JCO Clinical Cancer Informatics (JCO CCI)Development of a Dynamic Counterfactual Risk Stratification Strategy for Newly Diagnosed Patients with AML Treated with Venetoclax and Azacitidine

ELN 2017 and 2022 risk classifications were developed based on intensive chemotherapy outcomes and perform suboptimally for patients receiving venetoclax plus azacitidine. This machine learning study developed and validated treatment-specific risk models for newly diagnosed AML patients receiving ven-aza, using genetic and phenotypic features across a 316-patient academic cohort and an external real-world validation cohort of 971 patients at 87 community care sites.

Clinical Considerations

• ML-based models consistently identified three distinct risk strata — favorable, intermediate, and adverse — with highly significant survival separation (log-rank P<.001) across multiple modeling assumptions, feature sets, and both OS and event-free survival endpoints
• ML models outperformed ELN22 in the external real-world community cohort on OS prediction (median cAUC 0.62 vs 0.52), supporting generalizability beyond academic center settings where ELN22 was originally validated
• Allo-HCT confounding remains a key limitation — models reliably identified adverse risk transplant recipients but struggled to separate favorable and intermediate risk groups within the transplant-eligible population
• Binary simplification of genetic and flow cytometry features, and equivalence assumptions across composite mutation categories, may underestimate biologic complexity and limit individual-level precision

Practice Applications

• Familiarize with treatment-specific risk stratification tools as ven-aza becomes the dominant frontline regimen for older adults with AML ineligible for intensive chemotherapy
• Evaluate whether institutional risk stratification still relies primarily on ELN criteria calibrated to intensive chemotherapy populations rather than ven-aza outcomes
• Consider that community-based real-world validation strengthens the case for eventual clinical implementation beyond academic centers
• Monitor for prospective validation studies and emerging ML-based tools before applying this framework to individual treatment decisions

More in Prognostic Models