Predictive biomarkers validation of CD3+ cell apheresis yield in CAR-T manufacturing for diffuse large B-cell lymphoma: a machine learning approach Authors: Diego Carbonell, Alejandra Rodríguez-Sosa, Leticia Gómez-Serrano, Ana Pérez-Corral, Gloria Ruano, Marian Galayo, Nuria Panadero, Laura Rufián, Natalia Lorente, Pedro Rodríguez-Barquero, Diego Domingo-Merino, Cristina Muñoz-Martínez, Mi Kwon, Sudipto Das, Ramón García-Sanz, Javier Anguita Published: 2025 ABSTRACT AbstractChimeric antigen receptor (CAR) T-cell therapy has shown significant success in treating diffuse large B-cell lymphoma (DLBCL). The initial step involves collecting autologous CD3+ lymphocytes through apheresis, in which obtaining an adequate CD3+ cell yield is essential for therapeutic efficacy. Despite prior research, the factors influencing CD3+ cell apheresis remain poorly understood. Traditional statistical analyses offer limited insights, but machine learning (ML) approaches enable precision modeling of clinical predictors owing to their advanced pattern-recognition capabilities. In this study, we employed three ML algorithms, random forest classifier (RF), logistic regression (LR), and extreme gradient boosting (XGBoost) to analyze a homogeneous cohort of 98 DLBCL patients who underwent mononuclear cell (MNC) apheresis. The LR model, which achieved an area under the curve (AUC) of 0.824, identified four key predictive features: CD3+ cell absolute count, NK cell percentage, total blood volume, and CD3+ cell percentage. Among these, NK cell percentage and CD3+ cell absolute count showed the most significant negative impact on CD3+ cell apheresis yield. This study underscores the potential of ML approaches as a complementary analytical approach for identifying key factors that impact CD3+ cell apheresis efficiency, offering valuable insights for optimizing CAR-T therapy outcomes in patients with DLBCL.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-27061-2.Keywords: Chimeric antigen receptor t-cell, Apheresis, Data science, Artificial intelligence, Machine learning, B-cell lymphomaSubject terms: Biomarkers, Cancer, Computational biology and bioinformatics, Immunology Introduction Chimeric Antigen Receptor (CAR)-T cell therapy has revolutionized the treatment of hematologic malignancies, particularly those affecting B cells1. Approved by the U.S. Food and Drug Administration (FDA) and European Medicines Agency (EMA) in recent years, CAR-T cell therapy has shown remarkable efficacy in treating conditions such as diffuse large B-cell lymphoma (DLBCL) and acute lymphoblastic leukemia2–4. The initial step in CAR-T cell manufacturing is the collection of peripheral blood mononuclear cells (PBMCs) through a process known as apheresis. During apheresis, whole blood is drawn from the patient to isolate PBMCs, including T cells, which are subsequently genetically modified to express CAR5,6. The efficiency of apheresis is crucial, as the quality and quantity of the T cells collected directly impact the success of CAR-T cell therapy7. Identifying and optimizing the features influencing apheresis performance is crucial for improving therapeutic outcomes. Although various studies have attempted to improve the efficiency of T cell collection by identifying patterns and correlations that are not readily apparent through traditional statistical analyses, such as univariate and multivariate analyses, no consensus has yet been reached due to differences in patient cohorts and analytical methods8–12. Despite multiple attempts at designing predictive algorithms to optimize the T cell apheresis process13–15, these algorithms have faced criticism for not fully aligning with clinical requirements and lacking generalizability across diverse patient profiles16. Although the application of machine learning (ML) in clinical and laboratory practice is an evolving field, it offers a promising alternative for the development of new insights and a more effective analysis17. In this context, we aim to validate previously described predictive biomarkers of CD3+ cell apheresis yield in an independent, homogeneous cohort of patients with DLBCL undergoing CAR-T therapy manufacturing, and to explore the potential contribution of additional variables. By integrating an interpretable machine-learning framework using a Shapley Additive Explanations (SHAP)–based approach, we aimed to enhance model transparency and reproducibility. In doing so, this work complements prior publications by providing a clinically grounded, model-agnostic evaluation of predictive biomarkers relevant to CAR-T cell manufacturing14,15. This could lead to the development of more precise and effective strategies for optimizing CAR-T cell production, improving patient outcomes. Methods Patients A total of 98 patients diagnosed with B-cell lymphoma who underwent apheresis