Ashish Kamat: A warm welcome to all of you, to the UroToday studios. I'm Ashish Kamat, professor of Urologic Oncology at M.D. Anderson Cancer Center in Houston, Texas. And we're live in Las Vegas at AUA 2025. It's a pleasure to welcome you to the studio, Alex.

Alexandre Zlotta: Good morning. A pleasure to be with you, Ashish, for sure.

Ashish Kamat: So Alex Zlotta, needs no introduction. Professor of urologic oncology. Done a lot of work in bladder cancer. Today we're going to talk about your insights into AI and how it's evolving in the field of bladder cancer, both from diagnosis, prognosis, et cetera. So take us through that journey a little bit.

Alexandre Zlotta: So as you perfectly know, well, there are so many things we can improve in bladder cancer.

Ashish Kamat: Correct.

Alexandre Zlotta: And there's going to be a meeting on Monday where we'll be discussing the place of AI at different steps. It can be in the prognosis. It can be to improve pathology. It can be maybe to improve even surgery. And so although, in pathology there's definitely a need to improve things. Think about it, you give a slide to two pathologists, even world renowned. You ask them to decide whether this is low or high-grade, in 15% of the cases, they will disagree. It's unbelievable. And if you have grade 1, 2, and 3, they disagree in half of the cases.

So can I improve things? Absolutely, are we there yet? Absolutely not. In terms of surgeons, we perfectly know that when you look at NMIBC outcomes, you try to control for everything, but you can't control for the surgeons and the RESECT trials showed it. Can you have an AI that controls for the surgeons, nowadays? No. So what can we do? I think that we are at a time where the AI can basically improve our classical statistical nomograms and regression analysis.

And I think Jethro Kwong just won the best award this morning. And together with my partner in crime, Girish Kulkarni, in Toronto, with you and many other centers, in North America and Europe, we have assembled the largest data set of non-muscle invasive bladder cancers, contemporary from 2005 to 2023.

And we just asked the same question as the EORTC or the EAU Calculator, or the CUETO was asking, is there for someone showing up in your office that we can predict what the risk of progression? And you perfectly know that the EAU risk calculator is good, but is not applicable in patients who received BCG or doesn't really perform well. You perfectly know that in those patients, they were included only if they had not been treated with BCG. So it's just not standard of care.

The EORTC and the CUETO are using the older classification. Again, not applicable. And therefore, Jethro and 30 institutions worldwide, assembled a control set that was about 3,000 patients in Canada. Built a model on 14 parameters, and then validated it across 9,000 and something patients, both in US, North America, and Europe.

Ashish Kamat: Just to clarify for the listeners, these parameters are clinical parameters.

Alexandre Zlotta: Clinical parameters.

Ashish Kamat: Because we do have AI or machine learning that's looking at pathology slides. But this thing we're talking about is clinical parameters.

Alexandre Zlotta: Absolutely, you're talking, grade, stage, lymphovascular, age, you name it, number of tumors.

And ultimately what it showed using C index and also decision curve analysis, the AI model clearly outperformed the existing EAU risk calculator that everyone uses and the CUETO, by roughly 10 on the scale of the-- C index was 0.81 if I remember, in the training, and then 0.79 in the test sets. And the EAU was about 10 points less.

But more importantly, what it really is changing here is that all patients were included. These were some patients that luckily in EAU risk calculator, never saw BCG, although they should. Patients treated according to guidelines, patients not treated according to guidelines. And that truly is the big, big, big plus, is that it outperforms what the existing but is also applicable to all the patients. But to move a step forward, you and I perfectly know that we struggle to discriminate in the intermediate risk category.

And even in the high risk. The IBCG has designed the risk factors and it helps. But when we compared the IBCG risk model, with the new AI model, actually, the AI clearly outperformed it. And interestingly enough, it showed that there are three types of intermediate risk. One which is literally like a low risk, a 2%. One which is like an intermediate risk real, 7, 8% of progression, at 5-year estimated. And one which is at 17%, very close to the high risk.

