:19.000Welcome, everybody. Today we are here with Jeffrey Bloom and Melinda Aldridge to discuss lung cancer screening and what the science says about how many lives this testing could save. Jeffrey, tell us a little bit about your background.
3
00:01:19.000 -->
00:02:48.000So, I am a data scientist biostatistician by training. I guess we start at the beginning. I was born. Some things happened, and I got my Ph.D. at Hopkins in biostatistics. From there, I went to Brown University for eight years. Was on the faculty there in the department of biostatistics. Did a lot of work on lung cancer, well, general cancer screening, mostly through radiology in a cooperative group there. Then went to Vanderbilt, where I met Melinda. I was there for 13-14 years and then came to UVA in the summer of ’21. So, my area of expertise is broad in statistics. I study things like multiple comparisons, how do you measure evidence, things that are called foundations of statistics. I do a lot of clinical trials work, and now in my current role, I spend a lot of time trying to connect machine learning methods with traditional statistical methods and balancing the goals of various models with whether they actually want to just do prediction or whether you want to be able to do inference and learn from them about what the data say. And it turns out that models don't have to do both things very well. They can only do so. Most models only do one or the other very well. So that's what I do. I've had the pleasure of working with Melinda for a long time. I'll let Melinda talk now.
4
00:02:48.000 -->
00:03:50.000Oh, thank you. So yeah, Melinda Aldrich. I am a cancer epidemiologist, and my background is: I went to UC Santa Cruz for my undergraduate degree and got a degree in biology there. And then didn't really know what I wanted to do, and went and worked at Genentech and biotech industry for a number of years, about 10 years, and then discovered epidemiology and went to UC Berkeley for my master's as well as my Ph.D. Then ended up at UC San Francisco, did a postdoc in genetic epidemiology, and then from there moved to Vanderbilt, where I got my first faculty position and have been here now for 16 years, and really have a focus on lung cancer and research in its entirety across understanding what causes the disease through offering opportunities for screening and early detection, and it's here at Vanderbilt that I met Jeffrey.
5
00:03:50.000 -->
00:04:11.000Well, thank you for sharing the story about how you guys met at Vanderbilt and started working together. I think that leads great into talking a little bit about lung cancer screening and why is it so complicated to determine who is someone who should be screened?
6
00:04:11.000 -->
00:04:14.000It's complicated. You want to go first? You want me to go first? It's a big.
7
00:04:14.000 -->
00:06:09.000Yeah, I might even take us a brief step back into even why we started to focus on lung cancer screening. So, I mentioned I was at UC San Francisco before I landed here at Vanderbilt, and there was a group there, a tobacco research group, that really dug in deep about different patterns in smoking behaviors, why people are addicted to tobacco, and trying to understand the biology of it. So then I moved to Vanderbilt, and in the early—so around 2013—there were guidelines published for lung cancer screening. And so at that time, then became interested in how this might play. Out in different populations and knowing that there are differences in smoking behaviors across different populations, and it can be quite complicated because lung cancer is the only major cancer we screen for based on a behavior. So we screen based on your age as well as your smoking history, and there's no other major cancer—breast, colorectal, prostate, etc. That it's only really based on age and sex, and so that's in part what makes it so complicated because it includes this behavior that is incredibly difficult to capture in a very simple way, we try to do it as best we can with asking people questions about their smoking histories, but capturing that over a lifetime is challenging. And there are various blood tests you can do, but those are expensive and time consuming to quantify how much somebody has smoked, but it's complicated. And of course, not everybody that develops lung cancer has smoked. So that's my answer.
