2026-08-31 京都大学
移植後の経過を重ねるにつれて情報が蓄積し、その先の見通しが次第に明確になっていく様子を表した本研究のイメージ図。イラスト：Robin Hoshino
＜関連情報＞
- https://www.kyoto-u.ac.jp/ja/research-news/2026-08-31-1
- https://www.frontiersin.org/journals/immunology/articles/10.3389/fimmu.2026.1883360/full
同種造⾎幹細胞移植後の慢性GVHDリスク層別化のためのスタックド・アンサンブルモデルの開発：⽇本全国コホート研究 Development of a stacked ensemble model for risk stratification of chronic GVHD after allogeneic HSCT: Japanese nation-wide cohort study
Makoto Iwasaki ,Junya Kanda,Fumihiko Kimura,Sachiko Seo,Yoshiko Atsuta,Naoyuki Uchida,Noriko Doki,Takahiro Fukuda,…,Akifumi Takaori-Kondo
Frontiers in Immunology Published:25 August 2026
DOI:https://doi.org/10.3389/fimmu.2026.1883360
Abstract
Effective risk stratification is vital for donor selection and treatment strategies in allogeneic hematopoietic stem cell transplantation (HSCT). We developed a prediction model for mid- to long-term outcomes after HSCT using a stacked ensemble model (SEM). Using data from the Japanese Transplant Registry Unified Management Program, we analyzed 14,430 patients alive without chronic GVHD (cGVHD) or relapse on day 100 after their first HSCT for hematologic malignancies between 2010 and 2018, predicting 24-month outcomes from pretransplant variables and posttransplant acute GVHD (aGVHD) information, including its treatment, accrued by days 30, 60, and 100. Data were randomly divided into training (80%) and validation (20%) sets, with 14 pretransplant risk factors as input variables. SEM achieved the highest C-index across evaluated endpoints, cGVHD, non-relapse mortality (NRM), and all-cause mortality (ACM), significantly exceeding the weaker learners for all endpoints and, using pretransplant factors, the strongest learner for NRM and ACM as well (both p=0.01), with a smaller, non-significant margin for cGVHD (C-index for cGVHD/NRM/ACM—SEM: 0.574/0.652/0.642, Cox-PH: 0.553/0.635/0.615, Random Survival Forest: 0.564/0.638/0.613, XGBoost: 0.556/0.633/0.627, Dynamic-DeepHit: 0.508/0.607/0.564). The C-index increased as posttransplant aGVHD information accrued (day 30: 0.581/0.655/0.649; day 60: 0.602/0.688/0.656; day 100: 0.606/0.696/0.664). Grade III to IV aGVHD showed predictive contribution to NRM, ultimately impacting ACM. Our SEM-based model offers a useful framework for predicting post-HSCT outcomes and highlights the critical impact of early aGVHD events on long-term prognosis.
Highlights
- The SEM matched or significantly exceeded every individual learner, including the cause-specific Cox proportional hazards model and advanced machine-learning survival models.
- SHAP from SEM showed grade III–IV aGVHD impacts NRM/ACM, whereas grade II–IV aGVHD and treatment response affect cGVHD risk.
