Search

* AR and AN contributed equally to the abstract Background: When considering a high risk, but potentially curative procedure, such as an allogeneic hematopoietic stem cell transplantation (HSCT), prognostic models may help decision making. We have previously developed the Acute Leukemia (AL)-EBMT score for prediction of mortality following allogeneic HSCT in AL patients (Shouval et al., JCO, 2015). The score is based on 10 variables including: disease status,Karnofsky performance status, recipient age, diagnosis, interval between diagnosis to HSCT, conditioning regimen, donor type, donor and recipient CMVserostatus combination, HSCT year, and center experience (HSCT/year). External validation of the score on an independent cohort of AL patients is of importance. Methods: This was a retrospective validation study on an independent cohort of AL patients from the Italian national transplantation network (GITMO). Inclusion criteria included adult AL patients, receiving an allogeneic HSCTbetween 2000-2014. Competing risks analysis was used to calculate the 2 years non-relapse related mortality (NRM) cumulative incidence, using the Gray test to test differences between ADT score groups. Overall survival (OS), and leukemia free survival (LFS) at 2 years were estimated using the Kaplan-Meier method; hazard ratios were computed between subgroups using Cox regression for OS and LFS and competing risk regression for NRM. Competing risk regression was used to compute NRM cumulative incidence rates, considering non-transplant mortality as the competing event. The predictive ability of the AL-EBMT score was assessed using time-dependent receiver-operator curves (AUC) analysis. Results: A total of 1,848 patients from 61 Italian transplant centers were analyzed. The median age was 45.9 (IQR 35.2-55). Indications for HSCT were Acute Myeloid Leukemia (67.8%) and Acute Lymphoblastic Leukemia (32.2%). The majority of patients were in first complete remission (60.6%), and received myeloablative conditioning (81.3%). Median follow-up was 2 years (1.9-2.1, 95% CI). The AL-EBMT score was categorized according totertiles (low, intermediate, high) on the original ALWP-EBMT cohort. Increasing score intervals corresponded with decreasing probability for 2 year OS (95% CI), ranging from 75.52% (70.69-80.68) to 36.93% (32.81-41.56), and increasing probability of 2 years NRM ranging from 10.9% (7.97-14.93) to 27.39% (23.84-31.48) (Table 1, Figure 1). The highest scoring group was associated with a hazard ratio (95% CI) of 2.89 (2.01-4.16) for 2 year NRM and 3.66 (2,85-4.72) for death at 2 years (Table 2). The categorized score discrimination (AUC) for 2 year OS, LFS, and NRM was 66.42, 66.15, and 67, respectively. Conclusion: This is the first study externally validating the AL-EBMT score. The score identified 3 distinctive risk groups and was predictive of survival related outcome. It can be used as a decision support tool when considering an allogeneic HSCT in acute leukemia patients. Table 1 Cumulative Probability of NRM and OS (95% CI) at 2 Years Table 1. Cumulative Probability of NRM and OS (95% CI) at 2 Years Table 2 Hazard Ratios for Outcomes at 2 years Table 2. Hazard Ratios for Outcomes at 2 years Figure 1. Overall Survival Stratified by the AL-EBMT Score Figure 1. Overall Survival Stratified by the AL-EBMT Score Disclosures No relevant conflicts of interest to declare.

Background: Prognostic scoring systems for allogeneic stem cell transplantation (HSCT) are of clinical value when determining a leukemic patient's suitability for this curative, but risky, procedure. Several such scores have been developed over the years for HSCT from sibling or unrelated donors, but no predictive score has been developed specifically for umbilical cord blood transplantation (UCBT). Although individual parameters have been identified to be associated with UCBT outcomes in acute leukemia (AL) patients, integrative tools for risk evaluation in this setting are lacking. We sought to develop a prediction model for overall survival (OS) (primary objective) and leukemia free survival (LFS) (secondary objective) at 2 years following UCBT in acute leukemia patients. Methods: A retrospective, international registry-based study, of 3140 acute leukemia patients who underwent UCBT from 2004 through 2014. Inclusion criteria were patients with AL receiving single or double cord blood units transplantation. Median follow up was 30 months. The dataset was geographically split into a derivation (65%) and validation set (35%). A Random Survival Forest was utilized to identify predictive factors. Top predictors were introduced into a Cox regression model, and a risk score was constructed according to each variable's hazard. Results: The median age at UCBT was 21.9 years. The 2 years OS rate was 47.7% (95% CI: 45.8-49.6). After identifying the top predictive variables, the UCBT risk score (Table 1) was constructed using 9 variables (disease status, diagnosis, cryopreserved cell dose, age, center experience, recipient cytomegalovirus sero-status, degree of HLA mismatch, previous autograft and anti thymocyte globulin administration). Over the derivation and validation datasets, a higher score was associated with decreasing probabilities for 2 years OS and LFS, ranging over the validation set from 0.72 (0.64-0.8, 95% CI) and 0.68 (0.6-0.76, 95% CI) to 0.13 (0.06-0.27, 95% CI) and 0.14 (0.07-0.28, 95% CI), respectively (Figure 1). An increasing score was also associated with increasing hazard of the predictive outcomes (Table 2). The score's discrimination (AUC) over the validation set for 2 years OS and LFS was 68.26 (64.25-72.27, 95% CI) and 66.95 (62.88-71.02, 95% CI), respectively. Calibration was excellent. Conclusion: We have developed the first integrative score for prediction of overall survival and leukemia free survival in acute leukemia patients undergoing a UCBT. The score is simple and stratifies patients into distinct risk groups. Table 1 The UCBT Risk Score Table 1. The UCBT Risk Score Table 2 Association between the UCBT risk score and 2 years OS and LFS over the validation dataset Table 2. Association between the UCBT risk score and 2 years OS and LFS over the validation dataset Figure 1 Overall survival stratified by the UCBT risk score over the validation data set Figure 1. Overall survival stratified by the UCBT risk score over the validation data set Disclosures Bader: Servier: Consultancy, Honoraria; Neovii Biotech: Research Funding; Riemser: Research Funding; Medac: Consultancy, Research Funding; Novartis: Consultancy, Honoraria.

