Machine Learning Improves AML Relapse Risk Prediction Beyond ELN 2022: HARMONY Study

Machine learning helped predict the return of acute myeloid leukemia (AML) more accurately than the 2022 European LeukemiaNet (ELN) classification system. Researchers developed the HARMONY model using data from 2,550 patients in the HARMONY Alliance database. They then tested it in a separate group of 714 patients who took part in clinical trials in the United Kingdom.

The model uses standard medical information and genetic test results to estimate each patient's chances of relapse, survival without relapse, and overall survival after their first complete remission (CR1). These estimates apply to patients who received intensive chemotherapy and did not receive a donor stem cell transplant (allogeneic hematopoietic stem cell transplantation, or allo-HSCT) as follow-up treatment. For all three outcomes, the model performed better than the ELN 2022 system.

Why ELN 2022 May Not Give Enough Detail for Individual Patients

The ELN 2022 system places people with AML into favorable-, intermediate-, or adverse-risk groups based on chromosome changes and other genetic changes. Doctors use these groups to help decide whether a patient should receive an allo-HSCT after reaching CR1. However, several genetic changes can interact, so people in the same risk group may have very different outcomes. This is especially important for favorable-risk AML. Many of these patients stay in remission with chemotherapy-based follow-up treatment, but others experience a relapse even though their genetic results suggest a favorable outlook.

An allo-HSCT can lower the chance of relapse, but it also carries a risk of death from causes other than the leukemia returning. Researchers have suggested that an estimated relapse risk above about 40% without a transplant may be a reason to consider allo-HSCT, as long as the expected risk from the transplant itself is low enough. The researchers tested whether machine learning could provide more accurate estimates by looking at each patient's particular combination of genetic and clinical features.

Study Design and Model Development

The study included 2,550 adults aged 18–70 years from the HARMONY Alliance AML database. This database contains information from European clinical trials, national patient registries, and participating medical centers. All patients had reached CR1 after intensive chemotherapy without receiving an allo-HSCT during that remission. All also had chromosome test results and next-generation sequencing results from the time of diagnosis. Patients who received targeted medicines, including FLT3 or IDH inhibitors, were excluded. The median age was 52 years. According to ELN 2022, 53% of the patients were in the favorable-risk group, 25% were in the intermediate-risk group, and 22% were in the adverse-risk group.

The model uses a method called Bayesian Additive Regression Trees, which can identify complex relationships between clinical information and genetic features. It uses the patient's age, sex, type of AML, chromosome test results, and 76 genetic abnormalities. Each of these genetic abnormalities was a gene mutation found in at least ten patients. The predictions begin at CR1 and estimate overall survival, survival without relapse, and the cumulative chance of relapse at one, two, three, four, and five years. The researchers tested the model in a separate group of 714 patients from clinical trials run by the UK's National Cancer Research Institute.

How the Model Compared With ELN 2022

The model was better than ELN 2022 at distinguishing between patients with different outcomes at every time point studied. In the group used to develop the model, the area under the time-dependent receiver operating characteristic curve (AUC) was:

• Five-year overall survival: 0.741 compared with 0.700 for ELN 2022
• Five-year survival without relapse: 0.752 compared with 0.705
• Five-year cumulative chance of relapse: 0.742 compared with 0.708

The results from the UK validation group supported these findings. The AUC values were:

• Overall survival: 0.755 compared with 0.714
• Survival without relapse: 0.717 compared with 0.679
• Cumulative chance of relapse: 0.718 compared with 0.672

The predicted outcomes also closely matched the outcomes actually observed in patients. The model still performed better when it used only genetic information. This suggests that combining different genetic abnormalities was a major reason for the improvement. Although the improvement was consistent, it was moderate. The model was not accurate enough to guide treatment decisions on its own, without further clinical assessment.

