ASH 2025 | Data-driven hematology: next steps in the implementation of the CLL treatment infection model

So my research group spent a lot of time trying to develop data-driven hematology. Some people would call it AI. I actually call it pattern recognition because what we aim for is to use the huge amount of health data that we have available in a country like Denmark where we actually have a system where we collect the health data on a population-based level. So it’s a public health system and we can retrieve health data across the nation for at least the last 20 years. And using that kind of data now we have more than 3,000 different variables, just routine data per patient. We want to use that data to identify patterns for patients in specific need. We have developed what’s called the CLL treatment infection model, where we use this data-driven approach at the time of diagnosis of patients with CLL, identifying the patients who have a high risk, that’s more than 70% risk, of either having a serious infection or receiving CLL treatment within the first two years of that disease trajectory. And for these patients, we have included them in a clinical trial, randomizing them between three cycles of acalabrutinib and venetoclax as compared to standard of care, which would be wait and watch. And what we aim for here is not getting rid of the CLL disease, but turning the immune dysfunction or the natural history of infections for patients with CLL into a state of more indolent CLL without the increased risk of infections. So that’s one way of using AI or pattern recognition or data-driven hematology. We see, that was recently at the ASCO Congress, where we see a kind of disappointment in terms of developing these decision support tools because a lot of people have been working and doing that and been a bit frustrated that you need really to thoroughly test exactly in the same way that you do with new drugs you want to introduce to a patient population. You need to test that your new algorithm or pattern recognition is actually robust and that you would be able to also in a different health system, in a different hospital, maybe in a slightly different patient population, still use your algorithm. So I think this frustration that I’ve experienced discussing deployment of AI tools into the clinic to actually improve data-driven hematology is kind of turning our heads around to see that data-driven hematology needs to be thoroughly tested. Where this might bring us in the next five to ten years, I believe, is to actually, for a lot of clinical trials, use pattern recognition to identify different patient populations that might need slightly different management and we need to test that thoroughly as we do with new medications in clinical trials and we need maybe also when we deploy it into clinical practice and we have done that in an Epic-based health system covering half of Denmark now where we just automatically provide the prediction from the CLL treatment infection model in the patient context to the treating physician. We need to test these kinds of deployments in the real clinical setting but in a way where we actually monitor the performance and make it an integrated part where we can see if the prediction keeps giving the right prediction and giving the predictive performance that it was actually trained for.

This transcript is AI-generated. While we strive for accuracy, please verify this copy with the video.