How AI is Predicting Prostate Cancer Recurrence

Podcast Transcript

Dale Shepard, MD, PhD: Cancer Advances, a Cleveland Clinic podcast for medical professionals exploring the latest innovative research and clinical advances in the field of oncology. Thank you for joining us for another episode of Cancer Advances. I'm your host, Dr. Dale Shepard, a Medical Oncologist, Director of International Programs for the Cancer Institute and Co-Director of the Sarcoma Program here at Cleveland Clinic. Today I'm happy to be joined by Dr. Chris Weight, Director of Urologic Oncology here at Cleveland Clinic. He was previously a guest on this podcast to discuss the use of artificial intelligence to diagnose kidney cancer, and that episode is still available for you to listen to. He's here today to discuss using artificial intelligence to guide management of prostate cancer. So welcome back.

Christopher Weight, MD: Great. Thank you for having me back.

Dale Shepard, MD, PhD: Absolutely. So remind us what you do here at Cleveland Clinic.

Christopher Weight, MD: I'm the Director of the Center for Urologic Oncology, also the Vice Chair of the Department for Research, and we have a lab that we call the AIMHI Lab, Artificial Intelligence and Medicine and Health Imaging where we use artificial intelligence to help us better take care of our patients with cancer.

Dale Shepard, MD, PhD: Excellent. Well, previously, and people can go back and listen when we talked about kidney cancer. We're going to be talking about prostate cancer today. Give us a little bit of an idea. We're going to talk about predicting how to manage patients, and tell us a little bit about what's the current state of how we do that?

Christopher Weight, MD: Yeah. So when a patient is diagnosed with prostate cancer, and particularly in reference to this paper, we're looking at patients who were treated with their prostate cancer with a radical prostatectomy, so their prostate was removed. That's curative for the vast majority of patients, but there's a subset of patients who do recur, and we want to predict that because often our salvage therapies are more effective if we catch it early. If we wait too long, the salvage therapies may not be curative and they may eventually die from disease, but if we intervene at the appropriate moment, then hopefully we can cure this subset of patients. So the diagnostic dilemma is we have a group of patients who had cancer. Most of them will end up cured with the treatment we've already gave them, but a subset won't, and we want to capture this window of time where we can get in there and give a small group some extra treatment to prevent them from having their cancer recur or dying from it.

But we also don't want to treat everybody because most everybody's already cured and any subsequent cancer treatments gives them toxicity. So the current tools we have to help make this decision are a biomarker, a blood-based biomarker called PSA. We also have some clinical scoring systems that are based on patient demographics, some of the pathologic details. So we have a scoring system that helps us understand the aggressiveness of the cancer. Once that prostate comes out, our colleagues in pathology will look at that under the microscope and they'll give us a score, and this is called the Gleason Grade Group, and it's a scoring system that goes from 1 to 5. There's an old scoring system called the Gleason Score that's really a 6 to 10, but it's matched and moved to a 1 to 5 system. And so usually those details plus some pathologic details, biomarker details like the PSA, and then some patient details we use to predict the probability of recurrence.

Dale Shepard, MD, PhD: When we think about this subset that might ultimately recur either biochemical recurrence or metastatic disease, what percentage are we talking?

Christopher Weight, MD: About 30% will have a biochemical recurrence. Probably about half of those will have a metastatic or subsequent cancer progression.

Dale Shepard, MD, PhD: If we think about the current tools, what are some of the downfalls of what we have now that led us to look for some new ways to approach this?

Christopher Weight, MD: We use a test called an area under the curve that tries to predict how accurate we are in predicting whether someone's going to have a cancer recur. And so a really accurate classifier would be in the 0.9 or even 1.0 would be perfect. A flip of a coin would be 0.5. And so if we just use things like PSA or even our Gleason Score, usually we're only getting about in the 0.6, 0.65 range, which is better than flipping the coin, but it still leaves a lot of room for improvement.

We also have what is called a Genomic Classifier where it's another way to classify the probability of recurrence where you take some of that tissue that has been removed, we can go back and get it from the pathology block, and you get a gene expression profile, and then you take that gene expression profile, and there are 22 genes used in this test that give you a probability or at least a risk category on the probability of recurrence. And that has also been shown to be a predictor independent of PSA and their pathology, their Gleason Score. And that usually bumps up the score, especially if you use it into conjunction, maybe into the 0.6. 0.7 range. But we continue to have quite a bit of room for improvement in appropriately selecting patients who are going to recur.

