Latest / TeqTalk with Jas / Episode 53: Digital twins, imaging, and AI-driven treatment paths
Transcript
- Teqfocus: Welcome everyone to another episode of Tech Talk, where we bring together leaders shaping the future of business, technology, and society. Today's conversation sits at one of the most important intersections of our time: AI, data, and cancer care. Healthcare generates 30% of the world's data, and oncology may be one of the most complex environments of all, where you have genomics, imaging, pathology, treatment pathways, clinical trials, patient behaviors, life or death decisions, and whatnot. Yet, despite all this data, the real challenge isn't about collecting of information. It's turning complexity into clarity. On today's show, I would love to welcome Caroline Chang. She is a VP Chief Data and Analytics Officer at MG Anderson Cancer Center, one of the world's most respected cancer institutions. Take us through your journey, where and also the defining chapters that has led you where you are right now. I started my journey into medicine back in Vancouver. So I'm Canadian, I'm from Vancouver, Canada, and I did my medical degree at the University of British Columbia. in Vancouver and then went on to radiation oncology. The reason for that was that my undergraduate degree was in molecular biology and genetics. This was going to age me, but it was just as the Human Genome Project was coming to fruition. And so I was very fascinated with a disease like cancer that could continually escape treatment. And so wanting to pursue hard challenges, I went and got fascinated with the biology of cancer. And the reason why I ended up going into radiation oncology was that it was the a beautiful marriage between physics, technology, and the biology all coming together in a disease that really impacted patients in a major way, where I felt like I could make a difference from a research as well as care perspective to make huge strides in helping patients live longer, better lives ⁓ beyond the cancer diagnosis. And so I did my r residency in Vancouver, went to Toronto for my fellowship, and ended up getting grants in Toronto where I was starting to pursue quantitative imaging as an area that I was really fascinated in, as well as sub-specializing into brain tumors and the stereotypes technologies. And amidst that I started to pursue quantitative imaging more as a quantitative measurement science and built up an MR. program as well as ⁓ helped lead and develop the brain metastasis program. And that got me recruited down to MD Anderson in 2016. And so I ended up coming down from Canada to MD Anderson in 2016, where I started up an MR imaging program. We I helped the team build up an MR simulation program using multi-parametric MRI to really start to dissect and better understand what are we actually targeting, what are we measuring In these serial images. And we actually worked with a company to help them get through FDA approval for an MR-guided radiotherapy device. And we're the first to actually treat in North America. And so that was an exciting journey in itself. And then the pandemic hit, and everyone in the world was trying to get together their imaging data, their clinical data to make sense of how we can actually best deliver care in. the midst of a global pandemic. And when the clinics came to a halt, I was asked to step up and you know help out with clinical trial design. We were we had over 200 clinical IRB protocols submitted to say we want to research COVID and cancer. That should come as no surprise as a large academic institution. And we we would have taken years just getting through the IRB approvals of over 200 protocols. And so what happened was we all came together collectively, and this created the spirit of where my inaugural chief data officer role came up came about was really bringing together passionate clinicians, researchers, as well as the administrators who helped us develop out this protocol. And we actually designed got it reviewed and approved and activated within six days, which is probably record time anywhere. ⁓ and so that that was the umbrella protocol that brought us all together. But once the protocol was approved, now we had to figure out how are we going to this was this was 2020. This was this was this this was into the pandemic, well into the pandemic, but we you know got this protocol activated and then we needed to figure out How are we going to now coordinate a protocol that's going across what would have been 200 different protocols? And so all of a sudden, as much as we had a lot of multidisciplinary care and research happening across the organization, this was at another scale of saying every department needs to coordinate with each other. And how do we bring all of these teams together? And people who are interested in very specific subtypes of cancer. Those who are studying immunotherapy, those who are in lab sciences, pathology, as well as into the clinical space and beyond, all coming together to actually ask these questions. And with that lens in mind, and with a very, you know, we saw silos being broken down across the world during this pandemic, that spirit really helped us drive forward the way that we approach data as an institution and continues to live on. And so I think that. The culture that got grown out of that space. We were meeting three days a week, 7 a.m. And the the number of people attending this meeting was growing, which is an incredible testament to the passion of everyone at the institution. Wow, that's amazing. The whole trajectory of your journey is pretty inspiring. So was there any moment in your career when you realized that okay, data