Emerging Technologies Episode 2: Computation Medicine

FS: Yeah, so that's where it comes into play that I'm not a mathematician. (Intro Music Starts) SH: Hello everyone, I am Sam Hansen SW: And I’m Sadie Witkowski. SH: And you are listening to Carry the Two, a podcast from the Institute for Mathematical and Statistical Innovation aka IMSI. SW: The podcast where Sam and I talk about the real world applications of mathematical and statistical research. (Intro music ends) SH: Welcome back Sadie, and even more importantly welcome back to all of you our wonderful listeners SW: For sure, without y’all we wouldn’t be here. And Sam, why exactly is it important to know that someone is not a mathematician? SH: Well Sadie it just so happens that one of our guests today is a doctor - the medical kind - rather than the mathematical kind, which we have all the time (SW Laughs) as we are here for another episode of our series on emerging technology and this time we are going to talk about computation medicine SW: So, how mathematics and computational science techniques can be applied to medical research? SH: And clinical practice. You have already heard from our second guest, Fides, who is a radiologist - we’ll learn more from her in a bit. SW: Amazing SH: Glad you approve! But first, let’s hear from our first guest YD: First name is Yixiang, my last name is Deng. SH: And Yixian is YD: I'm currently assistant professor at the University of Delaware in Department of Computer and Information Sciences. SH: She is working on research into a disease that I bet you have heard of YD: Diabetes is a very globally a prevalent disease. The symptom is that a patient suffers from very unstable glucose fluctuations. SW: Oh yeah, I have definitely heard of diabetes. If I remember correctly there are two types of diabetes. Type one is where your body, specifically your pancreas, no longer produces sufficient insulin to manage blood sugar and type two is where your body develops a resistance to insulin. SH: Precisely. Though there is a third type of diabetes which is a complication of some pregnancies called gestational diabetes. In all cases these issues with insulin can lead to two dangerous states in the body: hypoglycemia, which means low blood sugar, and hyperglycemia, which is high blood sugar SW: And what happens when these happen? SH: Well hypoglycemia is a very immediate danger YD: They will suffer from like immediate coma or even A hospitalization which is very acute. SH: While hyperglycemia can cause severe long term effects if not managed YD: A long long time exposure to this high glucose level for patients' body will affect their in the well-being of their vascular system and causing, like, inflammations, like, chronic inflammations to the blood vessels and inducing, like, long term concerns. For example, if the inflammation happens in the eye will cause a disease called diabetic retinopathy and if it happens uh in your kidney will causes this nephropathy, I think. Yeah, so basically it will affect the integrity of your vascular system. SW: Ok, so we definitely do not want folks to be spending any time in these hypo or hyper glycemic states. SH: Not at all, and thankfully there are many treatments from insulin injections to drugs like GLP-1 inhibitors and lifestyle modifications that people with diabetes can use to help manage the disease. But there is definitely not a one size fits all approach for either type, it is always very personal. SW: Well it seems that it all comes down to the amount of glucose in people’s blood so I imagine that what people eat, their age, how active they are, and a lot of other things mean that everyone needs different treatments SH: Yes, and figuring out how much glucose is in people’s blood is something that has been worked on for a long time YD: Glucose prediction has been long been a very important and also like an inspiring task for a lot of computer scientists, bioinformaticians, and also, even electrical engineers. SH: And what do you think the first step is for something like predicting a person’s glucose level SW: I mean collecting data, specifically data on what their glucose levels look like over time SH: Exactly, and that was a problem for a long time because of the method for measuring glucose YD: You take a finger stick, but however, you cannot print the patient's finger so often to get a trajectory of the glucose. SW: Oh yeah, I remember seeing people prick their finger and then use a little strip to collect a tiny bit of blood to check their blood sugar before. I can see how that would not be an effective way to collect a large amount of data over time SH: Yeah, I for one definitely would not want to have to keep making my finger bleed every few minutes. Even in the name of science! Which is why it is good that there are now Continuous Glucose Monitoring devices. While they do not directly measure blood glucose, they are able to derive a good estimate of a person’s blood sugar levels using the body’s interstitial fluid. SW: Cool! So that gets them the data, but how do they make sense of it? SH: They applied some math of course. Specifically some