Dale Shepard, MD, PhD: So how are we using new technology like artificial intelligence to make improvements?

Christopher Weight, MD: Well, this is a really interesting study. So we're taking the pathology slide, the one that has already been read by a pathologist. And so the advantage of this approach is you don't destroy or use any other tissue. You don't have to go back to the block, you just upload a picture of that and it's at 40 x magnification. And we are now using computers to evaluate that slide image of the prostate cancer, and give us an additional way to predict biochemical recurrence or metastasis, or subsequent cancer treatments.

Dale Shepard, MD, PhD: And so what sort of, I guess maybe this is technical. But when it looks at the image, what kinds of things is it looking for?

Christopher Weight, MD: Well, that's one of the challenges of artificial intelligence is we're not entirely sure. It is similar to some of the mechanics is you take some images and you label them. So you get a lot of this work was done publicly available data sets like The Cancer Genome Atlas has an available data set. And so you took those slides of the prostatectomy specimen and then you said, "This group of slides, these guys recurred this group of slides, these guys didn't recur." And then you feed in hundreds of representations of recurrence and non recurrence. And then the artificial intelligence computer algorithm will start to make its own representations of trying to predict how it will correctly identify who will recur and who won't recur. And usually what they will do is they'll take that whole slide, which is a big image, and then they'll divide it into a whole bunch of small little tiles as we call them, small little bits of that image, and it will start to learn to classify them into a high risk or a low risk.

When we pull out those ones that it has classified into a high risk and look at them, we can see that it's clearly using something similar to what pathologists are using. So they look more aggressive. They would have a higher Gleason Score on average if they get classified as high risk. But it's also probably looking at something independent of a Gleason Score because when we compare it with Gleason Score in a model, they both are predictive still. And so it's looking at other features. RGU pathologists are actually very talented and have already identified a couple of features that don't technically fall in the Gleason Score system currently that have been independently associated with biochemical recurrence. So one of our interest is to try to see is it finding this unfavorable histology pattern that our team has also identified?

Dale Shepard, MD, PhD: So in some ways, I guess you don't necessarily know what it is, but that's the advantage that you don't have to have something you sort out and you're specifically looking for. Like in the old days, you'd have a particular marker, and you had to guess what that marker was, but here you're just looking and letting it figure out what that is.

Christopher Weight, MD: Right. It is the pro and the con that makes us a little uneasy because what if it's classifying it inappropriately, and do we have a way to validate it? And so that's why we do want to do things like, when it classifies it, do they on average have higher PSA at baseline? Do they on average have more aggressive looking pathology? Things that we already know correlates. But it's interesting. We already use black box, quote, unquote, "techniques," and the Genomic Classifier is a good example. I mean, this is now part of the NCCN guidelines. This is used by thousands of medical practitioners across the world.

And if you pin down, most of the practitioners who use these, they probably couldn't even tell you the 22 genes involved in the Genomic Classifier. And whether this gene profile expression is up or down has anything to do with the cancer outcomes. And quite honestly, our own brains are a little bit of a black box in how we make decisions. And so I think as we continue to study these, we need to study them, evaluate them in multiple different populations, multiple different ethnicities and locations. But as we see more and more reliable outcomes, we gain more and more confidence that even though we can't quite tell what's going on and how the decision is being made, we gain more confidence that it's a helpful decision.

Dale Shepard, MD, PhD: It makes sense. When you've gone through and you've used this approach, compared it to the traditional ways, what have you found?

Christopher Weight, MD: So in this study, we were really looking at three outcomes of interest. One is biochemical recurrence. So does the biomarker PSA become detectable after treatment? When the prostate is removed and there remains no cancer, theoretically the PSA should go to zero and be undetectable. As we mentioned at the beginning, 30% of the time that doesn't hold true indefinitely, and over the course of the follow-up of that man with prostate cancer, who has been treated with prostatectomy will see a rise in the PSA. So that was one outcome we were interested in predicting. The second one was the development of metastasis down the road. That's a really important outcome because that's closely associated with death from prostate cancer, subsequent treatments and disability. And then finally we had a combined endpoint that we called progression free survival, and that was an endpoint that included biochemical failure, any subsequent treatments and, or death from prostate cancer and, or metastasis. So that was a complex endpoint.