is not just a supportive factor? It is the center to the future of oncology. Was there any moment a part phase if you can define I I don't think it was a single moment, certainly. It was a number a series of events that occurred. As as I was pursuing the quantitative imaging space, what I realized was just how heterogeneous data was. Just even looking at imaging data in itself. Most people think that imaging data, because it comes off of very fancy, expensive machines, should be very consistent and all the measurements should be very precise. But the thing is that you can take a very expensive car and drive it well or not drive it well. And similarly, the imaging machines can be very fancy, have lots of bells and whistles, but each person and each hospital and each clinic can run their scanners with a different protocol. Right. And each of those protocols can affect what the actual measurement that you're trying to get can be. And so we have not yet calibrated all of our measurements. We are getting images and we're continuing to leverage those images. And when you use it as an image, it's an adjunct to your clinical decision making. I think we're just starting to scratch the surface of the rich data that we can actually pull out of this imaging. Data that we have. And so as I started to pursue that, I started to work with an international group. There are a number of different organizations. At the time when I was starting this, it was the Quantitative Imaging Biomarker Alliance under the Radiological Society of North America. Just this past year, we've broken out and made it into an independent 501c3 nonprofit organization to collaborate internationally. So We're now called the Quantitative Medic Medical Imaging Coalition. And we've we're working with the various different clinical groups, the imaging groups, the physics groups and the metrology labs, both in the US with NIST and in the UK with the National Physics Lab laboratory. And as I started to tease through just the heterogeneity and making imaging data more consistent and quantitative, I started to look at the other data. Because you know, imaging data alone does not answer all your clinical questions. And then you started to realize that a lot of the data in our medical charts are just as non-characterized or inconsistently captured. And so starting to harmonize those data sets is going to be a big challenge, but opportunity for us to really leverage a lot of the emerging technologies. Yeah, absolutely. And and like how you mentioned that data is so much right. So how you harmonize it, how you make it more sensible. And we all know that healthcare produces immense percentage of data overall. So I was reading it was more than thirty percent of overall data only comes from healthcare, which is pretty massive. So how you better utilize this with the help of AI and, you know, collaborate it, make the systems better, it's all up to the hands of organizations and experts like yourself. So When I talk to the leaders, experts, CIOs of different industries, right? Every sector wants to have AI now. And till last year they were, you know, experimenting. And now this year onwards, I've seen that they want AI implementation and speed. The adoption rate is going pretty high. But not every sector can afford AI mistakes. I think healthcare is one of them where we cannot afford to have anything wrong. So If Netflix recommends any wrong show, it's okay. Maybe there is some kind of loss, which is not the loss of life. If any retailer's algorithm misfires, the conversion rate can drop. If any ad platform makes weak recommendations, budgets are wasted. But in oncology, the cost of being wrong can be heavy, right? It can be delayed in the treatment, missed early detection window, which is so important. Wrong therapy path. overall loss in trust and most when people are in the most vulnerable state, right? Because the name itself is so scary. Now, when I talk to leaders like yourself, those who are working in this space, they have got common people in so much confidence that at least it it has turned from non-manageable disease to a manageable disease. Right. If it is detected early. So it is so important to detect it in the early window. So the if you see the data of oncology, every data is a real person behind the scenes. So if you're not able to manipulate the data properly or if the systems are detecting wrong recommendations or not working well, then overnight things can go worst. It is so sensitive. So given that level of consequences, how do you standardize your evidence or your validation, your deployment, which is actually different than other industries. Yes. And I think that's this is a very valid point in that, you know, when you're looking at your favorite song list or your movie list, the consequences are pretty minor and you may not be happy, but then you change the you change the station or change the song. They I think that there are different levels of using AI across healthcare though. I think everyone immediately jumps to AI helping make clinical decisions, but there are many other activities where we can actually take benefit of models that have been tried and true tested in other industries. For instance, process engineering is an area where you can actually use a lot of AI models very effectively, and we could probably learn from other industries to improve some of these processes. We have actually, even within the Institute for Data Science and Oncology, we've worked on. Projects with our surgeons to optimize surgical scheduling, to make sure that we're making the best use of our time to have as many of our patients treated in a timely fashion as