INTRODUCTION

Postremission treatment with allogeneic stem cell transplantation (alloSCT) for patients with acute myeloid leukemia (AML) or acute lymphoblastic leukemia (ALL) provides a potent graft versus leukemia (GVL) effect that significantly contributes to cure in acute leukemia.^1,2 Transplantation-related morbidity and mortality, considered as nonrelapse mortality (NRM) might compromise outcome after alloSCT. Therefore, a personalized approach weighing both the benefit (ie, GVL) and the risk of the transplant-procedure (ie, NRM) has been advocated for all patients before proceeding to alloSCT.^1,2

To support decision-making for alloSCT, prediction models have been developed to assess the risk of NRM of which the European Group for Blood and Marrow Transplantation (EBMT) score and the hematopoietic cell transplantation-comorbidity index (HCT-CI) are widely used. The EBMT score consists of patient- and transplant-related factors, and was originally constructed in a cohort of chronic myeloid leukemia (CML) patients to predict overall survival (OS),³ and subsequently validated for the prediction of NRM.⁴ The HCT-CI was developed by Sorror et al⁵ and consists of 17 comorbidities predictive of NRM and OS after alloSCT. Over the past decades, these NRM-risk models have been used for informing transplant decisions. Since then, the implementation of reduced-intensity conditioning (RIC) has allowed elderly, and younger patients with comorbidities to become increasingly eligible for alloSCT.⁶ A previous study demonstrated poor predictive performance of the EBMT score and HCT-CI in a large EBMT cohort of AML patients receiving RIC alloSCT, emphasizing the need to continually reassess and refine existing risk scores.⁷ That study proposed a different score for AML patients receiving RIC alloSCT, which integrated risk factors of the EBMT score and HCT-CI, which was validated by others.⁸

Recently, posttransplant cyclophosphamide (PTCY) has emerged as an alternative for GVHD prophylaxis after alloSCT. PTCY was developed by the Baltimore group in the context of haploidentical donor alloSCT,^9,10 but has successfully been extended to other donor types.^11,12 PTCY has been associated with in vivo depletion of alloreactive T cells while preserving the GVL-effect, which results in decreased incidence of severe GVHD and subsequent NRM.^11,13,14 Consequently, NRM-risk scores should be re-evaluated and refined in the setting of PTCY-based alloSCT. Therefore, we evaluated the discriminatory capacity of the EBMT score, HCT-CI, and integrated EBMT score for 2-year NRM in a cohort of AML and ALL patients receiving alloSCT, all followed by PTCY. We then set out to construct a novel risk score to better predict 2-year NRM, which was validated internally.

PATIENTS AND METHODS

Patients

A total of 1861 adult patients diagnosed with AML or ALL who received alloSCT in first complete remission (CR1) using PTCY-based GVHD prophylaxis between 2010 and 2018 were identified in the EBMT registry. Patients were included if comorbidities were scored according to the HCT-CI, both patient and donor CMV serology were available, alloSCT was administered within 2 years after diagnosis, and outcome measures were available.

A total of 76.4% (n = 1421) AML patients were included with a median age at alloSCT of 53 years (Suppl. Table S1). ALL patients (23.6%, n = 440) were significantly younger compared with AML patients with a median age of 39 years (P < 0.001) at alloSCT. The median time from diagnosis to alloSCT for patients with AML and ALL was significantly different (153 and 188 days, respectively, P < 0.001). Haploidentical donors were used in 55.8% of transplants, whereas matched-related, matched-unrelated, and mismatched-unrelated donors were used in 13.6%, 29.8%, and 0.8%, respectively. Conditioning was considered myeloablative (MAC) if it included either total body irradiation (TBI) ≥ 8.0 Gray, >9.6 mg/kg busulfan, or >42 g/m² treosulfan. Regimens were also considered MAC if these were associated with high scores according to the Transplant Conditioning Intensity score.¹⁵ MAC regimens were used in 54.5% of ALL patients, which was significantly higher compared with AML patients (22.6%; P < 0.001). Antithymocyte globulin was added to the conditioning regimen in 13.5% (n = 252) of patients. The majority (68.9%) of patients received additional GVHD prophylaxis with a calcineurin inhibitor and mycophenolic acid/methotrexate (Table 1).

At least one comorbidity (defined by the HCT-CI⁵) was present in 38.5% (n = 717) of patients (Figure 1). Pulmonary disease (moderate 12.2%; severe 7.0%), infection requiring treatment at transplantation (6.9%), and prior solid tumors (6.3%) were most frequently reported.