machine learning YD: We plan to use machine learning models, especially those machine learning models which have been proved to be capable of predicting time series analysis. And in our case, the glucose level over time is also a time series prediction task. SH: And just as important as applying math, are their collabors YD: In the projects I've uh worked on for glucose prediction we collaborate really closely with professors and especially clinicians from Harvard Medical School. SH: Which means YD: So we basically present our understanding of the mathematical question and we model them and then our clinician collaborators help us, you know, make it more realistic and feasible for this specific glucose prediction. For example, we know from the physiology that past thirty minutes of glucose from the CGM can be a powerful tool, and if it's longer than that, maybe it's less predictive. And then we present a prediction model, predict results, and they help us identify whether it's a capable one or we need to improve it further. And then we’ll work on that. And after this kind of iterative discussions, we achieve the mutual agreed best models. SW: I love to know that they are working directly with the clinicians instead of trying to do this all on their own SH: Definitely, no matter how good the math or stats or model is there is no way that it would not be able to be improved by bringing in the experts in treating diabetes. But speaking of math and stats, it is what we are here to talk about so how about we dive a bit deeper? SW: Let’s dive YD: The specific deep learning model to deal with the time series we have examined are recurrent neural network, convolutional neural network, and the transformer. All of them have proved to be very capable in solving time series or, like, sequential data. SW: Oh yeah, I did a whole class in grad school on these deep learning models! SH: Well, let’s see what you remember! SW: So for the first one he mentioned, that’s a type of machine learning method where at each step the result of the previous one is taken into account, unlike many methods which have independent results at each learning step. While the convolutional neural networks rely on multiple layers that can identify more and more complex structures the deeper into the network the layer is. Finally, Transformers are the network architecture behind tools like Chat GPT which rely on embedded data as vectors with weights for how much attention they should be given based off of inputs SH: Consider me impressed! That’s more than I knew SW: Hah, I’m sure you’ve got even more technical information to share that goes beyond my coursework. SH: True. So once they had the general models trained they had one more step YD: And then fine-tune the pre-trained model using the patient-specific training data set, which is smaller, to help mitigate data scarcity while the requirement of patient-specificity SW: Just to be sure I understand: fine tuning means taking a general model and then doing more training, but on data from a specific patient, so that it will be able to make predictions that are tailored for them? SH: That’s right, and now let’s just move on to the next things they did YD: We tested our models on a public data set which was measuring the glucose level for type 1 diabetic patients. And then we got the surprising results that our models in terms of the arrow met of predicting 30 minutes in the future, or 60 minutes in the future have outperformed the other models using the same features for the glucose prediction. And we found those neural network models specialized for dealing with sequential time series data actually outperformed the conventional machine learning, which does not take into consideration of the sequential characteristic among the glucose data. SH: Sadie? SW: Yes, Sam? SH: Would it be ok with you if we spend a minute talking about the data they used? SW: Of course SH: The dataset that they used was the OhioT1DM Dataset for Blood Glucose Level Prediction, which is an open dataset that was created with the help of the National Institutes of Health designed to support this exact sort of research SW: Oh, that is really cool! SH: It really is, and I can’t make it clear enough the critical role played by NIH, and other agencies, in creating datasets like this and then making them available to researchers. During my time as a librarian I worked with the University of Michigan’s Health Sciences Public Health Core team and quite a few of the questions we received were from researchers who were trying to find datasets in their area and were having no success doing so SW: And so those researchers were just out of luck? SH: Well sometimes we were able to find them datasets that could fit their needs, some ended up needing to change their timeline so they themselves could gather the data, and others, yes, did shift to different topics SW: Well that sucks SH: Yes, yes it does. But thankfully places like the NIH and our funders the NSF that have been supporting more and more of the work over time, as have private funders and state and local governments too. But then again, even with an