possible, making sure that we're finding efficiencies amidst that and looking at patterns in terms of turnover, et cetera. And so I think that it depends on the application where I think there's more wiggle room that if if you don't get that schedule optimization perfect, but better than before. You're still having a gain. And I think that that that's one piece. The other piece is that if we're really honest with ourselves, humans are not perfect and physicians are not perfect. And patients themselves don't remember everything perfectly. And it the the art of medicine is trying to get that whole story despite all of the imperfections. And so one of the things that I will call out is that. You know, in situations where there's a lot of uncertainty, we don't have all of the science figured out in medicine. That is the reason why we do scientific research. And so where in these areas of uncertainty, if we can leverage this technology as an additional source of information and a way to distill through large volumes of data, which AI is very good at doing, could we actually get better information to make clinical decisions? And I think that Creating that transparency of the accountability and how the tool is being used, making sure that the clinicians are trained up to understand where the weaknesses of any tool are, where the vulnerabilities may be, so that they utilize the outputs in the appropriate way. Those are all safeguards that we would want to use if it's being used in the clinical setting. And so those are all the cut all the pieces. Now, when we look at In an academic medical center where we're also developing models in-house, we do have different grades of validation and evaluation, as well as the type of implementation we would need. So we we've have a team that's in charge of life cycle management, knowing that a model that works on day one may not work on day 14, as early as day 14 or the next day, depending on what's happened with the data. And so having that continuous monitoring process. is something that we definitely want to put in place for something that's operational or will be used in clinical decision making. Whereas if we have models that we're using in the research space, we don't necessarily need that quite the same level of rigor, perhaps, depending on where in the research process it is. And if we're really doing exploratory work, you have more affordability to say, well, if something doesn't look right, we can go back and check again. Yeah. Yeah. I think I really like the statement where you said that nobody is perfect, right? Clinicians, patients do not remember everything. Nothing can be perfect. So in this imperfect world, if you are getting some additional help from AI to make even slight betterment, right? It's a winning situation for all of us. So, what has been the toughest organizational barrier to building a real data ecosystem, especially in oncology? Because when I talk to CIOs and different other leaders, they they say that technology is not a problem actually. The adoption is actually a problem when you have multiple verticals, right? Sales team, they want more speed, finance want control, IT want resilience, and legal, they want to be very safe side, security on. And healthcare, when we talk about healthcare, it becomes much harder, right? The version becomes much harder when you have when when clinicians they are like They want simplicity, they do not want complexity actually. And researchers need flexibility, compliance needs full control because it's it's so important part. Admin needs ROI, patients need continuity and a lot more. So, what are some of the toughest organizational barriers or challenges to build a data ecosystem, especially as you're leading the data site? You are in the forefront of you know having this data ecosystem built in your current organization. So Let's talk about that a little bit. I mean, I think I can share it with all the different rules that you just went through. I would say a fundamental first step was to realize that each of these different verticals speak a different language. And so and we love acronyms at our organization. And so, you know, even the same ac same letters could mean something completely different if you're in a different vertical and a different business unit. And so Number one that we actually started to learn, and this came out of this became very apparent when, as I said, the entire organization came together during the pandemic and all of a sudden everyone was talking past each other because their definitions were different. We suddenly hit pause and we said, okay, with each of these terms that we're saying, we have to make sure that my definition is the same as your definition. Yeah. What the context in which I've captured this data. Is different from the context in which you've captured your data, how do we capture that context? And so what we started to do was build out, and I've actually published a paper on this concept of a metadata supply chain that I think that, and I think that many organizations, as well as general AI approaches, would benefit from paying attention to that metadata. And I think the field is actually realizing this because as you've probably have heard. The term context engineering has become very popular. Our data management system, we named it back in, as I said, back in the early tw early in the pandemic. And so well before context engineering was a thing, we called our data management system and the overall framework of the people, processes, and technology, we called it context engine. So we ha we've had context engine and context engineering for a good number of years. The metadata