Background: Acute myeloid leukemia (AML) is a heterogeneous disease with high disease-related mortality and allogeneic hematopoietic stem cell transplantation (allo-HSCT) remains the only curative option for a large proportion of patients. However, allo-HSCT has a significant transplant-related mortality (TRM), so a careful evaluation between risk of disease relapse and TRM is needed for the challenging decision of whom to transplant in first complete remission (CR1). The European LeukemiaNet (ELN) provides a 3-tier static risk classification which is already a useful guidance, although dynamic risk assessment using measurable residual disease (MRD) information should also be taken into account. However, the intermediate-risk group is characterized by a wide heterogeneity and there are also some favorable-risk patients who require allo-HSCT because of early relapses after CR1. Therefore, more precise prediction tools are warranted to provide additional information valuable for clinical decision-making. Aims: To develop and validate a model that provides individualized outcome estimations in adult AML patients achieving CR1 with intensive treatment approaches. Methods: From the HARMONY Platform including 7467 AML patients at the time of the analysis, we selected patients aged 18-70 years old who received intensive treatment, achieved CR and were not consolidated with allo-HSCT in CR1. Patients who were refractory to first line chemotherapy or who had a relapse-free survival (RFS) of less than two months after achieving CR were excluded. A Bayesian Additive Regression Trees (BART) nonparametric machine learning model was used to predict the individualized likelihood of RFS, cumulative incidence of relapse (CIR) and overall survival (OS) in 1863 AML patients who were also selected based on the availability of comprehensive mutational information by next-generation sequencing analysis. Genomic aberrations that were present in at least 10 patients were included in the model. In order to test the accuracy of the model, a 10-fold cross-validation was applied using the area under receiver operating curve (AUC) at predefined time points (1, 2, 3, 4 and 5 years). Subsequently, the model was validated using an external publicly available database (Tazi et al, Nat. Commun. 2022) which included 722 patients based on identical inclusion criteria. Results: The study population of 1863 adult AML patients in CR1 included 51.2% of males and median age was 50 years. Regarding ELN2022 classification, 52% of patients were classified as favorable, 30% as intermediate and 18% as adverse risk. Most frequently mutated genes were NPM1 (35.9%), DNMT3A (25.6%) and NRAS (21.6%). The HARMONY model was able to estimate individualized likelihood of RFS, CIR and OS for all patients, providing confidence intervals for all estimations. Moreover, we could predict outcomes with increased accuracy over ELN2022 at all predefined time points. At 5 years, the AUC value was significantly better for the HARMONY model than ELN2022 in predicting RFS (0.729 vs 0.679, p=0.003), CIR (0.729 vs 0.680, p=0.002) and OS (0.733 vs 0.675, p<0.001) ( figure 1A). Validation in an external independent AML cohort confirmed superiority of the HARMONY model in predicting RFS (5-year AUC 0.714 vs 0.609, p <0.001), CIR (0.713 vs 0.610, p <0.001) and OS (0.737 vs 0.638, p <0.001) as compared to the ELN2022 risk stratification ( figure 1B), despite the fact that there was a different distribution of risk categories (28% favorable, 45% intermediate and 27% adverse risk). Conclusions: Analysis of large well-defined AML cohorts allows the development of increasingly precise predictive models for clinically relevant scenarios. The HARMONY machine learning model provides individualized outcome prediction for patients aged 18-70 years in CR1 consolidated without allo-HSCT and might be used in the future for clinical decision-making. The model can be accessed online via an interactive web calculator () and, while it has been validated using a large independent AML cohort, future studies are required focusing on prospective validation. Furthermore, the incorporation of additional variables such as MRD and the inclusion of patients treated with targeted therapies and non-intensive approaches will provide more accurate risk predictions.
Read full abstract
Sep 21, 2024
Dayamoy Mandal + 1
Open Access