A novel random survival forest model for adult NPM1-mutated Acute Myeloid Leukemia patients Free

Background:NPM1-mutated acute myeloid leukemia (AML) is the most common molecular subtype in adult AML, comprising 30-35% of cases. While the European LeukemiaNet (ELN) 2022 guidelines primarily assign “favorable” risk toNPM1-mutated AML patients lacking adverse cytogenetics and FLT3-ITDmutation, our previous research indicated that additional high frequency co-mutations — particularly DNMT3A or TET2mutations — abolish this survival advantage, whereas IDH1/2 or PTPN11 mutations suggested relatively better outcome. Thus, a stratification system based primarily on FLT3-ITD status is inadequate to capture the complex interactions of co-mutations, leading to imprecise stratification. Moreover, AML's molecular and clinical heterogeneity demands development of subtype-specific tools. Traditional prognostic models like Cox proportional hazards models (Cox models) have limited capacity to integrate complex biomarkers, whereas machine learning algorithms like random survival forest (RSF) excel in handling nonlinear correlations and complex feature interactions, potentially providing more precise risk stratification. To our knowledge, no studies have yet developed molecular-profile-driven RSF models for AML patients. In this study, we integrated co-mutation profiles in adult NPM1-mutated AML patients to develop a novel RSF model, aiming to overcome limitations of traditional risk stratification and provide more accurate risk assessment.

Methods: The study cohort comprised 349 newly diagnosed adult NPM1-mutated AML patients (acute promyelocytic leukemia excluded) from October 2018 to December 2024 in our center. Molecular aberrations were assessed by next-generation sequencing. The permutation-based feature importance evaluation was employed to rank mutation genes and chromosomal features. Further iterative variable selection based on importance ranking identified the optimal core feature set. To rigorously assess model performance and robustness, we conducted 20 independent trials, each randomly splitting the data into an 80% training set and a 20% internal validation set. We evaluated model discrimination and accuracy by C-index and AUC in both the average metrics across all 20 runs and the results of the single optimal-performing model. Based on the optimal trial, an improved RSF model was then constructed to generate individualized risk scores, which were further stratified into low-, intermediate-, and high-risk (LR, IR and HR) groups using a survivor-guided algorithm that leverages event-time distribution to define optimal cutoffs. Additionally, a Cox model was constructed for performance comparison. External validation is ongoing to assess clinical applicability and generalizability of the RSF model.

Results: The median age of 349 eligible patients was 60 (IQR 50-69) years. The median follow-up was 27.8 (95%CI, 24.2-31.3) months with the median OS not reached. Using the permutation-based feature importance evaluation and recursive modeling, we identified the optimal core feature combination: FLT3-ITD, TET2, DNMT3A, IDH2, PTPN11, FLT3-TKD, ASXL1, IDH1, SRSF2, RAD21, WT1, ZRSR2mutation, NPM1 mutation type, and chromosome character. Across 20 times of trials, the RSF model outperformed the Cox model in predictive performance, achieving higher average C-index (0.692 vs. 0.672) and better average dynamic AUC (0.707 vs. 0.680) in internal validation set. Using the optimal trial, the RSF model significantly stratified patients into LR (score ≤ 18.80), IR (18.80 < score ≤ 32.19), and HR (score > 32.19) groups, demonstrating improved discrimination between LR and IR compared with the traditional Cox model (P < 0.01 vs. P = 0.08), while both maintained good separation between IR and HR (P < 0.01 for each).

Summary/Conclusion: This study developed a novel RSF-based prognostic model to stratify risk in adult NPM1-mutated AML patients using molecular profiles. Preliminary results indicate its superior performance over traditional stratification models, with external validation ongoing.