of mononuclear cells (MNC) for CAR-T therapy between July 2019 and October 2022 at the Hospital General Universitario Gregorio Marañón (HGUGM) were selected for the analysis. Apheresis procedure Apheresis eligibility criteria included an absolute lymphocyte count of at least 100/µL in peripheral blood, absence of active severe infection or fever above 38.2 °C, no clinical signs of infection within 48 h of apheresis, platelet count greater than 20,000/µL, and a hematocrit of 0.24 L/L or higher. Collection was conducted using the Spectra Optia platform (Terumo BCT, Tokyo, Japan) in the cMNC collection program. The volume to be processed was estimated based on the whole blood count (WBC), T-cell percentage, and the platform’s minimum collection efficiency of 40%18. Acid citrate dextrose solution A was used as the anticoagulant at a 12:1 ratio to blood. All patients underwent apheresis using a central venous catheter to ensure consistent access and efficient blood flow throughout the collection process. The collection line was adjusted to 1%–2%, as per the white blood cell colorgram provided by the manufacturer. Sample analysis White blood cell (WBC) counts were performed on the patient samples before each apheresis procedure and in the product bag using a DXH800 automated cell counter (Beckman Coulter, California, USA). The patient status was assessed, hematological variables were obtained, CD3+ cell collection efficiency (CE2) was calculated using the appropriate formula19. Additionally, WBC counts were conducted on the apheresis product to determine the total nucleated cell (TNC) and total lymphocyte counts. Flow cytometry was used to quantify the relative and absolute lymphocyte populations based on the WBC counts. This included the analysis of total granulocytes, monocytes, lymphocytes, and their subpopulations (T, B, and NK cells). Monoclonal antibodies were used to detect surface markers CD3, CD19, CD56, and CD45. Samples were analyzed using 8-color flow cytometry on a Navios (Beckman Coulter) and a 10-color DxFlex cytometer (Beckman Coulter), and population analysis was performed using Kaluza software (Beckman Coulter). Data preprocessing Pre-apheresis features were selected, including demographic characteristics (sex, weight, height, and age), clinical features (transformed or de novo DLBCL), previous treatments (autologous stem cell transplantation [ASCT], bendamustine, methotrexate, gemcitabine, etoposide, polatuzumab, radiotherapy, and more than three lines of treatment), apheresis-related features (total blood volume [TBV], flow rate), laboratory results (leukocytes, erythrocytes, hemoglobin, hematocrit, platelets), and the absolute numbers and percentages of different cell subpopulations (granulocytes, lymphocytes, monocytes, T, B, and NK cells). TBV was calculated using the respective formula20. Patients were subsequently divided into two cohorts based on whether their CE2 was less than 50% or greater than or equal to 50%. The 50% threshold was selected because it represents a more stringent criterion than the conventional 40% and is commonly regarded as a standard benchmark of acceptable collection performance, as described by other authors14,21. In addition, sensitivity analyses were performed using alternative thresholds (CE2 ≥ 40% and CE2 ≥ median), to assess the robustness of the models. All variables included in the analysis were continuous or binary, therefore, no categorical encoding was required. No missing data were present in the analyzed variables; therefore, no imputation procedures were required. Statistical analysis Continuous variables were expressed as median and interquartile range (IQR), whereas binary variables were expressed as numbers and percentages. For univariable comparisons, the chi-squared test was used for binary variables, while the Shapiro-Wilk test assessed normality for continuous variables. Depending on the distribution of continuous data, Student’s t-test or Mann-Whitney U test was applied. For multivariable analysis, analysis of variance (ANOVA) or Kruskal-Wallis tests were used where appropriate. A two-sided p < 0.05 was considered statistically significant. Analyses and graphical representations were performed using Python Scikit-learn (version 1.6.1) library21. Machine learning modeling The ML model was constructed and interpreted using Python (version 3.12.3; Python Software Foundation, USA). Candidate variables were selected according to their statistical significance in the multivariable analysis and after excluding redundancy through correlation assessment (|r| > 0.70) and variance inflation factor (VIF) analysis (VIF > 5). This approach was chosen to minimize multicollinearity, reduce model complexity, and enhance clinical interpretability. The dataset was randomly divided into two subsets: the training set corresponding to 