Endpoints

The primary endpoint was the cumulative incidence of NRM at 2 years, for which relapse was considered a competing event. Secondary endpoints were OS, relapse-free survival (RFS), and cumulative incidence of relapse. The event for OS was death from any cause, and patients were censored at the last day of follow-up if alive. The events for RFS were NRM and relapse of leukemia. All time-to-event outcomes were analyzed from the date of transplantation until the first event or date of last contact.

Statistical analysis

Cumulative incidence curves were used to estimate the probabilities of NRM and relapse. These events are considered as competing events, which means that the occurrence of relapse excludes patients to be at risk for NRM and vice versa. The Kaplan–Meier method was used to determine OS and RFS. The inverse Kaplan–Meier method was used to calculate the median follow-up time with being alive at the end of follow-up considered as the event of interest.

Prediction of NRM using existing NRM-risk scores

The EBMT score,³ HCT-CI,⁵ and integrated EBMT score⁷ were calculated for all patients using the previously defined criteria. The predictive performance of these NRM-risk scores was assessed using competing risks analysis and calculation of the area under the receiver operating characteristic (ROC) curve (AUC). The AUC is summarized with the concordance statistic (c-statistic) reflecting the probability of predicting the outcome better than chance (c-statistic = 0.5). A c-statistic of 1.0 indicates no false positives, which means that no patients experiencing NRM were missed by the model.

Development of a novel NRM-risk score

We then reassessed the individual parameters of the EBMT score, HCT-CI, and integrated EBMT score to develop a new risk model. The dataset was randomly split into a training set (n = 1303; 70%) and a test set (n = 558, 30%). Multivariable regression using the Fine and Gray subdistribution hazards model was used to determine the predictive capacity of individual variables for the prediction of NRM in the training data. The subdistribution hazards model estimates the effect of covariates on the subdistribution hazard function and on the cumulative incidence of NRM. The multivariable model contained variables of the HCT-CI, EBMT, and integrated EBMT score, including comorbidities (yes versus no) if present in ≥15 patients, CMV serology (patient and donor positive versus patient or donor positive versus patient and donor negative), time from diagnosis to alloSCT (<6 months versus ≥6 months), and conditioning (MAC versus non-MAC). Age at alloSCT was included as previously defined in the integrated EBMT score.⁷ Disease stage was not included because all patients were transplanted in CR1. Next, covariates with a subdistribution hazard ratio (SHR) ≥1.2 were selected for the final model, similar to the approach used in the development of the HCT-CI and integrated EBMT score.^5,7 Patients were then assigned points based on the rounded SHRs of the selected covariates. We defined a 3-tiered risk score based on the aggregate of cumulative incidence of NRM for each score.

Using these 3 risk groups, the predictive performance was assessed in the training and test sets by competing risks analysis and estimation of AUC for 2-year NRM, and the Kaplan–Meier method for OS.

Statistical software

All analyses were performed using R (version 4.0.0 or higher) by the R Core Team. The survival and cmprsk packages were used for the time-to-event, competing risks, and competing risks regression analyses, whereas packages haven, magrittr, ggplot2, survminer, caret, and timeROC were used for data management, data visualization, data splitting, and AUC calculation.

RESULTS

Cohort outcomes

Overall, 608 (32.7%) patients died and 1253 (67.3%) patients were alive at the end of follow-up. The median follow-up was 28 ± 1 months in the full cohort, which was similar between the training and the test set. The 2-year OS and RFS were 66% ± 1% and 59% ± 1%, respectively. The cumulative incidence of NRM at 2 years was estimated as 18% ± 1%, which was similar among the training and test cohorts. It was also similar between AML and ALL patients (18% ± 1% versus 18% ± 2%, P = 0.90).

Prediction of NRM using existing NRM-risk scores

Validation of the established risk scores in the full cohort showed relatively poor predictive performance. Using the EBMT score, the majority of patients were assigned to the intermediate-risk group (n = 1182, 63.5%), whereas 1.8% (n = 33) and 34.7% (n = 646) of patients were in the low- or high-risk subgroups, respectively. The cumulative incidence estimates of NRM at 2 years were 13% ± 6%, 18% ± 1%, and 20% ± 2% in the low-, intermediate-, and high-risk subgroups, respectively (Figure 2A). Most patients in the HCT-CI were assigned to the low-risk subgroup (n = 1144, 61.5%). The low- and intermediate-risk subgroups of the HCT-CI were found to have a similar cumulative incidence of 2-year NRM, which was higher in the high-risk group (16% ± 1%, 18% ± 2%, and 27% ± 3%, respectively, Figure 2B). The integrated EBMT score classified 45.5% and 48.0% of patients in the low- and intermediate-risk group, respectively, whereas only 6.5% of patients were assigned to the high-risk group. The cumulative incidence of NRM in the integrated EBMT score estimated 13% ± 1%, 23% ± 2%, and 25% ± 4% in the low-, intermediate- and high-risk groups, respectively (Figure 2C).

