AI-Derived Digital Pathology-Based Biomarker To Predict the Benefit of ADT in Localized Prostate Cancer – Dan Spratt
March 4, 2022
Daniel Spratt, MD, Vincent K. Smith Chair in Radiation Oncology, University Hospitals Seidman Cancer Center, Case Western Reserve University School of Medicine, UH Cleveland Medical Center
Alicia Morgans, MD, MPH, Genitourinary Medical Oncologist, Medical Director of Survivorship Program at Dana-Farber Cancer Institute, Boston, Massachusetts
Alicia Morgans : Hi, my name is Alicia Morgans and I'm a GU medical oncologist at Dana-Farber Cancer Institute. I'm so excited to have here with me today, Dr. Dan Spratt, who is a professor and chair of the University Hospitals Seidman Cancer Center in Cleveland. Thank you so much for being here with me today.
Daniel Spratt : Thank you so much. It's a pleasure to be here.
Alicia Morgans : Wonderful. So I wanted to talk with you about a really exciting presentation you gave at GU ASCO, 2022. You presented information on an AI analysis of pathology and some clinical features to help clinicians understand which patients receiving ADT with radiation may actually benefit from the ADT. This is pretty exciting stuff and I'd love to hear you tell us a little bit more about what you did, how you did it and what you found?
Daniel Spratt : Absolutely. So this is a very exciting project that we undertook. And so what we did is as you know, there really is no predictive biomarkers that we use today in localized prostate cancer, predictive, meaning that it can selectively identify which patients will or will not benefit from a given therapy. So what we did is we took five phase three randomized trials from the RTOG or energy biobank of localized prostate cancer patients that received radiotherapy with, or without variable durations of hormone therapy. And instead of just using our typical clinical and pathologic features like T stage, Gleason score PSA, we imaged and we digitized all the histopathology slides from the patients that banked their samples. Very fortunately, the vast majority of patients of each of these trials actually did bank slides. So the slides were digitized and then fed through an algorithm with a very novel pipeline that you can extract features from the pathology that some of it, you and I or really a pathologist could interpret, but a lot of it is non-human interpretable data that a computer, that's why its artificial intelligence can extract.
And so those features combined with our clinical pathologic tools were then generated and linked together to create a predictive biomarker. What's very unique here is there's a lot of efforts. I've done things in the past, many have that have trained predictive biomarker, but none have had high level validation. You can train almost anything to do anything, but you have to have independent validation. And so we were very fortunate that one of the largest randomized trials called RTOG 9408, almost 2000 patients that randomized men to radiation versus radiation plus short term hormone therapy. We used and left that out as an independent validation. And so we applied that locked model, that biomarker to that, and what we ended up finding is that the biomarker positive patients, which is about one third of patients, selectively benefited from hormone therapy with hazard ratios around 0.3. Whereas those with biomarker negative disease, which is about two thirds of the patients on the trial had no benefit at all.
The hazard ratio was actually 1.00. And this is for the endpoint of distant metastasis. So one of the statistical tests that people always want to know that you're proving its predictive is this interaction test. Is there a differential benefit in those who have biomarker positive versus negative disease? And that was highly significant as well. So overall, this work with many, many people. We're very excited because we've trained and independently validated in a randomized trial, what I believe is the first predictive biomarker to help men with localized prostate cancer guide the use of hormone therapy.
Alicia Morgans : That is an incredible, incredible accomplishment. So kudos to you and a wonderful use of this data. So I wonder how might this model apply or are you thinking of applying it in other settings where I assume this is a, in maybe an intermediate risk population where you're using the short term ADT. What about higher risk patients who receive longer terms of ADT? And who might benefit even more if we didn't have to expose them to that?
Daniel Spratt : Yeah. And it's a great question. And so some of the patients on that RTOG 9408 trial about a third had high risk disease. It appeared to also identify patients who would and would not benefit. It starts to get I think, for many of us uncomfortable because we say, as you get to high risk, everyone gets hormone therapy. But we know from the trials, it's very clear that not all patients drive benefit. If you take the radiation alone arm on those historic trials, some are cured. So we know not everyone needs it. But I think that when you move into the high risk setting, we'll want more and more validation. And we have ongoing work actually in the trials that test short term versus long term, using a very similar AI based approach to identify patients that can safely just use short term hormone therapy.
But this really, I mean, your question spot on is, can we disrupt our brains enough to say, can we trust, can we rely on these biomarkers like you would, let's say in breast cancer. If you had ER negative breast cancer, whether it's node positive or node negative, we would not give them endocrine therapy. I think that we have worked to do, to get to that in our more aggressive disease states. But I think that in the intermediate risk state, I think that's where I would say this holds, we have the strongest evidence for this biomarker today.
Alicia Morgans : It is really exciting though. And I love to hear that there's a path forward to thinking through future steps and potentially tweaks to the model if needed and really continued validation, which is really, really exciting. I wonder if you could speak a little bit to how this model compares with use of clinical features, which we use every day in our practices. We can't quite use your model yet.
Daniel Spratt : Yeah.
Alicia Morgans : How did the two compare?
Daniel Spratt : Yeah, so actually just last year, one of my colleagues, Dr. Bob Des who's at University of Michigan, he looked across many trials, about 10,000 patients enrolled on randomized trials and localized prostate cancer. And he tested does PSA or Gleason score, do any of our clinical features alone, we know they're moderately prognostic, they help us risk stratify, but are they predictive for whether it's radiation dose, or benefit of hormones, or long term hormones? None of them are. So really, and that's why in localized prostate cancer we use, like the NCCN risk groups that really is just telling us the more aggressive patients disease are, we tend to give them more and more intensive therapy. And that's simply because we can't yet until hopefully, when this is published and becomes commercially used, predict the individual who's going to benefit from those therapies.
Alicia Morgans : It's really interesting. It's such crude and somewhat blunt approaches, which is the best that we have right now that seem to be so sharply focused when we integrate this, if we're able to. So,
Daniel Spratt : Yeah.
Alicia Morgans : Very, very exciting. So if you had to give a message to listeners about where we're going and what you expect from this work in the future, what would that be?
Daniel Spratt : Yeah, I think this is a disruptive time. I mean, I think we're entering a time that we are able to analyze mass amounts of data and extract features both whether it's from images from let's say MRIs, pathology data that I think were able similar in the early days of like genomic analysis, where people just had all these genes and features and we tried to figure out what they mean. And we didn't really know, but we sort of narrowed in to understand them better. I think we're at that era now where we're going to have thousands and thousands of features that are going to be very clinically and biologically relevant for our patients. We won't really understand what they are in the beginning, just like with genomics, there's all these different genes and non-coding genes.
But I think that this is an era that because it's so simple, like for a histopathology slide to digitize globally, right, this is far more, you could say cost effective, globally deployable, that I think this is going to really start to in prostate cancer and these exact same efforts we're doing in many other cancer types, breast cancer, head and neck cancer. I think it's going to be one of the most rapid, scalable, personalized precision medicine approaches. And so this for prostate cancer, we hope to actually be using at University Hospitals where I am by end of 2022 to be a pilot site for clinical deployment. So it's very exciting. And I think it's hope for patients, because I think any man who can safely avoid hormone therapy would love to. I think that's the key word safely, because we know it has potential benefits for many men.
Alicia Morgans : Well, I think that you would hear a million here, heres from the patients around the world who would absolutely agree with you. If we can do something like this safely, I think avoiding treatments that are unnecessary is absolutely what the patients are looking for. So I'm excited to hear that the future is now and I really look forward to seeing where this takes us. Thank you so much for your time and your expertise.
Daniel Spratt : Thank you so much for having me today.