AI-Powered Biomarkers and Precision Medicine in High-Risk NMIBC

Amanda Nizam, MD, of Cleveland Clinic, speaks with Vignesh Packiam, MD, of Rutgers Cancer Institute, about his innovative study on an AI-powered model designed to predict patient response to intravesical BCG in high-risk non-muscle invasive bladder cancer. They discuss the study’s rationale, findings, and implications for precision medicine, highlighting how AI evaluates histopathologic features to guide treatment decisions. Dr. Packiam explores the potential for broader validation through prospective trials like BRIDGE, the challenges of implementing AI in clinical practice, and future applications in tailoring bladder cancer treatment.

Transcript

Dr. Nizam: My name is Amanda Nizam. I am the GU Oncology Bladder Cancer Section Editor for GU Oncology Now. I’m joined by Dr. Vignesh Packiam this afternoon, and he’s going to discuss one of his studies that he presented at the Society of Urologic Oncology earlier in December 2024, talking about an artificial intelligence-powered model to predict response to intravesical BCG versus gemcitabine/docetaxel for high-risk non-muscle invasive bladder cancer. Welcome, Dr. Packiam.

Dr. Packiam: Thank you. Thanks for having me.

Dr. Nizam: I’ll have you start out, if you could tell us a summary of your study, what led, prior findings that led to this investigation and what your study aimed to show and what your study actually showed.

Dr. Packiam: Yeah, absolutely. So I’ll give some kind of rapid-fire background as to kind of the rationale for the study. One, we’ve had BCG for almost 50 years now as the primary kind if gold standard treatment for high-risk non-muscle invasive bladder cancer. There are some problems with BCG that have led to development of promising alternatives.

We’ve had a BCG shortage for more than 10 years, which can be severe at times. And we’ve had the development finally of some promising effective alternative treatments. So sequential gemcitabine and docetaxel was initially used for BCG failure, but in response to BCG shortage, it’s being used more and more in the frontline setting.

And we have a randomized controlled trial that is comparing those two treatments and soon GEMDOCE might also be a level one-backed standard first-line treatment. At the same time, we’ve had more and more improvements in predictive models to predict for response to BCG and for other therapies.

And I think as we’re moving on to 2025 and getting more into precision medicine, having more effective tools will help us to give the best treatment to different patients. So one of such tools that was recently developed was an artificial intelligence augmented model where essentially pathology slides from TRBT are digitized, zoomed in at 40x magnification, and then microscopically there’s an algorithm that looks for certain features that are predictive of better or worse response to treatments like BCG.

And what the AI model basically does is it looks at the tumor cells to look at nuclear characteristics, mitotic activity, and how the cells are organized. It also looks at the tumor microenvironment to look for immune infiltration, inflammation or how the stroma is organized. We recently published a study in Journal of Urology, which was a multi-institutional cohort, almost a thousand patients, where using this AI signature we were able to predict better or worse recurrence rates or progression after BCG.

In this current study, we essentially applied the same AI biomarker to a cohort of patients who either received BCG or GEMDOCE as their first treatment for high-risk NMIBC. And what we found was when this AI biomarker was present, patients did significantly better with GEMDOCE compared to BCG.

And when the AI biomarker was absent, patients had a very similar response to both BCG and GEMDOCE. So this suggests that this biomarker may actually be predictive in letting us prognosticate response to BCG versus other therapies rather than just being another prognostic biomarker that says this is an aggressive cancer that’s unlikely to respond to any therapy.

Dr. Nizam: And so when you developed this assay using the composite histopathologic features that you mentioned, when you did the Cox proportional hazard modeling, looking at the recurrence free survival, et cetera, what other clinical factors did you control for between the two groups?

Dr. Packiam: Yeah. When this was initially developed, it used the traditional clinical pathologic risk features, age, tumor stage, multifocality. In the ultimate model that was developed, the only additional pathologic feature that’s utilized is multifocality.

So this is kind of acting agnostically to the usual clinical pathologic features like T1, multifocal CIS, et cetera. And in this current report in one of our sensitivity analyses, we actually took multifocality out of the equation and the AI signature alone was able to predict response.

Dr. Nizam: Okay. So I guess, did you evaluate the model using those other traditionally used risk factors?

Dr. Packiam: Correct. Correct.

Dr. Nizam: Okay.

Dr. Packiam: And there wasn’t a big variation in the predictive capability of it when we took the pathologic features out.

Dr. Nizam: Okay. Okay. All right, and then so you mentioned the BRIDGE trial, the phase three trial, that’s looking at BCG versus gemcitabine and docetaxel for treatment naive, high-risk, non-muscle invasive bladder cancer. So that’s ongoing. Are you planning to do any work within the confines of that trial to look at AI and perspectively validate these findings? Or do you have additional prospective work that you’re doing to validate your findings?