These existing NRM-risk scores were found to have relatively poor discrimination of 2-year NRM (AUC of 51.7%, 56.6%, and 59.2% for the EBMT, HCT-CI, and integrated EBMT score, respectively). Consequently, we sought to develop a novel NRM-risk score specifically for patients receiving PTCY-based alloSCT.

Development of a novel NRM-risk score

Using a Fine and Gray multivariable regression model, we selected covariates with an SHR ≥1.2. A total of 22 covariates were considered of which 12 did not meet the cutoff and were thus not selected for the final model (Suppl. Table S2). Interactions were tested between variables (ie, age and conditioning, diagnosis and conditioning, and donor type and diagnosis), which were not statistically significant. The final model consisted of ten covariates, including age at alloSCT >60 years, patient and donor CMV positivity, female donor to male recipient, MAC, haploidentical donor, arrhythmia, diabetes mellitus, prior solid tumor, cardiac disease, and hepatic disease (Table 2). Moderate/severe hepatic disease, age at transplantation >60 years, and prior solid tumor were associated with the highest SHRs for 2-year NRM after alloSCT. The SHRs of these 10 covariates were rounded to the nearest integer and patients were attributed points for each covariate. The sum of these points resulted in the final score which ranged from 0 to 10 points, with a median of 3 points in the development cohort. These scores were collapsed into a 3-tiered risk model based on the 2-year NRM estimates for each point level (Suppl. Figure S1). Patients were considered low-, intermediate-, or high-risk if they scored between 0 and 2 points, 3–4 points, or ≥5 points, respectively.

This 3-tier EBMT-PTCY-risk score captured most patients in the low-risk (n = 518; 39.8%) and intermediate-risk group (n = 556; 42.7%). A smaller proportion of patients (n = 229, 17.5%) were allocated to the high-risk group.

The predictive performance of this novel EBMT-PTCY-risk score in the training data was assessed using competing risks analysis for 2-year NRM, and AUC. The cumulative incidence of NRM was 11% ± 2%, 19% ± 2%, and 36% ± 3% for each risk group, respectively (Figure 3A). The novel risk score demonstrated a relatively better predictive capacity of 2-year NRM, as measured by AUC (64.2%). The score also predicted 100-day NRM with an AUC of 66.6%. Furthermore, the risk score discriminated for 2-year OS, with survival estimates of 72 ± 2%, 68 ± 2%, and 51 ± 4% associated with the low-, intermediate-, and high-risk group, respectively (AUC 57.3%) (Figure 3B). The cumulative incidence of relapse was not different across the 3 risk strata (Suppl. Figure S2).

Model validation

The novel risk score was evaluated using the test data (n = 558) in which a median score of 3 points (range 0–11) was observed. The distribution of patients in the different risk strata was similar compared with the training set (Figure 4). The EBMT-PTCY-risk score yielded a cumulative incidence of 2-year NRM of 11% ± 2%, 18% ± 3%, and 31% ± 5% in the low-, intermediate-, and high-risk group, respectively (Figure 4A). In addition, the AUC for 2-year NRM in the novel risk score was 63.3% in the test data indicating relatively reasonable performance in an unseen dataset. An AUC of 64.7% for 100-day NRM was observed in the test data. Comparable to the training data, the novel risk score also discriminated for 2-year OS (low-risk: 72% ± 3%; intermediate-risk: 63% ± 3%; high-risk: 50% ± 6%) with an AUC of 59.3% (Figure 4B). The cumulative incidence of relapse in the different risk groups was similar in the test set (Suppl. Figure S3).

DISCUSSION

Weighing the risks of NRM and relapse is considered essential for risk-adapted strategies in the application of alloSCT in patients with acute leukemia.² Clinical care of transplant recipients is continuously evolving with strategies aiming to reduce relapse on the one hand, and reduce transplant-related toxicity and mortality on the other. With respect to the latter, GVHD prophylaxis with PTCY has been associated with lower rates of chronic GVHD and NRM after alloSCT and is increasingly being administered.^16,17 Given the reduced incidence of NRM and the specific risk–benefit ratio of cyclophosphamide, established NRM-risk scores need to be re-evaluated.

