Construction and evaluation of a prognostic model for patients with Acute Myeloid Leukemia

Abstract

Abstract Background： Acute myeloid leukemia (AML) is a group of aggressive, heterogeneous, malignant diseases and is the most common type of acute leukemia in adults. The European LeukemiaNet (ELN) genetic risk stratification, which is widely used in clinical practice, categorizes patients with AML into three groups based on fusion genes, genetic mutations, and cytogenetic abnormalities, while excluding other characteristics, such as age at onset, sensitivity to chemotherapy, and relevant biochemical indicators. The aim of this study was to identify the factors affecting the survival of patients with AML and to develop and validate a prediction model. Study Design and Methods: Clinical data of 603 patients with newly diagnosed AML treated at The First Affiliated Hospital, Zhejiang University School of Medicine, from January 1, 2019, to April 30, 2023, were collected as a training set, and included 158 patients from East China Leukemia Alliance (ECLA) as a validation set.Treatment strategies were based on factors such as ELN risk stratification, mutations, age, performance status, and comorbidity. OS was defined as the time from AML diagnosis to death or the last follow-up. Results: The median follow-up durations for the training and validation sets were 24.8 (95%CI 23.1–26.5) and 50.9 (95%CI 37.1–64.8) months, respectively. The 1-, 2-, and 3-year OS rates in the training and validation sets were 79.9%, 61.9%, and 51.9%, 89.1%, 74.2%, and 60.0%, respectively.In the training set,survival-related variables including age at onset, lactate dehydrogenase (LDH) level at initial diagnosis, CBFB::MYH11 gene fusion, KMT2A rearrangement, TP53 mutation (variant allele frequency [VAF] >10%), DNMT3A mutation, RUNX1 mutation, and achievement of CRc in the first course of treatment were independent predictors that were used to construct the ZJ-AML model. The prognostic model demonstrated excellent discriminative ability with the Harrell's concordance index of 0.748, 1- ,2-and 3-year area under the receiver operating characteristic curve(AUROC) of 0.802, 0.753, and 0.691,respectively. The calibration curve of the training set showed good agreement between the model prediction results and the actual observation results in terms of 1-, 2-, and 3-year survival probabilities We assigned values to the variables according to the coefficients of the variables in multivariate analysis in the training set, which were summed and served as the risk score. The risk score was calculated as: 0.033677 × age (years) + 0.000493 × LDH (U/L) + 0.553400 × RUNX1 mutation + 0.820775 × KMT2A rearrangement + 1.256855 × TP53 mutation + 0.311104 × DNMT3A mutation − 0.635064 × CBFB::MYH11 gene fusion − 0.930328 × achievement of CRc in the first course of treatment. A nomogram and an online calculator were generated. Based on the risk score, the patients in the training set were equally divided into low- (<1.169), intermediate- (1.169–1.986), and high-risk (>1.986) groups,with significantly distinct prognosis, and the model successfully identified candidates for haematopoietic stem cell transplantation.The ZJ-AML model outperformed both the ELN 2022 (C-index: 0.793[95% CI: 0.698–0.864] vs. 0.646 [95% CI: 0.542–0.738], P = 0.025) and the ELN 2024 (C-index: 0.793[95% CI: 0.698–0.864] vs. 0.614 [95% CI: 0.495–0.720], P = 0.012) classification with a higher C-index. The model was validated in the validation set,with the Harrell's concordance index of 0.685, 1- ,2-and 3-year AUROC 0.802, 0.753, and 0.691.According to the risk score-based grouping principle, in this study population, 48 patients were at low risk, 63 patients were at intermediate risk, and 47 patients were at high risk. The survival rates of the three groups were statistically significant. Analysis of subgroups stratified by induction regimen demonstrated that the ZJ-AML model showed comparable discriminative capacity to the ELN 2022 classification in the intensive chemotherapy cohort and superior prognostic performance in patients treated with B-cell lymphoma 2 (BCL-2) inhibitor-based regimens compared with the revised ELN 2024 classification. Conclusion: In conclusion, we developed the ZJ-AML model, which demonstrated excellent discrimination and calibration in both the training and validation sets. Notably, our model exhibited superior discrimination compared to the ELN 2022 and ELN 2024 classifications, particularly for patients with AML undergoing combination therapy with a targeted BCL-2 inhibitor.