Dr. Packiam: Yeah. I think that’d be a perfect setting to validate this. That’s a prospective study. It should be relatively straightforward to get images from all the pathology slides. The trouble with some of the other predictive tests that are out there, especially some of the genomic tests, is that they require the material from the TRBT, usually from the FFPE which is available, but you need contracting to get all those types of things analyzed with a research study.

Whereas for this, they can simply take a picture of the HNE slide. So that would be a perfect venue to try to validate these findings. We’re also assembling a larger multi-institutional cohort where we can, again validate this more broadly. This study was just done at two institutions.

Dr. Nizam: And yeah, so you’re referring similar to what Artera has done using the NRG trials in prostate cancer. So it’d be interesting to see that in bladder cancer. I know there’s some ongoing investigations across different stages of bladder cancer.

Dr. Packiam: The really nice thing about prostate cancer studies is that especially with radiation, they have a huge amount of prospective trials where they have this rich biobank to draw upon and we’re kind of catching up a little bit in bladder cancer.

Dr. Nizam: Yep, I agree. So how would you see this, if this is prospectively validated, how would you see this playing out into routine clinical practice?

Dr. Packiam: I think the first step is to validate that GEMDOCE is not inferior to BCG prospectively. I think a lot of people are anecdotally having a lot of success with GEMDOCE, especially if they’re dealing with a severe BCG shortage. But for the broader urologic community to get comfortable with GEMDOCE, I think it is going to need that level one evidence from the BRIDGE study.

So we need to wait for that study to read out. Assuming that that shows that it’s not inferior, which a lot of us are hoping that it does, then we’ll have another effective option for patients, which is great. And it’ll be nice to rationally predict which patient is going to benefit the most from either treatment.

BCG is nice because logistically it’s relatively easy. Patient comes into the office, they’re catheterized, BCG is instilled. They can leave 10 to 15 minutes later. GEMDOCE is a little bit more clunky in that it’s sequential administration of two agents. It’s a longer amount of time in the office, which can kind of be resource constraining in some settings.

Dr. Nizam: So what are the next steps that you see in your current study that you’re doing to validate your findings?

Dr. Packiam: So we’re working on assembling that broader multi-institutional cohort, and that’ll be additional retrospective data just to further validate and support these findings. We are working with a variety of different prospective trials. It would be great to validate this on BRIDGE, but it’s going to take time for even their primary endpoint to read out.

In the meanwhile, I know that the main group working with this is looking at other trials like the SWOG study, which is looking at different strains of BCG. And there’s a few other prospective studies as well. I think it’ll also be really interesting as some of these BCG combination trials get published. Because there’s I think three or four large prospective studies where they’re combining BCG with immunotherapy. And once we get those results, it’ll be great to again see can we rationally decide which patient’s going to benefit from the additional therapy.

Dr. Nizam: Yeah, I agree with you. And I think something, one other question I had is, so the vendor that you are working with who is reviewing these slides or developing the AI model, I’m guessing it’s a centralized vendor, and so all the… Well, when you do the multi-institutional study, it’s going to be all analyzed under the same model, correct?

Dr. Packiam: Correct, correct. Yeah. Valar Labs is the company that’s developed the test and they have a CLIA certified centralized lab that everything gets shipped over to.

Dr. Nizam: Very good. Very interesting. What other applications do you see for such models going forward, especially in the realm of bladder cancer, where our field is very rapidly evolving and we have a lot of treatment options and there’s such biological heterogeneity in the disease, whether it’s in the non-muscle invasive state or the muscle invasive state or the metastatic state. What other things can we augment in addition to our traditional risk models and now we have AI being developed? What other avenues do you see for AI in urethral carcinoma?

Dr. Packiam: Yeah. I think even in non-muscle invasive bladder cancer, we’re making so many decisions based on gestalt and just clinical experience, which I think could be a little more guided by real results. For instance, we decide to do a repeat TRBT on most patients with high-grade T1 disease, but there are some patients with just a little focal involvement of the lamina propria that maybe don’t need a repeat TRBT every single time. So having additional data to guide that would be helpful.

With all of these different new treatments that are coming out there, we are saving more and more bladders. So a lot of patients don’t end up getting cystectomy, but they definitely get more subclinical chemical cystitis over time. And a lot of these patients develop kind of end-stage bladder symptoms. So it would be nice to rationally decide how long certain patients need maintenance for. Can we do away with maintenance in certain patients that are getting more bladder symptoms over time? So there’s kind of a slew of different non-muscle invasive bladder questions I think that we can better answer.

Dr. Nizam: Yeah, I agree. So definitely similar to other fields in oncology, we are trying to personalize medicine with our new technology and it’s an exciting time for patients and then also for us investigators as well. Thank you Dr. Packiam for joining us, and thank you for your work.

Dr. Packiam: Absolutely. Thanks so much for having me.