Background: Umbilical cord blood transplantation (UCBT) is a potentially curative therapy acute leukemia (AL) patients. Transplantation benefit must be balanced against risks, such as transplant related mortality and relapse. The complex nature of hematopoietic stem cell transplantation data (HCT), rich in interactions and possibly nonlinear associations, has motivated us to apply machine learning (ML) for predictive modeling. ML is a field of artificial intelligence and is part of the data mining approach for data analysis. Our group has recently reported on a ML based prediction model for short term HCT outcomes (Shouval R et al; JCO 2015). Using a ML algorithm, the perspective of the current study was prediction of leukemia free survival (LFS) at 2 years after an UCBT, while exploring variables' importance and interactions. Patients & Methods: A cohort of 3,149 UCBT were analyzed. Inclusion criteria encompassed patients at all ages, undergoing an UCBT (single/double unit) in EBMT centers from the year 2004 to 2014, for AL, in all disease status. All conditioning and graft versus host disease prophylaxis regiments were included. A total of 24 variables were considered, including the number of total nucleated cell dose (TNC), donor and recipients HLA typing, as well as recipient, disease and transplant characteristics. The Random Survival Forest (RSF) ML algorithm was applied for model construction and data exploration. RSF is known to be adaptive to data, is able to automatically recover nonlinear effects and complex interactions among variables, and yields nonparametric prediction over test data. The analysis pipeline consisted of prediction model development, assessment of variable importance by their minimal depth from the tree trunk, and exploration of the top ranking variable with dependence plots. The latter promotes understanding of non-trivial associations between variables and outcomes. Results : The 2 years LFS was 49%, with a median follow up of 30 months. A RSF model of 1000 trees was developed, with each tree constructed on a bootstrap sample from the original cohort. A prediction error of 36.0% was calculated. The 10 most predictive variables (in ascending order) were disease status, age, TNC harvested and infused, recipient CMV serostatus, interval from diagnosis to UCBT, transplant year, previous autologous transplant, and use of anti-thymocyte globulin (ATG). Selected findings from exploration of variables-outcome relationship with dependence plots included a varying effect of TNCs in specific subpopulations. Increasing the number of infused TNCs had a positive effect on predicted LFS in patients receiving HLA mismatched (2 or more HLA mismatch) (figure) or single unit CB grafts, and patients in earlier disease status or older age. ATG administration was associated with worse LFS, whether unadjusted or adjusted to all other variables. However, there was an additional negative effect in advanced disease status patients, recipients of HLA mismatched or single CB units grafts, and older patients. Patients in 1st complete remission (CR) had higher predicted LFS as compared to those in 2nd CR. However, in patients receiving a HLA mismatched or a double CB graft, the difference in LFS between CR1 and CR2 was attenuated. Younger age had a favorable impact in early disease status, but lost its positive effect in advanced disease. Conclusions: A prediction model for LFS 2 years post UBCT was developed using the RSF ML algorithm. Variables were ranked according to their predictive contribution. Disease status, age, and TNC count were found to be the most important factors. Dependence plots revealed interactions and nonlinear associations between variables and the outcome, such as the effect of cell dose on HLA disparity. Apart from the study's clinical findings, it carries a methodological significance. A novel ML approach for prediction, variable selection and data exploration, accounting for long term time to event outcomes, has proved useful in the field of HCT. Figure 1. Variable marginal dependence coplot of predicted LFS at 2 years against TNC, conditional on HLA matching. Individual cases are marked with blue circles (alive or censored) and red `x's (event). Linear smooth (a linear extrapolation of the prediction function), with shaded 95% confidence band, indicates trends of variable dependence. Figure 1. Variable marginal dependence coplot of predicted LFS at 2 years against TNC, conditional on HLA matching. Individual cases are marked with blue circles (alive or censored) and red `x's (event). Linear smooth (a linear extrapolation of the prediction function), with shaded 95% confidence band, indicates trends of variable dependence. Disclosures Mohty: Janssen: Honoraria; Celgene: Honoraria. Sanz:JANSSEN CILAG: Honoraria, Research Funding, Speakers Bureau. Bader:Neovii: Other: Institutional grants; Medac: Other: Institutional grants; Riemser: Other: Institutional grants; Amgen: Consultancy; Novartis: Consultancy; Jazz Pharmaceuticals: Consultancy.

