AUTHORS
Carmelo Gurnari, MD,PhD, Adrian Mosquera, MD2, Luuk Gras3, Joe Tuffnell4, Linda Koster5, Francis Ayuk6, Henrik Sengeloev7, Didier Blaise, MD8, Victoria Potter9, Tobias Gedde-Dahl10, Stephan Mielke11, Thomas Schroeder12, Matthias Stelljes13, Jakob Passweg14, Igor Wolfgang Blau15, Francesca Kinsella, PhD16, Jan Vydra17, Lucia Lopez Corral18, Marco del Olmo Linares19, Carlos Perez Miguez, Bioinformatics Software/Biomedical Data Engineer20, Radovan Vrhovac21, Edoardo Benedetti22, Nicola Polverelli, MD23, Francesco Onida24, Christof Scheid, MD25, Joanna Drozd-Sokolowska26, Kavita Raj27, Marie Robin, MD28, Donal McLornan29
AFFILIATIONS
1 Department of Biomedicine and Prevention, University of Rome Tor Vergata, Rome, Italy, Translational Hematology and Oncology Research Department, Taussig Cancer Center, Cleveland Clinic, Cleveland, OH, USA, Cleveland, United States
2 University Hospital of Santiago de Compostela, Santiago de Compostela, Spain
3 EBMT Leiden Study Unit, Leiden, Netherlands
4 EBMT Leiden Study Unit, Leiden, Netherlands
5 EBMT Leiden Study Unit, Leiden, Netherlands
6 University Hospital Eppendorf, Hamburg, Germany
7 Rigshospitalet, Copenhagen, Denmark
8 Programme de Transplantation&Therapie Cellulaire, Marseille, France
9 Kings College Hospital London, London, United Kingdom
10 Oslo University Hospital, Rikshospitalet, Oslo, Norway
11 Karolinska University Hospital, Stockholm, Sweden
12 University Hospital | Essen, Essen, Germany
13 University of Muenster, Münster, Germany
14 University Hospital | Basel, Basel, Switzerland
15 Medizinische Klinik m. S. Hamatologie , Onkologie und Tumorimmunologie, Berlin, Germany
16 Birmingham Centre for Cellular Therapy and Transplant (BCCTT), Birmingham, United Kingdom
17 Institute of Hematology and Blood Transfusion, Prague, Czech Republic
18 Complejo Asistencial Universitario de Salamanca, Salamanca, Spain
19 Group of Computational Hematology & Genomics (GrHeCo-Xen), Health Research Institute of Santiago de Compostela (IDIS), Santiago de Compostela, Spain
20 Group of Computational Hematology & Genomics (GrHeCo-Xen), Health Research Institute of Santiago de Compostela (IDIS), Santiago de Compostela, Spain
21 University Hospital Center Rebro, Zagreb, Croatia
22 Azienda Ospedaliero Universitaria Pisana, Pisa, Italy
23 Unit of Bone Marrow Transplantation - Unit of Hematology, Fondazione IRCCS Policlinico San Matteo,, Pavia, Italy
24 Fondazione IRCCS - Ca’ Granda - Ospedale Maggiore Policlinico IRCCS, Milan, Italy
25 University of Cologne, Cologne, Germany
26 University Clinical Centre, Medical University of Warsaw, Warsaw, Poland
27 University College London Hospitals NHS Trust, London, United Kingdom
28 Assistance Publique - Hôpitaux de Paris, Paris, France
29 University College Hospital London, London, United Kingdom
ABSTRACT
Background: Allogeneic hematopoietic cell transplantation (allo-HCT) remains the only curative treatment for many patients with higher-risk myelodysplastic syndromes (MDS), but its benefit may be offset by non-relapse mortality (NRM), which competes with relapse. Conventional MDS prognostic scores, including IPSS-R and IPSS-M, were developed primarily to estimate disease-related outcomes and do not incorporate the full range of patient fitness, donor, graft, and conditioning factors relevant to transplant toxicity. Using the EBMT registry, we aimed to develop and geographically validate a machine-learning framework for pre-transplant identification of patients at high risk of NRM.
Methods: We analyzed 10,120 adults with MDS who underwent first allo-HCT between 2015 and 2022. Patients transplanted in Germany were held out a priori as a geographically independent validation cohort (n=2,276), and patients from all other countries formed the training cohort (n=7,844). Patients with >20% bone marrow blasts immediately before conditioning were excluded. All preprocessing was restricted to the training cohort. A performance-guided missingness filter selected variables with ≤7% missingness, followed by random-forest imputation and LASSO feature selection. Seventeen predictors were retained (Karnofsky performance status, donor age, donor relationship, HLA matching, haploidentical donor status, CMV serostatus combination, graft source, conditioning group, WHO disease classification, excess blasts, interval from diagnosis to HCT, and year of diagnosis). NRM and relapse were analyzed as competing events using Fine-Gray methods; Cox, penalized Cox, gradient boosting, random survival forest, and a stacked meta-learner were benchmarked for discrimination. A NRM threshold was derived exclusively in training and applied unchanged in validation. HCT-CI and IPSS-R cytogenetics, which exceeded the missingness threshold, were evaluated in complete-case sensitivity analyses.
Results: The training and test cohorts had broadly similar age and disease-risk distributions, although donor and graft practices differed. Competing-risk models achieved external validation c-indices of 0.623 for relapse and 0.621 for NRM. In the NRM benchmark, the stacked meta-learner had the highest validation discrimination (c-index 0.618), but performed only marginally better than elastic-net Cox, LASSO Cox, and conventional Cox regression (all c-index 0.616). The training-derived threshold classified approximately 14% of evaluable German patients as high-risk and 86% as standard-risk. High-risk patients were older and more frequently received HLA-mismatched grafts (43.9% vs 13.1%). HCT-CI ≥3 was more frequent in the high-risk group (38.0% vs 30.9%; P=0.020), whereas IPSS-R ≥3.5 was similarly distributed (24.2% vs 20.4%; P=0.145). The classifier identified patients with a significantly higher cumulative incidence of NRM and inferior overall survival (both P<0.0001). Adding HCT-CI or IPSS-R cytogenetics to the AI-derived score resulted in only modest improvements in OS discrimination (validation c-indices, 0.610 and 0.615, respectively, vs 0.601 for the score alone), while worsening model fit for NRM, supporting the score’s independent and parsimonious prognostic performance.
Conclusions: In this large EBMT registry study, a training-derived machine-learning framework reproducibly identified a subgroup at high risk of NRM in a geographically independent cohort. The moderate discrimination and small performance differences between algorithms suggest that rigorous validation and competing-risk modelling may be more important than algorithmic complexity alone. The NRM signal arose from a cross-domain combination of patient fitness, disease status, and donor/transplant characteristics rather than conventional MDS related variables alone, providing a clinically interpretable backbone for future models incorporating molecular and longitudinal data.
DISCLOSURES
Carmelo Gurnari, MD,PhD: GSK - Honoraria; Sobi - Research Funding; Alexion - Research Funding; Novartis - Research Funding; Genesis Therapeutics - Consultancy (Includes expert testimony)