AUTHORS
Brittany Salter, MD, PhD, Isidora Rovic, MD2, Youn Kyung Kim, MD3, Shreyash Dalmia, MD4, Joseph Aziz, MD5, Michael Radford, MD, BSc6, Wilson Lam, MD7, Kylie Lepic, MD8, Tobias Berg, PhD9, Alejandro Garcia-Horton, MD10
AFFILIATIONS
1 McMaster University, Hamilton, Ontario, Canada
2 McMaster University, Hamilton, Ontario, Canada
3 McMaster University, Hamilton, Ontario, Canada
4 McMaster University, Hamilton, Ontario, Canada
5 McMaster University, Hamilton, Ontario, Canada
6 McMaster University, Hamilton, Ontario, Canada
7 McMaster University, Hamilton, Ontario, Canada
8 McMaster University, Hamilton, Ontario, Canada
9 McMaster University, Hamilton, Ontario, Canada
10 McMaster University, Hamilton, Ontario, Canada
ABSTRACT
Risk stratification in AML informs prognostication, treatment selection, and clinical trial design. The ELN 2022 classification was developed in intensively treated populations but inadequately captures outcomes in those treated with lower-intensity regimens, such as hypomethylating agents plus venetoclax (HMA+VEN). The ELN 2024 non-intensive therapy (NIT) schema uses 4 genes (TP53; FLT3-ITD; NRAS; KRAS) but excludes cytogenetics. The Mayo genetic risk model integrates ELN 2022 adverse karyotype and molecular findings but has not been validated in a large real-world cohort. The PRISM score incorporates 17 clinical and genomic features specific to HMA+VEN. No study has comparatively evaluated all four models in a real-world population. We sought to determine which prognostic framework best stratifies OS and whether dynamic re-stratification at response assessment improves prognostication.
This is a single-center retrospective analysis of patients with newly diagnosed AML treated with AZA-VEN in Ontario, Canada. Patients aged ≥18 years with AML ineligible for intensive chemotherapy and alloHCT diagnosed between July 2022 and December 2025 were included. Each patient was classified by ELN 2022, ELN 2024, Mayo, and PRISM models. Model discrimination was compared using Harrell’s C-index. Reclassification was evaluated with transition matrices and net reclassification improvement (NRI). Predictive performance for CR/CRi was assessed using AUC and Cochran-Armitage testing. Multivariable Cox regression incorporated all four models simultaneously. Landmark analysis assessed prognostic discrimination for Mayo and PRISM schema at diagnosis and post-CR/CRi, resetting time origin to response achievement.
One hundred AML patients were included (63% de novo, 37% secondary/therapy-related; median age 77 [IQR 73-79]). Molecular profile: 18% NPM1, 9% IDH1, 21% IDH2, 15% FLT3-ITD, and 17% TP53. Cytogenetic ELN 2022 profile: 27% adverse (N=19 complex karyotype). During cycle 1, 89% received 14-day VEN. Median number of cycles was 5.5 (1-66). CR/CRi rate was 71% (median time-to-response 29 days [13-164]). At median follow up of 25.3 months (95% CI 22-30), 2-year OS was 32%; 36% relapsed (median time-to-relapse 12 months [1.5-61]). PRISM showed the highest discriminative accuracy (C-index 0.69) for OS, followed by Mayo (0.62), ELN 2024 (0.59), and ELN 2022 (0.55). ELN 2022 did not accurately predict OS (p=0.15), while PRISM (p<0.0001), Mayo (p<0.0001), and ELN 2024 (p=0.002) stratified OS in a statistically accurate manner.
PRISM and Mayo demonstrated the widest survival discriminations: PRISM low-risk median OS 61.4 (95% CI 27-NR) vs 3.7 months (95% CI 1.9-16) for high-risk; Mayo low-risk 61.4 (95% CI 18-NR) vs 6.7 months (95% CI 3.6-16.3) for high-risk. Transition matrix showed 65% and 64% of adverse and intermediate-risk ELN 2022 patients, respectively, were reclassified as ELN 2024 favorable. Conversely, 9% and 17% of ELN 2024 favorable-risk patients were reclassified as high-risk by PRISM and Mayo, respectively. ELN 2024 adverse-risk patients were concordantly classified as high-risk by both PRISM and Mayo models (100%). PRISM and Mayo demonstrated 82% concordance in high-risk classification. NRI confirmed PRISM improved reclassification over ELN 2024 (p<0.001), while PRISM and Mayo performed comparably (p=0.44). PRISM best predicted CR/CRi (AUC 0.75; p=0.0003), outperforming ELN 2024 and Mayo (AUC 0.64; DeLong p=0.03; p=0.012). ELN 2022 failed to predict HMA-VEN response (AUC 0.37; p=0.28). On multivariable regression, PRISM was the only independent predictor of mortality (HR 3.22; 95% CI 1.65-6.28; p<0.001). Mayo and PRISM maintained significant prognostic stratification at baseline (p<0.0001) and post-CR/CRi (Mayo p=0.001; PRISM p=0.002), confirming sustained risk discrimination beyond diagnosis and value of dynamic re-stratification.
HMA-VEN-specific models (PRISM, Mayo) outperformed ELN 2024 in discrimination, reclassification, and response prediction in a real-world older adult AML population. PRISM was the sole independent predictor of mortality. Dynamic re-stratification using Mayo and PRISM models at response assessment provided prognostic refinement. These findings support adoption of treatment-specific models over traditional ELN criteria for risk-adapted decision-making. Prospective validation of these models is warranted to enhance treatment selection and consider future evaluation of dynamic de-escalation strategies.
DISCLOSURES
Brittany Salter, MD, PhD: No relevant conflicts of interest to declare.