8
00:06:09.000 -->
00:08:43.000Yeah, no, I'll add to that in a sec. But I was thinking we met because we were working with this group together, basically, is how we got paired up. The group is still there. It's a fantastic group. It's led by a great guy, and actually one of the other principal investigators that is there was actually a student at the time who we helped get his degree. And there's fantastic staff there, and they were all focused in this area of lung cancer. And one of the guys who led with Melinda was a thoracic surgeon, and so I had students in the group, and we started talking about, you know, well, what do you do to find the people who need intervention, and so cancer, you know, is hard to cure, and so you want to do everything you can to find it early. So, there's this idea that by screening you can make your treatments better, or you can make it easier to treat the disease. You catch it earlier, and so we started talking about, well, how do you screen, and why do you screen? And who do you screen? And what's interesting about lung cancer is that there are real risks. Like, the screening isn't perfect. The people you choose to screen may or may not have the disease, and you'll miss people who have the disease, and you won't even screen them at the outset. And you can't do anything for them because you told them I don't want to screen you. Screening can be very expensive. It can also be very risky if there are downstream things that you need to do, like, sort of take a biopsy of a nodule or something like this, which we work very hard, actually, not to do in the in the lung. So, I think there's a lot of interesting big-stakes decisions for individuals who may be at risk for lung cancer. Lung cancer, you know, is a leading reason for death from cancer now. I think in the U.S., [it] might even be number one now. Worldwide now. So, you want to do what you can to catch it earlier, but it turns out it's not so easy, and it's very hard to play catch-up. So, the screening business is important to get right to the best that you can, and we don't have a ton of tools to do that, as Melinda was alluding to. [
00:08:35] Margaux: The complexity there sounds like a lot. So, how does data science factor into lung cancer screening and those decisions?
9
00:08:43.000 -->
00:10:49.000Well, yeah. That's a great question. So, besides being people who get lung cancer, data scientists get lung cancer. You want to figure out what tools work and what tools don't work for screening people, and the way you do that is you screen a lot of people with new tools. Some people who have cancer and some people who don't. That you determine in ways that we can talk about, and you want to see if they're effective because what you'd like to be able to do is, out of the general population, identify the people who are most likely to develop lung cancer and screen those people so that people who aren't likely to develop lung cancer don't have the burden of screening or the potential risks that go on with the screening procedures. And the only way to get at that is to collect a lot of data, and you follow people over time. You see who develops lung cancer, and then you look back in time and say, "Well, were there unique characteristics that could have told us they were at risk for lung cancer. The obvious being, well, okay, if you smoke, you're more likely to get lung cancer. And so, while that sounds simple, how much people smoke and how often people smoke turns out to be very, very, very big factors in what their risk is going to be for getting lung cancer. And now recently, we're seeing a lot of people who don't necessarily have a history of smoking who are developing lung cancer and they're younger, so the data sort of give you a preview of where you need to look, and then you have to understand how to handle those data, how to model the trends that are going on, and get a sense of what's going on? So, you can propose an intervention, and then you try it out in something of a thought experiment. You do some simulations, or you try to make it work. Melinda has been part of a coalition that has collected a lot of information on people who may be at risk for lung cancer, and we've tried out our ideas on that population in theory, essentially by running our data, their data, through the models, and that's how we've been able to get traction and make contributions to the field.
10
00:10:49.000 -->
00:10:58.000You spoke a bit about smoking as a possible risk. How do environmental or occupational exposures play into that?
11
00:10:58.000 -->
00:12:35.000Yeah, so smoking is the leading cause of lung cancer, but certainly there are other well-established environmental risk factors, causes. Radon is another, air pollution. So, there's a lot of interest in trying to understand whether or not we should be including these in some of our screening models to predict who's developing lung cancer. Other exposures of interest, particularly among Asian populations, are cooking fumes. So often, if you're cooking foods that are at really high temperatures, those are releasing smoke and some exposures that we think may also play a role. So, there are a number of environmental risk factors that are currently not considered in the current screening guidelines, but given that, as Jeffrey mentioned, we're starting to see some emerging populations developing lung cancer, more never-smoking individuals—younger individuals in their 20s, 30s, 40s—really outside of the age group that we expect. Usually, you know, on average, a typical lung cancer patient might develop lung cancer in their late 60s, maybe early 70s. So, to see someone developing lung cancer at such a young age is really shocking, and so we're trying hard to understand. Are there these other environmental risks that we should be considering for who should be screened? So, it's an incredibly complex puzzle to figure out, actually.
12
00:12:35.000 -->
00:12:43.000Does location or geography play a role? If you know, you think about some of the cancer clusters or things like that-is that a factor at all?
13
00:12:43.000 -->
00:13:26.000We think so. Again, I think it's, you know, maybe where urban populations might have greater exposure to, say, air pollution, for example. Rural populations tend to have higher smoking histories. We know there are geographic differences by continent as well, with some of the Asian populations, countries having very high burdens, say, particularly in China. So yeah, there are very strong geographic patterns that appear, whether that's due to environment or some genetic factors that are playing a role is probably a bit of all of the above.
