A revised prognostic model for patients with acute myeloid leukemia and first relapse Open Access

Disease recurrence is the most common cause of treatment failure in patients with acute myeloid leukemia (AML). Treatment of relapsed AML remains challenging, and outcome has been traditionally poor with median overall survival (OS) of <6 months.^1-5 Although outcome has been improved with novel targeted treatments,^6-9 allogeneic hematopoietic cell transplantation (allo-HCT) remains the only curative treatment for relapsing patients.^10,11 Risk models for these patients might support decision making for reinduction treatment and consolidation treatment with allo-HCT, whereas alternative experimental strategies might be considered for high-risk patients.

Prognostic factors associated with survival after relapse in AML have previously been identified.^1,2,5,12-14 These include the time between the first complete remission (CR1) and relapse, age at relapse, the presence of favorable and unfavorable karyotypes, the presence of FLT3 internal tandem duplications (ITDs), and previous allo-HCT.^1,2,5,12-14 Earlier developed prognostic model groups^1,2 have highlighted these factors as significant predictors associated with survival after first AML relapse. Since the development of these risk models, an increasing number of molecular and cytogenetic alterations has been recognized affecting outcome in patients with newly diagnosed AML,^11,15 which might also affect survival after first relapse. Models that allow for risk stratification after first relapse using cytogenetic and molecular variables established at initial diagnosis in a recently treated AML patient cohort are currently lacking.

We set out to study a large cohort of patients with AML who relapsed after intensive induction treatment and for whom baseline comprehensive clinical, cytogenetic, and molecular variables were available. We developed a revised stratification system that was validated in an independent cohort of older patients with AML.

The complete cohort consisted of 5086 patients aged 18 years and older enrolled in 8 consecutive HOVON-SAKK clinical trials (HO42, HO42A, HO43, HO81, HO92, HO102, HO103, and HO132; supplemental Materials) between 2000 and 2018 with an intensive induction chemotherapy backbone for newly diagnosed AML and high-risk myelodysplastic syndrome (MDS) with International Prognostic Scoring System of ≥1.5, Revised International Prognostic Scoring System risk score of >4.5, or excess blasts of ≥10% (see supplemental Figure 1A-H for trial details).^16-23 Postremission treatment with additional chemotherapy, high-dose chemotherapy followed by autologous HCT, or allo-HCT was based on AML risk in evolving classifications.^11,24 Patients were excluded because of no next-generation sequencing (NGS) data available at diagnosis (n = 2543), no CR or CR with incomplete count recovery after 2 cycles of induction chemotherapy (n = 416), and no relapse (n = 1184). The final development cohort consisted of 943 patients who developed a first relapse (Figure 1). The validation cohort included newly diagnosed AML and patients with high-risk MDS with excess blasts of ≥10% aged 60 years and older enrolled in the NCRI-AML18 trial who were treated with intensive induction chemotherapy between 2014 and 2018 (supplemental Figure 1I).²⁵ Patients were excluded because of no CR or CR with incomplete count recovery after 2 cycles of induction chemotherapy (n = 245) and no relapse (n = 354). The final validation cohort consisted of 377 patients who developed a first relapse (Figure 1). All trial participants provided a written informed consent in accordance with the Declaration of Helsinki.

High-molecular-weight genomic DNA was extracted from bone marrow samples or peripheral blood. Diagnostic samples at first AML diagnosis underwent gene resequencing using either a 97-gene panel (Oxford for UK-AML18) or a 54-gene panel (Erasmus MC, Rotterdam, for HOVON-SAKK), containing the most frequently mutated genes in myeloid malignancies. Targeted NGS used liquid-phase capture of molecule-barcoded libraries, with custom-designed probes targeting AML-associated genes for UK-AML18 (Roche, Basel, Switzerland) or the Illumina TruSight Myeloid panel for HOVON-SAKK (Illumina, San Diego, CA). Genomic libraries were sequenced on the Illumina platform, ensuring a minimum mean target coverage of 500×. Details of cytogenetic analysis, NGS, and additional molecular analyses have been previously published.¹⁵ 

