Patients who had received axicabtagene ciloleucel (axi-cel) as part of the ZUMA-1 trial (ClinicalTrials.gov number: NCT02348216) (1) and gave permission to explore the results in a post-hoc analysis represented the training cohort (n = 93). The validation cohort consisted of 121 patients treated with CD19-directed CAR-T cells (axi-cel, tisagenlecleucel/tisa-cel, or HD-CAR-1) at the University Hospital Heidelberg (n = 97) and Charité University Medicine Berlin (n = 24) from Oct. 1, 2018, through September 30, 2021.

Axi-cel and brexucabtagene autoleucel (Brexu-cel) are autologous anti-CD19 CAR-T cell products containing a second-generation CAR encoded by a retroviral vector with an scFy targeting CD19 with CD3ζ and CD28 intracellular domains that signal T-cell activation (1, 4). Tisa-cel is generated from autologous T cells transduced with a second-generation lentiviral vector to express an anti-CD19 CAR containing a CD3ζ domain and a 4-1BB (CD137) domain as costimulatory signal (2). In the HD-CAR-1 study, autologous T cells are transduced with a third-generation retroviral CAR vector encoding for anti-CD19 with CD3ζ for T-cell activation and CD28 and 4-1BB domains as costimulatory signals. HD-CAR-1 CAR-T cells were administered in variable doses ranging from 1 × 10⁶/m², 5 × 10⁶/m², up to 20 × 10⁶/m² body surface area after lymphodepletion (28). All patients of both the training and validation cohorts received lymphodepleting chemotherapy with fludarabine and cyclophosphamide in doses determined according to the manufacturer or study protocol (1, 2, 28).

The training cohort included 86 patients with DLBCL and 7 patients with PBMCL (1). In the Heidelberg cohort, patients were treated with standard of care Axi-cel, Tisa-cel, or HD-CAR-1 on the trial protocol. Axi-Cel was administered to 39 r/r DLBCL patients and three r/r PMBCL patients, while Tisa-Cel was administered to 15 r/r DLBCL patients and Brexu-Cel was given to three r/r MCL patients (1, 2). In the HD-CAR-1 study, CAR-T cells were administered to six r/r DLBCL patients, six r/r MCL patients, six r/r CLL patients, four r/r FL patients, and 15 ALL patients (28). All 24 patients (23 patients with DLBCL and one with ALL) from the Berlin cohort were treated with standard of care Tisa-Cel (2).

In the training cohort, patients received tumor lysis prophylaxis and prophylaxis for infection with pneumocystis pneumonia, herpes virus, and fungal infections according to NCCN guidelines or standard institutional practice (3). In the validation cohort, infection prophylaxis and supportive care were administered as described previously (29, 30). Specifically, all patients treated in Heidelberg received endothelial protection with pravastatin and ursodeoxycholic acid (UDCA) until day +14 (29).

According to the Declaration of Helsinki, written informed consent for all patients was obtained. Ethical approval and approvals from the local and federal competent authorities were granted from the Ethics Committee (Medical Faculty of Heidelberg University, reference number: AFmu-405/2017, October 2017).

Severity for both CRS and ICANS was graded according to Lee et al. (ZUMA1) and ASTCT consensus criteria (validation cohort) (5, 31, 32). The following dichotomization of severity of CAR-T cell therapy complications was used for the investigation of EASIX: severe (grade ≥3) CRS and/or ICANS versus neither severe CRS nor severe ICANS.

Concomitant high-grade ICANS and CRS cannot be distinguished, in particular in patients requiring mechanical ventilation. We therefore decided to combine both grade ≥3 CRS and ICANS as primary endpoint.

For the training cohort, EASIX was calculated at three different time points: EASIX before start of lymphodepletion (EASIX-pre), EASIX on the day of CAR-T administration (EASIX-d0), and EASIX on day 3 after application (EASIX-d3). In the validation cohort, EASIX-pre, EASIX-d0, EASIX-d3, and EASIX-d7 were available for analysis. Additionally, serum levels of four endothelial markers were measured in 47 patients of the Heidelberg cohort. Time of blood sampling was analogous to the EASIX measurement before lymphodepletion, as well as on day 7. Samples were stored at −80°C and ST-2, ANG2, sCD141, and CXCL8 were analyzed using commercial ELlSA (DuoSet ^® ELISA, Biotechne R&D, USA) according to the manufacturer’s descriptions.

EASIX was calculated by the formula: LDH (U/L) × creatinine (mg/dl)/platelets (10⁹ cells/L). A binary logarithm (log2)-transformed index, log2(EASIX-pre), was used for the primary statistical analysis. The EASIX derivatives were calculated by the formula: LDH (U/L)/platelets (10⁹ cells/L) (sEASIX), and by the formula: LDH (U/L) × CRP (mg/dl)/platelets (10⁹ cells/L) (mEASIX).