supply chain allows you to have data that's not looked at in a flat way, saying, you know, this column, this row, and it's a very 2D. It adds a third dimension to know when you should be combining data and whatnot when not to combine data and when it's appropriate to do so, but with appropriate harmonization. And that just mutual transparency and understanding is. Probably the very first step that I I think that we were able to start to overcome and journey through. Great. I think that's so important, right? We talk about context a lot right now. So tell me in your journey, did you face any challenges? A couple of interesting things, if anything comes to your mind when you were building this context layer, metadata layer and during the pandemic time, was there any kind of ⁓ disagreement between the teams or everything went without any issues, which would not be possible, I think. No, I I mean, ⁓ I don't know about disagreements necessarily, but different perspectives were brought to the table. And where we the the approach that we ended up taking was we try to find and prioritize the common needs. And so for instance in a cancer hospital, diagnosis is a very important field that everyone was interested in, whether you're in finance And you're trying to get your billings across the board, whether you're in your clinical space, whether you're a researcher, whether you're reporting to tumor registries, diagnosis was an important field. Yet capturing the details of treatment was also a very important piece. And in cancer, the staging and the risk factors of the patients in terms of staging was also an important piece. And so and finally, we're still working through outcomes measurement is A very complicated, nuanced piece that is that we're still continuing to work through. But we kind of tackled these sort of very high-level fields. And what we discovered was they're much more complex than you'd think they were. I mean, you'd think that a single cancer patient would come in and they'd have a single diagnosis, they'd get treated and they're done. And that's certainly not true when you actually start to dissect this down. A diagnoses can change over time. So we're discovering, as I said, research is continuing to be done. You know, what we thought was what we would categorize as cancers over the many years around anatomical locations. You have lung cancer, and then we separated lung cancer to small cell lung cancer to non-small cell lung cancer. That was based on what you could see down a microscope. And now today we've got a whole host of molecular and genomic tests to sub-subcategorize. each of these cancers. And so the diet the relevant diagnosis is evolving over time. And so if you have an organization that has decades of history, how do you actually resolve all of that across time? Some of these factors will be unknown because the test was never done 20 years ago. And some of them are known today. And how do you categorize and combine those data sets? And so even something like diagnosis, which everyone would think everyone should be able to resolve When it comes down to distilling the data, it is actually a it's a complex exercise. Yeah. Mm-hmm. Human body is complex. Each person will have different kind of cells, and I can imagine that how complex it could be. So I know that you have been working in this space from decades now. What are the interesting AI use cases you are implementing, which are really helping? Till last year we were talking a lot about ambient AI. And you know, scheduling and all of those manual things which were taking a lot of time off clinicians and you know, even ⁓ to increase their efficiency as well, all of these were introduced. But now I think things have gone a little beyond, people have got confidence. So, what are some of the interesting use cases you are working on? Yeah, so I guess in in the research space within the Institute for Data Science and Oncology, we're exploring everything from using AI and computational modeling for early drug discovery, early translation. Can you actually simulate out clinical trials before you actually run them and use the actual prospective trial as a validation exercise? Doing these sorts of activities could reduce the cost and accelerate the speed of discovery. We're we're we've heavily invested and and are continuing to develop out digital twin capabilities, both from a tumor perspective, patient perspective, drug discovery perspective. And that's a very exciting field that I think is going to continue to evolve and mature over time. And so these are some examples of where we're we we've been exploring and being at the cutting edge. Yeah, you talked about digital twin. It's pretty interesting when I was reading and I knew a lot of these things were happening in different industries, right? And aviation companies like Rolls-Royce and G, they use digital twins, digital twins of jet engines to predict the actual failure, right? Before actually it is getting fails. In manufacturing, factories are using this to reduce the downtime or production lines and all of that. I mean, they have the vision clear in the other sectors that create a live virtual replica, feed it into the real time, simulate the outcome, and make the better decisions faster. And now it's so good to see that. We are having the same kind of concept in healthcare as well. So I think this year itself I was reading the article early this year that heart digital twins are used to, you know, guide one of the treatment planning. And I I was also reading about the researchers are using organ twins for surgery simulations, right? And like how you mentioned in cancer care teams are exploring