Enhancing the prediction of CD34+ cells yield in autologous stem cell collection using complete blood count and clinical information: A quality improvement initiative spanning 4.5 years of data from a tertiary academic center Free

Background:Autologous stem cell transplant (ASCT) is an important treatment for various hematologic and non-hematologic diseases. It involves mobilizing CD34-positive (CD34+) peripheral blood stem cells (PBSCs) using granulocyte colony-stimulating factor (G-CSF) and plerixafor, known as the G+P regimen. Before collection, CD34+ cells are quantified by flow cytometry to guide the procedure and predict yield. At our institution, the use of pre-collection CD34 counts (pre-CD34) for the decision of initiating PBSC collection was phased out with the adoption of G+P, and the collection starts before the pre-CD34 result is available. Consequently, some patients with low pre-CD34 (<6 or 10 cells/µL) require additional days of collection or even remobilization. To enhance outcomes, we started a quality improvement initiative to evaluate our benchmarks and identify factors affecting successful PBSC collection (>3x10⁶CD34+ cells/kg).

Methods:We retrospectively analyzed consecutive autologous PBSC collections at the University of North Carolina (January 2021 - July 2025) to predict whether >3x10⁶ CD34+ cells/kg would be collected on day one (D1). Using only information available pre-procedure, a predictive modeling analysis included: age, gender, diagnosis, chemotherapy type/ number of lines/ number of cycles, and complete blood count (CBC) - white blood cell count (WBC), polymorphonuclear (PMN) relative count, monocyte relative count, hemoglobin, and platelet count. Diagnoses (plasma cell dyscrasias, lymphomas, solid tumors) and treatments (low, intermediate, and high intensity) were grouped by category. A linear regression model was trained with these covariates to predict D1 collection count. The ratio of patient weight/ liters collected was used to scale the laboratory value covariates. To evaluate out-of-sample prediction performance, we split the dataset into an 80/20% train-test set split model. To address test set variability, we performed Monte Carlo resampling by repeating the whole process (train/test split, model fitting, and test evaluation) 1000 times for more reliable estimates. We also compared our model to an established formula,¹ which would not be available before collection is initiated. We evaluated two additional models that include pre-CD34: a linear model fit to our data with the above covariates + pre-CD34, as well as the established formula (assuming 40% collection efficiency).

Results:We retrospectively reviewed 452 patients. Exclusion criteria included patients <18 years old, patients who did not receive G+P, leukemia or autoimmune diseases (n=13), resulting in an analysis cohort of 439 patients. The majority received G+P only (n=432, 98.41%), and seven (1.59%) received G+P in combination with another chemotherapy. Sixty-seven patients required a two- or three-day collection; however, we only analyzed D1. The mean age was 57.81 +/- 12.49; with a male-to-female ratio of 1.18. We analyzed correlations between collection counts on D1, and WBC (corr=0.408), PMN (corr=0.084), platelets (corr=0.322), and pre-CD34 (corr=0.896). Of the patients with pre-CD34 >10 cells/µL, 89.02% collected successfully on D1; and pre-CD34 >6 cells/µL, 87.53% collected on D1. The out-of-sample prediction accuracy of the three models was examined via two metrics: correlation and median absolute deviation (MAD). The correlation between the predicted and observed counts was 0.62 for the pre-initiation covariates model, 0.72 for the pre-initiation covariates + pre-CD34 model, and 0.79 for the established formula. The MAD between the predicted and observed counts for these models was 1.36, 1.16, and 1.15, respectively.

Conclusion:The predictive model using information available pre-collection (clinical and CBC) provided a degree of accuracy for predicting a successful collection; however, this model was less accurate than models that incorporate pre-CD34. Future analyses are needed to improve prediction with only pre-initiation data.

1. Pierelli L, Maresca M, Piccirillo N, Pupella S, Gozzer M, Foddai ML, Vacca M, Adorno G, Coppetelli U, Paladini U. Accurate prediction of autologous stem cell apheresis yields using a double variable-dependent method assures systematic efficiency control of continuous flow collection procedures. Vox Sang. 2006 Aug;91(2):126-34. doi: 10.1111/j.1423-0410.2006.00796.x. PMID: 16907873.