amazing dataset like the Ohio one nothing is ever perfect SW: Heh, imagine that. And what was the imperfection in this case? SH: Do you remember how dangerous hypoglycemia is? SW: It was just a few minutes ago, so yes SH: And do you think that diabetic patients also know this? SW: Of course? SH: So how often do you think in a dataset that is following the glucose level for people who are in treatment for diabetes will very low blood sugar show up SW: Ooooh SH: Yeah YD: You know, our patients, right, because they know they're diabetic and they actually follow very well through the guidance of their physicians, so they have really low low representation of the low glucose level. Right, because they take drugs and they try to avoid taking too much, say, insulin or drug to lower glucose because that is super dangerous. SH: This means that the out of range glucose in the data set is biased toward hyperglycemia, which is totally normal, but is a problem when it comes to training machine learning models called data imbalance. Yixiang and her collaborators used a technique called data augmentation to address this issue YD: It's a very straightforward idea is that we either synthesize more minority data, which is the low glucose samples by adding noise to the existing data or just repeating the data or actually use some modern technique of data generation or another technique in computer vision called mix up. But basically all of the techniques we have tried are trying the best effort to make the data set balance in terms of the training the model. SW: And the augmentation clearly worked since their models outperformed the conventional ones. So, now that they have the models what do they hope to do with them? SH: Well, there is an ultimate goal YD: Artificial pancreas is almost like the ultimate goal for Type 1 diabetic patients. SW: That is a hell of a goal SH: Right? And some parts of it are ready right now YD: Wearable technologies of CGM and ins pumps are really mature. SH: It is the prediction and the dosing that still need to be more completely developed. After all YD: The worst case scenario is that you lose a robot, but we cannot afford losing a patient, right? SW: (laughs) Yeah, patients are quite a bit less replaceable than a robot SH: Quite a bit, or not at all SW: It’s called understatement Sam SH: I know (laughs unsurely). And remember that each person is unique too, so for an artificial pancreas Yixiang and her collaborators want to use a fine-tuned data driven glucose prediction model alongside a mathematical model of the patient’s physiology driven by differential equations, which are tools that mathematicians often use to represent physical systems and their rates of change. They also want to incorporate some other data from the Ohio dataset to make things even more realistic YD: We employ a new technology called systems biology informed neural networks which can be really powerful in incorporating data-driven model and physiological model and wearable data. SW: Oh, including the data from wearable devices like smart watches and fitness trackers would help them incorporate things like physical activity SH: Exactly, though for Yixiang she doesn’t think they are quite there yet when it comes to the model for how activity, especially athletics, impacts on glucose yet YD: What we did in our work is using heart rate as a rough estimate of the impact of exercise on the glucose levels, but to do it more accurately, we may need to collaborate with a sports physiologists who know better about how different sports will affect the glucose levels for not only normal people, but also diabetic patients. SW: That makes total sense. And once they have that information their models should be able to predict and dose insulin? SH: Whoa, slow down. Remember humans not robots, which means at least for the foreseeable future YD: We don't automate, right? We recommend. SH: Or in other words YD: Humans collaborating with AI sounds like a promising intermediate point to achieve in the near future. For example, we can develop AI models to recommend insulin dosage and patients can be the decision maker, right? Judging from their own experiences or feelings to do the final optimal decisions. SW: Right. Keeping the patient in the loop, I like that. SH: So do I, especially as they will be able to feel things about their body and make predictions about their future actions that no machine could. Speaking of keeping humans in the loop, want to hear from a clinician about how computational and mathematical methods are impacting their work? SW: I would love to, but don’t you think you should tell our listeners about another amazing podcast from the University of Chicago Podcast Network? SH: Definitely! (Ad music) SH: If you're enjoying the discussions we're having on this show, there's another University of Chicago podcast network show you should check out. It's called Big Brains. Big Brains brings you the stories behind the pivotal scientific breakthroughs and research that are reshaping our world. Change how you see the world and keep up with the latest academic thinking with Big brains. Part of