Prognostic risk integration for survival modeling (PRISM) in newly diagnosed acute myeloid leukemia treated with venetoclax:  a multinational retrospective cohort study

Abstract

Abstract Background: The 4-gene molecular prognostic risk score (mPRS) improves overall survival (OS) prediction in patients (pts) with newly diagnosed AML (ND-AML) treated with hypomethylating agents (HMA) and venetoclax (VEN). Yet given the broad clinical and molecular heterogeneity of AML, risk stratification remains suboptimal. We consolidated a large international cohort of ND-AML pts treated with lower-intensity regimens including HMA or low-dose cytarabine (LDAC) plus VEN to improve OS prognostication. Methods: We analyzed 2,273 ND-AML adult pts who received frontline VEN with HMA (n=2,222) or LDAC (n=51) at an academic medical center in the US, UK, France, Germany, or Italy who had complete clinical, cytogenetic, and molecular data. An initial set of 1,974 patients were split into a training (TC, N=1,339) and internal validation (VC, N=635) cohorts, stratifying on key clinical variables (de novo vs. secondary AML, TP53 and IDH2 mutations, hematopoietic cell transplantation [HCT], death). An additional N=299 patients, acquired later, were utilized as an external validation set. OS was measured from therapy start, estimated with Kaplan-Meier curves, and compared across risk groups using the log-rank test. HCT was considered a time-dependent covariate. OS discrimination was estimated by Harrell's C-index (C). Clinical and genomic features (with a frequency ≥ 4% required for nominal features) were included in model development. Cox L1-penalized regression of OS applied to 1,000 bootstrap samples from the TC was used to identify stable features (selected with a consistent direction in ≥75% of bootstrap iterations). These features were entered into a robust Cox model. Non-significant genes “ResMut” were grouped according to their directional effect on OS to create “ResMut-fav” denoting the presence of ≥1 mutation in a favorable set of residual genes. The final robust Cox model beta coefficients were added to calculate patient-specific risk PRISM scores. Quartiles of PRISM scores defined four clinical PRISM risk groups: low (L), moderate (M), high (H), and very-high (VH). Results: In the TC, median pt age was 74 years; 61% were male. 41% had secondary AML (sAML), arising from an antecedent hematologic disorder (sAML-AHD) in 33% (11% with treated AHD) and from prior therapy in 8%. The mPRS distribution was 56% higher-, 22% intermediate-, and 22% lower-benefit. HCT rate was 12%. The final PRISM model included age, sex, sAML-AHD, ELN 2022 complex karyotype and other adverse-risk cytogenetic abnormalities, and adverse mutations in KRAS, PTPN11, FLT3-ITD, JAK2, ASXL1, and TP53. Favorable features included diploid karyotype, mutations in RUNX1 or IDH2, and ResMut-fav (comprising CEBPA, BCOR, IDH1, SF3B1). For TC pts, at median follow-up of 20.8 months (mo), median OS (mOS) was 12.2 mo (95% CI: 10.9–13.6). Higher PRISM scores were associated with worse OS (1-unit increase HR: 2.50; 95% CI: 2.21–2.82; C: 0.654). Median OS by PRISM risk group L/M/H/VH was 29.6, 17.6, 11.3, and 5.6 mo (C: 0.641; p<0.001). PRISM risk groups improved OS discrimination vs. mPRS (C: 0.641 vs. 0.583, p<0.001), with similar benefit after adjusting for HCT (C: 0.650 vs. 0.594, p<0.001). There were no significant differences in considered variables between the TC and VC. In the VC, median follow-up was 24.4 mo and mOS 13.0 mo (95% CI: 11.5–14.8). The PRISM score was significantly associated with OS (1-unit increase HR: 2.28; 95% CI: 1.91–2.71; C: 0.653). Median OS by PRISM group L/M/H/VH was 24.4, 15.0, 11.7, and 5.8 mo, similar to rates observed in the TC. PRISM risk groups outperformed mPRS (C: 0.648 vs. 0.613, p=0.002), including after adjusting for HCT (C: 0.659 vs. 0.629, p=0.018). The external validation cohort was marginally older (median age 75 years) with fewer HCT recipients (8%) compared to the initial TC/VC. Higher PRISM scores remained highly associated with OS (1-unit increase HR: 2.49, 95% CI: 1.92-3.24; C: 0.661). Median OS by PRISM group L/M/H/VH was 28.8, 13.9, 12.5, and 6.3 mo. PRISM risk groups had better OS discrimination than mPRS (C: 0.648 vs. 0.584, p=0.010), including after adjusting for HCT (C:0.662 vs. 0.607, p=0.004). Conclusions: Integration of prognostically relevant clinical and genomic features into the PRISM score and subsequent PRISM risk groups enables improved discrimination of survival following lower-intensity HMA/LDAC+VEN therapy in patients with ND-AML compared to current risk classification frameworks.