Artificial Intelligence Modelling to Predict the Risk of Cardiotoxicity Among Renal Cell Carcinoma Patients - Hesham Yasin
July 5, 2023
Hesham A. Yasin, MD, Genitourinary Oncology Fellow, Vanderbilt University Medical Center, Nashville, TN
Pedro C. Barata, MD, MSc, Leader of the Clinical GU Medical Oncology Research Program, University Hospitals Seidman Cancer Center, Associate Professor of Medicine, Case Western University, Cleveland, OH
Pedro Barata: Hi there, and welcome. I'm very, very happy to be joined by Hesham A. Yasin. Dr. Yasin is a GU oncology fellow at Vanderbilt Medical Center in Nashville. Welcome, Hesham.
Hesham Yasin: Yeah. Thank you so much, Dr. Barata. It's wonderful to talk to you and discuss my last project.
Pedro Barata: Absolutely. I mean, congratulations, first of all, because you have another outstanding presentation at IKCS about a topic that not a lot of us are really familiar with.
Hesham Yasin: Exact. Yeah.
Pedro Barata: Right? Artificial intelligence is something that we hear about being a hot topic, but it's hard for us to actually understand how much of that... We understand the potential, but-
Hesham Yasin: Yeah, exactly.
Pedro Barata:... a lot of us have this difficulty understanding how they can do it in their own EMRs, in their own system. And, of course, you have the study that you presented, the project you presented, over 2,000 patients with advanced renal cell carcinoma, was around cardiotoxicity and VEGF therapy. So, we'll get to that in a second. I guess the first question I might have for you is, since you are an expert in this, tell us, how was it to implement such a tool in EMR, or you already had it there at Vanderbilt, and you just use it to search for the kidney cancer patients. How did you actually make it happen?
Hesham Yasin: Yeah. Okay. Yeah, perfect. So basically, I was working with many scientists in the artificial intelligence team, and basically, they took most of the lead. And so, basically, they're using some algorithm. They're using to pick up data. So, basically, they look into all the data that we have at Vanderbilt, and basically, they de-identified the data. And basically what they do, they took these data, and they do this machine learning to train it to pick up certain pattern.
So, for example, in our project, we were to identify patients that do have higher risk for cardiotoxicity. Basically taking their risk factors, and see how many risk factors, and basically give them a score to see if they have high risk for cardiotoxicity, moderate risk for cardiotoxicity, or low risk for cardiotoxicity.
And the people who were working, we have a cardiologist, who held both a table, or basically guidelines, which risk factors should be implemented, and which risk factor are high or low or in the middle. So, basically, we took this data, and basically they trained this data to identify risk factors for risk for cardiotoxicity. It's a little complex. It's a theorem.
Many algorithms they use. They're used to algorithms. It's probably a little bit hard to explain. That's required for the reading. There's a random forest model, and there is an A&N model. So, basically, once they train these models to identify patients with a high risk factor, then they did validation. They picked up another with the data they got, 80% they train it, and 20% of the data left, they do validation to see if its algorithm is working.
And it performed well, like 95% picked up correctly patients with higher risk factors. So, that's a good thing. Now, we did that as a part of a study. Now, the next step will be is to implement these algorithms in the EMR. And now we use ABIC variables, and now we are doing a pilot project to implement these algorithms in the EMR.
So, once the clinician, he's reviewing a patient, it'll pop up to him that this patient, for example with a kidney cancer receiving VEGF therapy, this patient has major risk for cardiotoxicity and can be referred to cardio-oncology.
Pedro Barata: Sure. So, if I understand correctly, you work closely with the cardio-oncology and artificial intelligence teams there at Vanderbilt-
Hesham Yasin: Exactly. Yeah.
Pedro Barata:... and you are able to provide a number of risk factors. Able to select patients for to conduct this study. That's pretty cool.
Hesham Yasin: Yeah.
Pedro Barata: So, let's talk a little bit about the findings of your study, and I'll let you expand on it. To my understanding, you identify over 2,000 patients with renal cell carcinoma. Those patients were treated with target therapies, angiogenic therapies, to be precise. TKIs. And you basically analyze the data by the risk, by the cardiovascular risk, that was predefined. And so, you allow the AI tool to search those patients. And then you assess the outcomes of those patients. Is that what happened, and can you tell us a little bit about what your findings were?