14
00:13:26.000 -->
00:14:43.000There can be some very local effects as well. Nashville's built a lot on limestone. So, when you buy a house in Nashville, if your basement's in limestone, you usually have it checked for radon, and the solution is to basically put a fan in the basement to mix the air. But you know, people who have exposure to radon can obviously [be] high-risk for lung cancer, especially if you're breathing it all the time. And so, you just may have a house that's on a chunk of limestone. Nashville tends to have a lot in the surrounding area. Other places have less, but one side of the street may have this risk, and the other side may not. The hard part that we have—you were asking us about data—is getting that data. Like, we don't know people's that much exposure, and the same thing with cooking and fumes. Usually, the best we can drill down to, or get really solid data as occupational hazards because some occupations tend to have more exposure than others. But it's one of those things where if you imagine a world where you collect data on everything, you might eventually someday have information on this that you could use for a model. But today, we don't have it, and so it makes these predictions and identification of people at risk very hard.
15
00:14:43.000 -->
00:14:50.000How do grants and funding shape the data you're able to build and compare across models?
16
00:14:50.000 -->
00:14:54.000It's probably the biggest thing that shapes what we do.
17
00:14:54.000 -->
00:14:55.000Actually, it shapes everything. The reality of it.
18
00:14:48.000 -->
00:17:00.000I mean, for me, you need a team. To tackle these problems, you can't just do it by yourself. And even if you hand someone a data set, and someone's really good at modeling, they can't come up with all the answers on their own. The problems are just too complex, and there's too much going on. So, you need a team of people who maybe have helped collect the data, a team of people who can help you understand and parse that data and organize it in the right way. You need a team of people who can help you think about what data you need. You need someone who understands the ideology of the disease and what you should be looking for in the model. You need someone who can put the model together. You need some students to help you get all that work done, and then you need someone who can help you get that up on a website or somewhere, or get your risk models in use. And you need someone on the patient side who understands what the patients go through. So, you need a team, and the only way to really put that together is with a fair amount of funding. You need to, you know, you need 10% of this person's time and 15% of this person's time, and that adds up. So, you got to bring the team together and get them all aligned and working, working together. That was one of the great things about working with Melinda is. I think we have very complementary grant writing styles, which is good and bad, because it can be frustrating sometimes. But when we write these things—and we've written a lot of them—I feel like we've been actually quite successful in getting them, having just got a couple big ones. You need that partner that you can work with, that intellectual partner that you can work with, who can tell you, “Oh yeah, this idea is good.” That idea is good. Here's how you can help say that, so people understand it better. Or no, that's stupid. Go back to bed and try again tomorrow. I mean, all these things have to come together, and that's part of what makes a good team. And grants enable that.
19
00:17:00.000 -->
00:17:20.000I would like to ask Melinda another question about when we think about smoking as a risk factor. You kind of touched on [it] a little bit earlier, but I’d love you to walk us through a little bit more in depth around genetics and then how an individual metabolizes nicotine and how those differences could impact some outcomes.
20
00:17:20.000 -->
00:20:03.000Yeah, the genetics of smoking is incredibly complex, and I'm certainly not an expert in it, but I can certainly speak to some of it. It is a very complicated topic, as I said. So, there are well-known differences across racial populations in how different groups smoke, and whether that is the topography of smoking. Maybe it's how many cigarettes you smoke per day. Maybe it's how you smoke on that cigarette. What type of cigarette do you smoke? Is it menthol? Is it not? And then the genetics as well. There are slow metabolizers of nicotine and fast metabolizers of nicotine, and so it gets complicated very, very quickly. And in part, that's what led to the U.S. guideline changes that Jeffrey and I contributed to because we were aware that there are these strong differences in in smoking behaviors and histories across populations, and so, for example, one of the most well-known genes that plays a role is this nicotinic acetylcholine receptor on a certain region of your genome, and that plays a big role in what how tobacco smoke is metabolized, and then there are various components of tobacco smoke that then you can measure as output, say in your urine, for example, things like what we call total nicotine equivalents and other fancy things like that. But sort of biological ways that we can measure how much people are smoking, and we very definitely see differences across populations. For example, if you took also, say, lung cancer cases, people that had been diagnosed with lung cancer, and you simply looked at the distribution of their smoking history. So let's say how many cigarettes per day they smoked, or over how many years, and put that together, you see very different patterns emerge between black and white individuals, and that plays out in a real-world way of how that impacts our screening guidelines. So, we see there are differences across, say, Asian populations and Latino. So, it's not just one single way that we can capture it. And genetics plays into that, as well as probably environmental cues into how people smoke. And there's a lot of evidence around genetics and addiction to tobacco smoking.