The analysis adhered to the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis guideline for prediction model development and validation.²⁶ A total of 45 variables (supplemental Table 1) were considered for the association with 1-year OS in model development: age at relapse (≥60 vs <60 years), relapse-free interval (≤12 months vs >12 months) defined as the time between CR1 and relapse, number of treatment cycles needed to obtain CR1 (1 vs >1), sex (male vs female), white blood cell count (WBC) at diagnosis (≥10 × 10⁹/L vs <10 × 10⁹/L), previous allo-HCT (yes vs no), previous autologous HCT (yes vs no), 29 different gene mutations (presence vs absence), and 9 different cytogenetic abnormalities (presence vs absence) (see supplemental Table 1 for details of mutations and cytogenetic variables). Complex and monosomal karyotypes were grouped owing to their frequent co-occurrence in this cohort (94% concordant). During variable selection, we evaluated different thresholds for age, relapse-free interval, and WBC and found that the most optimal cutoffs were 60 years, 12 months, and 10 × 10⁹/L, respectively. Gene mutations and cytogenetic abnormalities were only evaluated if present in 10 or more patients (supplemental Table 1). OS was defined as the time from the date of first relapse to either the date of death from any cause or the date of last contact while still alive. OS curves were estimated using the Kaplan-Meier method and compared using the log-rank test. Model development consisted of 3 sequential stages, which included (1) variable selection using a random survival forest, (2) Cox regression analysis with backward selection (P < .10) to define the optimal model, and (3) assignment of weights and defining groups based on significantly different OS compared with a lower score to construct the final prognostic model. Further details are available in the supplemental Materials. The impact of allo-HCT after relapse was assessed using a time-dependent Cox regression allowing the allo-HCT covariate to change the state at the time of allo-HCT. Fine-Gray models were used to estimate the probability of an event accounting for competing risks, nonrelapse mortality (NRM), and relapse. Prognostic accuracy of existing models and the revised model was evaluated using Harrell’s C-index. Cytogenetic data were not available in 38 patients (4.0%) owing to failed karyotyping and 3 patients (0.3%) because of missing WBC at diagnosis in the development cohort. Missing cytogenetics data were uniformly imputed as absent, consistent with real-world clinical practice, and with the median WBC count for missing WBC. All analyses were performed in R version 4.4.1. The R script of the analyses can be found online (https://github.com/niekvandermaas/AML-relapse-model).

The development cohort consisted of 943 patients who relapsed after having obtained CR1 upon initial intensive treatment (Table 1). In this cohort, 1-year OS was similar across treatment periods (<2010 vs ≥2010; supplemental Figure 2) and disease type (AML vs high-risk MDS; supplemental Figure 3). In addition, no difference was observed for 1-year OS between missing and nonmissing NGS data (supplemental Figure 4). The median age of patients at relapse in this cohort was 58 years (range, 18-81 years; Table 1). At diagnosis, 22% of the patients were classified as having a favorable risk according to the European LeukemiaNet (ELN) 2022 AML classification, 31% as intermediate risk, and 48% as adverse risk (Table 1).¹¹ In this cohort, 33% of patients received postremission treatment with allo-HCT, whereas 14% underwent autologous HCT in first CR (Table 1). A total of 95% relapsing patients had at least 1 cytogenetic abnormality or genetic mutation at diagnosis (median, 3), with 62% of patients having molecular mutations only, 3% having karyotype alterations only, and 32% having both. Most mutations at diagnosis were found in the following genes: DNMT3A (30%), NPM1 (27%), and FLT3 ITD (22%). Mutated TP53 was found in 12% of patients. Complex or monosomal karyotype according to ELN 2022 was the most frequent cytogenetic abnormality in these patients (19%), whereas core-binding factor (CBF) abnormalities [t(16;16)/inv(16) (3%) and t(8;21) (2%)] were relatively infrequent (Figure 2).²⁷ The development cohort had a median follow-up from relapse of 49 months for patients alive (range, 0-128 months; Table 1). OS rates at 1 and 4 years after first relapse in the development cohort were 33% ± 2% and 16% ± 1%, respectively.