the University of Chicago Podcast Network. (Music Ends) SH: Sadie, please let me introduce you to FS: My name is Fides Schwartz. I'm a radiologist and I work at the Brigham and Women's Hospital in research mostly focused on CAT scans, so CT, computer tomography imaging. SW: Oh man, I did a smattering of research back in my grad school days on MRI - that’s the one that uses a magnetic field for imaging - but I never got any experience with CT’s. SH: Yeah, a CT scan is a bit different. Fides says that CT scans can be thought of as FS It's x-ray, but it's fancy x-ray, essentially. SH: CT scans do rely a lot on mathematics for the back end. Specifically they rely heavily on the Radon Transform, developed by Johann Radon, which uses integrals to construct three dimensional images from the data gathered by the detectors in the CT scanners. Though Fides did tell me that before, she did not learn any of this during med school itself SW: That makes sense, it's not like medical students are lacking for things they are expected to know and learn. Plus, the theoretics of their equipment is pretty far from the direct patient knowledge they need SH: (laughs) You have a point there. And while it was not top of the list for topics to cover in med school it did come up during Fides’ first board exam in Switzerland where she did her residency FS: That's heavily focused on physics and understanding the machines that you're using. So at that point, you do need to actually dive into at least a little bit of trying to understand what these things do, even if you don't know the matrix that it uses or something. SW: That is very fair, that’s what people like you are for after all SH: You know that I am a lot better at talking about math than doing it! SW: I said LIKE you SH: Ok, ok. (SW Laughs) So as Fides’s residency time continued on she kept seeing more and more instances of mathematics impacting her work. Such as a move from the standard Radon Transform approach, which is often referred to as filtered back projection, which often introduced a lot of noise to the final image, towards a more statistical based approach called Iterative reconstruction which uses ideal estimates and the existing image and produces clearer images over multiple steps. But it is important to remember that FS: We're very conservative in medicine. We don't like to use things unless we've really validated that they work and they don't introduce errors in some way. SH: And as with any statistical technique, it is possible that using iterative reconstruction may leave things out. It may introduce those errors FS: If you use what they call iterative reconstruction, you might smooth out things that you did want to see where there's a disease that you might miss because of the algorithm that runs in the background. So it took, I would say, the first five years of me being in radiology until it was fully accepted that iterative reconstruction might produce an image impression that some people don't like, but it still produces accurate data. So you're not losing any information by looking at those images. And then lung imaging was also switched to IR reconstructions instead of just filtered back. SH: And not that everyone thinks the clearer images this new technique produces are what they should be going for FS: The smoothing gets stronger the higher your IR strength is. And a lot of people don't like the strongest impression for most imaging because it kind of makes it feel too smooth, too much like plastic somehow, the image impression. And so most of the time, the people who decide on what they want to see, decide on the intermediate iterative reconstruction, where you still have some of the noise left, but you have the ability to see things more clearly. SH: Then after residency Fides moved to Duke where she did a research fellowship. Part of her time there was focused on work related to Fractional Flow Reserve Derived from CT. FS: Which are extremely mathematical by the way SH: And what FFRCT imaging tries to do is FS: It’s a way to simulate blood flow in the coronary arteries to see if a stenosis, so blockage is actually relevant for blood flow down below that stenosis. And so that's a hugely math heavy application, though I wasn't doing the maths part of things. I was doing the imaging part of things, of correlating what we see in the image with the output from the algorithm that comes back. SH: It was at Duke where Fides was introduced to a new type of CT Scanner called a Photon counting CT, which became her research focus. Once she started working with the Photon Counting CT scanners she was recruited away from Duke to Brigham and Women’s Hospital where she is working today SW: Still working with that new type of CT you mentioned? SH: Very much so FS: So, I’ve been working with this new CT technology. So it’s research but it is heavily translational. The idea is to actually use the scanner's capabilities to their full extent and not have essentially a Ferrari sitting in your garage or having this Ferrari and then driving it like a Honda Civic or something. SW: (laughs) Hey! Don’t dis my prius,! But yeah - if I had a