Relapse Risk Within ELN Risk Groups

In the group used to develop the model, the observed five-year cumulative chance of relapse was 38% for favorable-risk patients, 63% for intermediate-risk patients, and 75% for adverse-risk patients. However, the model's individual five-year relapse estimates varied widely within each group:

• 18%–60% in the favorable-risk group
• 46%–86% in the intermediate-risk group
• 52%–91% in the adverse-risk group

This means that patients in different ELN groups could have similar predicted chances of relapse. Among patients whose predicted five-year relapse risk was between 50% and 70%, the observed relapse rates were 66% in the favorable-risk group, 59% in the intermediate-risk group, and 59% in the adverse-risk group. The differences were not statistically significant. The separate validation group showed a similar range of risks within each ELN category.

Favorable-Risk AML Patients With a Higher Chance of Relapse

Patients with favorable-risk AML do not usually receive an allo-HSCT after their first complete remission. This makes the study's findings particularly important for this group. In the development group, 593 of the 1,357 favorable-risk patients had a predicted five-year relapse risk above 40%. Their observed five-year relapse rate was 48%, compared with 30% among those whose predicted risk was below 40%. In the validation group, 47.6% of favorable-risk patients had a predicted risk above this threshold. Their observed relapse rate was 54%, compared with 32% in patients below the threshold.

A predicted relapse risk above 40% does not prove that an allo-HSCT would improve survival. The study included only patients who did not receive a transplant during their first complete remission. The results instead support assessing each patient's need for a transplant more individually.

Genetic Differences Within Favorable-Risk Groups

Among patients with NPM1-mutated AML who did not have an FLT3 internal tandem duplication, about half had a predicted five-year relapse risk above 40%. Their observed relapse rate was 48%, compared with 29% among patients with a lower predicted risk. DNMT3A mutations were much more common in the higher-risk group. These patients were also more likely to have had a blood disorder before their AML diagnosis.

Among patients with t(8;21)/RUNX1::RUNX1T1 AML, 18.4% had a predicted relapse risk above 40%. Their observed relapse rate was 56%, compared with 27% in the lower-risk group. Mutations in RUNX1, DNMT3A, and TP53 were more common among patients in the higher-risk group. The validation group confirmed that relapse risks varied within both genetic subgroups, although the proportion of patients classified as higher risk was different.

Measurable Residual Disease

The study did not examine measurable residual disease (MRD), which is an important factor in predicting outcomes. MRD testing looks for small amounts of leukemia that remain after treatment and can show how well treatment has worked. The model does not include this information because it focuses on genetic findings from the time of diagnosis. The authors acknowledge that adding MRD results could make predictions more accurate for individual patients, especially when genetic risk does not reflect how well the leukemia responds to initial treatment and follow-up chemotherapy. However, a negative MRD result does not eliminate the chance of relapse. In some groups defined by their genetic features, the three-year cumulative chance of relapse can approach 40% even when MRD testing is negative.

Combining genetic information from diagnosis with MRD results could help distinguish between patients who have similar genetic changes but different amounts of leukemia remaining after treatment. A separate HARMONY study is examining MRD in another group of patients.

How Well the Model Applies to Current Treatments

The study looked back at existing patient data, and its selection rules limit how widely the results can be applied. The development group included adults who received intensive chemotherapy without targeted medicines and reached CR1. Some patients in the validation group received gemtuzumab ozogamicin or the FLT3 inhibitor lestaurtinib, but the model has not been fully tested across the range of treatments now used. Targeted medicines are increasingly used as initial treatment and may change relapse risks within groups of patients who share particular genetic features. The authors expect that the model will need to be updated as AML treatment changes.

The predictions should not be applied to patients older than 70, patients who are not well enough to receive intensive treatment, patients who do not reach remission, or patients who receive less intensive treatment unless the model is further tested in these groups.

Online Tool for Estimating Risk

The researchers created an online tool that estimates outcomes for individual patients using five types of information: age, sex, AML type, chromosome abnormalities, and gene mutations. It runs repeated simulations using patient data to estimate overall survival, survival without relapse, and the chance of relapse. It also provides confidence intervals, which show the uncertainty around the estimates. A sensitivity analysis shows how the predicted five-year relapse risk changes when different genetic abnormalities are considered.

The AML Outcome Predictor is available to the public. However, the study does not establish that it is a validated tool that can be used on its own to decide whether a patient should receive a transplant. It is best used together with a clinical assessment, molecular risk classification, MRD testing, and estimates of the risk of death related to transplantation.

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