Personalized prediction of overall survival in patients with AML in non‐complete remission undergoing allo‐HCT

Abstract Allogenic hematopoietic stem cell transplantation (allo‐HCT) is the standard treatment for acute myeloid leukemia (AML) in non‐complete remission (non‐CR); however, the prognosis is inconsistent. This study aimed to develop and validate nomograms and a web application to predict the overall survival (OS) of patients with non‐CR AML undergoing allo‐HCT (cord blood transplantation [CBT], bone marrow transplantation [BMT], and peripheral blood stem cell transplantation [PBSCT]). Data from 3052 patients were analyzed to construct and validate the prognostic models. The common significant prognostic factors among patients undergoing allo‐HCT were age, performance status, percentage of peripheral blasts, cytogenetic risk, chemotherapy response, and number of transplantations. The conditioning regimen was a significant prognostic factor only in patients undergoing CBT. Compared with cyclophosphamide/total body irradiation, a conditioning regimen of ≥3 drugs, including fludarabine, with CBT exhibited the lowest hazard ratio for mortality (0.384; 95% CI, 0.266–0.554; p < 0.0001). A conditioning regimen of ≥3 drugs with CBT also showed the best leukemia‐free survival among all conditioning regimens. Based on the results of the multivariable analysis, we developed prognostic models showing adequate calibration and discrimination (the c‐indices for CBT, BMT, and PBSCT were 0.648, 0.600, and 0.658, respectively). Our prognostic models can help in assessing individual risks and designing future clinical studies. Furthermore, our study indicates the effectiveness of multi‐drug conditioning regimens in patients undergoing CBT.


| INTRODUCTION
The prognosis of acute myeloid leukemia (AML) in noncomplete remission (non-CR) is poor and poses a challenge with respect to the selection of the optimal treatment for patients. Approximately 10%-20% of patients with refractory or relapsed AML exhibit long-term survival. [1][2][3][4] Although chimeric antigen receptor T-cell therapy 5 and several targeted therapies using FLT3 inhibitors, 6 IDH1/IDH2 inhibitors, 7,8 and CD33 antibodies 9 have been developed, survival outcomes have not been sufficiently improved. Consequently, allogeneic hematopoietic stem cell transplantation (allo-HCT) remains the most effective treatment to cure refractory or relapsed AML. Recently, it was reported that for acute leukemia or myelodysplastic syndrome, patients with minimal residual disease (MRD) who underwent cord blood transplantation (CBT) showed a more favorable prognosis than those who underwent bone marrow transplantation (BMT) or peripheral blood stem cell transplantation (PBSCT). 10 Despite the emerging importance of CBT in hematological malignancies with MRD, no large scale studies have been conducted on CBT in patients with AML in non-CR.
Here, we aimed to identify the prognostic factors and to develop and validate nomograms 11,12 and a web application for predicting the overall survival (OS) of patients with AML in non-CR undergoing allo-HCT, including CBT. Furthermore, we constructed and evaluated prognostic models for BMT and PBSCT. Therefore, our models can simultaneously simulate the prognosis of CBT, BMT, and PBSCT as per the clinicopathological characteristics of each patient and can be helpful in selecting an optimal treatment.

| Study design and population
In this multicenter, retrospective cohort study, three nomograms and a web application were developed to predict the OS of patients with AML in non-CR undergoing single-unit CBT, BMT, and PBSCT. We included consecutive patients undergoing allo-HCT with AML aged ≥16 years who had ≥5% blasts in the bone marrow or who had ≥20% blasts in the peripheral blood at transplantation. We excluded patients who underwent HCT within 90 days of the last HCT and those who had missing data for potential predictors. We retrieved the data for HCT outcomes from patients at the Transplant Registry Unified Management Program (TRUMP) [13][14][15] across >300 transplant centers in Japan. The data of patients who underwent allo-HCT between 2000 and 2014 were used to develop the prognostic models; the data of patients who underwent haploidentical transplantation were excluded. To validate the constructed models, we analyzed the data of patients who underwent allo-HCT between 2015 and 2016. Figure 1 shows the design of our study.

