Development of a dynamic counterfactual risk stratification strategy for newly diagnosed acute myeloid leukemia patients treated with venetoclax and azacitidine

Objective The objective of this study was to develop a flexible risk model (RM) stratification strategy for Acute Myeloid Leukemia (AML) that is specific for the new standard of care venetoclax plus azacitidine (ven/aza), captures disease heterogeneity, and addresses a range of real-world data issues. Materials and Methods A series of tunable RMs based on a dynamic counterfactual machine learning (ML) strategy that utilized next generation sequencing, cytogenetics, flow cytometry, and other features of the diagnostic AML samples were developed and tested on a single institutional cohort of 316 newly diagnosed patients treated initially with ven/aza. Results Favorable, Intermediate, and Adverse risk groups were identified in a series of novel RMs derived using ML models for overall survival (OS) and event free survival (EFS). Most, but not all models, demonstrated equitable patient distribution into the different risk categories (~20%-40% in each group) with significant separation between categories (Log-Rank based p-values <0.001), and with predictability computed by survival AUC values in the ~0.60-0.70 range. Discussion and Conclusion The general strategy employed here is specific for AML patients treated with ven/aza, is based on a wide range of diagnostic AML pathology features, considers feature interactions, addresses data missingness, sparsity, and the confounding effects of allogeneic hematopoietic cell transplant . It is also readily tunable through simple coding and context specific parameter updates, and adaptable to reflect different use case needs. This strategy represents a new approach to developing more effective RMs for AML and possibly other diseases as well.