Machine learning improved the prediction of relapse in acute myeloid leukemia (AML) beyond European LeukemiaNet (ELN) 2022 classification. The HARMONY model was developed in 2,550 patients from the HARMONY Alliance database and validated externally in 714 patients from UK clinical trials.
The model combines routine clinical and genomic data to estimate individual probabilities of relapse, relapse-free survival and overall survival after first complete remission (CR1) with intensive chemotherapy and no allogeneic hematopoietic stem cell transplantation (allo-HSCT) as consolidation. It discriminated better than ELN 2022 for all three outcomes.
Why ELN 2022 May Fall Short for Individual Patients
ELN 2022 stratifies AML into favorable-, intermediate- and adverse-risk groups by cytogenetic and molecular abnormalities, and it is central to deciding on allo-HSCT in CR1. Prognosis depends on interactions among multiple genetic alterations, so outcomes within a category can differ substantially. This matters most in favorable-risk AML, where many patients achieve durable remission with chemotherapy-based consolidation and others relapse despite favorable genetics.
Allo-HSCT lowers relapse risk but adds non-relapse mortality. A relapse probability above approximately 40% without transplantation has been proposed as a threshold for considering allo-HSCT, provided the expected transplant-related risk is low enough. Investigators tested whether machine learning could give more precise estimates from individual combinations of genomic and clinical features.
Study Design and Model Development
The study included 2,550 adults aged 18-70 years from the HARMONY Alliance AML database, which draws on European clinical trials, national registries and participating institutions. All had reached CR1 after intensive chemotherapy without allo-HSCT in that remission, and all had cytogenetic and next-generation sequencing results at diagnosis. Patients treated with targeted agents, including FLT3 or IDH inhibitors, were excluded. Median age was 52 years, and 53% were favorable-risk by ELN 2022, 25% intermediate-risk and 22% adverse-risk.
The model uses Bayesian Additive Regression Trees, which can capture complex relationships between clinical and genomic features. Inputs were age, sex, AML type, cytogenetics and 76 genomic abnormalities (gene mutations present in at least ten patients). Predictions start at CR1 and cover overall survival, relapse-free survival and cumulative incidence of relapse at one through five years. An independent cohort of 714 patients from UK National Cancer Research Institute trials served for external validation.
Performance Versus ELN 2022
The model discriminated better than ELN 2022 for every outcome and time point. In the development cohort, the area under the time-dependent receiver operating characteristic curve (AUC) was:
- Five-year overall survival: 0.741 versus 0.700 for ELN 2022
- Five-year relapse-free survival: 0.752 versus 0.705
- Five-year cumulative incidence of relapse: 0.742 versus 0.708
The UK validation cohort confirmed these findings, with AUCs of:
- Overall survival: 0.755 versus 0.714
- Relapse-free survival: 0.717 versus 0.679
- Cumulative incidence of relapse: 0.718 versus 0.672
Calibration showed close agreement between predicted and observed outcomes. The model kept its advantage when restricted to genomic variables alone, which suggests the gain comes mainly from combinations of genetic abnormalities. The improvement is consistent but moderate, and accuracy is not high enough to determine treatment without additional clinical assessment.
Relapse Risk Within ELN Categories
In the development cohort, observed five-year cumulative incidence of relapse was 38% in favorable-risk, 63% in intermediate-risk and 75% in adverse-risk patients. The model’s individual five-year relapse estimates ranged from approximately:
- 18%-60% within the favorable group
- 46%-86% within the intermediate group
- 52%-91% within the adverse group
Patients from different ELN categories could therefore share similar predicted risk. Among patients with a predicted five-year relapse probability of 50% to 70%, observed relapse was 66%, 59% and 59% in the favorable-, intermediate- and adverse-risk groups, with no statistically significant differences. The external validation cohort showed similar heterogeneity.
Favorable-Risk AML With Higher Relapse Probability
Favorable-risk patients do not usually undergo allo-HSCT in CR1, which is where the findings matter most. In the development cohort, 593 of 1,357 favorable-risk patients had a predicted five-year relapse probability above 40%. Their observed five-year relapse incidence was 48%, versus 30% in those predicted below 40%. In the validation cohort, 47.6% exceeded the threshold, with observed relapse of 54% versus 32%.
Exceeding the threshold does not show that allo-HSCT would improve survival, since the study included only patients who were not transplanted in CR1. The findings support more individualized transplant assessment.
Genomic Heterogeneity in Favorable-Risk Subgroups
In NPM1-mutated AML without FLT3 internal tandem duplication, about half of patients had a predicted five-year relapse probability above 40%. Observed relapse was 48%, versus 29% in patients with lower predicted risk. The higher-risk group was strongly enriched for DNMT3A mutations and more often had a previous hematologic malignancy.
In t(8;21)/RUNX1::RUNX1T1 AML, 18.4% of patients exceeded the 40% threshold. Observed relapse was 56%, versus 27% in the lower-risk group, and RUNX1, DNMT3A and TP53 mutations were more frequent among higher-risk patients. The validation cohort confirmed this heterogeneity in both subgroups, although the proportions in the higher-risk groups differed.
Measurable Residual Disease
The study did not assess measurable residual disease, a major prognostic factor that reflects treatment response and is not captured by genetics at diagnosis. The authors acknowledge that adding MRD could improve individualized predictions, particularly where molecular risk does not reflect response to induction and consolidation. MRD negativity does not remove relapse risk: in some molecularly defined subgroups, three-year cumulative relapse incidence can approach 40% despite an MRD-negative result.
Combining baseline genetics with MRD could separate patients with similar genomic abnormalities but different residual disease burden. A separate HARMONY investigation is evaluating MRD in another cohort.
Applicability to Current Treatment
The retrospective design and eligibility criteria limit generalizability. The development cohort consisted of adults who received intensive chemotherapy without targeted agents and reached CR1. The validation cohort included some patients who received gemtuzumab ozogamicin or the FLT3 inhibitor lestaurtinib, but the model has not been comprehensively evaluated across contemporary strategies. Frontline therapy increasingly includes targeted agents that may change relapse probabilities within molecular subgroups, and the authors expect the algorithm to need refinement as treatment evolves.
Predictions should not be extrapolated to patients older than 70, those who are medically unfit, those who fail to reach remission or those given non-intensive regimens without further validation.
Online Risk Estimation Tool
The investigators built a web-based tool that calculates individualized estimates from five inputs: age, sex, AML type, cytogenetic abnormalities and gene mutations. It uses repeated patient simulations to estimate overall survival, relapse-free survival and relapse, with confidence intervals, and includes a sensitivity analysis showing how predicted five-year relapse risk changes with alternative genomic abnormalities.
The AML Outcome Predictor is publicly accessible, but the study does not establish it as a validated standalone instrument for transplant selection. It is best used alongside clinical assessment, molecular risk classification, MRD testing and estimates of transplant-related mortality.
Reference
Hernández-Sánchez A, Martínez Elicegui J, Sträng E, et al. Machine learning provides individualized prediction of outcomes after first complete remission without allo-HSCT consolidation in adult acute myeloid leukemia-A HARMONY study. HemaSphere. Published August 30, 2026. https://doi.org/10.1002/hem3.70461