| Statistical analyses
OS was defined as the time from HCT to last contact or death from any cause. The OS rates were determined using the Kaplan-Meier method and analyzed using the log-rank test. We used a Cox proportional hazards model for multivariate analysis; the prognostic factors from potential predictors were identified by applying backward stepwise selection  and retaining the variables with p values <0.05. Nomograms and a web application were developed based on the results of the multivariate analyses. The accuracy of the prognostic models was validated through calibration (assessed by plotting the predicted vs. observed OS rates), discrimination (assessed by concordance probability estimate; c-index 34 ), and survival curves. A c-index of 1 indicated perfect discrimination, while a c-index of 0.5 indicated no discrimination. Internal validation of each prognostic model was performed using the bootstrap method with 1000 resamples for calibration and discrimination using the respective development cohorts. To validate each prognostic model, we used the respective validation cohort. Moreover, we applied a previously reported scoring system for patients with AML relapse or primary induction failure who underwent BMT and PBSCT 18 to our validation cohort (cases with missing values were excluded). Briefly, the scoring system was based on the response to chemotherapy, cytogenetics, HLA-match, circulating blasts, and Karnofsky score. Subsequently, the patients were categorized into four groups (scores of 0, 1, 2, and ≥3

| Identification of cytogenetic risk for allogenic hematopoietic stem cell transplantation in acute myeloid leukemia in non-complete remission
The Kaplan-Meier curve was plotted based on the cytogenetic risk classified by the NCCN Guidelines ( Figure S1A). However, patients with favorable risk did not have a better prognosis than those with intermediate risk.  (16), +8 alone, t(9;11), other non-defined] (Table S1). If cytogenetic risk was categorized into two groups, the worse risk classification was adopted. This grouping successfully stratified patients with non-CR AML who underwent allo-HCT ( Figure S1B).

| Conditioning regimen of ≥3 drugs including fludarabine in cord blood transplantation was associated with favorable overall survival and leukemia-free survival
Using the backward stepwise selection method in the Cox proportional hazards model, we identified the following significant prognostic factors for OS in patients in the development cohort who underwent CBT: age of the recipient at transplantation, sex, ECOG PS, HCT-CI, percentage of peripheral blasts, cytogenetic risk classification, response to chemotherapy, number of transplantations, and conditioning regimen (Table 4). Interestingly, compared with cyclophosphamide/TBI (conditioning regimen), the use of ≥3 drugs (including fludarabine) with CBT showed the lowest hazard ratio for mortality (0.384; 95% CI, 0.266-0.554; p < 0.0001). Among all conditioning regimens, the use of ≥3 drugs (including fludarabine) with CBT showed the best leukemia-free survival (LFS) and favorable OS (Figure 2), whereas the regimen with BMT or PBSCT did not show the best prognosis ( Figure S2). were identified as significant prognostic factors for OS in patients who underwent BMT ( Table 5); age of the recipient at transplantation, sex, ECOG PS, percentage of peripheral blasts, cytogenetic risk classification, response to chemotherapy, and number of transplantations were also identified as significant prognostic factors for OS in patients who underwent PBSCT ( Table 6). The conditioning regimen was a significant prognostic factor for OS in only patients who underwent CBT; the common significant prognostic factors among the three types of HCTs were age of the recipient at transplantation, ECOG PS, percentage of peripheral blasts, cytogenetic risk classification, response to chemotherapy, and number of transplantations.

| Development and validation of nomograms
Based on the results of the multivariate analyses, we constructed nomograms to predict the 1-, 3-, and 5-year OS of patients after CBT, BMT, and PBSCT (Figures 3-5). The point of each characteristic was determined by drawing an upward vertical line from the covariate to the points axis. The total points score was obtained by summing each point. The 1-, 3-, and 5-year overall survival probabilities were determined by drawing a downward vertical line from the total points score. Next, we validated the performance of the prognostic models. Figures 6A and B  is given in Table S3. These data indicate that our nomograms were at least as accurate as the previous scoring system. We also developed a web application (https://JSHCT -AMLWG. shiny apps.io/Predi ct-OS-non-CR-AML-post-HCT/) based on these prognostic models. This enabled us to simultaneously estimate the prognosis and construct survival curves after CBT, BMT, and PBSCT with ease ( Figure 9).

