Skip to main content
Premium Trial:

Request an Annual Quote

Machine Learning Model Improves Risk Estimates for AML Patients


NEW YORK – A machine-learning model provided individualized predictions of relapse and survival for patients with acute myeloid leukemia with greater accuracy than a widely used risk classification tool, according to results presented at the American Society of Hematology's annual meeting on Saturday.

In a presentation, Alberto Hernandez Sanchez of the University Hospital of Salamanca described the development and performance of a machine-learning model created using the HARMONY Alliance's big data platform. The HARMONY Alliance, which involves 80 public-private organizations from 22 European countries, aims to use machine-learning algorithms to translate big data into better treatment strategies for those with hematologic malignancies. The new model described at ASH is focused on helping to decide which AML patients should be considered for allogeneic stem cell transplantation.

Standard therapy for AML is determined by the phase of disease. In the remission or induction phase, the goal is to eliminate as many leukemia cells as possible, typically with intensive chemotherapy or targeted therapies. In the consolidation phase, treatment is geared toward mopping up any remaining leukemia cells. That treatment could consist of chemotherapy, or an allogeneic or autologous stem cell transplant. Of these, an allogeneic stem cell transplant is the only curative option, but there's a significant mortality risk associated with the procedure, which requires doctors to carefully consider the risks and benefits for each patient. Physicians typically recommend this type of transplant if there's more than a 40 percent chance that a patient's AML will relapse.

The recently updated European LeukemiaNet (ELN) guidelines from 2022 divide patients into three categories according to their risk of relapse: favorable, intermediate, and adverse. Using ELN2022, decision-making tends to be most difficult for patients in the intermediate group, but Sanchez noted that there is heterogeneity within all three risk groups. "Therefore, there is a need for individualized estimation of outcomes" beyond what ELN2022 can provide, he suggested.

Sanchez and his colleagues aimed to harness the HARMONY Alliance's big data platform to develop and validate a model that could provide those individualized outcome estimations for AML patients who have achieved their first remission. Within the platform, the researchers selected AML patients between 18 and 70 years of age who had achieved a complete response on intensive treatment, but excluded those who already had targeted therapy, non-intensive treatments, or allogeneic hematopoietic stem cell transplant.

After narrowing the group down to patients who had results from cytogenetic and next-generation sequencing panel tests available, the final dataset included 1,863 AML patients with a median age of 50 years, who were mostly determined to have favorable risk according to the ELN2022 criteria. Using this data, Sanchez's team developed a Bayesian Additive Regression Trees nonparametric machine-learning model that predicts outcomes based on five variables: age, gender, AML type, cytogenetic abnormalities, and genetic mutations. The model incorporated all mutations that were present in at least 10 patients.

The most frequently mutated genes were NPM1, found in 35.9 percent of the patients; DNMT3A in 25.6 percent; and NRAS in 21.6 percent. According to ELN2022 criteria, 52 percent of the patients had favorable risk, 30 percent had intermediate risk, and 18 percent had adverse risk of relapse.

Sanchez noted that the advantage of the type of machine-learning model his team used is that it accounts for uncertainty in predictions and can handle complex nonlinear associations. His team applied the model to make individualized predictions about relapse-free survival, the cumulative incidence of relapse, and overall survival.

When values for the area under the curve (AUC) were compared between the HARMONY model and ELN2022 criteria, the HARMONY model gave a superior prediction for each outcome, they found.

They then used an external public database comprising information on patients enrolled in the UK's National Cancer Research Institute trials to validate their algorithm. Focusing on the favorable-risk group, Sanchez said, "we were surprised to notice that our model predicted a five-year cumulative incidence of relapse of more than 40 percent for almost half of the ELN2022 favorable-risk patients."

The HARMONY model stratified the ELN favorable-risk patient group into two subgroups according to differences in their rates of AML relapse. Looking more closely at these subgroups, the researchers found that patients with certain mutations such as NPM1 and CEBPA had a higher risk of relapse. "There was heterogeneity in all ELN2022 favorable subgroups," Sanchez noted.

To predict whether a specific patient was at high risk, the HARMONY model factored in a combination of genomic alterations and other characteristics. For example, for patients with NPM1-mutant AML, a DNMT3A co-mutation was the most important predictor of relapse risk, resulting in a high-risk ranking for 87 percent of 93 patients with both mutations, compared to just 2 percent for those with only a NPM1 mutation. Similarly, according to the model, 78 percent of AML patients who were men and had a CEBPA mutation were in the high-risk group, while none of the female patients with a CEBPA mutation were.

In previous research presented at the European Hematology Association's annual congress in 2022, Sanchez developed a genetic stratification model using computational, not machine learning, methods for NPM1-mutated AML that predicted overall survival and relapse-free survival for three groups of patients. This model reclassified 33 percent of patients with NPM1-mutated AML that had been classified using the ELN2017 criteria.

At the ASH annual meeting this year, Sanchez said that the increased accuracy of the HARMONY model compared to ELN2022 stratification and the algorithm's validation in an independent cohort demonstrate that analysis of large AML cohorts can enable increasingly precise predictive models for clinically relevant patient scenarios. He concluded that the HARMONY machine-learning model can provide individualized outcome predictions for adult AML patients in their first remission following intensive chemotherapy who have been through consolidation therapy without an allogeneic stem cell transplant.

Still, there's room for improvement. "We are aware that the incorporation of additional variables such as [minimal residual disease] and the inclusion of patients treated with targeted therapies and non-intensive approaches will provide even more accurate risk predictions," Sanchez said.