Ferrari I would not drive it like a Civic either SH: Same, same. Maybe.. No, I would still follow the speed limit (SW laughs). To help explain this metaphor a bit it is important to know that the photon counting CTs can do everything that the previous generations of CT scanners can do. So it would be all too possible to get these fancy new tools and then just do everything the old way, because as both our guests have pointed out SW: Medicine is conservative, for very good reasons SH: Exactly, but FS: You can run it exactly the way every other CT on this planet has been run so far, but that would be wasting its capacity. SW: And knowing Fides she’s not wasting that capacity, right? SH: Not one bit. So the main difference between the photon counting ct and previous versions is how it measures the photons that are detected SW: How did the older ones do it? FS: The photons were translated into light and then into a cumulative electrical impulse. SW: And so the newer ones are detecting the impulses at the individual photon levels then? SH: That’s right, and this allows for much better noise handling FS: You can then bin the photons based on their energy levels into different energy bins, and you can cut out what we call electronic noise. You can just say, we're not using anything below 25 keV, so a certain energy level, for it to reconstruct our images at all, because we consider that noise. SH: Being able to filter out the noise at the individual photon level allows for a lot of things, especially better contrast with soft tissues. Which can mean both more accurate images and lower doses of things like contrasting dyes SW: Fides mentioned her research being very translational, that means she is using these CTs in the clinical setting as well? SH: In fact she is FS: One of the things that we've worked on is to detect pulmonary embolism, so where a clot goes into your pulmonary vessels and makes it hard to breathe, and it can be deadly if it's not detected and treated. SW: So not picking the easy stuff is she? SH: Not really, what Fides is hoping is that these new scanners will help lower the number of patients who have to be scanned multiple times to find these embolisms. Thankfully this is a well known CT task that uses a contrast dye which means that the improved contrast allowed by the filtering should be useful. But only because of another computational technique driven by math FS: We can only do that though because we can reconstruct the images with what we call a lower mono-energetic reconstruction. SH: Or in other words the photon counting CTs allow the algorithms that produce the images to work with a more clearly defined and sensitive set of low energy measurements. SW: And this helps? SH: It does FS: So we had a rescan rate up to 3.8% for these exams in our system, which is not high in comparison to many other places, but it's still You don't want to have 3.8% of patients have two chest CTs if you don't need to. And using this longer bolus and the lower monoenergetic reconstructions as a standard acquisition, we now have a scan rate of 0.1%. SW: 3.8% to 0.1% is a big change. It’s what 38 times better or so? But I have been talking math long enough to know that we should not just talk about relative change SH: You are right on point here. Being 38 times better at avoiding a rescan is not a big deal if only a handful of people get these scans in the first place. So I asked Fides about just this FS: So I looked at a patient period of eight months, and most of the scanners scanned up to 2000 patients during that time with only that protocol. Not all of the patients scanned, just the pulmonary embolism detection protocol. So that's for, we have 13 scanners that I looked at. And so that's almost at least, let's say, 8,000 patients that have had this protocol and 0.1 versus 3% of that is quite relevant. SW: I should say so. What would those absolute numbers be here Sam? SH: Well if there are 8000 people then with the old methodology around 304 patients would expect to be rescanned and only 8 people would need a rescan with the new CT scanners and methodolodu SW: Wow, that’s incredible SH: I know, it's so amazing SW: Are there other areas in radiology that Fides thinks that mathematics and computational science will be helping with soon SH: You can drop the soon, its happening now FS: Radiology is the field in medicine that has the most FDA approved algorithms available. SH: Caveats must be mentioned though FS: Does that mean they're actually ubiquitous, being used and useful in clinical practice? Absolutely not. SW: With those caveats understood are there cases where these algorithmic tools are being used SH: There are, one of the places where they are coming in very handy is not in the imaging itself but in the scheduling FS: A lot of departments would just go by when did something arrive in our system and then just read by time. But it does make more sense to make sure that, I don't know, if somebody scanned a stroke CT to read that first before you read something where maybe an appendix is not doing so well, but that's not something that needs to be addressed within the next