CAR-T toxicity prediction is maturing, but bedside use still outpaces proof
CAR-T cell therapy has changed the treatment of several hematologic malignancies, yet toxicity prediction remains an unfinished part of the clinical workflow. Cytokine release syndrome, immune effector cell-associated neurotoxicity syndrome, and immune effector cell-associated hematotoxicity still complicate care before infusion, during the first inflammatory wave, and well into recovery. A growing number of scoring systems now stratify risk using routine laboratory and clinical variables, but most were built in retrospective cohorts and only a small subset has been externally validated across independent populations.
The central message is straightforward: prediction has improved, but prediction alone has not yet been shown to change outcomes.
Endothelial, inflammatory, and marrow-based scores drive most current risk models
The best-established tools for predicting CRS and ICANS are the EASIX-based scores. The original EASIX formula combines lactate dehydrogenase, creatinine, and platelet count, and was developed as a marker of endothelial stress in transplantation settings before being adapted to CAR-T recipients. Simplified variants remove creatinine or replace it with C-reactive protein, allowing calculations at several time points, including lymphodepletion, the day before infusion, day 1, and day 3 after infusion. In early retrospective cohorts that included diffuse large B-cell lymphoma and B-acute lymphoblastic leukemia, these indices predicted severe CRS consistently across time points, while later assessments, especially day 3, tracked more closely with severe ICANS.
The signal is biologically plausible. Endothelial injury, systemic inflammation, and evolving vascular activation sit close to the center of CAR-T toxicity biology, and the EASIX family captures that state with markers already available in routine care. In a later single-center study of large B-cell lymphoma treated with axicabtagene ciloleucel, EASIX variants incorporating ferritin or both ferritin and C-reactive protein also separated patients at risk for higher-grade CRS and ICANS. External validation in a population-based cohort of relapsed or refractory large B-cell lymphoma reinforced the score’s performance, though the predictive strength varied by toxicity endpoint and timing.
A separate m-EASIX validation extended the utility of this framework beyond lymphoma alone. In patients with non-Hodgkin lymphoma and multiple myeloma receiving anti-CD19 or anti-BCMA CAR-T therapy, the modified score outperformed standard EASIX for severe toxicity and intensive care unit admission and even helped distinguish CRS from sepsis at symptom onset. For clinicians facing fever and inflammatory instability in the early post-infusion window, that distinction can affect escalation, antibiotics, and the use of immunosuppression within hours.
Registry-derived scores sharpen risk stratification for severe CRS and ICANS
Two additional prognostic systems emerged from the DESCAR-T registry, which included 925 patients with large B-cell lymphoma treated with axicabtagene ciloleucel or tisagenlecleucel. The CRS-PSS combines bulky disease, platelet count below 150 G/L, C-reactive protein above 30 mg/L, and either no bridging therapy or stable/progressive disease after bridging. A score above 2 predicted grade 3 or higher CRS. The companion ICANS-PSS used female sex, platelet count below 150 × 10^9/L, axicabtagene ciloleucel use, and the same bridging-related variable, again with a threshold above 2 indicating higher risk of grade 3 or higher ICANS. Both tools were externally validated in United Kingdom and Spanish cohorts.
These models are not elegant abstractions. They translate disease burden, marrow reserve, and prior disease control into a near-term inflammatory forecast. Bulky disease and inadequate control before infusion appear to prime a more explosive cytokine response, while lower baseline platelets likely reflect both marrow compromise and broader physiologic fragility.
CAR-HEMATOTOX remains the most extensively validated hematotoxicity tool
Among all available scores, CAR-HEMATOTOX has the broadest validation footprint for ICAHT. It incorporates platelet count, hemoglobin, absolute neutrophil count, C-reactive protein, and ferritin, and classifies patients as high risk at a score of 2 or higher. Rejeski and colleagues first derived it in large B-cell lymphoma, then validated it in independent European and US cohorts. Subsequent studies expanded its use into mantle cell lymphoma and multiple myeloma, showing that the score travels better than many toxicity models across disease groups and CAR-T products.
Its value extends beyond neutropenia itself. High CAR-HEMATOTOX scores have been linked to more infections, particularly bacterial infections, across large B-cell lymphoma, mantle cell lymphoma, and multiple myeloma cohorts. They have also tracked with higher non-relapse mortality and inferior overall and progression-free survival in later analyses. In practical terms, a high score identifies patients who may need more intensive blood count surveillance, antimicrobial vigilance, and a lower threshold for supportive care after infusion.
An ALL-specific adaptation, ALL-HEMATOTOX, addresses a limitation of the original model in acute lymphoblastic leukemia. Because many ALL patients already have elevated ferritin at baseline, the original score loses discriminatory power. Replacing ferritin with bone marrow disease burden restored clinical separation, and disease burden correlated strongly with post-infusion neutropenia. In pediatric and adolescent/young adult cohorts, the ALL-specific model performed well across multiple validation sets and CAR constructs.
Cytokine panels can predict severe CRS, but routine use remains limited
One externally validated cytokine nomogram has also shown strong performance for severe CRS. Built from five cytokines — CX3CL1, GZMB, IL-4, IL-6, and PDGF-AA — and measured on days 1 to 2 after infusion, the model performed well in multiple cohorts, including multiple myeloma and mixed acute lymphoblastic leukemia or non-Hodgkin lymphoma populations. The limitation is obvious at the bedside: cytokine panels are not yet available in most routine clinical settings, which restricts adoption despite impressive statistical performance.
Prospective risk-adapted treatment has not yet been proven
The broad problem across the field is not a lack of models. It is the absence of prospective interventional trials showing that assigning patients to different preventive or pre-emptive strategies based on predicted risk improves outcomes. Most scores were created in retrospective datasets, often with disease- and product-specific cohorts that may not generalize cleanly to other CAR-T constructs or indications. Even where validation exists, performance can vary by endpoint, time point, and disease context.
New toxicities add another layer of uncertainty. Immune effector cell-associated hemophagocytic lymphohistiocytosis-like syndrome, tumor inflammation-associated neurotoxicity, movement and neurocognitive treatment-emergent adverse events, and local immune effector cell-associated toxicity still lack mature predictive tools. Some of these entities are emerging precisely because broader CAR-T use has exposed biological effects that older models were never designed to capture.
Clinical practice already uses these scores for triage, closer monitoring, and earlier supportive care. The next step is harder: proving that prediction-guided intervention changes the course of toxicity rather than merely describing it more accurately.