Session
31
Oral presentations Clinical hemato-oncology
Nov. 20, 2024,
3:45 p.m. - 5:15 p.m.,
Shanghai 1-2
Abstract
3
Challenging MDS and AML separation with covariate-aware unsupervised learning: introducing the MDS-AML Aggregative Risk Classification System (MARCS)
M. Roncador1, 2, F. Bayer2, J. W. Deuel1, J. Kuipers2, K. Takahashi3, M. G. Manz1, S. Böttcher1, N. Beerenwinkel2, S. Balabanov1, Presenter: M. Roncador1 (1Zurich, 2Basel, 3Houston)
Objective
Myelodysplastic syndromes (MDS) and acute myeloid leukemia (AML) share clinical and genetic features. As understanding of genomic drivers increases, molecular criteria are partially replacing morphology-based definitions in MDS and AML classifications. Increasing evidence suggests that at least in some cases MDS and AML may represent a continuum of disease evolution, rather than distinct entities.
Methods
We analyzed 7,480 patients diagnosed with MDS (n=3,729) and AML (n=3,751) between 2008 and 2019, collecting data on blood parameters, mutations in 32 genes, and cytogenetic abnormalities. Using CANClust, a covariate-aware clustering method, we developed the MDS-AML Aggregative Risk Classification System (MARCS). This new system was validated on an independent cohort of 1,035 patients (MDS: n=489, AML: n=546), confirming the results.
Results
MARCS categorized patients into nine risk groups, accurately reflecting the genetic understanding of MDS and AML. It captured distinct groups like NPM1-mutated AML and identified an ultra-high-risk category with TP53 mutations. The system also aligned with favorable prognostic markers, such as SF3B1 mutations and del(5q). MARCS outperformed the 2022 European LeukemiaNet (ELN2022) and the Molecular International Prognostic Scoring System (IPSS-M) in predicting disease progression and outcomes, even when both scores were combined (likelihood ratio [LR] = 501.8; Wilks test p-value < 10⁻⁴, n = 7,480, df = 19). MARCS was particularly effective for patients in the new International Consensus Classification (ICC) MDS/AML subclass, outperforming IPSS-M (log-rank p-value < 0.0001). When applying separate clustering processes for MDS and AML, predictive accuracy was inferior compared to the combined analysis (LR = 444.5 vs 502.6; Wilks test p-value < 10⁻⁴, df = 11 vs 10).
Conclusion
In conclusion, MARCS integrates genomic mutations and clinical data to predict outcomes in MDS and AML more accurately than current systems, effectively dissecting the MDS-AML continuum. It is particularly valuable in identifying high-risk patients and in guiding treatment decisions for aggressive disease. The classification is publicly accessible at https://marcs.ethz.ch.