Validation of the ELN2024 and Mayo Genetic Risk Models in an External Cohort of 352 Patients With Newly Diagnosed AML Receiving Less-Intensive Therapies

Abstract

Risk stratification is critical for guiding therapeutic decisions and predicting outcomes in acute myeloid leukemia (AML). Current widely used risk models, such as the European LeukemiaNet (ELN) genetic risk classification, were predominantly developed based on cohorts treated with intensive chemotherapy [1]. However, when applied to patients receiving less-intensive regimens, these systems have often demonstrated limited discriminatory power, frequently skewing patients toward adverse-risk categories [2].

Recently, Gangat et al. [3] proposed a new Mayo Genetic Risk Model in the American Journal of Hematology, specifically designed to predict outcomes in newly diagnosed AML patients treated with venetoclax combined with hypomethylating agents (Ven-HMA). Alongside the updated ELN2024 classification, which was adapted to better stratify patients receiving non-intensive therapies, these models represent a new generation of tools tailored for patients ineligible for intensive chemotherapy. Nevertheless, their applicability to other non-intensive treatment regimens, such as low-dose cytarabine, remains unclear.

In the present study, we conducted an independent validation of the Mayo Genetic Risk Model and the ELN2024 classification outside the context of hypomethylating agent (HMA)-based treatments, using a publicly available dataset from Tazi et al. [1], and compared their prognostic performance with that of the established ELN2022 model.

We analyzed a cohort of 352 AML patients treated with non-intensive regimens, enrolled across several UK-based NCRI clinical trials, including AML11, AML12, AML14, AML15, AML16, and AML LI1. Specifically, the cohort comprised 162 patients from the AML16 trial and 190 from the remaining studies, with the majority of the latter originating from AML15 and AML LI1 (exact numbers unavailable). Available data included age, sex, complete blood counts, bone marrow blast percentage, genetic and cytogenetic profiles, and WHO 2016 [4] diagnostic classifications. Treatment details at the individual level were not accessible. All patients in the AML16 trial (n = 162, 46%) and AML LI1 received low-dose cytarabine [5], either alone or in combination with investigational agents. Non-intensively treated participants in AML15 received gemtuzumab ozogamicin monotherapy. Risk scores were reconstructed using R software (v4.3.1), and survival differences between groups were assessed with Cox proportional hazards models and visualized via Kaplan–Meier plots.

Table 1 summarizes the patient characteristics: the median age was 75.6 years, with a slight male predominance (57.1%) and a modestly decreased ability to perform activities of daily living (ECOG 1 in 52.6% of patients). The most frequent diagnoses (WHO 2016 classification) were AML not otherwise specified (AML NOS; 37.8%), AML with NPM1 mutation (23.0%), and AML with RUNX1 mutation (20.2%). The most mutated gene in the cohort was TET2 (111, 31.5%), followed by SRSF2 (97, 27.6%) and RUNX1 (88, 25.0%); in comparison with the cohort employed to develop the Mayo Risk Score, we observed a significantly lower frequency of TP53 mutations and increased prevalence of myelodysplasia-related mutations such as DNMT3A, TET2, and SFRSF2 (Table S1). A complex karyotype was identified in 16.4% of patients. Most patients were originally stratified as high risk according to the ELN2017 (56.5%) guidelines.

We reconstructed the Mayo Genetic Risk, ELN2022, and ELN2024 scores for all patients. The Risk group assignments according to the two classifications explicitly designed for non-intensively treated patients (ELN2024 and Mayo Genetic Risk) overlapped in 166 subjects: 51 were classified as adverse risk, 87 as intermediate risk, and 28 as favorable risk. Notably, the Mayo Genetic Risk Model tended to classify patients toward higher-risk categories: 153 patients (84.5%) initially deemed favorable by ELN2024 were reclassified as intermediate or adverse by the Mayo model (McNemar's Chi-squared p-value < 0.001, degrees of freedom [df] = 3). We considered including VEN-PRS [6], a third risk score designed for non-intensively treated AML patients; however, we ultimately excluded it from the comparative analysis due to the lack of critical clinical information such as extramedullary AML involvement and prior AML treatments in our cohort.