66.3% of the samples (65/98) and the testing set consisting of the remaining 33.6% (33/98) of the samples. Three machine learning algorithms were implemented to predict CD3⁺ cell apheresis yield (< 50% or ≥ 50%) based on the selected features. Logistic regression (LR), random forest (RF), and extreme gradient boost (XGBoost) were chosen to represent complementary modeling paradigms: linear interpretability, non-linear ensemble learning, and gradient-based optimization, respectively. All models were trained on a tabular, homogeneous dataset derived from a single-center cohort, with consistently measured and preprocessed variables to ensure comparability across samples. This approach allowed for a robust comparison of performance and interpretability within a controlled framework22,23. All numerical variables were normalized using the StandardScaler function from the scikit-learn library for use in the LR model. Normalization was performed after data splitting, within each training fold during cross-validation, ensuring that scaling parameters were derived exclusively from the training data to prevent data leakage22. Hyperparameter tuning was performed using the GridSearchCV tool from scikit-learn library, and 10-fold cross-validation, employing various metrics to select the best model, including area under the receiver operating characteristic (ROC) curve (AUC), average precision, weighted F1 score, Matthews correlation coefficient, and balanced accuracy (Supplementary Information)21. The effectiveness of each model was evaluated using several metrics, including accuracy, recall, F1 score, positive predictive value, and negative predictive value. In addition, the ROC AUC was calculated for the test sets to assess the performance of each model. To demonstrate stability of our LR model and identify any overfitting issues, we applied a permutation test of the AUC using the permutation_test_score function from scikit-learn and plotted the accuracy of the model over 2,000 permutations22. Finally, SHAP method, using shap package (version 0.49.1) was employed to interpret the contribution of each predictor within the optimal model. Pairwise SHAP interaction analyses were performed to explore potential non-linear dependencies between the main predictors. For each pair of variables, SHAP dependence plots were generated using the final model. SHAP values were computed on standardized features and then mapped back to their original scale for interpretability. (Fig. 1)24. Diagram of the study’s experimental design to identify key factors affecting CD3+ cell apheresis yield for CAR-T therapy in high-grade B-cell lymphoma patients. The CAR-T therapy cycle starts with leukapheresis, followed by genetic modification, CAR-T cell expansion, and infusion back into the patient. The inner diagram represents the analysis of the leukapheresis process, detailing the filtering and selection of factors, the generation of machine learning models, and the identification of key factors that optimize CD3+ cell apheresis yield. Model discrimination was compared using the DeLong test for paired receiver operating characteristic (ROC) curves, with 95% confidence intervals for area under curve (AUC) derived from the DeLong variance estimate. To ensure unbiased performance estimation, we applied a nested cross-validation (CV) framework in the selected model. In the inner loop, recursive feature elimination with cross-validation (RFECV) was performed to identify the optimal subset of predictors. In the outer loop, predictive performance was evaluated in held-out test folds. A 10-fold outer CV was used, with stratification to preserve outcome balance. All feature selection and model training steps were performed exclusively within the training partitions of each fold to prevent information leakage. The area under the receiver operating characteristic curve (AUC) was used as the primary performance metric. AUC distributions across outer folds were summarized by mean, standard deviation, median, and interquartile range (IQR). Variable selection frequencies were also recorded to evaluate predictor stability across folds. Results Patients’ characteristics Data were analyzed by dividing the CD3⁺ cell apheresis yields into two subsets: <50% and ≥ 50%, with a median yield of 53%. Among the 98 patients included, 57 (58.2%) achieved CE2 ≥ 50%, while 41 (41.8%) had CE2 < 50%. Statistically significant differences were observed in TBV, WBC, percentage of granulocytes, as well as in T and NK cells, both in absolute number and percentage (Table 1). These results were confirmed by multivariate analysis. Table 1Pre-apheresis characteristics (demographic, clinical, apheresis-related, laboratory, and flow cytometry) of the two patient cohorts, with < 50% and ≥ 50% CD3+ cell apheresis yield. Statistical analysis for qualitative variables