In the study that we performed this time with a collaboration with the European Bone Marrow Transplantation Society, we tried to ask the question, if, in 2022, using more advanced statistical methodologies such as AI-based method, now referring to machine learning methods and also deep learning algorithms, it is possible to improve the risk of death and also risk of non-relapse mortality for patients receiving allogeneic stem cell transplantation...
In the study that we performed this time with a collaboration with the European Bone Marrow Transplantation Society, we tried to ask the question, if, in 2022, using more advanced statistical methodologies such as AI-based method, now referring to machine learning methods and also deep learning algorithms, it is possible to improve the risk of death and also risk of non-relapse mortality for patients receiving allogeneic stem cell transplantation.
In this study, we collected the data of almost 50,000 patients from the Registry. All the patients were older patients affected by the most prevalent type of blood cancers, referring to myeloid diseases, lymphoma, and non-Hodgkin and Hodgkin lymphoma, and also plasma cell dyscrasias. In our study, we tried to consider all the variables which had prognostic factor, which was already known by the transplant community. However, what we show is that even using this AI-based methodology, the advantage over a multivariate logistic regression was not so great, but also, the area under the curve and the C-index of all these new scores were not better than logistic regression. In absolute terms, the C-index of this type of test was poor in the sense that the trend is from 0.57 up to 0.64, which means that also trying to use this AI-based methodology, we didn’t improve our risk of stratification and prognosis for our patients.
The conclusion of this study is that, maybe, of course, AI methodology represents a very powerful tool today and in the future to stratify patients. However, the quality of the data, the type of the data that we are using is fundamental. Otherwise, this algorithm won’t improve anything. In this case, the EBMT Registry has a lot of clinically significant variables, but maybe they’re not sufficient to improve our capacity to predict the risk of death of such patients and their overall mortality. Maybe in the future it would be ideal to use and integrate clinical variables with biological variables, and also maybe a dynamic evaluation of such factors, not just in one time point. And that is probably the way we should go in the next years.