| DISCUSSION
We developed three nomograms and a web application to predict the 1-, 3-, and 5-year OS of patients with AML in non-CR after CBT, BMT, and PBSCT. We validated the nomograms showing adequate calibration and discrimination despite the diversity in patient characteristics, leukemia subtype, and treatments.
In this study, we revealed the common significant prognostic factors for the three types of HCTs. These factors were attributed to patient characteristics and tumor characteristics and not to treatment. Intriguingly, the conditioning regimen that physicians selected was a significant prognostic factor only in CBT. A previous single-arm study showed excellent survival outcomes (2-year OS rate =54.9%; 2-year progression-free survival rate =54.9%) of patients with myeloid malignancies in non-CR who underwent CBT and were treated with FLU/BU/MEL. 32 Notably, we demonstrated that the use of a ≥3 drug regimen, including fludarabine, such as the combination FLU/BU/MEL, resulted in a favorable prognosis, but the conditioning regimen was not a significant prognostic factor for the OS of patients undergoing BMT or PBSCT. It was reported that cyclophosphamide/TBI supplemented with high-dose cytarabine was effective for patients undergoing CBT but not for those undergoing BMT or PBSCT, 19,36 which is in accordance with the findings of our study. The distinct difference may be due to differences in the composition and properties of cord blood and bone marrow or peripheral blood. 37 Our data suggests the importance of selecting appropriate conditioning regimens for each donor source.
The ≥3 drug regimen such as FLU/BU/MEL had a positive impact on prognosis. This is because the respective chemotherapy drugs may have different anti-tumor mechanisms. For example, fludarabine inhibits DNA/RNA synthesis by incorporating the drug into DNA or RNA. 38,39 Melphalan and busulfan are alkylating agents, but melphalan is classified as nitrogen mustards and busulfan as alkyl alkane sulfonates. 40 Melphalan reacts with N7-guanine, N3-adenine, and O6-guanine in DNA to form covalent alkyl lesions. 41 Whereas, busulfan reacts with not only N7-guanine and N3adenine in DNA, but also with proteins. 40 Furthermore, busulfan does not elicit toxicity via alkylation of O6-guanine. 42 Thus, the combination of drugs with different mechanisms may be useful in enhancing the anti-tumor effect and eradicating leukemia cells. Actually, a previous study showed that fludarabine and double alkylating agents (busulfan and thiotepa) could enhance the anti-tumor effect compared with fludarabine and a single-alkylating agent (busulfan). 30 It was previously reported that circulating blasts, cytogenetic risk, duration of first CR, and Karnofsky or Lansky score significantly affected the OS of patients with relapsed AML or failure in primary induction who underwent BMT or PBSCT. 18 In our study, they were also selected as prognostic factors for CBT as well as for BMT and PBSCT. Moreover, we found that an increase in the number of transplantations F I G U R E 3 Nomogram to predict the overall survival after cord blood transplantation. This nomogram predicts the 1-, 3-, and 5-year overall survival probabilities of patients with acute myeloid leukemia undergoing cord blood transplantation in non-complete remission. BU, busulfan; CA, cytarabine; CBT, cord blood transplantation; CY, cyclophosphamide; FLU, fludarabine; HCT-CI, hematopoietic cell transplantation comorbidity index; MEL, melphalan; PIF, primary induction failure; Relapse ≥6 months, the duration of the first complete remission was ≥6 months; Relapse <6 months, the duration of the first complete remission was <6 months; TBI, total body irradiation

Conditioning regimen
HIRABAYASHI et Al.

