A total of 757 de novo AML patients were eligible for training cohort, including 401 males and 356 females, with a median follow-up of 30 (range, 1–91) months after allo-HCT. The median age at allo-HCT was 40 (range, 14–69) years. According to the ELN2022 classification, 34% (n = 257) were favourable, 42% (n = 316) were intermediate, and 24% (n = 180) were adverse. Compared to the ELN2017 categories, 95% of favourable, 80% of intermediate, and 84% of adverse patients in the ELN2022 categories remained in their previous risk stratification. The redistribution of patients and major mutation types from the ELN2017 to ELN2022 is shown in Fig. 1B and Table S1, and the detailed baseline demographic is provided in Table S2.

According to the ELN2017, the 3-year cumulative incidence of relapse (CIR) for the favourable, intermediate, and adverse groups was 13%, 18 and 40%, respectively, with corresponding relapse-free survival (RFS) rates of 81%, 75 and 52%, and overall survival (OS) rates of 85%, 81%, and 59%, all showing statistically significant differences (P < 0.001) (Fig. 1C, E, G). In the ELN2022, the 3-year CIR were 11%, 19%, and 40% (P < 0.001); the 3-year RFS were 84%, 74%, and 52% (P < 0.001); and the 3-year OS were 88%, 79%, and 59% (P < 0.001) across the favourable, intermediate, and adverse groups, respectively (Fig. 1D, F, H). Similar results were observed when survival analysis was confined to patients who achieved complete remission (CR) at allo-HCT (Fig. S1). Furthermore, in multivariate model, the ELN2022 risk classification at diagnosis was an independent prognostic factor for CIR, RFS and OS (Table S3). Receiver operating characteristic (ROC) analysis was performed to compare the prognostic prediction power of the ELN2022 and ELN2017 risk systems in our training cohort. The C-statistics (area under the curves [AUC]) for predicting relapse (AUCELN2017 = 0.660 vs. AUCELN2022 = 0.668, P = 0.530), RFS (AUCELN2017 = 0.648 vs. AUCELN2022 = 0.658, P = 0.372), and OS (AUCELN2017 = 0.641 vs. AUCELN2022 = 0.653, P = 0.318) were not distinct between the ELN2022 and ELN2017 (Fig. 1I–K and Table S4).