21
00:20:03.000 -->
00:22:25.000I would add to that that the hard part is figuring out how much those factors contribute to your risk. So, the way we currently try to make these decisions is we try to calculate a risk score for each person, and if your risk is high over some threshold, okay, you should. We say we you should go get screened. We can talk about why that works, but that's the general framework. What makes part of this hard is that, okay, I've smoked 10 cigarettes, you smoke 10 cigarettes. It turns out that that doesn't raise your risk the same amount. It depends on: if you're male or female, it depends on lots of other different things that you might not think. It might depend on family history; might depend on the other amount of exposure you're getting. Might depend on whether you have, you know, some other health issues at the time, and so they're lung diseases. So, all these things turn out to be really important. The model just can't, you know … it's not like you just stack up. Oh, I did this risky thing and this risky thing and this risky thing. So, my risk goes up the same; [it] goes up differently for different individuals. And we're trying to figure out, well, what are [we] trying to figure out, well, mathematically? How do we write that down? So, when you tell me those pieces, I can tell you your risk, and then that's pretty good, because I don't really care what you did to raise your risk. I only care that your risk went up, and if your risk goes up enough, then I'm going to screen you. And, you know, the smoking. And the best you can do is look at other people who are similar to you in some sense. There's lots of ways to define how people are similar, and then you want to say, well, okay, they tend to be at risk when they have these factors. So, then you kind of back-calculate the risk. Understanding how those guidelines work and how the risk is attributed, and what factors play into that, turns out to be really important. Because if you miss important things, like, say gender, you know, or then you don't get the right risk simply because you know you're assuming the wrong thing in the model, so those pieces are very important.
22
00:22:25.000 -->
00:23:43.000Yeah, and it's also, I just might add, that it's complex. Another complexity is we screen. The rule of thumb is you screen people that are asymptomatic or asymptomatic, do not have symptoms, and the whole goal is to have early detection. So, you can intervene when you have really good outcomes and long-term survival. But it's also a balance between some of the harms you don't want to over-screen. I think as we touched on, and because it does involve certain risks trying to reach the lung if you do find something, and there's a lot of other sort of mental stress that might come along with screening. So it's also important to weigh, or some financial considerations if you are diagnosed with some. We also call these incidental findings. If maybe you find something else you weren't looking for, so yes, we want to screen people if they have sufficiently high risk. But that's just one part of the equation that we have to weigh. There are the benefits and harms that have to be considered in this equation.
23
00:23:43.000 -->
00:24:39.000Yeah. So, what is sufficiently high becomes very hard to figure out because you want to balance those pieces out, and those are real. I mean, just to give you a sense, if you took everybody over 40 and you said, okay, we have to give everyone a CT scan to screen them, like that is a huge amount of money. It's not clear that we have the money as a country to do that. So, you could find it, but it's got to come from somewhere else. It's not a simple add-on. Hospitals may love that. They would get some revenue from that. They would probably also be totally overwhelmed. Their scanner would be totally overwhelmed with screening, and they would need more scanners and all sorts of things. So, these balances are really important, and they're hard. They're hard. The trade-offs are hard to make.