was conducted using the Chi-squared test. For quantitative variables, student’s t-test and the Mann-Whitney U test were used for parametric and non-parametric analysis, respectively. IQR: interquartile range. DLBCL: diffuse large B-cell lymphoma. ASCT: autologous stem cell transplantation. TBV: total blood volume. WBC: total white blood cell count.CharacteristicCD3+ cell apheresis yieldp-value< 50% (n = 41)≥ 50% (n = 57)Demographic variablesSex, n (%)0.117Female21 (51.22)19 (33.33)Male20 (48.78)38 (66.67)Age (years), median (IQR)62 (54–69)58 (50–65)0.166Weight (kg), median (IQR)68 (58–75)71 (62–83)0.157Height (cm), median (IQR)166 (158–172)169 (163–177)0.069Clinical variables, n (%)Type of DLBCL De novo 34 (82.93)45 (78.95)0.816Transformed7 (17.07)12 (21.05)Treatment≥ 3 lines of treatment13 (31.71)27 (47.37)0.178Bendamustine9 (21.95)12 (21.05)1.000Etoposide28 (68.29)44 (77.19)0.452Gemcitabine5 (12.20)9 (15.79)0.834Methotrexate5 (12.20)6 (10.53)1.000Polatuzumab6 (14.63)9 (15.79)1.000Radiotherapy9 (21.95)16 (28.07)0.652ASCT7 (17.07)21 (36.84)0.056Apheresis variablesFlow rate, n (%)0.8360.8 mL/min29 (70.73)38 (66.67)1 mL/min12 (29.27)19 (33.33)TBV (mL), median (IQR)4350 (3528–4886)4629 (3991–5211) 0.046 Analytical variables, median (IQR)Erythrocytes (1012/L)3.4 (3.1–3.9)3.43 (3.0–3.8)0.463Hemoglobin (g/L)104 (93–123)106 (98–119)0.976Hematocrit (L/L)0.31 (0.279–0.361)0.309 (0.293–0.348)0.809Platelets (109/L)182 (133–231)170 (99–221)0.237WBC (109/L)6.0 (3.5–7.3)4.2 (3.2–5.9) 0.009 Flow cytometry variables, median (IQR)Cell subpopulations (absolute) (109/L)Granulocytes3.538 (2.276–4.635)2.747 (1.803–4.371)0.235Monocytes0.420 (0.305–0.625)0.370 (0.283–0.553)0.200Lymphocytes1.319 (1.001–2.358)0.822 (0.517–1.252) < 0.001 T lymphocytes (CD3+)1.044 (0.815–2.039)0.655 (0.392–0.963) < 0.001 B lymphocytes (CD19+)0.001 (0.001–0.004)0.001 (0.000–0.007)0.331NK lymphocytes (CD56+)0.105 (0.079–0.230)0.059 (0.036–0.096) < 0.001 Cell subpopulations (% over WBC)Granulocytes61.10 (51.215–70.720)68.14 (57.695–78.480) 0.013 Monocytes8.67 (5.765–13.485)9.13 (6.785–12.135)0.519Lymphocytes30.59 (19.900–38.615)20.92 (11.880–31.700) 0.010 T lymphocytes (CD3+)24.58 (14.282–31.514)17.41 (9.675–26.949) 0.041 NK lymphocytes (CD56+)2.34 (1.415–4.149)1.59 (0.947–2.518) 0.012 B lymphocytes (CD19+)0.02 (0.013–0.066)0.027 (0.008–0.191)0.818 Pre-apheresis characteristics (demographic, clinical, apheresis-related, laboratory, and flow cytometry) of the two patient cohorts, with < 50% and ≥ 50% CD3+ cell apheresis yield. Statistical analysis for qualitative variables was conducted using the Chi-squared test. For quantitative variables, student’s t-test and the Mann-Whitney U test were used for parametric and non-parametric analysis, respectively. IQR: interquartile range. DLBCL: diffuse large B-cell lymphoma. ASCT: autologous stem cell transplantation. TBV: total blood volume. WBC: total white blood cell count. Post-apheresis In terms of post-apheresis features, the average processed volume in the entire cohort was 12,455 mL, with significant differences between the two subsets (11,827 vs. 12,921 mL). Differences between the two subsets were also found in the absolute number of red blood cells (36 vs. 43 × 109 cells) and hematocrit in the collection bags (0.021 vs. 0.028 L/L; Supplementary Table S1). Feature selection and ML models For the development of the models, only those features that were statistically significant when comparing the CD3+ cell apheresis efficiency cohorts in both univariate and multivariate analyses were selected, after excluding redundancy by correlation (Supplementary Figure S1) and VIF analysis (Supplementary Figure S2). Because the CE2 ≥ 50% dataset contained a near-balanced distribution of outcomes (58.2% vs. 41.8%), no resampling or weighting adjustments were required. Following the construction of the initial models, the test set showed that LR achieved the best discriminative performance compared with RF and XGBoost (Table 2; Fig. 2). LR was therefore retained as the final model. The LR logit equation can be consulted in Supplementary Information. Model performance metrics shown in the main Table 2 correspond to the test partition from a fixed train–test split, which served as the reference model for interpretability analyses. Table 2Performance comparison of the three models on the testing set using a fixed random seed. Reported metrics correspond to a single train–test split and include the area under the ROC curve (AUC), accuracy, recall (sensitivity), F1 score, positive predictive value (PPV, precision), and negative predictive value (NPV). LR, logistic regression; RF, random forest; XGBoost, extreme gradient boosting.ModelAUCAccuracyRecallF1 scorePPVNPV LR0.8240.7880.8130.7880.7650.813 RF0.7100.6670.6880.6670.6470.688 XGBoost0.7320.6060.7500.6490.5710.667 Performance comparison of the