Background Allo-HSCT has been shown to increase survival and improve cure in acute leukemia (AL). However, this procedure is accompanied by high rates of morbidity and mortality. Several risk scores based on conventional statistical methods may aid decision regarding whom and how to perform allo-HSCT, but these methods carry inherent limitations, which may lead to sub-optimal candidate selection. Machine learning (ML) is a field in computer science stemming from artificial intelligence and is part of the data mining approach for data analysis. ML algorithms are commonly applied in technological and commercial settings. They allow for coping with complex data scenarios and thus may be suitable for outcome prediction in allo-HSCT. With this background, and using a ML prediction method- the alternating decision tree (ADT) algorithm, we developed an interpretable model for overall mortality (OM) and treatment-related mortality (TRM) at day +100 after allo-HSCT in AL. Patients and Methods 28,995 adult allo-HSCT recipients from the registry of the ALWP of EBMT were analyzed. Twenty two variables were available including year of transplant (range, 2000-2011), diagnosis (Acute Myeloid Leukemia and Acute Lymphoblastic Leukemia), disease status, Karnofsky performance status, conditioning regimen (myeloablative or reduced- intensity conditioning), graft type (peripheral blood, bone marrow or cord blood), donor and recipient HLA compatibility, CMV serostatus, GVHD prophylaxis regimens, etc. Per study definitions, the primary outcomes to be predicted were OM and TRM at day +100 days after allo-HSCT. The complete dataset was split into 3 sets: training set (n=11,600), testing set (n=8,688) and validation set (n=8,707). The ADT prediction model was tested and optimized according to the first 2 subgroups and validated on the last one. Output from the ADT model, included variables selection with assigned scores in a tree-based structure and area under the ROC curve (AUC), a measure for model discrimination Results Each of the ADT models selected 12 out 22 variables for prediction of OM and TRM at day +100. Ten variables were mutual for both prediction models, although different weights were assigned. These included: age, diagnosis, disease status, Karnofsky performance status (all at time of transplant), donor-recipient HLA-matching, number of transplants in each center per year, year of transplant, conditioning regimen and the donor's and patient's CMV serostatus. Variables selected exclusively by the OM prediction model were graft type and donor-patient CMV serostatus match, whereas the TRM model selected time from diagnosis to transplant and donor-recipient sex match. Applying the models on the validation set yielded AUCs of 0.701 (95% confidence interval [CI] 0.691-0.710) for OM prediction and 0.67 (95% [CI] 0.66-0.68) for TRM. The ADT prediction models assigned scores correlating with patient outcome. Patients in each of the validation sets were grouped according to their score range and the prediction success. A Higher score was correlated with higher rate of the measured outcome in both models (figure 1 and figure 2). Conclusions We present two new models, based on the ADT ML algorithm, for prediction of OM and TRM at day +100 after allo-HSCT. The models are robust as they rely on a high number of samples and a large validation set. As shown in the figures, higher scores correlated with a poorer outcome, reaching more than 50% mortality for a score range of 5.76-7 in the OM prediction model. The AUC performance measure was better for OM than TRM, possibly due to a higher event rate in former, making it easier to predict. Improving the predictive ability will probably necessitate evaluation of more variables, as the limitation of the maximal predictive performance is most likely in the information gained from the variables and not from the sample size or algorithm used. This is currently under progress, especially combination with other risk scores linked to comorbidities. In summary, our models can aid candidate selection for allo-HSCT, by providing a measurable score that correlates with transplant success. Disclosures: Schmid: Novartis: Honoraria, Research Funding, travel grant Other; Roche: travel grant, travel grant Other; MSD: Honoraria.