Machine Learning Identifies Gene Mutations and Variant Allele Fractions That Refine the 2022 European Leukemianet Risk Stratification for Acute Myeloid Leukemia

Introduction: Incorporation of cytogenetic and molecular alterations helps risk-stratify patients of acute myeloid leukemia (AML) patients. The 2022 European LeukemiaNet (ELN) recommendations incorporate cytogenetics and next-generation sequencing (NGS) derived molecular data for risk stratification. Using a machine-learning (ML) approach, we identified heterogeneous outcomes in the favorable and intermediate ELN22 groups by incorporating gene mutations and variant allele fractions (VAF).

Methods: A total of 508 patients of adult AML (>18 years) were accrued over 10 years (2012 – 2023). Patients were uniformly treated with "3 + 7" induction and 3 doses of HiDAC. Diagnostic samples were sequenced using a 50-gene myeloid panel (till 2020) based on single molecule molecular inversion probes and subsequently using a 135-gene hybrid capture-based panel. Based on cytogenetics and genomic information, cases were risk-stratified per ELN22 recommendations. In each ELN category, we incorporated commonly occurring (>2%) mutations and their variant allele fractions (VAF) as features. Features in each ELN risk were selected using recursive feature elimination. Multicollinearity was assessed using variance inflation factor. A logistic regression-based model with k-fold cross-validation was developed for each ELN category to predict overall survival (OS). Top features in each ELN category were used to derive a cumulative score based on which subcategories were created. Results of the subcategories were analyzed for their impact on OS and RFS using log rank test.

Results: Median age of the cohort was 35.0 years (M:F, 1.6:1) with median follow-up of 29.5 months. The median OS was 80 months (95% CI 34.8-87.3) and median RFS was 52.3 months (95% CI 35.2–87.9) months. Based on ELN22 recommendations, patients were classified as favorable (n=275), intermediate (n=157), or adverse risk (n=75). Patients classified as ELN22-adverse had inferior OS [(HR 3.2; 95% CI 2.0 to 5.3;(p <0.0001)] and RFS [(HR 2.7; 95% CI 1.3 to 5.4;(p=0.0004)] as compared to ELN-22 favorable risk. Similarly, patients classified as ELN22-intermediate risk had inferior OS [(HR 1.8; 95% CI 1.1 to 3.0] and RFS [(HR 1.4; 95% CI 0.9 to 2.0] as compared to ELN22 favorable risk. Based on supervised ML in the ELN-favorable group, high NPM1 (≥37.8%), RAD21 (≥14.74), WT1 VAF (≥42.87), presence of ETV6 and KIT exon17 mutations were each assigned a negative point whereas NRAS mutations were assigned a positive point. For ELN-intermediate risk, high FLT3-ITD VAF (≥18.57%) and WT1 mutation were each assigned a negative point. No significant feature could be identified in ELN22 adverse group. In each ELN22 risk, the presence of a cumulative negative score was assigned to a subgroup (ELN22 Favorable II (n=81), ELN22 Intermediate II(n=43)) and those with positive scores to another subgroup (ELN22 Favorable I (n=194), ELN22 Intermediate I(n=114)). The median OS of ELN22 Favorable II [42.2 months, (95% CI 22.1 to 87.3 months) HR: 0.96 (95% CI 0.6-1.53)] was similar to ELN22 Intermediate I risk group [34.8 months, (95% CI 23.8 to 86.0 months); HR: 1.04 (95% CI 0.7-1.7)] but greatly different from ELN Favorable I group [median not reached, HR: 2.01 (95% CI 1.3-3.1)]. Similarly, the median OS of ELN22 Intermediate II [15.9 months, (95% CI 12.8 to 70.9 months) HR: 0.72 (95% CI 0.4-1.4)] was similar to ELN22 adverse risk group [16.6 months, (95% CI 12.4 to 20.7 months); HR: 1.4 (95% CI 0.7-2.7)]. The median RFS of ELN22 Favorable II [42.4 months, (95% CI 23.6 to 71.8 months) HR: 1.04 (95% CI 0.6-1.8)] was similar to ELN22 Intermediate I risk group [38.4 months, (95% CI 23.4 to 52.3 months); HR: 0.96 (95% CI 0.6-1.6)]. However, the median RFS of ELN22 Intermediate II [35.2 months, (95% CI 11.9 to 87.9 months) HR: 0.65 (95% CI 0.2-1.8)] was superior to ELN22 adverse risk group [21.1 months, (95% CI 10.0 to 63.8 months); HR: 1.5 (95% CI 0.6-4.2)]. This classification proposal was independently validated on the BEAT-AML2 dataset (Bottomly et al, 2022), where the ELN adverse group had different survival rates when compared to ELN favorable I and II risk groups (HR 2.3; 95%CI 1.66-5.0 & HR 4.8; 95%CI 3.5 to 6.5 respectively). Such observations were not made for the ELN intermediate risk group in the BEAT-AML2 cohort.

Conclusion: We demonstrate the applicability of ELN2022 risk stratification for AML and suggest improvements based on the incorporation of additional genes and mutational VAF.