A multistage model instructed by a large dataset (knowledge bank [KB] algorithm) has recently been developed to improve outcome predictions and tailor therapeutic decisions, including hematopoietic stem cell transplantation (HSCT) in acute myeloid leukemia (AML). We assessed the performance of the KB in guiding HSCT decisions in first complete remission (CR1) in 656 AML patients younger than 60 years from the ALFA-0702 trial (NCT00932412). KB predictions of overall survival (OS) were superior to those of European LeukemiaNet (ELN) 2017 risk stratification (C-index, 68.9 vs 63.0). Among patients reaching CR1, HSCT in CR1, as a time-dependent covariate, was detrimental in those with favorable ELN 2017 risk and those with negative NPM1 minimal residual disease (MRD; interaction tests, P = .01 and P = .02, respectively). Using KB simulations of survival at 5 years in a scenario without HSCT in CR1 (KB score), we identified, in a similar time-dependent analysis, a significant interaction between KB score and HSCT, with HSCT in CR1 being detrimental only in patients with a good prognosis based on KB simulations (KB score ≥40; interaction test, P = .01). We could finally integrate ELN 2017, NPM1 MRD, and KB scores to sort 545 CR1 patients into 278 (51.0%) HSCT candidates and 267 (49.0%) chemotherapy-only candidates. In both time-dependent and 6-month landmark analyses, HSCT significantly improved OS in HSCT candidates, whereas it significantly shortened OS in chemotherapy-only candidates. Integrating KB predictions with ELN 2017 and MRD may thus represent a promising approach to optimize HSCT timing in younger AML patients.
© 2021 by The American Society of Hematology.