24
00:24:39.000 -->
00:27:57.000You know, when we think about some other cancers and some genetic findings like abracca [BRCA], is there an equivalent out there for lung cancer? And if not, where do you think the research stands in that space? Yeah, so lung cancer is very different from BRCA. And that, there, is not an equivalent as with breast, in terms of the BRCA gene. There just really aren't these sort of big-picture type genes where you have these really big effects like you do with BRCA. These are very low-frequency genes, so they're not really there. There's EGFR, is one that comes to mind with lung cancer, but it doesn't have the frequency that the BRCA genes do. So there is not currently any screening based on that in the, you know, terms of including and screening models. There are a lot of sort of low-level genes that are being considered, and we think about them and summarizing them, and another type of a risk score, if you will. If you add them all up, do you have a sufficiently high enough of a genetic risk? So that is one aspect that we are considering in the field, because we do know family history plays a role, for example, in your risk of lung cancer. There is definitely a genetic component to the disease, but it's a lot of little effects. And can we add up enough of those little effects that it has enough of an impact in the risk model? Is where we're at. And then there's a bunch of work happening in these—what we call the biomarker space—sort of a biological marker that we could measure in your blood, and could we use that as a way to identify who should be screened for lung cancer? And there's lots of different technologies that are being considered, and different components of the genome or products of that genome, say with proteins, et cetera. But again, it's complex. A lot of these might be things that diagnose or identify lung cancer close to the time of a diagnosis, and really, what you want is something that detects a signature of lung cancer years earlier, before it really has developed into disease stage, so that then you can, you know, have more localized disease within the lung, and perhaps just do surgery, and then you're done. For example, and not need treatment throughout the body, for example, so I see this playing out in really exciting ways over the next several years. Again, a huge amount of interest in the field is very competitive. I expect we will probably see a blood test at some point here in the very near future for lung cancer, but I think a lot remains to be seen about [what] is the right approach here? Is it cost-effective? Does it reach all populations, or is it only reaching the well-resourced populations? So, lots of important questions I think that the field will be facing in the near term.
25
00:27:57.000 -->
00:29:00.000That problem is an exceptionally hard problem. You can have blood tests for specific types of cancer, but we generally don't have blood tests for cancer. Period. And if the blood test really worked, it would be great to know that there's cancer. But then yeah, I’ve got to be able to find it. There's this Goldilocks zone of having some cancer, and then it's got to be able to scan the body and sort of see it, right? So that you can go in and do something about it, or you can make the argument, no, maybe I'll just do a whole-body type of chemo and try and kill it while it's microscopic. But there's lots of things to learn, and folks doing the blood tests are sort of at the front lines of those. But what Melinda points out is important is you do want something that gives us some lead time, so we want something that works a little bit early rather than okay. Well, it's already there and fairly advanced, and so now we can detect it in the blood. Well, there's not much we can do, and that's a really hard that's a really hard game to play.
26
00:29:00.000 -->
00:29:58.000Yep, I was going to say too. The other thing I think that's exciting is this integration with imaging. I think as well into the risk prediction, and that's, I think, a clear area where data science will play into this and how we can extract features from, say, those images. And you know it's important not only for people to come in and get their first scan, but they need to come in for their annual scan. So then, you've built up a data set, essentially a database of repeat images. And how does that information maybe use AI to extract important information from that? And, I'm sure Jeffrey might want to weigh in here, but there's, I think, lots of interesting ways that the models are sort of real-time live data sets that are constantly changing over time, where we will begin to put together these multiple different types of data sets for predictions.
27
00:29:58.000 -->
00:33:19.000Yeah, so that's another really important point is, you know, mammography is great, but what is most valuable about mammography in many cases is your past, the change in the image from your past mammography, and so those changes can be very prognostic, and we would want to do something similar with lung cancer. You want to look for those changes now. How do we do that? I mean, do we just take your scan and shove it into a model and have the model give you your risk, or do we take the scan and compartmentalize it into things that we see and their shapes and their, you know, their makeup and all these sorts of things are things we're still trying to figure out with the newer technologies? It's very easy to take an image now and put it into a model and let the model make some predictions based on the image. The hard part of that is that there's not very much for someone to go on if you want to say, okay, go find the cancer. Where is it? So, then you have to ask the models where is it and what's going on, and that can also be sometimes a little bit more complicated. Whereas the more traditional way has been to have someone take an image and then have a physician read it, and then give you characteristics of what they see, and you know, do they see, you know, a lesion and all these, and what's the shape, and was it there before, and all these different things. So, we're still working a lot to put those things together. But again, for screening, that would mean someone would have to actually have an image, so they would have to take a CT and regularly. So that means you have everyone getting regular images, and I don't know, you know, for something that is still on the rarer side, whether or not it makes sense for the whole population to be doing that. So, you really want to whittle down the population to the people who are at risk, and then have them get some annual screening in some sense, and then use that screening to sort of find out what's going on. So, there's lots of lots of different stages, and so we're largely at the very first stage—which is [where] everyone's out there—who has to go get a scan, you know, and I think that's a big, it's a big still open question. We have some models that do okay; they're not great. They're not terrible. They really guide what's going on, and we've had some models now that have stood the test of time in some sense, and we've been noodling on them and changing them and saying, "Well, you need this, or you need to change that, or you need to do this for that population. But in some sense, they've worked well on their own, and now it'll really be to see whether the new technology comes along and those things need to be changed, or whether or not they are as good as that technology. Because if you imagine, if you have a blood test and you don't have disease, then the blood test should show you don't have the disease. That doesn't necessarily mean you're not at risk for the disease, because you've been smoking and maybe you get screened anyhow, so they're there. It gets complicated, which I feel like we keep saying. But that's how I think about it. I feel like every step is like, oh, this is complicated too, and this is complicated too, and what can you do here? I can do something, but that's complicated, you know. [00
33:16] Melinda: Just get to keep our job, hopefully.