tumor digital twin, right? Combine imaging and genomics, pathology, and different other ⁓ factors. So I think sooner we'll have a time when the doctors will be able to say that okay, this person having cancer will be cured by immunotherapy and not by chemotherapy or something like this, right? The prediction can happen in that way. So talking about those lines, what do you think? When is that time going to come? Is it someone Next decade or a couple of years, or what is the rate of discovery which is happening? I think the rate of discovery is incredibly fast. At the same time, I would say that some of the I feel like my career is going full circle and I'm even I'm even more motivated today to make the precision imaging measurements an implemented activity across medicine because of this promise. So if you think about If you're going to use something like digital twins and not have humans looking at the image and making a clinical decision, but having these measurements feed into a model, the precision and the consistency of the imaging becomes so much more important and critical to allowing that prediction to be meaningful. When you were using it as a clinical adjunct and you were reading the image and taking it as a picture. And you were saying this looks bigger. I think that there's response here, we can continue on with this treatment. That was one level of decision making. Right. When you say, I'm going to feed in the measurement out of this image, and it's going to be considered with all these other parameters in a model, and it's going to simulate out predictions of response to a thousand different treatments to say which is best for you. This measurement that you're feeding in better be precise. And if you want to iterate over time. It better be consistently measured every single time. And so I do feel ironically my my journey has come full circle where it's still coming and landing and saying we need to get the precision measurements out in a meaningful way and scaled across the healthcare system because most cancer patients don't start and end their journey in one hospital. They you know go to various different clinics amidst their care. And I think it is so personal, right? And like how you mentioned this keeps you motivated because ultimately you're touching everybody's life in a who are suffering from this disease. So I'm hoping at least in in half of a decade, like in coming five years, there will be a lot of developments done. And Caroline, also I wanted to understand from your perspective, what is actually trustworthy AI or any AI model looks like to you? And the second part of the question would be. Probably how do you balance between this innovation and the compliance factor? So I think I'll take the that one first and then I'll come back to the what is trustworthy AI, because that's a loaded term. In terms of the compliance versus the speed of innovation, I think that you know, the way that people have aggregated data up until this point, in terms of Until recently, now people are talking more and more about how can we actually facilitate federated systems. I think was limited by the technology at the time. And so historically, the way that data collaborations happened is data got pooled into one place. Well, the requirements from a compliance perspective for privacy reasons, very appropriately so, data had to be de identified. But When you lose again, you lose the context and you lose the metadata and data's completely de-identified, you can gain certain level of insights, but you can't go deeper because the data is de-identified. I think that the regulations need to come because tools like digital twins, once you've got that much granular detail, truly is identifiable. If you have a good digital twin of you or a good digital twin of your organ, that's an identifiable asset. And so if you create these and you create many of these without proper regulations, there is a risk that it you're putting out there. And so I think that that is the tension that is there. Having said that, I think that the technology is also emerging and developing very quickly to allow us to maintain privacy while allowing us to train models and work within this space. So I think that it's a pressure on us to, I think everyone wants to do the pieces that are fun of just developing the models, but I think there's a lot of research to be done in facilitating meaningful collaborations that allow you to maintain the privacy preserving ⁓ methods of collaboration is just as much of an important development space that we all need, I think, globally. And it's not only in healthcare. Obviously the ⁓ some of the most sensitive data is in healthcare, but it is true across the board. And so I think these developments as I see them come is actually very exciting to see. In terms of going back to the trustworthy piece, I think, you know, the trustworthy AI is a nice catchphrase term that makes everyone feel warp and fuzzy and feel like this is this is it. I mean, trust, if we're honest with ourselves, is an emotional term to some degree. You build up a certain level of confidence. In healthcare, I think that trust can be both a good and a ⁓ source of risk because you can have people who become very trusting and over-reliant and become more susceptible to automation bias because you start to rely and depend on these sorts of algorithms. This may not be a conscious decision. And so there was a recent publication of Colonoscopists who are using AI and they had introduced this AI tool into colonoscopy. And these were experts in their field, trained for many, many years. And after weeks of using the tool, they took away the tool and they