ASH 2025 | AI in myeloid neoplasm risk models: addressing the “black box” problem

We connect back again to the problem of the variability in the different institutions. So to start with, currently we’re working on different models. The idea to solve the black box problem, so the fact that you cannot understand how an artificial intelligence is working behind the scenes, is to try to attach to it an explainability framework. We are working towards that as much as we could, as much as we can. Not always is that possible, but we really think that is needed for a clinician to understand how the model is reasoning behind the scenes, because these are probably probabilistic tools so in certain cases they might have problems and so a clinician needs to understand if clinically what the model is deciding is relevant or if it’s wrong so our solution is to use explainability frameworks again in order to create these tools you need to have a lot of data and unfortunately again with images there is a huge variability so we’re working again with federated learning in collecting data from multicentric institutions in order to and homogenizing them and working with all this data, we want to create more robust models. This is also something that we need to do if we want to integrate this kind of features into a new score such as EPSSN. We need to be sure that the features extracted from the different centers are robust and they’re extracted in the same way without this variability impacting the result. So again it’s still a bit early but the results and also the fact that we are in a consortium called Patroplus where we want to work in a federated way with different institutions on this problem hopefully will give us results in the next few months. Thank you very much for the questions.

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