In this study, we found a relatively poor predictive value for 2-year NRM in 3 previously developed NRM-risk models, including the HCT-CI, EBMT score and integrated EBMT score.^3,5,7 Using the individual parameters of these models, we developed a novel NRM-risk score in patients with AML and ALL in CR1 receiving alloSCT with PTCY-based GVHD-prophylaxis. The risk model consisted of 10 variables which were assigned points and then used in a 3-tiered risk score. The EBMT-PTCY-risk score captures patients in risk groups that are associated with distinct estimates for NRM. The associated distinct outcome and discriminatory capacity of the novel risk score were preserved in an independent test set, indicating a robust prediction model. We observed higher performance of the novel NRM-risk score in both development and test sets for the prediction of 2-year NRM compared with the HCT-CI, EBMT score, and integrated EBMT score. Finally, the new risk score was also associated with differential 2-year OS.

NRM is predominantly caused by severe GVHD and its related complications, which has encouraged the development of strategies that reduce GVHD. Intensified GVHD-prophylaxis with PTCY to deplete proliferating alloreactive T cells has been developed in the setting of haploidentical donor alloSCT by the Baltimore group.⁹ Effective GVHD prevention after haploidentical donor alloSCT was shown with relatively low rates of severe acute and chronic GVHD, which resulted in promising NRM and overall outcome.^9,12 PTCY has been subsequently administered as GVHD prophylaxis for matched donor alloSCT with substantially reduced rates of chronic GVHD and low NRM.^14,18–21 Recently, the prospective randomized phase III HOVON-96 trial demonstrated improved GRFS for patients receiving HLA-matched alloSCT with PTCY+cyclosporine versus cyclosporine+mycophenolic acid, and confirmed a low NRM after PTCY-based GVHD prophylaxis.¹⁴ Among other improvements in transplant care, the increasing application of PTCY and its favorable association with NRM warrants an evaluation of existing NRM-risk scores in that patient group for whom PTCY is considered as GVHD prophylaxis.

The EBMT score and HCT-CI are commonly used to assess transplant-related morbidity and mortality.^3,5 These risk models have limitations including reproducibility in other cohorts of patients over time with different transplant strategies or disease types.^22–25 As a result, the integrated EBMT score, combining elements from both EBMT score and HCT-CI, has been developed for patients with AML in CR1 who received RIC alloSCT.⁷ That risk model was associated with better predictive capacity of 2-year NRM compared with the EBMT score and HCT-CI in a specific subgroup of older RIC alloSCT recipients. Similarly, several studies to further improve existing risk scores or to develop novel models have been performed.^22,24,26,27 These studies have shown that validation of prediction tools in different, more recent cohorts results in less predictive power, possibly explained by the continuous improvement in transplant strategies. Here, we were also not able to validate the EBMT score, HCT-CI, and integrated EBMT score for the prediction of 2-year NRM in AML and ALL recipients of alloSCT with PTCY-based GVHD-prophylaxis, emphasizing the need for continuous re-assessment of existing scores in changing patient populations and in new treatment modalities. This should be ideally performed regularly allowing clinicians for the most updated and accurately informed transplant decision.

In this study, we have developed a novel risk model for the prediction of 2-year NRM in patients receiving PTCY-based GVHD prophylaxis. Patient- and transplantation-related variables were included in the risk model, with different weighing compared with previous studies which included various hematological diseases.^3,5 Acute leukemia patients in CR1 constitute a specific patient population receiving intensified induction chemotherapy with attributable toxicity including cardiac morbidity, pulmonary function decline, and infection requiring prolonged treatment.

Pretransplant cardiac comorbidity might be of concern as PTCY could induce cardiac toxicity in a cumulative-/dose-related fashion,^28–30 further worsening cardiac function after alloSCT. In our PTCY cohort, pretransplant cardiac comorbidity was indeed associated with a two-fold increased risk for NRM. Data are not available to assess whether patients with cardiac comorbidity experienced cardiac-related mortality induced by PTCY. Recently, pretransplant cardiac toxicity was investigated in a single-center study of 585 patients receiving HLA-matched alloSCT combined with both PTCY and non-PTCY-based GVHD prophylaxis. Pretransplant cardiac toxicity was reported to be associated with age >55 years, history of hypertension, arrhythmia, and diabetes, but this association was independent of GVHD prophylaxis used, contradicting earlier reported PTCY-based cardiac toxicity.³¹ These findings are in contrast to the retrospective observation that patients treated with PTCY developed more cardiac events during the first 100 days posttransplant than patients in the non-PTCY group.³² Using a composite risk model of cardiovascular comorbidities might further improve prediction of NRM, which has been reported recently.²⁷