three models on the testing set using a fixed random seed. Reported metrics correspond to a single train–test split and include the area under the ROC curve (AUC), accuracy, recall (sensitivity), F1 score, positive predictive value (PPV, precision), and negative predictive value (NPV). LR, logistic regression; RF, random forest; XGBoost, extreme gradient boosting. ROC curves comparison from the analysis of the three applied models, of the train set (A) and the test set (B). Included models: logistic regression (LR), random forest (RF), XGBoost. Y-axis indicates the true positive rate; X-axis indicates the false positive rate. ROC: receiver operating characteristic curve. AUC: area under the curve. To assess the statistical significance and variability of model performance, 95% confidence intervals were derived via 2,000-iteration stratified bootstrapping, and pairwise AUC comparisons were performed using the DeLong test (Supplementary Tables S2 and S3). Nested cross-validation of this model further confirmed its performance and generalizability, with a mean outer test AUC of 0.799 (SD 0.092; median 0.812, IQR 0.719–0.833). Feature stability analysis demonstrated that CD3 absolute count, NK percentage, TBV and CD3% were selected in ≥ 90% of folds. Sensitivity analysis and AUC comparison The permutation test for overfitting testing showed a gaussian distribution with the AUC of the training set at 0.7957 and a permuted mean of 0.4977 (Supplementary Figure S3). To evaluate the robustness of the predictive models across different definitions of collection efficiency, we re-ran the analysis using alternative thresholds (CE2 ≥ 40% and CE2 ≥ median) in addition to the primary cutoff (CE2 ≥ 50%). LR consistently achieved the highest discriminative performance across all definitions, with AUCs of 0.80 (95% CI, 0.71–0.89), 0.72 (95% CI, 0.59–0.84), and 0.73 (95% CI, 0.63–0.83) for CE2 ≥ 50%, ≥ 40%, and ≥ median, respectively (Supplementary Table S2). Across these outcome definitions, class distributions remained balanced, 57/98 (58.2%) for CE2 ≥ 50%, 73/98 (74.5%) for CE2 ≥ 40%, and 49/98 (50.0%) for CE2 ≥ median, indicating that no resampling or weighting adjustments were required. Stratified cross-validation was applied throughout to preserve class proportions across folds. Pairwise DeLong tests confirmed that LR significantly outperformed RF and XGBoost for CE2 ≥ 50% (p = 0.002 and p = 0.006, respectively) and for the median-split definition (p = 0.0002 and p = 0.010), while no significant differences were found for CE2 ≥ 40% (Supplementary Table S3). Model interpretation and impact of features SHAP analysis was applied to evaluate the importance of the features within the selected LR ML model. The selected features involved in achieving a CD3+ cell apheresis yield greater than 50% were TBV, percentage and absolute number of CD3+ cells, and percentage of NK cells. The absolute number of CD3+ cells had the highest (0.996) mean SHAP impact, followed by NK percentage (0.845), total blood volume (0.414), and CD3+ cell percentage (0.076; Fig. 3A). The individual SHAP impact of each case of the test set, including whether the impact was positive or negative, is presented in greater detail in the beeswarm plot (Fig. 3B). Regarding the interpretability of the results, the force plots illustrate the contributions of individual features to the predictions for two representative samples from the dataset (Fig. 4). To illustrate the quantitative impact of key predictors, SHAP force plots were generated from the selected logistic regression model (Fig. 4). These plots display how individual variable values contribute to the predicted probability of achieving ≥ 50% CD3⁺ apheresis yield. The largest negative contributions come from higher TBV (4,522 mL) and higher CD3% (28.33), both pushing the prediction downward. In contrast, lower NK% (6.394) and lower CD3 (abs) (0.9632 × 109 cells/L) contribute positively, partially offsetting the drop, but not enough to overcome the downward pull (Fig. 4A). Conversely, lower NK percentage and absolute CD3⁺ counts were associated with higher probabilities of efficient collection (Fig. 4B). These quantitative effects provide an interpretable link between immunological composition and model output. Shapley Additive Explanations (SHAP) summary plot for the logistic regression model. (A) The summary plot displays the ranking of feature importance based on the mean absolute SHAP values. CD3+ (abs): CD3+ absolute number (cells/L). TBV (mL): total blood volume. (B) The beeswarm plot shows the impact of each feature on the model, with red indicating higher feature values and blue indicating lower feature values. Dots represent the impact (location in the x axis) and the value (color) of the feature in each case of the test set. Shapley Additive Explanations (SHAP) force plots