Determination of the required apheresis volume in stem cell apheresis is a challenge. Prediction using the CE2 formula assumes a certain collection efficiency, which is often deliberately underestimated, in order to reliably reach the target CD34+ cell number. However, this exposes most donors longer to apheresis than necessary. The aim of this work is to use machine learning to generate a model that predicts apheresis volumes more accurately with the same reliability as the CE formula.

We divided data from 3322 collections, which included CD34+ cells in the product, apheresis volume as well as 16 additional parameters, into a test and a main data set. We applied the main data in an optimization process to determine which machine learning model, which parameters and hyper parameters and which split of the main data set into training and validation data sets would generate the best model and which safety margin would lead to acceptable certainty to reach the desired CD34+ cell target. Finally, we tested the model with the test data set.

The generated model has a probability of 98% to predict an apheresis volume to reach the desired CD34+ cell target. If only collections with true apheresis volumes of < 15 L are considered, the mean standard error of the model is 606mL below that of the CE2 formula, with a collection efficiency that results in the same probability of 98% to collect the target amount of CD34+ cells. If all collections are considered, the difference is 3233mL.

The generated model calculates apheresis volumes closer to the real throughput volumes needed with the same reliabilty as the CE2 formula. It therefore enables to shorten apheresis of stem cell donors without compromising safety or collection yield. The machine learning approach thus offers an advantage over previous prediction methods.

There is no conflict of interest