An integrative prognostic model for transplant outcomes in myelofibrosis Free

Introduction

Allogeneic hematopoietic cell transplantation (allo-HCT) is the only curative option for patients with myelofibrosis, yet transplant outcomes remain heterogeneous, with survival influenced by a complex interplay of disease-related, molecular, and transplant-specific factors. Existing prognostic tools were primarily designed for pre-transplant risk stratification while transplant-specific systems lack molecular granularity and thus may not fully capture variables critical to post-transplant outcomes. Furthermore, most models were based on primarily patients without JAK inhibitor exposure. To address this gap, we conducted a large international multicenter study using up-to-date machine learning approaches to develop a novel, comprehensive prognostic model in the modern era of molecular analysis and JAK inhibition. The model aims to more accurately predict overall survival and transplant-related risk in patients with myelofibrosis undergoing allo-HCT, thereby guiding clinical decision-making and improving individualized risk assessment.

Methods

We analyzed 1,258 patients with myelofibrosis in the era of JAK inhibition undergoing allo-HCT across multiple international centers (60% primary, 40% secondary myelofibrosis). Clinical, molecular, and transplant-related variables were collected. The dataset was randomly split into a 60% training cohort and a 40% independent validation cohort. Recursive partitioning was used to identify potential hierarchies of factors such as age and leukocyte counts. To develop an integrated prognostic model for overall survival (OS) and non-relapse mortality (NRM), we applied both penalized Cox regression and ensemble survival methods. Variable selection and model tuning were performed via cross-validation in the training set. Final model performance was assessed in the validation cohort using concordance indices and calibration metrics. The best-performing model was selected based on predictive accuracy and clinical applicability.

Results

Older age (consistent cutoff identified by non-supervised recursive portioning was 60 years, HR 1.39; p<0.001) and poor performance status (HR 1.52; p<0.001) were independently associated with reduced OS and NRM. Anemia, thrombocytopenia, and both severe leukocytosis and leukopenia predicted OS and NRM. Other disease-related risk factors were shorter time to secondary myelofibrosis transformation (continuous HR 1.05; p<0.001). Looking at cytogenetics, regularization models identified complex karyotype as the only predictive factor (HR 1.51; p<0.001). On a molecular level, CALR/MPL mutations were protective, while ASXL1, TP53, U2AF1, and RAS pathway mutations conferred high-risk features. In terms of donor relations, haploidentical (HR 1.49; p<0.001), and mismatched unrelated donors (HR 1.67; p<0.001) consistently showed worse OS and NRM compared with matched related or unrelated transplants.

An ensemble survival model was trained on 700 patients and validated in an independent cohort of 558 patients. Model performance was strong (C-index: 0.74 for OS, 0.68 for NRM in training; 0.70 and 0.64, respectively, in validation), with good calibration. The model stratified patients into 4 risk groups based on predicted OS. Five-year OS and NRM rates were 90% and 7% (low risk), 71% and 20% (intermediate), 46% and 34% (high), and 25% and 52% (very high risk) (p<0.001). The 4-tiered model outperformed existing prognostic tools (disease-related: DIPSS, MIPSS70 and MIPSS70v2.0; and transplant-related: MTSS, EBMT scores) by incorporating transplant-specific and molecular risk features. This improved and integrative myelofibrosis transplant scoring system (iMTSS) supports more personalized transplant risk assessment and informed clinical decision-making.

Conclusion

We developed and validated a robust, integrative myelofibrosis transplant scoring system (iMTSS) undergoing allo-HCT, incorporating clinical, molecular, and transplant-specific variables. The model effectively stratifies patients into four distinct risk groups with significantly different survival outcomes, outperforming existing pre-transplant scoring systems. This tool enables more precise, individualized risk assessment and has the potential to guide transplant decision-making, donor selection, and post-transplant management strategies in clinical practice. A web-based calculator will be developed and presented at the meeting.