for the logistic regression model. All values are represented in the log-odds (logit) scale. The base value (0.5199 log-odds) represents the expected output of the model across the training data. (A) In the first plot the predicted value is −1.31 log-odds, below the base value, indicating a lower probability of achieving a CD3+ cell yield greater than 50%. The negative contributions (blue) include NK percentage [NK (%)] = 6.394% and CD3+ cell absolute number [CD3 (abs)] = 0.9632 × 109 cells/L, outweighing the positive contributions (red), such as total blood volume [TBV] = 4,522 mL and CD3+ cell percentage [CD3 (%)] = 28.33%. (B) In the second plot the predicted value is 1.78 log-odds, significantly above the base value, suggesting a higher likelihood of achieving a CD3+ cell yield greater than 50%. Here, the positive contributions of NK (%) = 0.8494% and CD3 (abs) = 0.2066 × 109 cells/L) counteract the negative effects of total blood volume TBV = 4,010 mL and CD3 (%) = 4.216%. Exploratory SHAP interaction plots showed no evident nonlinear dependencies between the evaluated features in the test set. The direction and magnitude of SHAP values for each variable appeared independent from those of the others, leading to predominantly additive effects on the model output. Discussion In this study, three machine learning models were applied to identify the features that influence CD3+ cell apheresis yield for CAR-T therapy, along with the SHAP method, to further enhance the interpretability of the results obtained. Previous studies have explored the identification of relevant features affecting T-cell apheresis using traditional statistical methods9,10,15,25. Although a classic approach is useful for identifying features that impact yield in a broad sense, it is important to account for the effects of small and/or heterogeneous cohorts. These factors can influence key features such as the number of circulating cells, thus introducing potential biases and reducing the model’s applicability. Some examples can be found in previous studies, where differences in apheresis yield between patients with non-Hodgkin lymphoma and those with lymphoblastic or lymphocytic leukemia have been reported, and studies using different apheresis platforms presented inconsistent results due to performance variations between collection systems9,26. The use of a highly homogeneous cohort of DLBCL patients in this study ensured a focused understanding of the CD3+ cell apheresis yield, minimizing variability due to patient heterogeneity. In this cohort, the apheresis data reported a median CD3+ cell yield of 53%, which is consistent with the typically observed ~ 55% from literature27. Comparing the post-apheresis data of the subsets for < 50% and ≥ 50% yields, no significant differences were observed in apheresis features, such as procedure time or final product volume. However, there were significant differences in the processed volume variable, which was notably higher in the group with better apheresis yield. This can be explained by the higher estimated volume required for processing in patients with a lower pre-apheresis absolute CD3+ cell count, which is also more prevalent in the subset with better apheresis performance18. Differences were also found in the absolute number of erythrocytes and the hematocrit in the final product, which were higher in the group with better apheresis yield. With a reduced quantity of these cells, it is possible that the mononuclear layer in the centrifugation gradient was thinner, thereby facilitating the collection of a greater number of erythrocytes from the lower layer28. There were no differences in the total amount of each of the analyzed cell subpopulations. This confirms that a better apheresis yield is not linked to a higher quantity of collected target cells. Regarding the results of the modeling analysis, the application of machine learning methods introduces a new paradigm over traditional statistical methods previously used, offering the ability to perform more complex, insightful analyses29–31. The robustness of the predictive framework was supported by multiple complementary validation strategies. First, a permutation test with 2,000 random label shuffles confirmed that the logistic regression model learned non-random patterns, with a statistically significant separation between real and permuted AUC distributions (p = 0.001). Second, 2,000 bootstrap iterations were performed to estimate 95% confidence intervals for all classification metrics, confirming the stability of the results across resampled subsets. Finally, pairwise DeLong tests were applied to compare AUCs between models, showing that the logistic regression significantly outperformed random forest and XGBoost in the primary dataset (p < 0.01), reinforcing the model’s statistical robustness and