�㽶��Ƶ

Challenges and opportunities: The future of AI in oncology (20:39)

Johnathan Tward: Yeah, it's precision medicine at its finest. And you know, AI is excellent at this, but I don't want to suggest that there aren't other methods that are also good at this. For example, we use genomic profiling to do something very similar, and so you can look at a genomic profile of a tumor and take in some clinical information and also get to a similar conclusion. But what's kind of interesting is running genomic profile tests might cost thousands of dollars, and so even though you can arise to a similar precision medicine conclusion with a lot of expensive molecular testing. 

What is so remarkable about the AI, especially this particular example we've been discussing, is that theoretically, the cost to run the test is whatever it costs to take a digital image of a slide and the amount of electricity to run the computer algorithm. So I like to think of these AI based approaches that use pattern recognitions on existing data as one that can really democratize cancer care, because now people from around the world, maybe from lower resource countries, or people who don't have a lot of assets, or even where distance is a disparity, and it's very challenging for them to come to the cancer center. You know, they might literally be able to just upload things from their computer. And you know, even though I've spent my own research career looking at the use of AI in prostate cancer, specifically, you can utilize this technology in any kind of cancer. 

There's nothing particular to prostate cancer that is lends itself to artificial intelligence use what really matters is the availability of data, and so it doesn't matter. You could be a breast cancer patient, a head and neck cancer patient, any kind of cancer patient. We could utilize these same techniques that we have validated in the prostate cancer space to now extend to almost any cancer on the planet. In fact, that is an area that we are actively working on. My collaborators who originally published this research in the prostate cancer space now have access to lots of other clinical trials in different disease settings, and so we're going to see this kind of personalization percolate all the way through the entire spectrum of the various cancers out there.

Heather Simonsen: Was there any resistance when you presented this research? I mean, did people have their doubts?

Johnathan Tward: When I was first presenting this research to an expert community. Actually, I remember I was at an RTOG meeting, which is now called the NRG cooperative group. It wasn't that there was resistance to the research. You could see that it was so new, that people were struggling with understanding it, and the first questions people routinely asked were, what is it that the AI is seeing? And we had to explain to them that you cannot explain to a human all the patterns it is seeing. And that created a lot of sort of confusion on people's faces and concern like people like to know what is behind the scenes. They don't like black boxes. So that created, I don't know about resistance so much as concern.

And the other interesting thing, thinking back at that initial meeting, was when we were telling the scientists that we didn't lead the AI at all. In other words, we just gave it digital images and outcomes. We didn't point to the cancer in the biopsies. We didn't try to say, this is the ugly part. Like, this was what's called an unsupervised learning you know, people kept coming back and saying, well, what if you were to show it what you thought was the worst part of the cancer. Would it do better? Or what if you were to exclude the pieces of the biopsy tissue that you knew were irrelevant to which, I'd say the whole point is, is that what we're calling irrelevant, the computer is finding relevance in. 

So, it is challenging our perception of what is relevant and what is irrelevant. These AIs are outstanding at identifying things that we were discarding as irrelevant. And in fact, my most recent research, looking at prognostic models with AI, has taught me that there are lots of easily obtainable pieces of clinical information that we are discarding that are hugely relevant to people's prognosis. These are free pieces of information that we collect on people routinely in the medical record. And there will be some surprise at how important these things are.

Heather Simonsen: Well, the future is here, but I can imagine there's much more to come with AI and cancer care and what do you think that looks like?

Johnathan Tward: It's so revolutionary a technology that it is almost impossible to forecast where we will be in the next 20 years. It's such a fundamental change. Almost like when you when you go back at the course of human evolution, really, and you start thinking about like, what were massive shifts in our ability to become civilized and advanced. You have things like the Bronze Age, the Iron Age. You know, these things used to take millennia to shift, and then Industrial Revolution, and then computers. But what we're now doing is compressing the pace of discovery from millennia to 100 years to years, to advances that are now coming in weeks to months, because the technology is so powerful. 

I think that we are going to have tremendous tools that help us, you know, really tailor therapies to the patient, as we've kind of discussed, but the same time, we'll introduce a lot of efficiencies and inefficiencies. And we're going to struggle with how we deal with all the additional information that an AI can lead us to ponder. But on balance, even though there's always good and bad to all new forms of technology, I'm highly optimistic that there is far more good, especially in the world of cancer care and AI, that will come than the downsides.

Heather Simonsen: And let's address that. What are the potential drawbacks?

Johnathan Tward: Yeah, I think the drawbacks of AI and cancer care is that the conclusions an AI can draw are really only as good as the data it is trained on. And just like people, it can be highly biased. And so we have to be very careful with trusting the AI implicitly. So, I think it is still critical that humans are ultimately the people who are watching over this technology and ensuring that the conclusions AIs are drawing are accurate. I think that the AIs, although, can create a lot of efficiencies, can also create a lot of inefficiencies. 

So, for example, a lot of attention has been placed on using AI to help radiologists interpret imaging studies like CAT scans and MRIs. And at first someone might think, well, they're really good at identifying diagnoses based on imaging. Does this mean we don't need as many radiologists? I think it's the opposite, which is that AIs can be so good at identifying possibilities, including rare ones, that it might suggest, for instance, to radiologists to consider additional diagnoses that they hadn't thought of before. And now that's going to require more time and energy for that human radiologist to really go through the data and do things that, like look through the medical record that the AI may not have access to, to kind of ensure that they're coming to the correct conclusion.

Heather Simonsen: Well, and then the ability with more oncologists and AI to reach more people. There's really a shortage of care currently, so I could see the potential there being great.

Johnathan Tward: There's so much enthusiasm for that right now. And this kind of dovetails into other technologies, like people are constantly using their phones, and the phones themselves are gathering data. And it's kind of interesting to ponder, is there unlocked potential on just certain behaviors that people routinely do when they're using their phones that maybe we can use to help in the oncology space? Now, of course, that brings up a lot of privacy issues. Or what about wearables, you know, like people are wearing, you know, smart watches that can monitor their pulse rate and oxygen saturation levels. 

And so, the availability of all that data is very enticing to AI researchers in medicine, because we think that we can use all of that data to potentially head off problems the past with people's health. But there's so much that we have to work toward, especially with the privacy issues and things of that nature. Before we can really exploit that data instead of using all this data to market to you, you know, we'd like to explore using this data to actually make you healthier.