In each of those three endpoints, we looked at the traditional tools that we have, and all of the traditional tools were significantly predictive of those outcomes. But when we added in this AI histologic classifier, we improved our ability to predict, and it was a significant independent predictor. The one variable that we didn't have a lot of events was metastasis. We only had 15 or 16 metastasis, I think, in this cohort it was a relatively short follow up. For prostate cancer is almost five years, but often prostate cancer metastasis develop many, many years down the road. So it precludes our ability to draw strong conclusions there, but it was also independently predictive in that group as well. And so in each of these clinical scenarios, it was an independent predictor of recurrence of the cancer.

Dale Shepard, MD, PhD: This sort of testing, is this something that we're envisioning is going to be done right after surgery in a serial manner to look for changes that occur? I mean, you're looking at the tissue from the prostatectomy, but is there a thought that this could develop into a blood test that you're following over time?

Christopher Weight, MD: Well, this particular test is primarily... It is based on the pathology at the point of surgical treatment. So that won't change over time. But to be totally transparent in this, the histologic AI computer vision analysis of the pathology slide was one component of this histologic classifier. There were also some clinical details that went into that. Those clinical details, especially the PSA postoperatively could change. So I would envision this test in particular would be just used one time, but it might be used in conjunction within subsequent time variations of PSA, again, trying to maximize our opportunity of not over treating patients unnecessarily. So not giving them cancer treatments that may not benefit them, but identifying the cohort that's extremely high risk for recurrence and being able to jump onto treatment very early in those folks, adjuvant treatments to try to maximize their chance for cure.

Dale Shepard, MD, PhD: And you mentioned previously things like salvage therapies like radiation, and right now we don't have great systemic therapies in an adjuvant setting, so we'd mostly be thinking about radiation. How do we think this might change surveillance in terms of frequency of surveillance, the way we surveil, what might be... If you have someone, this is a great test, someone's high risk, then what?

Christopher Weight, MD: Yeah. So it might help us in both ways. So if someone is very low risk, that's the best case scenario. You may do far less surveillance. You may say, "Hey, we don't..." Instead of doing a PSA every three months, for example, we might get them every six months or a year and remain really quite comfortable that that patient is unlikely to develop problems down the road. Interestingly, in the 16 metastatic patients, this classifier identified 15 out of the 16 correctly. So if you had a high risk score on this classifier, all of them except for one were identified, so that helps us be reassured. If you come out with a low risk in the following five years, very unlikely you're going to get a metastatic disease. Conversely, if you do get a high risk, then you probably would go into a more intense follow-up and a lower trigger for subsequent adjuvant treatments because we know that the risk would be higher for you to recur. And we know, again, with adjuvant radiation, the results seem to be better if you get to it earlier rather than waiting too long.

Dale Shepard, MD, PhD: And certainly that'd be particularly with the low risk, that'd be very helpful because men get very anxious around their PSA measurements. And if you can prevent that part, that would be huge.

Christopher Weight, MD: Yeah. I always try to counsel my patients to... I say, "Cancer has already taken something from you. You've either had to go through some toxic treatments or surgery. You've already worried you may be canceled plans, changed vacations, etc. Don't let it take any more than you can," although it's still a mind game, as you indicated. Every time they come up for some kind of evaluation of whether there's any recurrence, whether that's imaging or blood test, etc, the anxiety goes off the charts. And so if we can more accurately reassure them, the chance of recurrence is really low, then we can cut down on the number of those visits and cut down on that anxiety.

Dale Shepard, MD, PhD: How have patients been accepting of these sorts of AI technologies? And you mentioned the black box component, and we don't really know what we're measuring. Are patients pretty accepting of this or are they a little skeptical?

Christopher Weight, MD: Yeah. This particular test is still just research great. It's not been approved into clinical practice yet. But it's interesting you asked that. I used to be at University of Minnesota. Minnesota has a really big state fair every year where 2 million people go. And we set up a booth there and we did a research project, and one of our focuses was how do people think about AI technology? And much to my surprise, actually, I have a healthy degree of skepticism for AI. I think we need to really put it through the ringer.