And so that you can think that in the future, instead of just having patients in the intermediate high-risk, we will basically guesstimate by the AI that risk and then they can be included in trials. That's another way to look at things. And last but not least, even in the high-risk category, we know about the very high risk. And you look at the EAU risk calculator, you have what 40% to 45% progression huge. But in the high risk, the AI was able to tease out a risk around 11%, and the risk around 25%.

And I'm sure that we as surgeons, will have a different conversation if we tell a patient that your risk is around 10% or your risk is close to 30%. You don't start the same conversation with patients. And that's why I really believe that, thanks to this, I really think we should thank absolutely all the centers who participated. The EAU guidelines, NMIBC brought many of the patients. M.D. in your group, of course, brought patients and throughout the Canadian consortium.

And we really should thank Jethro and Girish, to have also spearheaded this amazing work, which hopefully will be available to everyone, anywhere on Earth. It's going to be an app and just an online where any patients will walk in the office of your surgeon. You have an intent to treat with BCG or not, a recurrence in the past, a guideline because you have or you don't have access to the ultimate management that is according to guidelines. And then you can plug and play everything. And this will give you truly a risk, real time for and real world.

Ashish Kamat: That's great. And again, congratulations on the award that your group got this morning. A couple of questions if I might drill in a little bit, because it's important for us to understand that obviously all these models are built upon existing data and existing models that we have. And as you recognize, of course, the International Bladder Cancer Group, intermediate risk classification, has been shown to be beneficial in selecting patients for active surveillance.

And then recategorizing those patients. Same thing with the model that you've generated. But explain to me a little bit about how this is different from fancy statistics or fancy nomogram? Why-- where is the AI element that we should be looking at, and how are you going to develop that further?

Alexandre Zlotta: So first of all, most likely would say that Jethro would be the best person to explain in detail, those things. But the model is also evolving. And it's a real time model. And so there's also a way to integrate data all the time so that the models are updated all the time, which is very different from what has been done in the past.

The second thing that Jethro and the group will do is, to also look at the model that follows longitudinally. We only have models at one time point, but we don't take what's going on longitudinally. And so to answer what's the difference, that's going to be the difference. That the model will go longitudinally.

But then last but not least, at the end of the day, AI is a kind of statistics, one way or never. But it's very similar to the way that you train a child. You teach him 1, 2, 3, 4, 10 times, and then you get to the first step, the next step, and the third step, in which is still different from a pure regression analysis. And on pure stats where you have the key elements, you kick out anything which is related in the non-statistically significant, still the way to design it and to build and validate is slightly different.

But overall in the study, we also did a Lasso. And the Lasso was actually using the same parameter, reasonably performing what the difference was in terms of decision curve analysis, net benefit, and calibration. And the ability to be valid throughout all the subgroups of patients. So usually you're going to get improvement not necessarily overall, but for each of these specific subgroups. That's where the uniqueness in addition to the 12,500 patients, which is the largest, to my knowledge, that has ever been assembled.

Ashish Kamat: So Alex, let me ask you a provocative question. Do you think AI will replace Professor Zlotta?

Alexandre Zlotta: I really hope so. I think it's a great question and I can't really answer. I think the future is always what you think is unthinkable today. That's what the future will be in 10, 15 years. That's truly. And I know that from lawyers who are basically most of the reports are now generated by AI to a certain level. The senior lawyers will actually focus on very specific points where you absolutely need the expertise.

Where I think AI will not be able necessarily to replace us is that, my favorite line is that you can google everything, but you can't google reasoning. And I do believe that one way or another Ashish Kamat, Peter Black, Girish Kulkarni, and the others, will need their reasoning power to implement in a better way AI. That's what I really think.

Ashish Kamat: Well said. And I really think one of the reasons I don't want AI to replace you is because then we can have this talk again at future AUA meetings. So thank you again for coming. This was a great discussion.

Alexandre Zlotta: Thanks, Ashish. Thanks, everyone.