28
00:33:19.000 -->
00:33:24.000But it's the interesting problems are the ones that are hard and that are complicated.
29
00:33:24.000 -->
00:33:33.000So, when we think about guidelines and screening and such, how do other countries' guidelines compare to ours?
30
00:33:33.000 -->
00:34:49.000There are some subtle differences. The U.S. is really kind of setting the stage for what other countries are doing. We've been the one kind of out in front, and other countries such as France are beginning to bring this online. Australia, et cetera. Other countries where it does differ are in Asia, for example. For example, there's a study, large study in Taiwan where they are screening people that have never smoked, and again, there may be these differences across populations, whether they're exposures that differ, risk factors, genetics that may differ across different continental populations, to really make that a good strategy. Certainly, there are those models [that] are incorporating other risk factors such as maybe a family history or something that might add to their elevated risk, but there are differences across different countries, and it might also be how much you've smoked can differ when you quit. If you've quit smoking, that can differ as well. So, but really, the U.S. is the one that's, in large part, set the standard for the rest of the world to follow.
31
00:34:49.000 -->
00:35:04.000That's great. And then when you think about the future and lung cancer screening, like what would be a big change that you envision within the next five years, or what type of new data? That could play a role in trying to figure out people's risk.
32
00:35:04.000 -->
00:36:50.000I think it's some of the things we were touching on, which is I expect imaging will start to play a role in prediction models. Obviously, if you're scanning somebody for the first time, you won't have that, so you've got to have some sort of decision threshold about who even gets that first scan, and that's again the space that Jeffrey and I are largely working in. But I do see imaging will likely be included in risk models after you get that first scan, and there's all kinds of other information about maybe your cardiovascular risk, et cetera, that's included in that scan, so it can have other beneficial health information that's carried along with it. But hopefully, we'll have other technologies where we can incorporate environmental risk factors. I think an exciting area could be wearables. You know, we carry around our watches that monitor a lot of our health, and it could measure the things that you're exposed to. It knows your sleep patterns. Maybe it can measure the ambient air around you. So, whether that will play out, I think, remains to be seen. But that could be another exciting area. Now, keeping in mind, lung cancer is a chronic disease, as we like to refer to it. It develops over many, many years, as opposed to an acute disease. So, you know, measuring this over a sufficiently long period of time, of years, is going to be important. Wearing, say, your wearable watch for, you know, several months is not going to probably capture enough of an exposure period, but I do think that we'll start to see some of these technologies being integrated in some interesting ways.