reproducibility. Interestingly, some features identified as significant in univariate and multivariate statistical analyses were excluded from the ML models because of their lack of impact on apheresis yield. The exclusion of the WBC feature from the final ML model suggests that other features or a combination of features may play a more significant role in the prediction of apheresis yield. Notably, lymphocyte count was excluded from the final model. As lymphocyte levels can directly influence the success of CD3+ cell apheresis, it would have been expected to see this variable also included in the final model. The exclusion of these two features exemplifies the differences between traditional statistical analyses and machine learning-based models, where the pattern detection of the latter enables the discovery of nonlinear interactions between features that are much harder to detect otherwise. The key features associated with higher CD3+ cell apheresis yield (≥ 50%) identified by the best machine learning model were the absolute number and percentage of CD3+ cells, percentage of NK cells, and the patient’s TBV. Although the percentage of CD3+ cell features had some impact on apheresis yield, it did not appear to carry significant weight in the final outcome. Some studies have reported a proportional relationship between CD3+ cell absolute counts and the final CD3+ cell count, although not necessarily with yield12,25. Other authors have described an inverse relationship between pre-apheresis CD3+ cell levels and CE2, which is consistent with the results of the present study14, which can be explained by a low collection efficiency due to the inability of the platform to capture all available cells when pre-apheresis CD3+ cell counts are higher. The negative correlation between the percentage of NK cells and CD3+ cell apheresis yield has been previously reported8,32. This could be explained by competition during collection, as both cell types occupy the lymphocyte layer in the centrifuge of the apheresis platform. The positive effect of TBV on the CD3+ cell apheresis yield has not been previously reported, but it may be attributed to the enhanced efficiency of blood processing and separation. A larger TBV can stabilize the interface during centrifugation, providing greater flexibility in configuring the apheresis platform settings, resulting in more consistent CD3+ cell separation. Additionally, a higher blood volume may improve flow efficiency through the apheresis system, reducing interference from other cell types and increasing the overall yield. Further SHAP interaction analyses were performed to explore whether the effects of these predictors were modulated by nonlinear dependencies or feature interactions. Pairwise SHAP interaction plots for the four main variables did not reveal consistent interaction patterns in the test subset, likely reflecting both the additive nature of the logistic regression model and the limited sample size. This suggests that each variable contributes independently to the probability of achieving ≥ 50% apheresis yield, reinforcing the interpretability and stability of the predictive framework. Our results confirm the relevance of previously reported predictors of CD3+ cell apheresis yield. However, unlike prior studies that were conducted in heterogeneous cohorts or with limited sample sizes, and where these predictors were often identified alongside variables, our analysis highlights a consistent and more restricted set of robust factors. This strengthens the evidence for their reproducibility and suggests that some of the previously described associations may have been context-dependent or less stable. To further assess the robustness of our findings, we examined whether model performance remained consistent across alternative definitions of collection efficiency. When applying more lenient thresholds (CE2 ≥ 40%) or a median split, LR continued to outperform the other algorithms, achieving AUCs of 0.72 and 0.73, respectively, compared to 0.80 for the primary CE2 ≥ 50% definition. Pairwise DeLong tests confirmed that LR maintained significantly higher discrimination compared with RF and XGBoost for the 50% and median thresholds (p = 0.002–0.010), while no significant differences were observed for CE2 ≥ 40% (p > 0.47). Other performance metrics showed comparable trends across CE2 thresholds. These results demonstrate that the model’s relative performance and the key predictive features were not driven by an arbitrary cutoff value but remained stable across outcome definitions. This consistency supports the use of CE2 ≥ 50% as a stringent yet clinically relevant indicator of optimal apheresis yield, in line with previous literature14,21. Some limitations of this study include the relatively small sample size, which may not be optimal for a more robust