Machine Learning Model Predicts Overall Survival in Metastatic Prostate Cancer - Ali Sabbagh

June 17, 2024

Ali Sabbagh presents a novel study on using machine learning to predict overall survival in clinical trials for metastatic prostate cancer. Dr. Sabbagh explains that traditional endpoints like overall survival often require long follow-ups, delaying access to potentially life-saving treatments. His research leverages baseline clinical characteristics and PSA changes over four months to accelerate predictions of trial outcomes. By analyzing data from major trials, including TITAN and LATITUDE, his model accurately forecasts both positive and negative results, potentially transforming the speed and efficiency of clinical trials. Dr. Sabbagh highlights key metrics, such as PSA slope, which significantly impact the model's predictability. Moving forward, the focus will be on validating the model across diverse patient cohorts to ensure its broader applicability.


Ali Sabbagh, MD, Radiation Oncologist, Postdoctoral Scholar, Radiation Oncology, University of California, San Francisco, San Francisco, CA

Phillip J. Koo, MD, Division Chief of Diagnostic Imaging, Banner Health MD Anderson Cancer Center, AZ

Read the Full Video Transcript

Phillip Koo: Hi, I'm Phillip Koo and welcome to today's coverage of ASCO 2024. We have with us today Dr. Ali Sabbagh, who's a postdoc at the University of California, San Francisco, who'll also be starting his radiation oncology residency this summer. So welcome and happy to have you.

Ali Sabbagh: Thank you. Thank you so much for having me.

Phillip Koo: Congratulations. And also congratulations on presenting a very novel abstract at this meeting that talks about using a machine learning model to predict overall survival in clinical trials in patients with metastatic prostate cancer. Can you tell us more about this project?

Ali Sabbagh: Of course, yeah. So, with clinical trials in oncology, specifically in metastatic prostate cancer, the gold standard endpoint is overall survival. But the issue with overall survival is that it's going to take a long time for trials to actually achieve median survival in patients. And I'm sure you can think of a few trials where median survival took several years to be achieved. So what we wanted to do was to see whether machine learning or artificial intelligence could predict the outcome of those trials in terms of survival without having to go through the entire follow-up and using just four months of baseline clinical characteristics and PSA changes in response to treatment.

Phillip Koo: I think that's wonderful to be able to accelerate, I guess, the results and be able to obviously get quicker access, for patients to have quicker access to some of these drugs that are really impactful. So tell us more about the characteristics or the metrics you looked at and tell us about some of those that seem to have a bigger influence on that prediction of overall survival versus others.

Ali Sabbagh: Yes. So mainly we used six baseline clinical characteristics that were already associated with overall survival in the literature. And I can go through them: LDH, disease site, ECOG status, albumin, hemoglobin. So, there were mainly six baseline clinical characteristics as well as PSA changes. We had around 18,000 PSA values for all patients. So what we did was we took those values and we engineered specific features, some that were known in the literature to be associated with overall survival like PSA50 and PSA90 and some that we wanted to engineer to capture that change more granularly in a way. And that includes the PSA slope, which is essentially the decay in PSA to be able to describe it in the first four months. And we found that that slope actually had a huge impact on the model's predictability.

Phillip Koo: I think that's wonderful. I think PSA is something that, it's amazing how valuable and how good a biomarker it is and it's very cheap as well, which is nice. So can you tell us about what trials you looked at for this study?

Ali Sabbagh: We had some trials of patients with metastatic prostate cancer, six of which were phase three, and one was phase two. We had TITAN, MAGNITUDE, LATITUDE, GALAHAD, COUGAR-301 (COU-AA-301), COUGAR-302 (COU-AA-302), and ECIS. So with these trials, we split them into a training and testing set. And in the training set, we had TITAN, COUGAR-301 (COU-AA-301), and ECIS. And in the testing set, we had MAGNITUDE, LATITUDE, COUGAR-302 (COU-AA-302), and GALAHAD.

Phillip Koo: So something like this, obviously, validation's important. Can you tell us about how you go about validating the results that you found with this new model?

Ali Sabbagh: So the model actually predicts the outcome of overall survival at a trial level, right? And the way it does that is it simulates the trial a thousand times to be able to see what the hazard ratio is likely to look like at the end. So we plot that distribution and then we see whether the hazard ratio is going to be less than one or whether that distribution is going to cross one. And then we can tell where the hazard ratio is likely to fall and whether the trial is likely to be positive or negative. And actually, in three of our testing cases, in three of our trials, two of which were positive, one of which was negative, we were able to accurately predict the outcome of that trial, both in direction, whether it's going to be successful or not, but also the distribution included the hazard ratio that was seen in the trial itself, and that was essentially how we validated those results.

Phillip Koo: I think that's great to be able to predict positive results and negative results, which are just as important. So something like this obviously could have a huge impact on the landscape of trials, patient care, drug development, a lot of different areas. What are the next steps when it comes to a project like this?

Ali Sabbagh: Yeah, so far we're very excited with our results. We're very happy to see that the model works well on positive and negative trials and also on trials in patients with hormone-sensitive and castrate-resistant prostate cancer. I think the next steps would be to be able to see how well the model generalizes to new situations and new cohorts of patients because the trials that we validated on are excellent and they show a lot of promise for the model. But we do feel that we want to be able to validate on even more scenarios and more patient cohorts. And I think those are the next steps of the project for us.

Phillip Koo: Great. Wonderful. Thank you for sharing your time with us and best of luck to you as you start your residency and your training. And I'm sure we'll hear a lot of great things from you during your career. So thank you.

Ali Sabbagh: Thank you so much.