Machine Learning Prediction Model for Neutrophil Recovery after Unrelated Cord Blood Transplantation.

ABSTRACT

Delayed neutrophil recovery is an important limitation to the administration of cord blood transplantation (CBT) and leaves the recipient vulnerable to life-threatening infection and increases the risk of other complications. A predictive model for neutrophil recovery after single-unit CBT was developed by using a machine learning method, which can handle large and complex datasets, allowing for the analysis of massive amounts of information to uncover patterns and make accurate predictions. Japanese registry data, the largest real-world dataset of CBT, was selected as the data source. Ninety-eight variables with observed values for >80% of the subjects known at the time of CBT were selected. Model building was performed with a competing risk regression model with lasso penalty. Prediction accuracy of the models was evaluated by calculating the area under the receiver operating characteristic curve (AUC) using a test dataset. The primary outcome was neutrophil recovery at day (D) 28, with recovery at D14 and D42 analyzed as secondary outcomes. The final cord blood engraftment prediction (CBEP) models included 2991 single-unit CBT recipients with acute leukemia. The median AUC of a D28-CBEP lasso regression model run 100 times was .74, and those for D14 and D42 were .88 and .68, respectively. The predictivity of the D28-CBEP model was higher than that of 4 different legacy models constructed separately. A highly predictive model for neutrophil recovery by 28 days after CBT was constructed using machine learning techniques; however, identification of significant risk factors was insufficient for outcome prediction for an individual patient, which is necessary for improving therapeutic outcomes. Notably, the prediction accuracy for post-transplantation D14, D28, and D42 decreased, and the model became more complex with more associated factors with increased time after transplantation.