Predicting algorithm for GRFS after allogeneic hematopoietic stem cell transplantation: A study by the SFGM-TC

Abstract

Abstract Context : Allogeneic stem cell transplantation (alloHCT) is used in majority of acute myeloid leukemia and myelodysplastic syndromes to obtain complete remission. Graft-versus-Host disease (GVHD) and relapse are is the most important complications after alloHCT. Numerous prognostic scores are used pre-transplant to guide therapeutic strategies, but none has been sufficient to predict graft-versus-host disease/relapse-free survival (GRFS). In this study, we studied clinical, biological and therapeutic factors associated with GRFS and attempted to build a predictive score. Material & Methods : Patients over 18 years of age who received a first hematopoietic stem cell transplant between 01/01/2015 and 31/12/2022 at the University Hospitals of Saint-Etienne, Lyon, Grenoble-Alpes and Clermont-Ferrand for acute myeloid leukemia or myelodysplastic syndrome were included. Data was extracted from the allogeneic EBMT registry and patient medical records. Several parameters about clinical and biological characteristics of patient, disease, conditioning regimen and type of donor were studied. Statistics were performed with Kaplan-Meier method and Cox analysis. Regression models and supervised machine learning models were explored by the Henri Fayol Institute of the Ecole des Mines de Saint-Etienne in order to establish a predictive model. Results : Analysis included851 patients. The cohort consisted of 59% men, with a median age of 58 years. More than 70% of the patients received alloHCT for acute myeloid leukemia. The median follow-up was 938 days, and median overall survival was not reached. Median GRFS was 210 days. At the end of follow-up, GRFS was estimated at 30%. Log-rank tests found 15 significant variables. Concerning pre-transplantation characteristics, Performance Status ≥ 2, neutropenia, lymphopenia, low creatinine level, elevated lactate dehydrogenase or c-reactive protein, and pre-transplant hypoalbuminemia were associated with lower survival. Therapy-related disease or secondary to a predisposing hematologic disorder was also predictive of GRFS, as well as adverse 2022 European LeukemiaNet risk group. Patients with refractory disease or positive residual disease before transplantation had poorer survival. A mismatch unrelated donor or a graft CD3+ cell count greater than 10^7/kg were associated with reduced GRFS. Finally, alloHCT performed more than 12 months after diagnosis was linked to improved GRFS. We used these variables to develop predictive models for GRFS using linear regression and machine learning algorithms (Alternating Decision Trees, Random Forest, Support Vector Machine, Association Rules). The highest predictive performance power was achieved by association rules. Non-redundant rules, with a lift greater than 1, a minimal confidence of 0.8 and a minimal support of 0.011 (9 patients) were kept to predict GRFS at 6 months, 9 months, 1 year, 18 months and 2 years. More than 800,000 rules were generated. Accuracy was respectively 0.95, 0.96, 0.90, 0.80, and 0.73 for each timepoint. An online tool is under development to simplify the calculation of the score and will be presented at the meeting.Conclusion : Our findings are consistent with existing literature on GRFS and overall survival following alloHCT. We identified significant clinical, biological and therapeutic parameters for development of a predictive algorithm for GRFS using association rules algorithm.