Zachary Klaassen: Hi. Welcome to AUA 2025 in Las Vegas, Nevada. My name is Zach Klaassen. I'm a urologic oncologist at the Georgia Cancer Center in Augusta, Georgia. I'm pleased to be joined on UroToday by Dr. Jethro Kwong, who is a urology resident at the University of Toronto. Jethro, thanks so much for joining us on UroToday.

Jethro Kwong: Thank you so much for having me. It's a pleasure to be here.

Zachary Klaassen: You've done some incredible work in the early part of your career, looking at artificial intelligence in bladder cancer. And I know we're going to talk about Progression, which is your guys' AI marker, particularly in the intermediate risk setting and some of the work you presented at AUA looking at a validation in an international cohort. So before we get into all that, maybe just bring our listeners up to speed about this intermediate risk group, which is kind of a catch all for everything that doesn't land in low grade and doesn't land in high grade.

Jethro Kwong: Yeah, for sure. I think you hit the nail on the head how intermediate risk really is whoever doesn't qualify as low-risk or high-risk patients.

Zachary Klaassen: Right.

Jethro Kwong: And even when you look at all the major urological guidelines that we have, there's still some subtle variations in how people define and treat intermediate-risk disease. And so a few years ago, the International Bladder Cancer Group had put forth these recommendations of using a risk factor approach. So counting the number of risk factors that a patient has, whether that's multiple tumors, tumor greater than three centimeters, prior intravesical or failure of intravesical therapy, early recurrence within a year, or frequent recurrence more than once a year. And then you can substratify these patients into either 0, 1 to 2, or 3 or more risk factor groups.

And there was a recent study from the Young Academic Urologist Urothelial Working Group that very nicely showed that this approach can separate out the patients in terms of recurrence. But in terms of progression, the 0 and the 1 to 2 risk factor group, the progression curves kind of overlap. And so in the ideal world, we would want to have a model or a system that can substratify patients into, I think, three subgroups.

Zachary Klaassen: Sure.

Jethro Kwong: One where, potentially, it behaves similar to low-risk patients, and then your intermediate, and then one that behaves like high-risk patients. And so this was kind of a clinically unmet need that we were hoping to address with the study.

Zachary Klaassen: Excellent. Before we get into the study at AUA, just tell us about how progression was developed, some of the previous work that led up to this international validation.

Jethro Kwong: Yeah. So to take a step back, we actually started with a systematic review looking at all of the AI studies predicting NMIBC recurrence or progression. And so this review was published in npj Digital Medicine. And here we used a tool called Appraise AI. And this is a tool that we've developed to methodologically and quantitatively evaluate the methodological and reporting quality of clinical AI studies.

Zachary Klaassen: Right.

Jethro Kwong: And we found that, unfortunately, a majority of the AI studies in this space were actually low quality. And some of the issues were things like small, unrepresentative data sets; inconsistent outcome definitions; poor methodological conduct; reproducibility issues; and then inadequate model evaluation. And so last year, together with my supervisors, Dr. Kulkarni and Dr. Zlotta, we had developed our AI model PROGRxN-BCa. And we evaluated it in a Canadian setting. And we showed that it performed quite well. It had some promising results. And this year, at the AUA, we presented our international evaluation using what we believe is the largest NMIBC cohort in the world. It's over 12,000 patients, and it's actually over 100 times larger than the median cohort size of prior AI studies in this space.

Zachary Klaassen: Wow. That's great. Tell us a little more about the design and maybe the centers that contributed. And you mentioned over 12,000 patients. That's incredible.

Jethro Kwong: Yeah. So this was an international retrospective cohort study. And certainly a study of this type would not be possible without the collaboration with all of our international partners. So certainly want to give a special thanks to Dr. Girish Kulkarni, Dr. Alexandre Zlotta, Dr. Andrew Feifer from Trillium Health Partners. And then with the Canadian Bladder Cancer Research Network, Dr. Wes Kassouf, Dr. Peter Black, and Dr. Rod Breau. We had Dr. Ashish Kamat's group from MD Anderson Cancer Center.

We had Dr. Romain Diamand from Jules Bordet Institute, and then also Drs. Paolo Gontero, Bas van Rhijn, and Dr. Richard Sylvester from the EAU group. And so this is a truly international study. What we did was we had our locked AI model. So PROGRxN-BCa was trained on over 3,000 patients from four academic and community-based institutions from Canada.

It's trained to predict progression using 14 clinical-pathological features. And this was an intentional choice because we wanted our model to be generalizable. And at the same time, we recognize that not every institution may have access or the ability to upload digital histopathology slides, or they may not have access to advanced imaging. And so this was an intentional choice. And then we externally validated our model on over 9,000 patients from 30 institutions across Canada, the US, and Europe.

Zachary Klaassen: That's great. And what were the key results that you guys found in your study?

