Age-Stratified Game-Theory-Informed Machine Learning of Molecular Alterations Unveils Prognostic Divergence in 3062 Pediatric and Adult Acute Myeloid Leukemia Patients

Age plays a major role in both AML biology and patient management. Current treatment algorithms function as decision trees, assuming that once a genetic variable is identified, its impact on personalized risk is context-insensitive. This implies that an alteration has a defined effect irrespective of patient context. Given the simplicity of this model, we hypothesized that incorporating patient age with genetic alterations could offer additional insights into risk stratification. We pooled adult patient data including cytogenetics, molecular genetics (next-generation sequencing of over 50 genes commonly altered in myeloid neoplasms), and outcome from previous trials of the German Study Alliance Leukemia with pediatric patients from Fred Hutchinson Cancer Center and publicly available data. Patients were assigned to six age groups: infants (0-2 years), children (3-14 years), adolescents and young adults (AYA, 15-39 years), adults (40-64 years), seniors (65-74 years), and elderly (75+ years). We trained three machine learning (ML) models, Random Forest (RF), XGboost (XGB), and Logistic Regression (LR), to predict complete remissions (CR) after intensive chemotherapy and 2-year overall survival (OS). To explain model decisions, Shapley (SHAP) values were used that are derived from game theory and quantify impact on decision-making. We found the frequency of alterations affecting signaling pathways to be highest in infants and children and decline with age while conversely alterations affecting epigenetic regulators or the spliceosome as well as TP53 alterations increased with age. The lowest median for age was found for KIT alterations at 18 years (IQR: 11-51) while alterations of SRSF2 had the highest median age at 68 years (IQR: 59-74). For CR prediction, RF outperformed other models with an average area-under-the-receiver-operating-characteristic (AUROC) of 0.801, while XGB had the highest average AUROC at 0.791 for 2-year OS prediction over 100 runs each. Automatically selected features associated with favorable CR predictions were alterations of NPM1, CEBPA, FLT3-ITD, KIT, GATA2, inv(16)/t(16;16) and t(8;21) while alterations of TP53, RUNX1, ASXL1, SRSF2, U2AF1, TET2, PHF6, and SF3B1, as well as -7, del(5q), -17, and trisomy 8 were found to predict treatment failure. For 2-year OS, t(8;21), inv(16)/t(16;16), and -Y as well as alterations of NPM1, CEBPA, IDH2, GATA2, and STAG2 were prognostic for favorable outcomes, while -7, del(5q) and alterations of TP53, DNMT3A, NRAS, WT1, KRAS, ASXL1, U2AF1, and FLT3-ITD predicted unfavorable outcomes. Quantification of feature impact on model decisions via SHAP values showed an age-dependent pattern where feature impact varied substantially depending on which age group was affected by which alteration. For instance, NPM1 alterations had the largest impact on model decisions amongst genetic variables reflected by high mean SHAP values. However, we found a schism in model impact, where for CR and 2-year-OS 28.8% and 29.9% of the impact on model decisions came from patients in the adult group, while only a minority of the impact could be traced back to patients at both ends of the age spectrum. We found this pattern to repeat across all genetic alterations and age itself and confirmed it in univariable analyses of model-selected genetic features as age substantially modified prognostic effect sizes of each alteration. Interestingly, for some alterations such as TP53, RUNX1, ASXL1 as well as -7, -17, and del(5q) age-modified effect sizes were disproportionally disassociated from an alteration's age distribution, i.e. younger patients with gene alterations that frequently affect older patients exhibited higher risk disease. Automatic feature selection for highly accurate ML models in an explainable manner underscores that genetic risk stratifications are context-sensitive and dependent on patient age. Hence, current risk stratification models underestimate the complexity of disease biology as an alteration does not necessarily constitute a prognostic impact merely by its absence or presence. This may reflect patterns related to aging being associated with higher risk disease especially if they occur in younger patients warranting closer clinical monitoring.