Defining the Limits of Pre-Transplant Risk Prediction in AML: Evidence From Machine Learning and Regression Models

Allogeneic hematopoietic cell transplantation (allo-HCT) for acute myeloid leukemia (AML) is associated with considerable morbidity and mortality. Machine learning (ML) techniques are increasingly applied to predict outcomes in medicine.

To evaluate the role of ML in predicting overall survival (OS) after allo-HCT for AML and compare ML with traditional Cox regression.

Using an internal cohort of 2253 patients and 14 pre-allo-HCT variables, we developed three models: Cox regression with time-varying coefficients (Cox-TVC), Elastic-net Cox Regression (Cox-EN), and Random Survival Forest (RSF). Performance was evaluated using multiple metrics including C-index, net reclassification improvement (NRI) and decision curve analysis (DCA). Patients were stratified into tertiles of model predicted 24-month mortality. External validation was performed in 252 single-center patients with uniform measurable residual disease (MRD) assessment.

Model-derived risk scores strongly correlated (r = 0.886 to 0.963). Across models, age ≥ 60 years, MRD positivity, and adapted European LeukemiaNet (aELN) adverse risk were the strongest predictors. Effects of age attenuated over time (HR for age ≥60: 2.59 [95% CI: 1.54 to 4.35] at 1 year, 1.92 [95% CI: 1.04 to 3.56] at 5 years). Compared with Hematopoietic Cell Transplant Comorbidity-Index (HCT-CI) and aELN, models improved risk stratification (NRI: 31% to 44%, p < .001). Discrimination remained modest but was higher in external cohort compared with internal cohort (0.69 to 0.71 versus 0.60 to 0.61), likely reflecting uniform MRD assessment. At a 25%, risk threshold for clinical decision-making, models identified approximately 1 additional high-risk patient per 100 versus HCT-CI, or aELN. At 2-years, 25% to 27% of patients categorized as low-risk had died, while 45% to 48% categorized as high-risk were alive.

ML approaches improved risk stratification over HCT-CI and aELN but performed comparably with Cox model. Individual outcome prediction using static pre-transplant models remained modest. Progress will require MRD standardization, richer data, and dynamic peri- and post-transplant modeling.