Hesham Yasin: Yeah, yeah. I mean, as I mentioned, they were like... The model itself, it's performed well in validation. So, after the training was done, we did validation, and the validation was kind of near, like 95% or 94%, for both models on renal cell carcinoma. That means these models will help identify patients with a high risk for cardiotoxicity from renal cell carcinoma in 95% of the cases.
And also, what we did after that, so we did something else. We did external validation. This means that we go to the actual records for patients, let's say, who have high risk, and see how many of those patients were actually referred to cardiology. So, we want to see how we are doing now in terms of clinicians. Are we sending these people who were identified by this algorithm with the high risk cardiotoxicity to cardio-oncologist?
And we found that around, I think, 58% of patients that we are seeing in the clinic with a high risk of cardiotoxicity, they were referred to cardio-oncology. So, it means that we do some good work, but we need some improvement for that in the future. Yeah.
Pedro Barata: Got you. So, I guess that's my follow-up question on that. I mean, sounds to me it's huge. Sounds like almost half of the patients who should be, sounds like, based on the formal risk factors that they have, they should be referred to specialized care. And for whatever reason they weren't, I'm sure the numbers can be easily extrapolated to other sites, right? So, it's not specific to your site.
I have to tell you, a lot of us don't refer patients to cardio-oncology for a lot of reasons. Sometimes the primary care doctors are doing a great job. Were you able to do with this project, to assess whether or not those cardiovascular risk factors were being addressed by other specialties, like primary care doctors?
Hesham Yasin: Yeah.
Pedro Barata: Or is that something for future steps, I guess?
Hesham Yasin: Yeah. We took that into account. So, the 58% including the patient who has been seen by another cardiologist or who has follow-up from before. So, the remaining patients, they never been seen by a cardiologist.
Pedro Barata: Got you. So, I guess my final question would be, I mean, sounds to me this is huge, but I'll ask you that question. I think, how relevant this is, and do you see these kind of tools to be applicable in other topics around kidney cancer?
Hesham Yasin: Absolutely. Yeah.
Pedro Barata: I mean, you have a great kidney-cancer powerhouse at Vanderbilt.
Hesham Yasin: Exact. Yeah.
Pedro Barata: So, what are your thoughts and insights about future steps for this work, and also other areas within the kidney cancer world that you can apply AI?
Hesham Yasin: Yeah, absolutely. I mean, we have these tools, the machine learning, absolutely we can apply it to many other areas. For example, let's say for example, we have identified patients who have a high risk for disease, or have a high risk for progression to metastatic disease. Maybe, for example, we can easily identify patients who may be good candidates for adjuvant treatment. For example, pembrolizumab. That's one example I was thinking about.
And for example, maybe I identify, maybe assist, for example, survival. It would give us good estimation for survival. Or maybe in the future, give us an indication of how much the patient will respond to immunotherapy. Although now, we don't have good predictors of which patients will respond to immunotherapy or not, but maybe that's something we can use in the future.
Pedro Barata: Got it. I promise you, that's my last question now. I said that a while ago. Are you able with this technology, let's say, or this methodology, better said, are you able to go beyond your own institution and apply this tool to multiple institutions, and turn this into a multi-institutional effort? Is that feasible?
Hesham Yasin: Yeah. I think it's feasible. Basically, it's just, if we have the same... So basically, using the ABIC, I think any system that has ABIC, we can apply the same algorithm. It did not require sophisticated tools to extract the data, just the machine learning process, and how you pick up the algorithm, how you do programming to train your data that you have. So, I think yeah, it can be applied to other areas. For example, any institution that has a good EMR, and also they have good documentation of all diagnoses, because that's very important to know what cardiac risk factor they have.
Pedro Barata: Got it. That's very informative. Thank you so much.
Hesham Yasin: Absolutely. Thank you so much.
Pedro Barata: Actually, this was great.
Hesham Yasin: Yeah, that's wonderful.
Pedro Barata: You know a lot about it, which is amazing.
Hesham Yasin: Yeah, thank you so much.
Pedro Barata: I might pick your brain at the next meeting where we're together.
Hesham Yasin: Absolutely.
Pedro Barata: Thank you for taking the time, and congratulations for your fantastic presentation. Thank you.
Hesham Yasin: Thank you so much. Absolutely. Wonderful. Nice meeting you. I'll talk to you.
Pedro Barata: Thanks, take care.