Jethro Kwong: Yeah. So in the international evaluation set, our AI model PROGRxN-BCa achieved a C-index of 0.79.

Zachary Klaassen: Wow.

Jethro Kwong: And essentially, this is about 10% better than our comparator, which was the EAU risk calculator. And we selected this as our baseline comparison because this is currently guideline endorsed, and it's the only guideline-endorsed tool that uses our current WHO 2004-2022 grading system. And so it's generally about 10% better.

We also looked at performance across clinically relevant subgroups—age group, sex, socioeconomic status, tumor history, whether or not patients had just papillary-only disease, or papillary with concomitant CIS, as well as prior intravesical therapy. And another thing that was very important to us was looking at whether or not patients received what we consider as guideline-concordant care—so repeat TURBT for T1 tumors and then BCG when it's appropriate.

Zachary Klaassen: Right.

Jethro Kwong: This is important, because at the end of the day, we want our model to be generalizable.

Zachary Klaassen: Yeah.

Jethro Kwong: And we realized that there are probably many different reasons why patients may not receive guideline-concordant care, whether that's patient comorbidities, BCG shortage, or just variations in clinical practice. And we found that, reassuringly, the performance between those that received guideline-concordant or non-guideline-concordant care was essentially the same.

Zachary Klaassen: Right.

Jethro Kwong: The second big takeaway in terms of results is the substratification.

Zachary Klaassen: Right.

Jethro Kwong: So we've talked about the importance of substratifying the intermediate-risk group. And so here, we've actually assembled the largest intermediate-risk disease cohort to date. And we showed that our AI model is able to separate out these patients into distinct risk tertiles.

So you have the lower tertile, and that accounts for about 17% of intermediate-risk patients. Their five-year progression risk is 2%, so very close to low-risk patients. Intermediate had a progression risk of 7%. And then the upper tertile, which accounts for 9% of patients in the intermediate-risk group, they had a five-year progression risk of 17%.

Zachary Klaassen: Wow.

Jethro Kwong: We were able to do something similar with high-risk patients as well, separating them into two subgroups, one with 10% and one with 26% five-year progression risk.

Zachary Klaassen: It's a great summary. Those results are really exciting. I mean, I think we were talking offline a little bit about that IBCG paper and how it's a beautiful paper, European Urology paper. But to show that you guys are now looking at progression in this substratification, maybe just talk a little bit about the clinical relevancy and how we can maybe apply this in the future.

Jethro Kwong: Yeah, for sure. So, at the end of the day, PROGRxN-BCa was designed to predict progression.

Zachary Klaassen: Right.

Jethro Kwong: It hasn't been tested for recurrence. And at the same time, the IBCG recommendations, it's been validated, and it shows very nicely that it can separate out patients in terms of recurrence.

Zachary Klaassen: Yeah.

Jethro Kwong: And so I think, at the end of the day, PROGRxN-BCa can be used in conjunction with the IBCG recommendations, because now you have two well-validated systems—one that works for recurrence, one that works for progression.

Zachary Klaassen: Yeah. Well said. It's been a great conversation. You have two phenomenal mentors in Kulkarni and Zlotta. I know them both very well. You've done some fabulous work. And I'm excited to see where the next steps go in your career. But maybe just to wrap this all up with this exciting work at AUA, just give our listeners a couple take-home messages.

Jethro Kwong: Yeah, for sure. So, again, this is the largest NMIBC prognostic study to date. We've showed that our AI model, PROGRxN-BCa, works well. It works well for patients that received or did not receive BCG. It works well for patients who did or did not receive guideline-concordant care.

Zachary Klaassen: Right.

Jethro Kwong: It's publicly available on our website, progrxn.ca, and users or patients can upload their clinical-pathological information, similar to a nomogram, and they can get a personalized progression risk assessment. The advantage is you can also toggle on and off between whether or not patients receive BCG. And so you can see what the impact of this treatment can have on your progression risk. So I think this can be certainly something that's very usable in the clinical setting.

Zachary Klaassen: Yeah.

Jethro Kwong: The second takeaway I think is in terms of substratification. We've talked about this at length and how it has implications for future guidelines on how it can inform patient counseling, management, and so forth. And then the last thing I would say is for those working in the clinical trial space—

Zachary Klaassen: Absolutely.

Jethro Kwong: I think this potentially can broaden eligibility criteria for enrollment. So we've shown how for intermediate risk, we've identified 9% of patients that actually behave very similarly to high-risk patients. And so there's room for expanding these patients for clinical trials there. And the other thing that we're very interested in looking—especially timely now, now that the space has absolutely exploded with new trials—is whether or not our tool can predict response to these new treatments. So those are some of the things kind of on the horizon.

Zachary Klaassen: Fantastic. Great presentation of your data. Really exciting stuff. I mean, the AI in this disease space is just incredible. So thanks, again, for joining us on UroToday.

Jethro Kwong: Thank you so much for having me.