Predicting Patient Death after Allogeneic Stem Cell Transplantation for Inborn Errors Using Machine Learning (PREPAD): A European Society for Blood and Marrow Transplantation Inborn Errors Working Party Study

Abstract

Allogeneic hematopoietic stem cell transplantation (HSCT) is a curative treatment for many inborn errors of immunity, metabolism, and hematopoiesis. No predictive models are available for these disorders. We created a machine learning model using XGBoost to predict survival after HSCT using European Society for Blood and Marrow Transplant registry data of 10,888 patients who underwent HSCT for inborn errors between 2006 and 2018, and compared it to a simple linear Cox model, an elastic net Cox model, and a random forest model. The XGBoost model had a cross-validated area under the curve value of .73 at 1 year, which was significantly superior to the other models, and it accurately predicted for countries excluded while training. It predicted close to 0% and >30% mortality more often than other models at 1 year, while maintaining good calibration. The 5-year survival was 94.7% in the 25% of patients at lowest risk and 62.3% in the 25% at highest risk. Within disease and donor subgroups, XGBoost outperformed the best univariate predictor. We visualized the effect of the main predictors-diagnosis, performance score, patient age and donor type-using the SHAP ML explainer and developed a stand-alone application, which can predict using the model and visualize predictions. The risk of mortality after HSCT for inborn errors can be accurately predicted using an explainable machine learning model. This exceeds the performance of models described in the literature. Doing so can help detect deviations from expected survival and improve risk stratification in trials.

Keywords: Hematopoietic stem cell transplantation; Inborn errors; Machine learning; Pediatrics; Survival modeling.