33
00:36:50.000 -->
00:41:50.000So I totally agree with Melinda. I think, for me, on the data science side, what's most interesting is the types of data that are coming in, the problems you have with those data. So, for example, with imaging data, when you run them through machine learners, and you don't ask someone to interpret them, and say whether there's cancer or not, but you just shove the whole image in there, so you turn it into a certain type of data and you put it in there. People build these, what are called multimodal models, which takes the imaging and takes things like your age and your family history and all that stuff, puts it all together in one model and spits out a risk score. The machines turn out very smart, and so they find things in the images, like say the quality of image or a particular property of an image that might indicate why images from that site might be different. I'll give you an example. So, when you're doing these studies, you get images from comprehensive cancer centers. More people tend to have cancer on those images because people are going there to get screened, for example, or there are larger populations, and so people are showing up there rather than others. Whereas if you get images from, say, [a] hospital that's not a comprehensive cancer center, and so the machines learn that, oh this is actually an image from this hospital, they tend to have a little bit higher prevalence of cancer so I'm going to give you a little bit more extra risk and we know that this happens with the machine learners is that they use information that's in there that we don't necessarily think about in the image. In this case, where the image is from, and trying to figure out what the machine is using to determine that risk when you just break the image down into a bunch of zeros and ones and spit it through and ask it for an answer is a very, very hard problem. But for me, one that's really important and really will open up the machine learning toolset for medicine, and in a way that I don't think is quite open now. In that you'll be able to sort of understand why some decision? Why [has] the risk went up? What were the properties of the image that made the risk go up? Without having to break it down yourself, because when you break it down yourself, you can only break it down by the things that you happen to think of, and that is part of the exciting part of the next type of data that we'll get and trying to understand how you attribute risk to various things, and and what people can do to get you know screened based on their attributable risk. I think that the other piece is, for me, the really interesting pieces. Melinda says you start without the scan. Like, what do you know about the person? Like, it would be great to get their environmental exposure, but to do that, you would have to have Big Brother-type information: [a] little drone that follows you your whole life and samples the air to know really what your exposure is, and that's really, really, really hard. And so, then there becomes a question of, okay, if you're in a city, and you're exposed to more pollution? But you only know sort of whether you're in a city or not in a city. Is that enough information for you to say that the person's really at higher risk? How do you know how long have they lived there? How much are you measuring? And certainly, for things like asthma and like that, I think there's some preponderance of evidence in the literature that yes, those can be contributing factors. But when you stick them in risk models, it begins to be get a little bit harder to see whether or not those are those are you would think they'd be obvious contributors, but measuring that exposure turns out to be turns out to be harder, and because the disease has a long process and it's a chronic disease, you need a lot of data to figure that out, and so how do you do that? You know, in the framework that we have, there are data sets of daily air pollution in various cities and places. You know, so then maybe all you need is somebody's address, but then, you know, it kind of goes back and forth. Before you get screened, you might have to say, "Well, where are all the places you've lived in the last 30 years? Well, that's a hard question, and then people may not get screened because they don't want to fill out the questionnaire. So, figuring out how you do that, so that you could be most effective with these risk predictions, is important. And I think that's where a lot of these risk prediction models are going. If you walk into a hospital, we'd like to sort of know, you know, in your chart, what your risks are and how you put it all together. And these are all the sorts of questions that we're dealing with. You know, what do we collect, and does it contribute enough information for us to bother collecting it? So that for me is the exciting part, and that goes along with how you attribute that risk in the model, and there's lots of exciting things going on with how machine learning, machine learners work versus these traditional statistical models that are a lot more transparent about what those pathways are.
34
00:41:50.000 -->
00:42:12.000Well, this has been a really great conversation with you guys. So, appreciate your time and all the work that you're doing in the space. It's hugely important, and people's lives are impacted by the research that you're doing. So yeah, thank you so much. Anything you want to add [that] you didn't get to say?
35
00:42:12.000 -->
00:42:55.000I think you know lung cancer impacts a lot of people. I know Melinda's had some personal impact, and of course the school of data science just lost, you know, our founding dean to lung cancer, and that was a very, very, very painful process to go through as a friend of his. And so you want to help solve these problems. It's hard to do it. I think the awareness is important, but there needs to be continued investment and interest in these problems. The machines aren't going to solve them for us on their own. We can use them as tools, but it's going to take smart people to work these things out.
36
00:42:55.000 -->
00:42:56.000No, I'm glad you said that, Jeffrey. I mean, there are real people behind these numbers, real stories, real lives impacted, and as you said, Margaux, it's important to remember that.
37
00:42:07.000 -->
00:43:10.000Agreed.
38
00:43:10.000 -->
00:43:34.000Thanks for listening to UVA Data Points. More information can be found at data science dot virginia dot edu. If you're enjoying UVA Data Points, be sure to give us a rating and review wherever you listen to podcasts. We'll be back soon with another conversation about the world of data science.
