Yair Lotan, MD, on the development of an AI algorithm to predict BCG response in NMIBC

In this interview, Yair Lotan, MD, highlights the development and validation of an AI algorithm designed to predict response to BCG in patients with non–muscle-invasive bladder cancer. Lotan is the chief of urologic oncology and a professor of urology at UT Southwestern Medical Center in Dallas, Texas.

Video Transcript:

There is a company that I work with, Valar [Labs], that is helping develop an AI technology for predicting outcomes in patients with non–muscle-invasive bladder cancer, specifically patients who are getting BCG for high-risk disease. They're also working on algorithms for intermediate risk disease as well, but for now, they're this initial research was looking at patients who are BCG naive, with high-risk features.

Whenever you are developing an AI algorithm based on pathology, there are several steps. First of all, you have to teach the AI program what they're looking at. So, for example, what does a cancer cell look like? What does an immune cell look like? What's a blood vessel? What's a fibroblast? What's a stroma? You have to make sure that you test that over and over to make sure that the AI program recognizes the cells with a very high rate of accuracy. The next step is, if you're looking at scan slides, you don't want to necessarily specify which platform is used to scan the slide. You want to test the accuracy of this identification across different slide scanning platforms. Those 2 steps were done. The good news is that any given slide from a bladder cancer patient has thousands of cells, so you can quickly take a small number of patients and give information to the AI program for hundreds of thousands of cells, so you can make sure that the algorithm works in a robust way. Those were the first steps: Can the AI program identify what cancer looks like vs benign elements, and can you use different platforms?

The next step was putting together a cohort of almost 1000 patients who had high-risk bladder cancer and were getting BCG, and then taking 300 or so patients and telling the program what the outcomes were. And [we] said, "Okay, look at these slides. We taught you what cancer looks like. We taught you what the other cells look like. We're telling you who recurred or progressed, and we told you who didn't. Identify a pattern." We also looked at clinical features, since you already know if a patient has multifocality, carcinoma in situ, a large tumor, a small tumor, etc. We found that certain clinical features matter. Multifocality, for example, increases your risk for recurrence, and invasion to the lamina propria, T1 disease, increases your risk for progression. We took those clinical features and added them to our AI algorithm. Then we said, "Okay, this is locked now. This is what the program is going to use."

Then we took over 600 patients, and we didn't tell the AI program what happened with them. We asked it to predict based on this algorithm. We were able to identify a group of patients who are at high risk for progression, 3 times higher risk for progression and 3 times higher risk for recurrence, compared to other patients using this algorithm. Then we compared it to some of the known algorithms from the EORTC, or the European Association of Urology, which just uses clinical features. We found that this artificial intelligence program actually performed better than either of those algorithms did just using clinical features. We think that this is a potentially useful algorithm, because if you know a patient is at higher risk, it [may] help you select more intensive treatments or more intensive surveillance.

There are several trials right now looking at BCG plus checkpoint inhibitors. Those combination trials haven't been published yet; they may show an increased efficacy, but they definitely will show an increased cost and toxicity of adding a checkpoint inhibitor. So, maybe you don't want to use it for everybody, but if you could identify a patient who is 3 times more likely to progress, maybe that's a good patient to consider. There are other treatment options, like intravesical gemcitabine/docetaxel, that you could theoretically use. There's a bridge trial comparing gemcitabine/docetaxel to BCG. If you knew a patient wasn't going to respond to BCG, maybe they would do better off using a different combination treatment.

It's still relatively early, but this study in close to 1000 patients validated this algorithm, and it's now being studied prospectively as well. Hopefully these tools, which have some advantages of not using up tissue and potentially not being as expensive as genetic tests, might actually be easy tools for a pathologist to incorporate into their standard practice when they give you your results and they say, "You have a high-risk patient, and by the way, this algorithm says they're at a higher than average risk for recurrence of progression."

This transcription has been edited for clarity.