Previous prognostic models were assessed for their predictive performance in the development cohort. The HOVON-SAKK model¹ was associated with a C-index of 0.65 ± 0.016 for 1-year OS. Most patients were classified as adverse risk in that model (684 of 943, 73%). One-year OS after relapse was clearly distinct estimating 74% ± 7% for the favorable group, 54% ± 3% for the intermediate group, and 24% ± 2% for the adverse group, respectively (Figure 3A). Similarly, the GOELAMS model² was associated with a C-index of 0.64 ± 0.017, outcomes of which were distinct in the 3 risk groups associated with survival probabilities at 1 year: 49% ± 3% for the favorable group, 29% ± 2% for the intermediate group, and 11% ± 3% for the adverse group (Figure 3B).

To determine the optimal patient-specific and genetic variables for risk stratification of OS, we used a random survival forest algorithm, which revealed the hierarchy of each clinical, genetic, and cytogenetic variable for stratifying 1-year OS (details in the supplemental Materials). A total of 45 candidate variables were considered (supplemental Table 1), of which 18 were identified as significantly affecting (P < .01) 1-year OS (supplemental Figure 5). These were included in a multivariable Cox regression analysis using a stepwise backward selection, which resulted in a final model of 9 predictors. The 3 most important variables for 1-year OS after relapse were the absence of CBF abnormalities (hazard ratio [HR], 2.07; 95% confidence interval [CI], 1.38-3.12; P < .001), mutated TP53 (HR, 1.99; 95% CI, 1.48-2.67; P < .001), and relapse-free interval of ≤12 months (HR, 1.76; 95% CI, 1.46-2.12; P < .001) (Figure 4A). Notably after adjustment for covariates, including mutated TP53, complex or monosomal karyotype was independently correlated with adverse OS (HR, 1.40; 95% CI, 1.08-1.82; P = .01; Figure 4A). Each of the 9 variables was assigned points based on their rounded HRs (Figure 4A) resulting in 3 points for either no CBF abnormalities or a TP53 mutation. Similarly, relapse-free interval of ≤12 months, previous allo-HCT, t(v;11q23)/KMT2A rearrangement, and age at relapse ≥60 years were each assigned 2 points, whereas WBC of ≥10 × 10⁹/L, complex and/or monosomal karyotype, and FLT3 ITD were assigned 1 point. These points were subsequently used to derive the total score for each patient. The median total score in the development cohort was 7 points (range, 1-14). OS decreased with increasing scores (supplemental Figure 6). Based on significantly different OS, we collapsed the development cohort into 3 groups: favorable (≤6 points; n = 389, 42%), intermediate (7 points; n = 197, 21%), and poor (≥8 points; n = 357, 38%; supplemental Figure 7).

Favorable-risk patients in the revised model had a 1-year OS of 51% ± 3% (Figure 4B). Intermediate-risk patients had a 1-year OS of 29% ± 3%, which was 14% ± 2% in the poor-risk subgroup (Figure 4B). The C-index of the prognostic model in the development cohort was 0.71 ± 0.016 with excellent calibration (supplemental Figure 8). Four-year OS was also significantly different among the 3 subgroups (29% ± 2%, 11% ± 2%, 5% ± 1%, respectively; supplemental Figure 9A). The model restratified 57% and 49% of patients from the previous HOVON-SAKK and GOELAMS risk groups, respectively. Most patients in the large HOVON-SAKK poor-risk group were reclassified as favorable (26%) or intermediate (24%), whereas 30% and 42% of patients in the large GOELAMS intermediate-risk group were reclassified to the favorable- and poor-risk group of the revised model, respectively (Figure 4C-D).

