Development and validation of predictive models of early immune effector cell-associated hematotoxicity (eIPMs)

• Disease type, baseline cytopenia, lymphodepletion intensity, and systemic inflammation were strongly associated with severe early ICAHT.

• We validated two prediction models of grade 3-4 ICAHT with near-perfect calibration and high discrimination (AUROC: 0.87-0.88).

Immune effector cell-associated hematotoxicity (ICAHT) is associated with morbidity and mortality after chimeric antigen receptor (CAR) T-cell therapy. To date, the factors associated with ICAHT are poorly characterized, and there is no validated predictive model of ICAHT specifically. Therefore, we performed comprehensive univariate analyses to identify factors associated with severe (grade 3-4) early ICAHT (eICAHT) in 691 patients who received commercial or investigational CAR T-cell therapy for hematologic malignancies. In univariate logistic regression, pre-infusion factors associated with severe eICAHT included disease type (acute lymphoblastic leukemia), pre-lymphodepletion (LD) blood counts including absolute neutrophil count (ANC), lactate dehydrogenase (LDH), and inflammatory (C-reactive protein [CRP]; ferritin, interleukin-6 [IL-6]), and coagulopathy biomarkers (D-dimer). Post-infusion laboratory markers associated with severe eICAHT included early and peak levels of inflammatory biomarkers (CRP, ferritin, IL-6), coagulopathy biomarkers (D-dimer), peak cytokine release syndrome (CRS) grade, and peak neurotoxicity grade. We trained (n = 483) and validated (n = 208) two Early ICAHT Prediction Models (eIPM): eIPMPre including pre-infusion factors only (disease type and pre-LD ANC, platelet count, LDH, and ferritin) and eIPMPost containing both pre- (disease type and pre-LD ANC, platelet count, and LDH) and post-infusion (day +3 ferritin) factors. Both models generated calibrated predictions and high discrimination (area under the receiver operating characteristic curve [AUROC] in test set: 0.87 for eIPMPre and 0.88 for eIPMPost), with higher net benefit in decision curve analysis for eIPMPost. Individualized predictions of severe eICAHT can be generated from both eIPMs utilizing our online tool (available at https://eipm.fredhutch.org).