By adding the post-transplant progress, future risks become clearer - Chronic GVHD/mid- to long-term prognosis prediction using 14,430 cases nationwide and machine learning -

Researcher information

overview

A research group led by Professor Akashi Takaori, Lecturer Junya Isada, Research Fellow Iwasaki Atsushi, and Director Yuko Atsuta of the Japan Hematopoietic Cell Transplant Data Center at Kyoto University Graduate School of Medicine has developed a machine learning model to predict post-transplant chronic graft-versus-host disease (chronic GVHD) and medium- to long-term prognosis using allogeneic hematopoietic stem cell transplant registries across Japan. Previous predictions were mainly based on patient, disease, and donor information before transplantation, but in this study, we sequentially added the onset, severity, treatment, and treatment response of acute GVHD obtained up to 30, 60, and 100 days after transplantation. Using a stacked ensemble model (SEM) that combines multiple statistical and machine learning techniques, the ability to predict chronic GVHD, non-relapse mortality, and all-cause mortality within 24 months improved as post-transplant information accumulated. Severe acute GVHD was important in predicting non-relapse mortality and all-cause mortality, and acute GVHD and response to initial treatment were important in predicting chronic GVHD. This result is expected to become the basis for ``dynamic prognosis prediction,'' which updates future risks not only at a single point in time before transplantation, but also according to the progress after transplantation. The results of this research were decided to be published in the international academic journal "Frontiers in Immunology" on August 5, 2026, and published online as an accepted paper.

image