The model was validated in an independent data set derived from the NCRI-AML18 trial, consisting of 976 older patients with newly diagnosed AML aged 60 years and older who also received intensive induction chemotherapy. A total of 377 relapsing patients were identified with a median age at relapse of 69 years (range, 60-81) (Figure 5A). Sixty-two percent were classified as poor or very poor risk in the AML60+ classification at the time of diagnosis,¹⁵ whereas 60% were adverse according to the 2022 ELN risk classification. Allo-HCT in CR1 was applied in 14% as postremission treatment (Figure 5A). A total of 97% of patients had at least 1 cytogenetic or molecular abnormality (median, 4), with 77% of patients having molecular mutations only, 1% having karyotype alterations only, and 21% having both. The most frequently mutated genes identified were DNMT3A (37%), NPM1 (27%), and ASXL1 (26%). TP53 was mutated in 11% of patients. According to ELN 2022 criteria, the most common cytogenetic abnormality was a complex or monosomal karyotype (14%), whereas CBF abnormalities, specifically t(16;16)/inv(16) (1%) and t(8;21) (3%), were rare (Figure 5B). Median follow-up from relapse was 21 months for patients alive (range, 0-59 months; Figure 5A). OS rates at 1 and 4 years after first relapse year in the validation cohort were 30% ± 2% and 9% ± 2%, respectively. The 3 groups defined in the development cohort were used to classify the validation cohort: favorable (≤6 points; n = 137, 36%), intermediate (7 points; n = 85, 23%), and poor (≥8 points; n = 155, 41%), which were associated with distinct 1- and 4-year OS (1-year OS, 51% ± 4%, 26% ± 5%, and 14% ± 3%, respectively; Figure 5C; supplemental Figure 9B). The performance of the revised prognostic model as measured by the C-index was 0.71 ± 0.028 with excellent calibration (supplemental Figure 10). The GOELAMS and HOVON-SAKK models had C-indices of 0.69 ± 0.027 and 0.62 ± 0.027 in the validation cohort, respectively. Similar to the development cohort, the revised model restratified 59% from the original HOVON-SAKK risk groups and 35% from the GOELAMS groups (Figure 5D-E). HOVON-SAKK poor-risk patients were considered favorable or intermediate in 24% and 26%, respectively, in the revised model.

In the development cohort, reinduction treatment was used in 690 patients (73%), which resulted in a second CR (CR2) in 348 of patients (50%) (supplemental Figure 11; supplemental Table 2). Across subgroups, CR2 after reinduction treatment was obtained in 65% and 44% of the favorable- and intermediate-risk patients, respectively, which was higher than the poor-risk group (32%; C-index, 0.66; supplemental Tables 2 and 3). In addition, the revised model showed a distinct OS and event-free survival for patients who received reinduction treatment and relapse-free survival for patients in CR2 (supplemental Figure 12; supplemental Table 3). The revised model compared favorably in terms of accuracy to these outcomes with the previous risk stratification systems (supplemental Table 3). Allo-HCT was used as consolidation treatment of a CR2 in 159 patients (46%), and 38 (11%) received a donor lymphocyte infusion after CR2 (supplemental Figure 11; supplemental Table 2). A total of 342 patients (50%) did not achieve a CR2 despite reinduction treatment consisting of intensive chemotherapy (n = 279, 82%), upfront allo-HCT (n = 43, 13%), or donor lymphocyte infusion (n = 20, 6%) (supplemental Figure 11; supplemental Table 2).

A time-dependent analysis with allo-HCT as a time-varying covariate was performed for patients who had attained a CR2. OS was improved by allo-HCT compared with no allo-HCT (HR, 0.54; 95% CI, 0.37-0.79; P = .001; supplemental Figure 13), and cumulative incidence of relapse was lower after allo-HCT than no allo-HCT (HR, 0.58; 95% CI, 0.38-0.88; P = .011; supplemental Figure 13). In contrast, the cumulative incidence of NRM was higher in the allo-HCT group, although not significantly different (HR, 1.30; 95% CI, 0.61-2.80; P = .499; supplemental Figure 13). Low patient numbers precluded an analysis of the impact of allo-HCT per risk group.

