Background: Haploidentical (Haplo) stem cell transplantation (SCT) provide a curative option for nearly all Acute Myeloid Leukemia (AML) patients lacking an HLA matched donor. However, outcomes following Haplo-SCT vary and are dependent on a number of individual features. Integrative prognostic models for decision support towards a Haplo-SCT are lacking. We sought to develop a prediction model of Leukemia-Free Survival (LFS) for AML patients undergoing a Haplo-SCT.

Methods: A total of 1,804 de-novo (80%) and secondary (20%) AML patients who received a non-T-cell depleted Haplo-SCT between the years 2005-2017 were included. All patients were reported to the registry of the Acute Leukemia Working Party (ALWP) of the European Society for Blood and Marrow Transplantation (EBMT). To account for non-linear associations, violation of the proportional hazard assumption, and to reduce bias associated with feature selection, a non-parsimonious non-parametric machine learning algorithm, Random Survival Forest (RSF), was used. RSF provides a continuous probabilistic estimation of LFS by fitting an ensemble of decision trees. Variables included in the model were reflective of patient, disease, and transplantation characteristics. Since RSF models are not readily interpretable (i.e., "black box" models) variable importance (VIMP) of covariates included in the model (Xv), were assessed by calculating the difference in prediction error before and after permuting Xv. The model's generalizability and accuracy were tested through repetitive bootstrapping (5000 iterations) and calculation of the C-index.

Results: The median age of the patients was 53 years. The majority had an early disease status (complete remission [CR] 1[44%]) with intermediate cytogenetic risk (43%) and were undergoing allogeneic transplantation for the first time (93%). Reduced-intensity conditioning (RIC) was used in 57% of cases, and grafts were from peripheral blood in 54% of transplants. For graft-versus-host disease (GvHD) prophylaxis, 82% of the patients received post-transplant cyclophosphamide (PTCy) and 18% anti-thymocyte globulin (ATG). The median follow-up duration was 2.0 years. In the RSF prediction model, the top-ranking variables (Figure A) were disease status, GvHD prophylaxis, time from diagnosis to transplantation, and age. Bootstrapped C-index of the prediction model was 0.66. Prognostic discrimination was assessed by dividing the predicted LFS probabilities into quartiles that were then used to plot Kaplan-Meier curves, demonstrating LFS ranging from 24.8%-60.1% at 2-years (Figure B). Differing features of the four prognostic groups are listed in the Table.

Conclusions: Our group has developed the first prediction model for LFS in AML patients treated with a Haplo-SCT. The model is based on a machine learning technique and provides an individualized estimation of LFS probability. It is conceivable that once this model is verified, it could serve as an important clinical tool when considering a patient to Haplo-SCT.

Figure.
Figure.
View largeDownload slide
Disclosures

Angelucci:Vertex Pharmaceuticals Incorporated (MA) and CRISPR CAS9 Therapeutics AG (CH): Other: Chair DMC; Jazz Pharmaceuticals Italy: Other: Local ( national) advisory board; Celgene: Honoraria, Other: Chair DMC; Novartis: Honoraria, Other: Chair Steering Comiittee TELESTO Protocol; Roche Italy: Other: Local (national) advisory board. Tischer:Jazz Pharmaceuticals: Other: Jazz Advisory Board. Mohty:MaaT Pharma: Consultancy, Honoraria.

Topics:
hematopoietic stem cell transplantation, leukemia, leukemia, acute, leukemia, myelocytic, acute, graft-versus-host disease, transplantation, antithymoglobulin, bone marrow transplantation, clinical diagnostic instrument, complete remission

Author notes

*

Asterisk with author names denotes non-ASH members.

© 2018 by The American Society of Hematology
2018
Sign in via your Institution