(P-BT-2) A Machine Learning Model for Unstimulated T-cell Collection in The Cell Therapy Starting Materials For CAR-T Cell Manufacturing

Background/Case Studies: Chimeric Antigen Receptor (CAR) T-cell has emerged as a revolutionary treatment for relapsed/refractory B-ALL and mature large B-cell lymphomas. A critical step involves the collection of T cells through leukapheresis. However, variability in leukapheresis can lead to collection failures, overcollections, and inefficient utilization of resources.
This study aims to investigate the impact of different variables on CD3 collection and propose a machine-learning model that guides the leukapheresis protocol.

Study

Design/Methods: A retrospective analysis of leukapheresis procedures conducted at one center on patients approved for Tisagenlecleucel between November2020 -June2022. The analysis was performed on datasets using GraphPad Prism software(Prism 7) to evaluate the impact of different variables on the collected CD3 in the CAR-T starting material.A machine-learning model was applied.

Results/Findings: Our collected data included 95 leukapheresis procedures for 90 patients, 61 males and 29 females, median age of 34.3 years (range, 3 to 71 years). The primary diagnosis was B-cell Acute Lymphoblastic Leukemia (ALL) in 38 patients and Diffuse Large B-cell Lymphoma (DLBCL) in 58 patients. The median total blood volume (TBV) processed by leukapheresis is 12.3 L (range, 3 to 19 L), and the median CD3+ cells harvested was 8.6 × 10^9 (range, 0.51 × cells to 34.9 10^9).
In multivariate analysis, (TBV) processed, (ALC), and (BMI) correlate with the collected CD3+ cells collected, R (TBV)=0.4, R(ALC)=0.5, R(BMI)= 0.5 (P-value < 0.001) respectively (Figure 1). Other variables showed weaker correlation, R(Age)=0.2, R(Diagnosis)=0.2, R(WBC)=0.2, R(Time)=0.1.
To assess the potential for predicting collected CD3+, we employed various machine learning algorithms, encompassing Linear Regression, MultiLayer Perceptron (learning rate: 0.3), Simple Linear Regression (utilizing solely BMI), and Sequential Minimal Optimization for regression (SMOreg), which implements the support vector machine methodology for regression tasks. Each of these models was evaluated through a cross-validation approach with 10 folds. Linear Regression yielded most favorable results, obtaining highest CC of 0.6, and lowest values for ME and RMS. The derived model from this analysis is represented in equation (1). This equation was validated for the subsequent 9 patients approved for the therapy with 70% sensitivity.
CD3 count = 0.1029 * BMI + 0.0005 * TBV processed (mL) + 3.3478 * Pre ALC +(-4.9943)(1)

Conclusions: BMI, ALC, and processed TBV by leukapheresis strongly correlate with the collected CD3 in the CAR-T cell starting material. Among different machine-learning models, Linear Regression exhibited the most promising for predicting the collected CD3. This model can be incorporated through a web-based or mobile app to guide apheresis collections.

Importance of research: Apheresis collection of T cells through leukapheresis is essential to successful treatment outcomes. However, the leukapheresis process still needs to be standardized. This is a proposal for an AI-guided method to predict the best-processed apheresis volume which can avoid variability and minimize collection failures, overcollections, and inefficient utilization of medical and laboratory resources in clinical practice.