Relapse in AML is frequent with low response rates to reinduction treatment and dismal OS.^1-5,14 Recent prognostic models for patients with relapsed AML incorporating comprehensive genetic data are currently lacking. Therefore, it is necessary to reassess the prognostic value of clinical, cytogenetic, and molecular AML characteristics in patients with relapsed AML after intensive induction treatment. We analyzed the genomic landscape and clinical outcomes of 943 patients with AML with a first relapse aiming to develop a simple prognostic classification system. Using a machine-learning method for variable selection, 9 variables were identified, which were used to stratify patients into 3 risk groups. Although age, relapse-free interval, CBF abnormalities, FLT3 ITD, and previous allo-HCT were confirmed from previous risk models,^1,2,14 we additionally found WBC count, mutated TP53, t(v;11q23)/KMT2A rearrangement, and complex/monosomal karyotype. The prognostic model was associated with highly distinct OS in 3 risk groups. The revised prognostic model was validated in an independent cohort of older patients with AML treated within the NCRI-AML18 trial indicating similar prognostic accuracy. It classifies patients into other risk groups allowing for better discrimination of OS outcomes for patients with AML experiencing their first relapse than existing prognostic models, such as those from the HOVON-SAKK and GOELAMS groups.^1,2 

Salvage treatment for patients with relapse AML includes high-dose chemotherapy, hypomethylating agents, or targeted treatments for patients with specific mutations followed by allo-HCT as consolidation treatment in patients who obtain CR2.^10,11 High-dose chemotherapy regimens (eg, intermediate-dose cytarabine with or without anthracycline or the combination of fludarabine, cytarabine, idarubicin and G-CSF) are associated with remission rates of 20% to 65%, but also considerable toxicity and mortality (6%-22%).^28-33 Although the type and intensity of chemotherapy-based reinduction treatment strategies were not available in our data set, we observed a relatively high CR2 rate of 65% and 44% for favorable- and intermediate-risk patients, which was 32% for poor-risk patients. Alternative salvage treatment approaches might be considered such as targeted treatments with or without hypomethylating agents. Gilteritinib has been approved for patients with relapsed FLT3 AML with CR rates of 34%.⁸ Similarly, both ivosidenib (IDH1 inhibitor) and enasidenib (IDH2 inhibitor) offer 23%-33% CR rates in refractory or relapsed patients with IDH-mutated AML,^6,7 whereas menin inhibitors show a CR rate of 30% in KMT2A-rearranged or NPM1 mutated leukemia.⁹ Alternatively, combining venetoclax, a BCL2 inhibitor, with hypomethylating agents presents as a potential alternative to attain CR2, despite the absence of trial data in the relapsed setting.^34-37 Combination treatments of targeted drugs, hypomethylating agents, and venetoclax or high-dose chemotherapy with venetoclax are currently investigated, and early results have been encouraging.^38-41 Nevertheless, it remains largely unknown whether patient outcomes will significantly improve with these novel treatment modalities. These recent advancements may transform the therapeutic landscape of relapsed AML significantly in the next years, making it a necessity to further validate this prognostic index to reflect emerging treatment modalities in the future.

Historically, allo-HCT in CR2 has been the preferred approach for long-term survival. Our analysis confirms that allo-HCT in CR2 provides a survival benefit with reduced relapse incidence compared with non-allo-HCT treatments. Nonetheless, the benefit of allo-HCT in terms of relapse reduction can be compromised by NRM, particularly in older patients or those with underlying comorbidities.^42-44 Although the risk of a second failure without allo-HCT is high (85% in our study), it needs to be balanced against the risk of NRM as assessed by risk scores.^42-44 Ultimately, the decision to proceed with allo-HCT after AML relapse requires a personalized approach, depending on the specific characteristics of each patient and their disease. Of note, data from the recent ASAP trial suggested that immediate allo-HCT might be an alternative for fit patients with nonproliferative refractory/relapsed AML who have a stem cell donor available.⁴⁵ However, the small number of relapsing patients in this study precludes robust conclusions.

