2026 ASH®
Annual Meeting & Exposition December 12-15, 2026 ▪ New Orleans, Louisiana

Paper Details

Abstracts, authors, affiliations, and disclosures.

SESSION

628. Aggressive Lymphomas: Cellular Therapies: Poster I

Program: Poster Abstract Session
Saturday, December 12, 2026
05:30 p.m. - 07:30 p.m.
MCCNO - Hall F
Access, Representation, Population Health, Technology and Procedures, Clinical Procedures, Treatment Considerations, Therapy Sequence, Biological Therapies, Transplantation, Bispecific Antibody Therapy, Chimeric Antigen Receptor, (CAR)-T Cell Therapies, Immunotherapy, Infusion, Non-Biological Therapies, Chemotherapy, Adverse Events, Artificial Intelligence (AI), Machine Learning, Biological Processes, Immune Mechanism, Clinical Practice (Health Services and Quality), Clinical Guidelines, Patient Safety, Quality of Care, Diseases, Lymphoid Malignancies, Lymphoid Leukemias, Lymphomas, CLL, Non-Hodgkin Lymphoma, B Cell Lymphoma, Aggressive Lymphoma, Research, Clinical Research, Clinical Trials, Health Outcomes Research, Real-World Evidence, Registries, Survivorship, Study Population, Human, Adult
PAPER

1914 Predicting early CAR T-cell toxicity in DLBCL using machine learning: An internally validated CIBMTR registry study

Shatha Farhan, MD, Deniz Akdemir2, Reid Shaw3, Sowjanya Vuyyala4, Nasheed Mohammad Hossain, MD5, Loretta Nastoupil6, Roni Shouval Jr, MD, PhD7, Temitope Oloyede, MPH8, Brent Logan9, Amy Moskop, MS MD10, Cameron J. Turtle, MBBS, PhD11, Christine Phillips, MD12, Sairah Ahmed, MD13, Adrian Mosquera, MD14
1 Henry Ford Health System Stem Cell Transplant & Cellular Therapy Program, Detroit, United States
2 CIBMTR® (Center for International Blood and Marrow Transplant Research), NMDP, Minneapolis, United States
3 Loyola University, Chicago, United States
4 Henry Ford Health System Stem Cell Transplant & Cellular Therapy Program, Detroit, United States
5 Department of Medicine, Division of Hematology/Oncology - Cell Therapy and Transplantation Program, University Of Pennsylvania, Perelman School of Medicine, Philadelphia, United States
6 M.D. Anderson Cancer Center, Houston, United States
7 Memorial Sloan Kettering Cancer Center, New York, United States
8 CIBMTR® (Center for International Blood and Marrow Transplant Research), Medical College of Wisconsin, Milwaukee, United States
9 CIBMTR® (Center for International Blood and Marrow Transplant Research), Medical College of Wisconsin, Milwaukee, United States, Division of Biostatistics, Data Science Institute, Medical College of Wisconsin, Milwaukee, United States
10 CIBMTR® (Center for International Blood and Marrow Transplant Research), Medical College of Wisconsin, Milwaukee, United States, Division of Pediatric Hematology/Oncology/Blood and Marrow Transplant, Department of Pediatrics, Medical College of Wisconsin, Milwaukee, United States
11 Fred Hutchinson Cancer Center, Seattle, United States, Royal North Shore Hospital, St. Leonards, Sydney, Australia, University of Sydney, Sydney, Australia
12 Division of Bone Marrow Transplantation and Immune Deficiency, Cincinnati Children’s Hospital, Cincinnati, United States
13 M.D. Anderson Cancer Center, Houston, United States
14 Hospital Clinico Universitario de Santiago de Compostela, Santiago de Compostela, Spain

Background: Early toxicities after CD19-directed CAR T-cell therapy, including cytokine release syndrome (CRS), immune effector cell-associated neurotoxicity syndrome (ICANS), and prolonged cytopenia, remain difficult to predict before infusion. Existing risk tools have limited granularity for patient-, disease-, and CAR T-specific factors. Accurate pre-infusion risk stratification could support patient counseling, prophylaxis planning, trial enrichment, and resource allocation.

Objective: To develop and internally validate supervised machine-learning models using exclusively preinfusion variables to predict five early toxicity endpoints in patients with diffuse large B-cell lymphoma (DLBCL) receiving commercial CD19-directed CAR T-cell therapy: any CRS, severe CRS (grade >=3), any ICANS, severe ICANS (grade >=3) per ASTCT consensus grading and prolonged cytopenia defined as the lack of neutrophil engraftment (absolute neutrophil count [ANC] < 500/mm3) and/or platelet recovery (platelet count < 20 × 109/L) through Day 30 post-infusion.

Methods: We analyzed 5,256 CIBMTR registry patients with DLBCL treated with commercial CD19-directed CAR T-cell therapy from 2017-2023. A contemporary 2021-2023 subset of 2,935 patients was analyzed in parallel to assess performance in a more recent treatment era and to incorporate variables with lower missingness in recent years. Models used only pre-infusion patient, disease, laboratory, and CAR T-related variables. Variables with >40% missingness were excluded from the full-cohort analysis and reintroduced in the recent-era subset when missingness was substantially lower. Nine prespecified machine-learning algorithms, plus BART and AutoGluon comparators, were evaluated. Data were split into 75% training and 25% held-out test sets. Hyperparameter tuning, class-imbalance handling, and multiple imputation were performed within nested cross-validation to minimize data leakage. Post-hoc calibration and threshold selection were based on out-of-fold training predictions and then evaluated on the held-out test set. Model interpretation used SHAP-based feature attribution when available.

Results: Outcome prevalence in the full cohort and 2021-2023 subset was 75.7% and 74.8% for any CRS, 8.1% and 5.7% for severe CRS, 35.8% and 35.6% for any ICANS, 16.6% and 14.9% for severe ICANS, and 20.5% and 18.6% for prolonged cytopenia, respectively. Discrimination was modest to moderate for most endpoints. In the full cohort, the best-performing models achieved AUROC values of 0.686 for any CRS, 0.654 for severe CRS, 0.711 for any ICANS, 0.692 for severe ICANS, and 0.697 for prolonged cytopenia. In the 2021-2023 subset, corresponding best AUROC values were 0.686, 0.687, 0.675, 0.674, and 0.783, respectively. Using thresholds selected from out-of-fold training predictions, negative predictive value was high for severe CRS (96.9%) and moderate for severe ICANS (88.0%) on held-out testing. Across endpoints, CAR T product, bridging therapy, baseline LDH, hemoglobin, platelet count, and timing-related variables were among the features contributing most strongly to model predictions. Although SHapley Additive exPlanations analyses and descriptive patterns suggest potential CAR-T product differences, none of these interactions remain statistically significant after correcting for multiple comparisons.

Conclusions: In this large real-world CIBMTR registry study, machine-learning models based exclusively on pre-infusion variables showed reproducible, internally validated performance for early CAR T-cell toxicity prediction in DLBCL. Although overall discrimination was modest to moderate, the models demonstrated potentially useful rule-out performance for severe CRS and identified clinically plausible predictors across toxicity endpoints. These findings support the feasibility of pre-infusion, data-driven toxicity risk stratification, but external validation in independent cohorts is required before clinical implementation as a decision-support tool.

Shatha Farhan, MD: No relevant conflicts of interest to declare.