A Pragmatic Monitoring Model for Non-Relapse Mortality Risk Based on Machine Learning in Patients Undergoing Allogeneic Hematopoietic Stem-Cell Transplantation

Background：Prognostic biomarkers play a crucial role in assessing non-relapse mortality (NRM) following allogeneic hematopoietic stem cell transplantation (allo-HSCT).The concentrations of biomarkers fluctuate with the progression of complications and corresponding clinical interventions. However current predictive models are commonly built and applied based on biomarker concentrations measured at a single fixed time point, which may limit their predictive validity and utility for clinical applications.

Methods: A prospective, multicenter study (NCT04284904) involved 621 allo-HSCT patients from 7 Chinese transplant centers. Concentrations of biomarkers (ST2, REG3α, TNFR1, IL6, and IL8) were measured at day 7/14/28/60/90 post-HSCT; patients with acute GVHD were monitored at onset and weekly during the first treatment month.The time point of relative maximum concentration for each patient (MC) was determined based on the time at which the maximum metric was reached. A pragmatic monitoring model employing the relative maximum concentrations of ST2 and TNFR1 within three months post-transplantation was developed to forecast 6-month NRM in allo-HSCT patients: log₁₀[-log₁₀(1 - p)] = 1.4798(log₁₀ST2) + 3.2733(log₁₀TNFR1) - 21.632，where p = predicted probability of 6-month NRM.

Results：The 460 patients in the modeling cohort were randomly split into a training (n=310) and a test (n=150) set, with model validation performed on an independent cohort (n=159). In both the modeling and validation cohorts, the combination of ST2 and TNFR1 exhibited the highest AUC and C-index among the models (training set: 0.85, 0.80 ; test set: 0.85, 0.80; validation set: 0.95, 0.93). Patients were categorized into three risk groups based on model-derived thresholds: low risk (LR, ≤0.0567), moderate risk (MR, 0.0567–0.1309), and high risk (HR, ≥0.1309). The HR group experienced significantly higher 6-month NRM than the MR and LR groups in both the modeling cohort (40.62% vs. 13.11% vs. 1.57%, P < 0.001) and the validation cohort (47.37% vs. 9.76% vs. 0.00%; P < 0.001). Further, our investigation revealed that our predictive model effectively gauged NRM risk independently of HLA matching, recipient age, donor type, ATG-based GVHD prophylaxis, and other clinical considerations. Multivariate analysis identified model score as significant NRM predictors (P < 0.0001). The model consistently identified high-risk NRM patients throughout post-HSCT time points. The model effectively stratifies NRM risk in patients with or without acute GVHD.

Conclusions: A pragmatic monitoring model for predicting NRM in allo-HSCT patients was established, showing potential for identifying high-risk individuals and improving personalized treatment.