Our study has several limitations. The optimal external validation cohort may be debated, but Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis guidelines have recommended using a cohort that differs in key characteristics (eg, age), treatment approaches, or time period.²⁶ The revised model was validated in an independent cohort, in which patients received different intensive chemotherapy regimens and were older than the development cohort. The model was associated with strong performance across both cohort populations, without any limitations imposed by narrow age restrictions, highlighting broad applicability in patients with relapsing AML after first-line intensive treatment. Second, intensive chemotherapy and targeted therapies might exert a selective pressure to the clonal landscape of AML, leading to both treatment-sensitive and treatment-resistant cells. For example, RAS-pathway mutations that are acquired at relapse may confer a particularly therapy-resistant disease.^46,47 Leukemic transformation by RAS mutations exhibits resistance to the BCL2 inhibitor venetoclax, driving clinical resistance, relapse, and worse OS after relapse after venetoclax-based therapy.^37,46 This clonal evolution of the AML is not accounted for given that molecular and cytogenetic data were only available from the time of initial diagnosis. The predictive accuracy and discriminatory power of the model, as measured by the C-index, indicate that the model performs relatively good in distinguishing among patients with different risk levels. The C-index quantifies the risk classification performance based on the predicted risk, with values ranging from 0.5 (no better than random chance) to 1.0 (perfect discrimination). A higher C-index indicates better model performance in correctly identifying which patients are more likely to experience an event. The revised relapse model improved upon previous risk classification systems, and its predictive capability may be further enhanced by incorporating molecular and cytogenetic data at the time of relapse. In addition, the model was developed in cohort of patients who did not receive FLT3 inhibitors as part of the first-line induction regimen. Although 26% of patients (22 of 85) with FLT3 ITD received quizartinib added to the intensive induction backbone in the validation cohort, the model needs further validation in the era of FLT3 targeted combination therapies. Finally, there is an inherent bias regarding treatment decision making in the patient population after first relapse. High-risk patients might not have received treatment because of their adverse risk factors (eg, mutated TP53, short relapse-free interval, older age, other adverse cytogenetics). For example, in the poor-risk group, only 215 patients (60%) received reinduction treatment. Among those, response rates were very low (32%), with only 20 patients proceeding to allo-HCT in CR2, suggesting that this is a particularly difficult AML population to treat.

In conclusion, the revised prognostic classification system for adult patients with AML in first relapse offers a useful and distinctive model for clinical practice. It identifies favorable- or intermediate-risk patients who may benefit from reinduction strategies and consolidation with allo-HCT taking into account the risk of NRM. Conversely, patients classified within the poor-risk group have dismal survival after first relapse and might be considered for novel treatment strategies, experimental treatments, or even best supportive care. To facilitate the integration of this prognostic tool into clinical practice, an online calculator has been developed.⁴⁸ Based on a patient-specific hematologic, cytogenetic, and molecular profile, the tool provides a personalized prediction score, assigns an associated risk group, and offers estimated OS projections. This tool may aid clinicians in evaluating the relative benefits of salvage and experimental treatments, balancing these against potential treatment-related risks.

This work was supported by a grant from the Erasmus Medical Center Foundation (D.d.H.).

Contribution: N.G.v.d.M. and J.V. designed the study, analyzed the data, conceived of the statistical plan, performed the statistical analysis, had full access to all data, and wrote the manuscript; D.B., T.P., A.T., B.J.B., J.K., C.H.M.J.V.E., O.V., M.-C.V., C.G., J.M., S.K., M.D., S.F., I.T., G.H., C.C., P.V., N.R., G.O., and B.L. provided patient data; N.G.v.d.M., P.G., H.B.B., P.J.M.V., and J.V. collected and assembled clinical, laboratory, and genetic data; N.G.v.d.M., J.J.C., and J.V. interpreted data and contributed to research discussion; and all authors reviewed the manuscript and approved the submission.

Conflict-of-interest disclosure: The authors declare no competing financial interests.

Correspondence: Jurjen Versluis, Department of Hematology, Erasmus Medical Center Cancer Institute, University of Rotterdam, Dr Molewaterplein 40, 3015GD Rotterdam, The Netherlands; email: j.versluis.1@erasmusmc.nl.