F I G U R E 4
Nomogram to predict overall survival after bone marrow transplantation. This nomogram predicts the 1-, 3-, and 5-year overall survival probabilities of patients with acute myeloid leukemia undergoing bone marrow transplantation in non-complete remission. BMT, bone marrow transplantation; FAB, French-American-British; HCT-CI, hematopoietic cell transplantation comorbidity index; PIF, primary induction failure; Relapse ≥6 months, the duration of the first complete remission was ≥6 months; Relapse <6 months, the duration of the first complete remission was <6 months   was associated with a poor prognosis for any stem cell source. This might be attributed to the condition of patients with AML and an increase in leukemic stem cell frequency and heterogeneity after unsuccessful treatment. 24,25 Commonly used risk scores to predict the OS of patients with AML in relapse or with primary induction failure undergoing BMT and PBSCT have been developed using a large cohort. 18 In the commonly used risk scores, each prognostic factor has an equal prognostic weight in the outcome despite having a different hazard ratio, which results in a reduction of the predictive accuracy of the prognostic model. 43 However, each hazard ratio in this study was accurately represented in the prognostic model. Various studies have documented the superiority of the method used in this study over risk categorization. 43,44 We selected candidate predictors that they have not been previously included, such as HCT-CI, FAB classification, and number of transplantations. Moreover, the model included data from pediatric AML patients; however, recent studies have indicated a distinct difference in biological and molecular profiling between pediatric and adult AML. 45,46 Therefore, to develop a prognostic model suitable for adult AML patients, we focused only on data from adult patients. These reasons could have resulted in the improved performance of our prognostic models compared with that of the previous scoring system. 18 Furthermore, our prognostic models can compare the prognosis of different types of transplantations. They can be useful because there have been no randomized trials to determine appropriate donor sources. 47 Recently, the use of haploidentical transplantation has been increasing for refractory AML. However, there are a few retrospective studies comparing haploidentical transplantation with other transplants for refractory AML, and there are no published randomized clinical trials. Suitable situations for haploidentical transplantation are not yet fully understood. It was reported that haploidentical transplantation for refractory/relapsed AML was associated with shorter GVHD-free relapse-free survival, inferior LFS, and shorter OS than transplantation from an HLA-identical sibling, mainly due to infections, 48 whereas another report showed no differences in GVHD-free relapse-free survival, LFS, or OS between haploidentical transplants and transplants from HLA-identical siblings for AML in first CR with high-risk cytogenetics. 49 As our data could be used to estimate OS adjusted for the characteristics of patients after allo-HCTs, except for haploidentical transplantation, it may be useful for a reference when haploidentical results are evaluated. It is important to note the limitations of this study. First, the regimens for haploidentical transplantation were heterogeneous in our cohort because of limited previous evidence, and the number of transplantations was insufficient to build an accurate prognostic model. Therefore, haploidentical transplantation was excluded. Second, in this study, we used a Japanese cohort, which differs from other populations in some aspects. For example, in the US, most CBTs in adults are performed with double-unit cord blood grafts, whereas in Japan, CBTs in adults are performed with a single unit. Moreover, for unrelated transplantations, in the US, most grafts are derived from peripheral blood, whereas in Japan, most grafts are derived from bone marrow. Such differences may limit the generalizability of the findings and prognostic models. Therefore, our findings must be validated using data from other countries. Third, comprehensive genomic studies on AML using next-generation sequencing have recently revealed the relevance of clinical outcomes. 23,50,51 However, data on somatic mutations were not available. Thus, in future studies, genomic information should be incorporated for developing effective prognostic models.
In conclusion, we designed and validated novel nomograms and a web application to predict the OS of patients with AML undergoing allo-HCTs in non-CR, indicating that the performance of our models was at least as favorable as that of the previous scoring system. These prognostic models can be helpful in estimating the benefits and risks of a patient and can provide clues as to whether to conduct transplantation when encountering a patient with AML in non-CR. Furthermore, the web application enables us to easily compare the OS in F I G U R E 9 Web application to predict the overall survival following three types of transplantations. The web application is available at https://JSHCT -AMLWG.shiny apps.io/Predi ct-OS-nonCR -AML-post-HCT/. BMT, bone marrow transplantation; BU, busulfan; CA, cytarabine; CBT, cord blood transplantation; CY, cyclophosphamide; FLU, fludarabine; HCT-CI, hematopoietic cell transplantation-comorbidity index; FAB, French-American-British; MEL, melphalan; PBSCT, peripheral blood stem cell transplantation; TBI, total body irradiation