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

Abstract

Objective: Mobilization and collection of peripheral blood stem cells (PBSCs) are time-intensive and costly. Excessive apheresis sessions can cause physical discomfort for donors and increase the costs associated with collection. Therefore, it is essential to identify key predictive factors for successful harvests to minimize the need for multiple apheresis procedures.

Methods: We retrospectively analyzed 88 PBSC donations at our hospital. Mobilization involved disease-specific chemotherapy plus human recombinant granulocyte-colony-stimulating factor (G-CSF; lenograstim) or G-CSF alone for 5 days, followed by apheresis on day 5. The baseline characteristics of donors, pre-apheresis complete blood counts, and CD34+ cells were evaluated. Univariate logistic regression, the eXtreme Gradient Boosting algorithm, and multivariate logistic regression were applied to select significant predictive variables. The multivariate logistic regression results were integrated into various machine learning models to assess predictive accuracy.

Results: The percentage of pre-collection monocytes (Mono%), age, and CD34+ cell percentage (CD34+ cell%) were identified as significant independent factors that could accurately predict the success of an initial PBSC harvest.

Conclusions: We used machine learning methods to identify and validate Mono%, age, and CD34+ cell% as significant factors predictive of successful PBSC harvest on the first attempt, offering important insight to guide the clinical harvesting of PBSCs.

Keywords: Peripheral blood stem cell; apheresis; logistic regression; machine learning model; mobilization; prediction.

Figure 1.

Identification of important factors influencing success of initial PBSC collection using the XGBoost Algorithm. (a) Mono%, age, and CD34⁺ cell% emerged as the three leading variables in terms of importance and (b) results of the confusion matrix suggested that the three leading variables selected using XGBoost can effectively differentiate between positive and negative outcomes of the first PBSC collection. XGBoost, eXtreme Gradient Boosting; PBSC, peripheral blood stem cells; Mono%, percentage of pre-collection monocytes; LDH, lactate dehydrogenase; WBC, white blood cells; BMI, body mass index; HGB, hemoglobin; PLT, platelets.

Figure 2.

Machine learning models to validate the accuracy of independent factors in predicting successful PBSC collection. (a) Accuracy values obtained from the 30 leading cases predicted using Mono%, age, and CD34⁺ cell% with multiple algorithm combinations are presented in the bar plot. (b) ROC curve showing the predictive ability of the GBM model in the training set and (c) ROC curve showing the predictive ability of the GBM model in the test set. PBSC, peripheral blood stem cells; Mono%, percentage of pre-collection monocytes; GBM, ROC, receiver operating characteristic; AUC, area under the ROC curve; KNN, K-nearest neighbor; LR, logistic regression; RR, ridge regression; ENR, elastic net regression.