iwMDS 2025 | Implementing successful AI-driven models into clinical practice in the MDS/MPN setting

We’ve done two big projects in Spain with the Spanish group of MDS and myelofibrosis. And nowadays we have risk stratification methods supported by these entities being recurrently used and recommended for risk stratification within our country. Okay, so this is a good starting point. But more recently we’ve provided an AI-driven prognostication tool for myelofibrosis patients who underwent allogeneic stem cell transplantation, and this was a project supported by the EBMT. So we trained the model within the EBMT database, which encompasses 5,000 myelofibrosis patients treated with allogeneic stem cell transplantation. And what this project enabled us was to better identify high-risk patients, to identify more high-risk patients who, after transplantation, sadly performed very poorly. And this is important because it helps guiding which patients should be prioritized for allogeneic stem cell transplantation compared to others who might face more difficulties, And I think this was a very well accepted project because it’s not only machine learning, it’s not only interesting data, it’s not only a high impact journal publication but it was also supported by the EBMT. 

Actually there is some black box effect with AI and people are concerned about that. Well, it’s interesting because if you make the same question to people or you reverse the question and tell people, do you know how a Cox regression model works? Well, they will probably tell you, oh, I have no idea about the statistics between Cox regression, and we all accept that as a standard. So I have to admit that there is some black box effect on all machine learning models, and that may limit their interpretability somehow, but there are tools which are readily applicable that enable you to deconvolute the decision-making that the AI did and tell you, you know, I paid a lot of attention to this parameter and this parameter in this patient to make this prediction, whereas in this other patient, these other parameters were more important. And that’s relatively easy to do. It needs more computational power, so it needs obviously more support, more funding. 

Another thing that we have done in our projects is, you know, you can make a web-based calculator with the model, and rather than using this complex methodology to identify which are the important variables the model is paying attention to, you can also always just select a standard patient and make some modification in one of the variables and see if the risk changes a lot, because if you change one variable and the risk changes a lot then the model is paying a lot of attention to that variable. 

Regarding diagnosis, not prognostication, I think there are some tools that enable you to create heat maps indicating in which parts of the images are the models focusing more to make that decision. And in some cases it has been surprising to see how the machine learning model paid more attention to some elements that morphologists do not care so much about to make their predictions.

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