ASH 2022 | Using AI to predict patient response to CAR-T cell therapy: analysis of the JULIET trial
Margot Jak, MD, University Medical Center Utrecht, Utrecht, The Netherlands, describes a project using artificial intelligence (AI) to analyze PET/CT data from the JULIET trial (NCT02445248) and identify patients with relapsed/refractory (R/R) diffuse large B-cell lymphoma (DLBCL) that will respond to CAR-T cell therapy. Early results have shown that this AI model can identify a subgroup of patients at very high risk of poor outcome. Future studies will be needed to clarify whether this AI model can improve patient selection for CAR-T therapy, and Dr Jak also shares some insights into the potential use of explainable AI in hematology. This interview took place at the 64th ASH Annual Meeting and Exposition congress held in New Orleans, LA.
These works are owned by Magdalen Medical Publishing (MMP) and are protected by copyright laws and treaties around the world. All rights are reserved.
Transcript (edited for clarity)
Artificial intelligence is a very new technique, especially artificial intelligence on PET/CT scans, which we performed in our study. And it’s a whole new field, actually. A lot of things have to be worked out, I think, in the future. For my project, we analyzed PET/CT scans for CAR-T cell patients. So we did AI techniques on the PET/CT and find out that the AI technique can predict which patient will respond to CAR-T cell therapy and which patient will not respond...
Artificial intelligence is a very new technique, especially artificial intelligence on PET/CT scans, which we performed in our study. And it’s a whole new field, actually. A lot of things have to be worked out, I think, in the future. For my project, we analyzed PET/CT scans for CAR-T cell patients. So we did AI techniques on the PET/CT and find out that the AI technique can predict which patient will respond to CAR-T cell therapy and which patient will not respond. So there is absolutely an unmet need in CAR-T cell therapy to improve patient selection, because it’s a very expensive treatment and it’s a burden for patients. It’s a very heavy treatment with a lot of toxicity, so it’s very important that we improve patient selection. And there, AI can be very helpful, I think. For every CAR-T cell patient, a PET/CT is made, so that’s standard of care.
And with our AI model, we can predict with a very high certainty which patient will not respond to the therapy. So of course, this is only one dataset that we used, so we have to validate this in other datasets and see if we can move this AI model forward to the clinic and see if this can help us improve patient selection. And of course, there are a lot of other AI techniques that we could use, not only on PET/CT scans, but, for instance, also on the biopsies, on molecular biology. For instance, circulating tumor DNA. There’s a lot of things you can think of. So yeah, a long way to go, but this is, I think, the first step towards implementing AI in hematology in CAR-T cell therapy.
For doctors, it’s always difficult to accept an answer and you don’t know why you got this answer because this is a big problem of AI, it’s a big black box. So you got an answer out of the AI. So in this case, the patient responds or the patient will not respond. But why you got this answer, that is not known. So there’s an alternate technique that’s called explainable AI, so you need a lot of data. So actually, we need a lot of real-world data, for instance, to find out why this AI is telling us that this patient won’t respond. Because if we know why, then we maybe can do something about it before we give the CAR-T cell. So this is actually something that we would like to do, to do this explainable AI technique. So there are techniques to do this, it’s known, but the problem always with AI is that you need a lot of bulk of data.
Read more...
