BIOPREVENT: Machine-learning algorithm using post-transplant biomarkers to predict cGvHD and NRM

Results from an analysis of the BIOPREVENT machine learning (ML) algorithm, predicting chronic graft-versus-host disease (cGvHD) and non-relapse mortality (NRM) after allogeneic hematopoietic stem cell transplantation (allo-HSCT), were published in The Journal of Clinical Investigation by Martens et al. Predictive risk models were developed by analyzing seven previously validated plasma biomarkers at Day 90/100 post-allo-HSCT and nine clinical variables in 1,310 allo-HSCT recipients. Predictive accuracy of ML models was evaluated using the time-varying area under the receiver operating characteristic curves (AUCt) on Days 180–540 after allo-HSCT.

Key data: Across evaluated ML methods, models incorporating biomarkers outperformed clinical-only models for predicting cGvHD, with Bayesian additive regression trees (BART) and Cox regression with extreme gradient boosting (CoxXGBoost) achieving AUCt >0.65 at 1 year. For NRM, biomarker-inclusive models achieved AUCt of 0.76–0.91 across early timepoints. BART consistently demonstrated high predictive performance and was selected for the final BIOPREVENT model. In variable importance analysis, matrix metalloproteinase‑3 (MMP3) and C-X-C motif chemokine ligand 9 (CXCL9) were among the top predictors for cGvHD, while interleukin‑1 receptor-like 1 (IL1RL1) and soluble CD163 (sCD163) best predicted NRM.

Key learning: The BIOPREVENT ML algorithm accurately predicts individual risk of future cGvHD and NRM using biomarkers at 3 months post-allo-HSCT, supporting further research to validate and incorporate ML-based risk stratification into clinical practice.