The research, led by Minglu Li and colleagues at the Department of Hematology of Aerospace Center Hospital in Beijing, set out to build what statisticians call a nomogram — a visual scoring instrument that converts a handful of patient characteristics into a personalized probability estimate. What makes this nomogram distinctive is its statistical engine: rather than treating death from any cause as a single uniform outcome, the team used a competing-risk framework that explicitly accounts for the fact that a patient who relapses and dies of leukemia is no longer at risk of dying from transplant complications. In survival analysis, ignoring such competing events can badly distort risk estimates, inflating apparent mortality rates and muddying clinical decision-making. The Fine-Gray model, the framework the researchers adopted, was developed precisely to handle this subtlety, modeling the cumulative incidence of the event of interest while treating relapse as the competing event.

The study population consisted of 587 patients with high-risk or refractory acute leukemia who underwent their first allogeneic transplant between January 2015 and December 2021 at the single center. To test the model honestly, the team randomly split the cohort into a training set of 411 patients, used to build the model, and a validation set of 176 patients, held back to see how the model performed on data it had never seen. This kind of internal validation is a critical safeguard against overfitting — the statistical sin of building a model that memorizes the quirks of one dataset rather than capturing genuine biological patterns that generalize to new patients.

Selecting which variables to include was itself a methodical process. The researchers turned to least absolute shrinkage and selection operator regression, better known as LASSO, a technique that shrinks the coefficients of weak predictors toward zero and effectively eliminates them, leaving behind only the most informative variables. To ensure the selection was stable rather than a fluke of the particular data, they employed ten-fold cross-validation, repeatedly partitioning the training data so that the variable selection process was tested across many different subsets. Out of this rigorous filtering emerged six predictors of non-relapse mortality: patient age, disease status before transplantation (whether the leukemia was in complete remission or not), the Hematopoietic Cell Transplantation-specific Comorbidity Index score, the time to white blood cell engraftment, the time to platelet engraftment, and the grade of acute graft-versus-host disease.

Each of these predictors tells a clinically meaningful story. Age and comorbidities reflect the body’s baseline resilience going into the procedure — an older patient with significant organ dysfunction simply has less physiological reserve to withstand the assault of conditioning chemotherapy, immune suppression, and potential infections. Pre-transplant disease status matters because patients transplanted in complete remission generally fare better than those transplanted with active, refractory disease. The engraftment timings are particularly interesting because they are early post-transplant events: if the donor’s stem cells take longer than 21 days to produce white blood cells, or longer than 14 days to produce platelets, that sluggish recovery signals a fragile marrow environment associated with higher risk. And acute graft-versus-host disease — the notorious complication in which donor immune cells attack the recipient’s tissues — remains one of the most feared drivers of transplant-related death, so its grade from II to IV carries substantial weight in the model.

The performance numbers are where the tool earns its keep. In the training cohort, the time-dependent area under the receiver operating characteristic curve — a standard measure of how well a model separates those who experience the event from those who do not — reached 0.786 at one year, 0.764 at two years, and 0.746 at three years. In the validation cohort, the model actually performed slightly better, with AUCs of 0.818, 0.799, and 0.806 at the same time points. Values in this range indicate useful, though not perfect, discrimination. The concordance index, a related measure of predictive accuracy for survival data, came in at 0.715 after bootstrap correction in the training set and 0.767 in the validation set, where the 95 percent confidence interval spanned 0.703 to 0.831.

Discrimination alone is not enough, however; a model must also be calibrated, meaning its predicted probabilities should match observed reality. A model that tells every patient they have a 30 percent risk when only 10 percent actually die is discriminating poorly or calibrating poorly, and clinicians acting on such numbers could make harmful decisions. The team assessed calibration using optimism-corrected calibration slopes derived from 200 bootstrap resamples, obtaining values of 0.94, 0.92, and 0.97 at one, two, and three years in the training cohort — figures close to the ideal value of 1.0. In the validation cohort, the slopes of 1.30, 1.14, and 1.16 suggested the model was, if anything, slightly conservative in that patient group, a direction of error that is generally less dangerous than overconfidence.

Perhaps the most practically important result came from the head-to-head comparison with the HCT-CI, the comorbidity index that has long served as a standard pre-transplant risk assessment tool. The new nomogram significantly outperformed it, with a C-index difference of 0.114 (P = 0.020) in the training cohort and 0.191 (P = 0.008) in the validation cohort. The advantage makes intuitive sense: the HCT-CI captures only the patient’s condition before the transplant, whereas the nomogram folds in early post-transplant events like engraftment speed and graft-versus-host disease, which carry fresh, dynamic information about how the transplant is actually unfolding. Decision curve analysis reinforced the point, showing that across clinically relevant threshold probabilities from 5 to 50 percent, the nomogram delivered positive net benefit and consistently beat both the treat-all and treat-none default strategies — meaning that clinicians using the tool to guide decisions would, on average, make better calls than following blanket policies.

The study’s limitations are worth keeping in view. It was a single-center, retrospective analysis, and all 587 patients were treated at one institution in Beijing, so the model’s generalizability to other populations, donor types, and conditioning regimens remains to be demonstrated. External validation on independent cohorts at different centers is the natural next step, and the authors themselves frame the tool as internally validated rather than externally proven. Still, the work represents a meaningful advance in a field where risk prediction has often relied on static, pre-transplant snapshots. By integrating baseline characteristics with the early dynamics of recovery, the nomogram gives clinicians a living risk estimate that can be computed in the crucial weeks after transplantation — precisely the window when interventions such as intensified monitoring, prophylactic treatments, or early escalation of care might tip the balance for the patients who need it most. For the growing population of high-risk and refractory leukemia patients whose only curative option is an allogeneic transplant, turning the black box of non-relapse mortality into a quantifiable, actionable number could ultimately mean the difference between anticipation and reaction.

Subject of Research: A competing-risk nomogram for predicting non-relapse mortality after allogeneic hematopoietic stem cell transplantation in high-risk or refractory acute leukemia

Article Title: Development and internal validation of a competing-risk nomogram for predicting non-relapse mortality after allogeneic hematopoietic stem cell transplantation in patients with high-risk or refractory acute leukemia

Article References: Li, M., Zhang, W., Zhang, S., Fei, X., Zhao, J., Luo, R., & Wang, J. (2026). Development and internal validation of a competing-risk nomogram for predicting non-relapse mortality after allogeneic hematopoietic stem cell transplantation in patients with high-risk or refractory acute leukemia. Annals of Hematology. https://doi.org/10.1007/s00277-026-07294-5

Image Credits: AI Generated

DOI: 10.1007/s00277-026-07294-5

Keywords: non-relapse mortality, allogeneic hematopoietic stem cell transplantation, acute leukemia, nomogram, competing-risk model, Fine-Gray model, LASSO regression, HCT-CI, graft-versus-host disease, engraftment, risk prediction, hematology