Predicting apheresis yield and factors affecting peripheral blood stem cell harvesting using a machine learning model

Area of Science:

• Hematology
• Transplant Immunology
• Biostatistics

Background:

• Peripheral blood stem cell (PBSC) collection via apheresis is vital for hematopoietic stem cell transplantation.
• Current mobilization and collection protocols are resource-intensive, often requiring multiple apheresis sessions.
• Minimizing apheresis sessions is essential to reduce donor discomfort and healthcare costs.

Purpose of the Study:

• To identify key predictive factors for successful initial PBSC harvests.
• To reduce the need for repeat apheresis procedures in PBSC collection.
• To optimize the efficiency and cost-effectiveness of PBSC mobilization and collection.

Main Methods:

• Retrospective analysis of 88 PBSC donations.
• Mobilization using chemotherapy with G-CSF (lenograstim) or G-CSF alone.
• Evaluation of donor characteristics, pre-apheresis blood counts, and CD34+ cells.
• Application of univariate logistic regression, eXtreme Gradient Boosting, and multivariate logistic regression.
• Integration of findings into machine learning models for predictive accuracy assessment.

Main Results:

• Monocyte percentage (Mono%), donor age, and CD34+ cell percentage (CD34+ cell%) were identified as significant independent predictors of successful initial PBSC harvest.
• Machine learning models incorporating these factors demonstrated accurate predictive capabilities.

Conclusions:

• Established machine learning methods successfully identified and validated Mono%, age, and CD34+ cell% as key predictors for first-attempt PBSC harvest success.
• These validated factors provide valuable insights for guiding clinical PBSC harvesting strategies.
• The findings can help optimize PBSC collection, minimize multiple apheresis sessions, and improve donor experience.