We found no association of pretransplant pulmonary comorbidity with 2-year NRM, although the frequency of moderate and severe pulmonary disease was 12% and 7%, respectively. Since PTCY predominantly reduces chronic GVHD, we hypothesize that patients with pulmonary comorbidities receiving PTCY harbor no apparent increased risk of NRM because of reduced severe pulmonary GVHD.

Similarly, infections requiring treatment at time of alloSCT were not associated with NRM in our PTCY-specific risk score. The overall reduction of higher grades of GVHD using PTCY might again prevent NRM due to pre-existing infections.

Current risk-adapted strategies are weighting risk factors for NRM, as assessed by predictive scores, with the risk of disease relapse.^1,2 Acute leukemia relapse is primarily determined by genetic characteristics of the underlying leukemia, age of the patient, and response to treatment, especially whether or not residual disease has been eradicated before proceeding to transplant.^33–36 In personalized transplant approaches, patients at low risk of relapse have limited absolute benefit of a GVL-effect, and are therefore not considered for alloSCT because of the counterbalancing risk of NRM. We found a cumulative incidence of NRM of 11% in the low-risk group, in which the estimate was only 3% in patients with 0 points. These very low 2-year NRM estimates might identify subgroups of patients with favorable disease characteristics for whom the GVL effects of alloSCT might still be beneficial with respect to overall outcome. For these subgroups, alloSCT might be considered as consolidation treatment instead of additional chemotherapy.

The development of predictive models is influenced by the modeling approach applied. In general, clinical prediction research has been hampered by the selection of predictors based on statistical significance, categorization of predictors, and inadequate sample sizes,³⁷ along with selective evaluation of model performance.³⁸ The TRIPOD statement^39,40 is a framework for clinical prediction modeling to address these analytic difficulties. Here, we present the results of a novel prediction tool that was developed with predefined primary and secondary outcomes and did not rely on statistical significance for model development. Furthermore, the risk score was developed in a training cohort, and validated in a dedicated, independent test set. All of these decisions are consistent with the TRIPOD statement.^39,40 Some comorbidities were relatively infrequent, which limits predictive capacity. However, this statistical limitation reflects the current transplantation practice and highlights the need for external validation. Additionally, the score may be improved by evaluating composite predictors, such as composite cardiac comorbidity (valvular disease, arrhythmia, and congestive/ischemic heart disease), or by investigating the additive effect of previously identified biomarkers, such as lactate dehydrogenase,⁴¹ serum ferritin,⁴² albumin,⁴³ or circulating endothelial cells.⁴⁴ The EBMT database does not routinely report these factors, precluding their validation into EBMT-based risk scores.

In conclusion, ongoing improvements in transplant care require continuous reassessments of NRM prediction scores. GVHD-prophylaxis with PTCY has demonstrated a substantial reduction of GVHD and subsequent NRM, which encourages the need to reassess and refine existing NRM-risk scores. Here, we have presented a novel risk score for NRM prediction in patients with AML or ALL in CR1 receiving PTCY, that better predicted 2-year NRM compared with existing models, and which might be preferential in the setting of specific toxicities of high-dose cyclophosphamide. The risk score was validated in an independent EBMT test set with preserved predictive capacity. External validation of this EBMT-PTCY risk score is needed to further establish the clinical applicability.

ACKNOWLEDGMENTS

Following EBMT publication rules, co-authorship was offered to centers contributing the highest number of patients. Nevertheless, the authors highly appreciate the contribution by many physicians and data managers throughout the EBMT, who made this analysis possible.

AUTHOR CONTRIBUTIONS

SJFH and JV designed the study, analyzed the data, and wrote the article; YN and ML provided statistical oversight and guidance. JV, JJC, and MM supervised the study. All other authors provided clinical data and carefully reviewed and approved the final version of this article.

DISCLOSURES

RS served as a consultant to MyBiotics and Medexus. All the other authors have no conflicts of interest to disclose.

Supplemental Digital Content

• HS_2023_01_12_VERSLUIS_HemaSphere-2022-0259_SDC1.doc; [Word] (215 KB)