five minutes. SH: So these algorithms can help better prioritize the order of scans for the radiologists. Some even prioritize by examining the scans directly FS: There are a lot of sub-specialized algorithms essentially that, for example, detect the likelihood of a bleed in the brain. So it'll alert if it thinks, if it thinks, if it's found something that's higher density in the brain than should be there and gives the radiologist the option to review that scan earlier SH: Caveats continue to apply here too of course FS: Is that always correct? No. But there are quite a few of those applications that are either already available or being used and some of them can save time. SH: And saving time is crucial FS: We have a massive shortage of radiologists everywhere, not just in the United States, absolutely everywhere, just because of the volumes of imaging that are supposed to be read. If there was any technology that could help reduce the workload for the few people that are available. It's not going to improve either. There's going to be fewer radiologists in the future. We know that. That would be great. SW: Did she provide any examples of what one of these workload reduction tools could look like? FS: And hopefully there will be things that automatically segment all lung nodules in a chest CT and then compares it to the priors and gives you something that you can just plug into your report and say, yes, none of these lung nodules have grown in the last five years. We're good. That would be one of the things that we would love. We're not there yet, but fingers crossed. SW: Somehow I can already sense that you are somehow going to caveat this even though it doesn’t exist yet? SH: You know me too well at this point. This time the caveat is about making sure that we are not expecting the algorithms to do all the work alone. There will still need to be radiologists FS: And if it meant that it was just checking a report that's been generated by the AI, I think up to the fact that somebody still has to check it and be responsible for it will not change for the longest time. SW: Just like with the diabetes case I really think that building these algorithmic tools with a human in the loop makes total sense. SH: I agree, there is real expertise and skill that can’t be automated away. Even if some others may not always see it that way FS: A lot of the non-specialty people who have talked about what AI can do in medicine and especially radiology, we're not aware of the level of complexity that's involved with actually reading these things. And a lot of things are actually weirdly easier for a human to learn because we can transfer our knowledge more easily (outro music) SH: Don’t forget to check out our show notes in the podcast description for more Yixiang and Fides, including links to their work we discussed on this episode SW: And if you like the show, give us a review on apple podcast or spotify or wherever you listen. By rating and reviewing the show, you really help us spread the word about Carry the Two so that other listeners can discover us. SH: And for more on the math research being shared at IMSI, be sure to check us out online at our homepage: IMSI dot institute. We’re also on Bluesky at IMSI dot institute, as well as instagram at IMSI dot institute! That’s IMSI, spelled I M S I. SW: And do you have a burning math question? Maybe you have an idea for a story on how mathematics and statistics connect with the world around us. Send us an email with your idea! SH: You can send your feedback, ideas, and more to sam AT IMSI dot institute. That’s S A M at I M S I dot institute. SW: We’d also like to thank Blue Dot Sessions for the music we use in Carry the Two. SH: Lastly, Carry the Two is made possible by the Institute for Mathematical and Statistical Innovation, located on the gorgeous campus of the University of Chicago. We are supported by the National Science Foundation and the University of Chicago. SW: My mom was like, sugar stat! I was like, Jesus. SH: I need to make a change here because I'm not saying something factually correct. SW: You're like, um, actually. SH: I need to fact check myself. SH: Predicting a glucose's person's. SW: Glucose's person's level. SH: Glucose's person's level, yes. SW: As a glucose, my person's level is about a three. SH: That that was that was appropriately dismissive right SW: Words are hard SH: Words are so hard I hate them so much i don't make my living on them at all SW: I was gonna say it's a good thing that's not what i have to do for life SH: I just see words and SW: I'm like i'm gonna put them in a new order SH: Yeah as long as all of them are there it's fine SW: Just like we can't kill patients to study their brains. SH: You know that I'm a lot better about talking about talk…(laughing) SW: Talking about talking? You are a lot better about talking about talking. SW: Statistically significant. SH: I don't know the R value, so I can't comment on that. SW: You're like, I actually don't know. SW: I'm going to try it again. I'll try to be surprised in a new, different way. SH: Laughs SH: You're going to music us out? SW: (vocalizing)