analysis using machine-learning techniques. Future studies should validate these findings in larger and independent cohorts to confirm the identified predictive features. Additionally, the analysis was intentionally restricted to a homogeneous cohort of patients with DLBCL to minimize confounding and ensure internal validity. While this design strengthens the consistency of the observed associations, it may limit the generalizability of the model to other hematologic malignancies or manufacturing settings. To address this, retraining and external validation across multicenter and mixed-histology datasets, ideally encompassing different apheresis platforms, would be valuable to test the reproducibility and translational applicability of the proposed framework. Moreover, the range of clinical variables included was limited to routinely available pre-apheresis parameters, and the study did not incorporate additional biological or disease-related variables such as tumor burden, cytokine levels, or immune functional markers, which could further enhance predictive performance. Finally, downstream validation linking CD3⁺ apheresis yield to CAR-T cell expansion and clinical outcomes was beyond the scope of this work but represents an important next step to confirm the biological and clinical relevance of the identified predictors. Nevertheless, the selected patient population represents not only the most frequently treated group with CAR-T therapy but also one of the few indications currently approved for commercial use, underscoring the clinical relevance of this analysis32–34. A recent systematic review has emphasized the multifactorial nature of apheresis efficiency, highlighting patient-specific factors such as baseline lymphocyte counts, disease burden, and prior therapies as significant predictors of collection success7. Incorporating these evidence-based predictors into clinical workflows can optimize timing and patient selection, thereby enhancing the overall efficiency and success rates of CAR-T cell manufacturing. Conclusions Our findings confirm that previously reported factors, including CD3+ cells absolute number and NK percentage, are reproducible predictors of CD3+ cell apheresis yield in patients with large B-cell lymphoma undergoing CAR-T manufacturing. By validating these variables in an independent cohort, we provide additional evidence supporting their robustness and clinical relevance. Notably, we also identified blood volume as a novel predictor, which remained stable across multiple resampling strategies and warrants consideration in future studies. Taken together, these results suggest that CD3+ cell yield is influenced by a limited but consistent set of clinical parameters, and that incorporating them into pre-apheresis evaluation could improve patient selection and optimize collection strategies. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements We are deeply grateful to the patients and their families, for their courage and generosity. Author contributions DC and JA participated in the conception or design of the work. DC, LG, APC, GR, MG, NP, LR, NL, PRB, DDM, CMM, MK, SD, RGS did the data collection. DC and ARS did data analysis, interpretation, and drafting of the article. All authors did critical review and final approval of the article. Data availability The datasets generated and/or analyzed during the current study are not publicly available due to ethical restrictions and institutional policies regarding patient confidentiality but are available from the corresponding author upon reasonable request to corresponding author under GDPR compliance. Declarations Competing interests The authors declare no competing interests. Ethics approval and consent to participate This study was conducted in accordance with institutional and national ethical standards. The protocol was reviewed and approved by the Research Ethics Committee of Hospital General Universitario Gregorio Marañón (protocol name: LINFOR, contract number: 82/25, institutional code: 71/25, date of approval: March 3rd 2025), which granted an exemption from obtaining individual informed consent due to the retrospective nature of the analysis and the use of routinely collected clinical data. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. *Diego Carbonell and Alejandra Rodríguez-Sosa have contributed equally to this work. Contributor Information Diego Carbonell, Email: diegocarbonell3@gmail.com. Alejandra Rodríguez-Sosa, Email: aroso491@gmail.com. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The datasets generated and/or analyzed during the current study are not publicly available due to ethical restrictions and institutional policies regarding patient confidentiality but are available from the corresponding author upon reasonable request to corresponding author under GDPR compliance.