were missing polyps. So they had become dependent on this tool relatively quickly. In my opinion, I don't think they lost their skill, but they clearly lost some level of attention. Because they had become reliant on the tool, putting little red boxes around where the the suspected polyps were. And so, you know, if you took the tool away for long enough, I would, my hypothesis would be that they would regain those skills because they had been trained for many, many decades to do this and they would be able to do this again. My question though is if we're training a whole new generation who has been trained using these tools the entire time. Will they would they have ever gained that skill or are they completely dependent on these tools? And I think so this whole education piece is actually an interesting one. Yeah, it it's pretty interesting. And being a mom, I think the same thing because everywhere so much of automation is there, right? And even when you drive the car, right, nowadays, Tesla, FSD, full service drive, sometimes I wonder that what's gonna happen to my kids. Will they never learn because by the time they're grown up, they will not even have that we have already seen driverless cars, right? Going on. And sometimes even even we feel that okay, we are not used to driving cars anymore because of this automation. So I think it's happens everywhere. So your question is so valid that what is the fallback mechanism? What if this doesn't work? Where do we start from? And I think it's a great, great question. And ⁓ An exciting time to see what's going to happen to the next generation. Yeah. But in terms of the immediate application, in my mind, of trustworthy AI, I think you need at least two big factors. One is transparency. So transparency and what is this model intending to do? What was it, what data was it trained on? Can physicians understand? Perhaps, I mean, there was a whole area where everyone wanted ⁓ I think we've kind of gotten past a little bit of that now. But I think that even if it if even if it's not fully explainable, understanding where the vulnerabilities are is a critical piece. And I think that, you know, if you have the transparency in combination with accountability, then you can start to build up confidence and trust. And so I think it's that combination, but I think that just as I showed in just as I shared with that example, is that just putting humans in the loop is not the ultimate filter for distilling any AI risk. And so we're just about to, you know, we're I've been presenting on this concept of creating a high reliability organization framework around a an institution with AI. And you know, the high reliability organization framework was designed well before AI existed. And it was because we wanted to admit that humans were fallible. We weren't going to point fingers at that individual, but say, what processes can we put in place to safeguard and mitigate for errors that could pop up? And if you block all the holes of the Swiss cheese with different layers. And so we need to redesign that high reliability organization framework with AI embedded because AI could potentially help de-risk, but it could also introduce new risks and it will change that human in the loop activity and their behavior, just as we were just discussing. Right, right. Very well said, Caroline. And coming back to the last segment of our podcast, tell me what you would wish more leaders to understand about AI in healthcare, not from the technology point of view, but from more from a human one. Yes. In healthcare, I mean I think that I'll go back to the context piece in that the I think AI will actually help us embrace the complexity of medicine. And healthcare is incredibly complex and context dependent. For instance, even in cancer, although many you know the primary outcome of cancer studies is survival, there are when I speak with patients and some patients who are 90 and they say I've you know I've lived a full life I just need to make it to this major event. And then I'm perfectly happy not going through any more treatment. That is a very human decision to make. And so I think that, you know, making sure that we're capturing the full picture of what that individual patient desires as their end goal. Some people are more afraid of, I say this to my patients and I have an honest conversation. I said, Are you more afraid and will it keep you up at night that the cancer might come back? Or are you more afraid of all the side effects of this treatment? And when weighing those two things out is an honest question for you to consider as you're making these decisions. Because some people will say, I don't care if I potentially die of the side effects of this treatment, I want to fight this. Other people will say, If that means I could die from the side effects of the treatment, I may make a different decision. And but that's a very personal decision. I don't know that there is ever going to be an AI model that could actually distill those emotions and anticipate that kind of decision making. Perhaps someday, but I I don't think it's going to be anytime soon. And I think that I don't know that we necessarily want that human component of medicine to necessarily And that human choice to necessarily ever be replacing the decisions that we're making. Absolutely. And that's why we need to have humans to know the emotions, because a lot of things can be controlled through the emotions. So, Caroline, thank you so much for joining me today. This felt bigger than just healthcare conversation. We discussed trust function systems, responsible AI and and leadership under pressure, a lot of things. Thank you so much.