An AI model classifies risks of early relapse post–CAR T-cell therapy in a multicenter real-world population with DLBCL
Blood Adv 2025, Michelle Wang et al.
Comment by Nicola Polverelli and Giulia Losi, Unit of Bone Marrow Transplantation and Cell Therapies, Fondazione IRCCS Policlinico San Matteo, Pavia, Italy.
Wang et al. present a large, multicenter real-world (RW) analysis of 416 adult patients with diffuse large B-cell lymphoma (DLBCL) treated with axicabtagene ciloleucel (axi-cel) across the University of California Health System from 2017 to 2024. The study evaluates RW effectiveness and toxicity, and develops a machine-learning (ML) model to predict early relapse (≤6 months). Median progression-free survival (PFS) was 10.1 months and overall survival (OS) 54.4 months, with 18-month PFS and OS rates of 41.6% and 65.5%, respectively, comparable to or slightly improved over pivotal trial results. Severe cytokine release syndrome (CRS) and immune effector cell-associated neurotoxicity syndrome (ICANS) occurred in 18.8% and 32.5% of patients, respectively. Prolonged hematologic toxicity was frequent, particularly thrombocytopenia and anemia.
The ML model is a simple, interpretable decision tree using only age and six post-infusion laboratory values (LDH, ferritin, CRP, hematocrit, platelets, and prothrombin time). It achieved robust predictive performance (AUROC 0.82 in out-of-sample testing) and stratified patients into groups with significantly different PFS. Biomarker analysis also identified post-infusion LDH, ferritin, CRP and cytopenias as correlates of severe CRS/ICANS. The authors propose this model as a scalable tool for early identification of DLBCL patients at high risk of relapse following axi-cel infusion.
This study provides one of the most comprehensive RW evaluations of axi-cel to date, with several features of particular relevance for clinical practice and for transplant/CAR-T working groups across EBMT. First, the cohort is large, heterogeneous, and includes 57.5% of patients who would have been ineligible for ZUMA-1, thus reflecting the true population currently treated in Europe. Despite this, survival outcomes were comparable to clinical trial data, underscoring the robustness of axi-cel across diverse settings and supporting its continued expansion beyond ideal-trial candidates.
The detailed toxicity profiling confirms the high burden of prolonged cytopenias, an increasingly recognized determinant of morbidity, transfusion requirements, infectious risk, and timing of subsequent therapies. These findings reinforce the need for structured post-CAR-T survivorship pathways, an area where EBMT initiatives such as hematological toxicity (ICAHT) grading and long-term follow-up guidelines are particularly relevant.
The most innovative element is the ML-based decision tree. Its strengths lie in its transparency, limited variable set, and use of universally available laboratory parameters, making it potentially implementable in any CAR-T center without additional resources. Early relapse remains the major cause of CAR-T failure; a tool able to risk-stratify patients within 24 hours of infusion could help prioritize intensified monitoring, early imaging, or even accelerated consideration of consolidation/bridging strategies, including allogeneic stem cell transplantation in selected cases.
Limitations include the retrospective nature of the dataset, center-specific variability, and the fact that the model was developed using only one commercially available CAR-T product, which may limit its broader applicability. Prospective validation—ideally within large cooperative clinical networks such as the EBMT—remains essential. Moreover, the model predicts early relapse but does not incorporate key biological and kinetic parameters such as tumor burden, CAR-T expansion and persistence, or dynamic response assessments, all of which could further refine risk stratification. Overall, this study represents an important step toward integrating artificial intelligence (AI)-driven decision support into CAR-T practice and highlights actionable RW insights for clinicians treating aggressive lymphomas.