But I was very surprised to find that we purposefully set up some conflicts. So we said, "In this particular scenario, your doctor tells you one thing, but AI tells you something else. What do you choose?" And to my horror, many times they were good to go with AI despite the fact that we haven't really demonstrated clear scenarios where AI in a randomized perspective way really makes substantial difference in most disease states still, so I think they're unfortunately quite accepting of it. I think that makes it incumbent on us to make sure we demonstrate that it really adds value in a meaningful perspective way in multiple cohorts in different patient populations.

Dale Shepard, MD, PhD: So too often in clinic, we are up against what we say and then what they say.

Christopher Weight, MD: Yes.

Dale Shepard, MD, PhD: Now, it's going to be also what AI says.

Christopher Weight, MD: Yes, we'll be in the minority, the significant minority.

Dale Shepard, MD, PhD: Nice. So this is research as you said. What is it going to take to take this on to the next step and maybe eventually get this into clinical practice?

Christopher Weight, MD: I think continued validation in larger cohorts with more events, so we can really do true multivariate analysis will give us the impetus, in my opinion, I think sometimes these things are getting approved with too low of a burden of evidence. They can just show it in multiple retrospective cohorts. I think we really should see them perform well in prospective cohorts, but that hasn't been mandated.

In my mind, I would envision the optimal scenario would really be large physician led collaboratives where we have collected pools of data from multiple institution, multiple countries where all of these tests have to come into those multiple collaboratives to be tested. I think that a perspective to really demonstrate this, this particular test makes a difference really would be a prospective study where people will then get subsequent treatments or how frequently they get monitored, and then compare that to the standard of care, and if it shows that it really makes a difference, then that would really be the ultimate burden of proof that it adds value to our current system. But I think we plan to look at this in a broader cohort and more institutions to really see if it's bearing out in multiple different cohorts.

Dale Shepard, MD, PhD: Yeah, that's fantastic. So certainly interesting to be able to use technology to maybe improve what we're doing.

Christopher Weight, MD: Yeah. I think we're at an exciting time in biomedical research where we have a lot of tools available. I think we really hope that we can still have support and funding these kind of research advances because these come a lot of work and a lot of effort. And so I think it's an exciting time because we have computing tools that we've never had before. We have genomic tools that we've never had before. We have ways to combine them in ways that we haven't had before. And I think we have an opportunity to really personalize cancer care, which is important to you and I, and a lot of people on this podcast because it helps us really to tailor the treatment to the particular patient and the cancer in front of us.

Dale Shepard, MD, PhD: And particularly prostate cancer because there's been so much concern about overtreatment. So if we can minimize that, then that'd be fantastic.

Christopher Weight, MD: Yeah. Prostate cancer has been, I think, a poster child both from overtreatment, but then also maybe throwing the baby out with a bath water because there was such a concern about over diagnosis and over treatment that there was about 10 years where there was very little diagnosis happening and very little looking for and screening for prostate cancer. Now, we've seen an uptick in metastatic prostate cancer and a number of men in the United States dying from prostate cancer. So we've taken a little bit of a step back in that. And I think if we can get exactly what you were alluding to, if we have the right tools to help us identify the appropriate men to screen and, or to diagnose and then know how to apply the right treatments versus watching, which is what's something we do in a lot of men. And then once we start down that treatment pathway applying just the minimum amount of treatment needed to get the cure, I think we can be in a much better place where we see that number of metastatic deaths starting to decline again without treating a bunch of men unnecessarily.

Dale Shepard, MD, PhD: Well, that's fantastic. So you've been with us to talk about AI and prostate cancer, and kidney cancer. We'll look forward to the next AI topics. Appreciate your insights.

Christopher Weight, MD: Yep, I'll keep working on it. Thank you.

Dale Shepard, MD, PhD: To make a direct online referral to our Cancer Institute, complete our online cancer patient referral form by visiting clevelandclinic.org/cancerpatientreferrals. You will receive confirmation once the appointment is scheduled.

This concludes this episode of Cancer Advances. For more podcast episodes, visit our website, clevelandclinic.org/canceradvancespodcast. Subscribe on Apple Podcasts, Spotify, or wherever you listen to podcasts.

Thank you for listening. Please join us again soon.