All three tested models, Mayo, ELN2022, and ELN2024 classifications significantly distinguished groups with different overall survival (OS) (Cox proportional hazards p = 0.0035 for Mayo Genetic Risk Score; p < 0.001 for ELN2024 (Figure 1); p = 0.0048 for ELN2022). However, risk groups derived from ELN2022 exhibited inconsistent behavior: patients classified as favorable had slightly shorter median OS (5.44 months) than those classified as intermediate (5.62 months), and their survival curves closely resembled those of adverse-risk patients (Figure S1). For these reasons, ELN2022 was not further investigated.

Median OS estimates according to the Mayo and ELN2024 prognostic models are summarized in Table 1. Remarkably, observed OS in the current cohort was considerably shorter across all risk categories compared to the outcomes initially reported in the development studies of these models. For the Mayo model, median OS was 6.28 months for the favorable risk group (originally not reached), 4.99 months for intermediate risk (originally 19.1 months), and 2.92 months for the adverse risk group. Similarly, using the ELN2024 classification, median OS was 6.21 months versus the originally reported 23–39 months for favorable risk, 3.61 months versus 12–13 months for intermediate risk, and 2.46 months versus 5–8 months for adverse risk groups.

Finally, we evaluated model performance using the area under the receiver operating characteristic curve (AUC), concordance index (C-index), the log-likelihood ratio, and corrected Akaike Information Criterion (AIC). Both models demonstrated limited discriminatory power. The Mayo model achieved an AUC of 0.544 and a C-index of 0.544. The ELN2024 classification demonstrated slightly improved predictive performance compared to the Mayo model, with a C-index of 0.572 and an AUC of 0.555. This superior predictive ability of the ELN2024 model was further supported by a significantly higher log-likelihood value (log-likelihood = 27.81, p < 0.001) compared to the Mayo model (log-likelihood = 10.56, p = 0.005). Model parsimony was comparable between the two approaches, as indicated by similar AIC values of 3397.7 (Mayo) and 3380.5 (ELN2024), suggesting that the inclusion of additional variables in the Mayo Genetic Risk Model—such as KMT2A rearrangements, IDH1 mutation status, and karyotype—did not lead to increased model complexity or overfitting.

The survival outcomes observed in our cohort were notably poorer than those originally reported during the development of both the Mayo Genetic Risk Model and the ELN2024 classification. This discrepancy likely arises from the lower efficacy of treatment regimens lacking venetoclax in our patient cohort. Venetoclax, when combined with HMA, has demonstrated a strong synergistic effect, substantially improving response rates and survival outcomes compared to HMA monotherapy [7].

Despite the shorter median OS observed in our cohort compared to the original studies, both the Mayo Genetic Risk Model and the ELN2024 classification consistently stratified patients into prognostic groups with decreasing OS. This risk-aligned survival gradient reinforces the internal validity of both models, even in a population not treated with HMA. Their overall discriminatory ability was, however, limited, as reflected by low C-index and AUC values. The marginally better performance of the ELN2024 classification compared to the Mayo model may be attributed to its broader validation across different non-intensive regimens, but even this advantage was modest. The ELN2022 model performed poorly, with paradoxical survival patterns among favorable and intermediate-risk groups, reaffirming that older risk systems derived from cohorts of patients treated with intensive chemotherapy are unsuitable for patients receiving low-intensity treatments.

In conclusion, our findings support the clinical utility of both the Mayo Genetic Risk Model and the ELN2024 classification beyond their originally intended context of HMA and venetoclax-based therapy and suggest that these models may be preferable to the ELN2022 classification for risk stratification in patients receiving non-intensive regimens. As anticipated, survival estimates within risk categories may not be directly comparable to those derived from cohorts treated with hypomethylating agents. Further independent validation studies in bigger, well-characterized patient cohorts, stratified by specific nonintensive treatment regimens, are necessary to refine the clinical applicability of these scores and support a broader implementation in clinical practice.

M.R. and S.B. conceived and designed the study. M.R. acquired the available data and performed statistical analyses. M.R., M.M., J.W.D., and S.B. interpreted the results, drafted the manuscript, and critically revised it. All authors approved the final version for submission.

All the studies referred to in the present work were conducted in accordance with the Declaration of Helsinki.

The authors declare no conflicts of interest.