For the combination of both types of events (CRS and/or ICANS), an optimized univariate cutoff of EASIX-pre was estimated both by Generalized Maximally Selected (Chi-Squared) Statistics and a Conditional Inference Tree with EASIX-pre as the only splitting variable (33, 34). This cutoff was graphically confirmed by a partial dependence plot using multivariate conditional inference forests that besides EASIX-pre also includes established confounders [age, gender, disease (aggressive B-cell lymphoma vs. others) and disease stage (refractory disease and progressive disease vs. others)].

Distributions of baseline data were described by standard statistical measures, number, frequency, and Fisher tests for categorical variables and median, range (minimum/maximum), interquartiles with 25th and 75th percentiles (IQ), and Kruskal–Wallis tests for continuous data. Uni- and multivariable binary logistic regression analyses provided odds ratios (OR) for the prognostic effect of EASIX-pre (continuous log2 transformed or with cutoff) on risk of grade ≥3 complications. Covariables were age, gender, disease (aggressive B-cell lymphoma vs. others), and disease stage (refractory disease and progressive disease vs. others). As only Axi-cel was given in the training cohort, the product could not be included as confounder. These multivariate prognostic effects were validated in the validation cohort by comparing the prediction accuracy of the full model, a nested reduced model (without EASIX) and a reference model (intercept only), all fitted based on the training cohort, via Brier scores estimated on the validation cohort. Explorative multivariate models in the validation cohort used the Firth correction because of rare events in the gender distribution. ROC curve analyses (including area under the curve, AUC) were used to assess the prediction performance of EASIX scores as univariate predictors. Additionally, Pearson correlations were used for associations of (log)EASIX with (log) endothelial serum markers. Univariate logistic regression was used to assess the effect of endothelial serum markers on the risk of grade ≥3 associations.

All calculations were performed using R version 4.0.4 (R Foundation for Statistical Computing, Vienna, Austria, 2021). In all tests, a p-value < 0.05 was considered significant without corrections for multiple testing.

Training and validation cohorts were comparable in terms of age and gender. There was also no significant difference of disease status before start of the procedure, with the vast majority of patients entering lymphodepletion with unresponsive disease. Due to the inclusion of HD-CAR-1 study patients and those treated with SOC tisa-cel in the validation cohort, however, distribution of underlying diagnoses was significantly different. Detailed patient characteristics are given in Table 1 .

In the training cohort, 83 (90%) experienced CRS while ICANS was observed in 62 (67%) patients. Grade ≥ 3 CRS was detected in 9 patients (10%) with a median onset of 2 (1–12) days and high-grade ICANS occurred in 28 patients [30%; median onset 6 (1–17) days]. A total of 87 patients (93%) developed either CRS or ICANS or both (grade ≥3: 32 patients, 34%).

Of the 121 patients of the validation cohort, 70 patients (58%) developed CRS grades 1–4 and 29 patients (24%) developed ICANS grades 1–4. Higher-grade CRS (grade ≥ 3) affected 10 patients (8%) with a median onset of 4 (0–14) days, while grade ≥ 3 ICANS occurred in 14 patients [12%; median onset 6 (3–17) days]. A total of 70 patients (58%) experienced either CRS or ICANS or both (grade ≥ 3: 21 patients, 17%) ( Table 1 ).

There were no significant differences between training and validation cohorts regarding EASIX scores measured at any time point. Similar EASIX trends were measured over time in subgroups of patients with and without aggressive B-cell lymphoma ( Supplementary Figure 1A ). Age groups below and above 60 years differed in EASIX-pre in the training cohort, and in EASIX-d3 and d7 in the validation cohort only, with higher values in older patients ( Supplementary Figure 1B ). Patients with progressive or refractory disease had higher EASIX values than patients with stable or sensitive disease at any time point in the validation cohort, but not in the training cohort ( Supplementary Figure 1C ).

Prognostic value of EASIX before chemotherapy was tested in the training cohort by multivariate logistic regression including age, gender, diagnosis, and disease status as confounding variables ( Table 2 ). EASIX-pre associated with grade ≥3 CRS/ICANS after CART infusion with an odds ratio of 1.71 (p-value < 0.001). This model was validated in the independent cohort using the Brier score. The model including EASIX-pre had an improved prediction performance (Brier score 0.151) in comparison to the reference model (0.173) and the multivariable model without EASIX (0.166).

Time courses of EASIX in patients who have experienced or will experience grade ≥3 CRS or ICANS events throughout the observation period are visualized in Figure 1 . EASIX values were increased in patients of both cohorts with (future) complications at any time points. A similar picture was observed in patients with progressive or refractory disease status at lymphodepletion ( Figure 2 ), as well as in patients in remission/stable disease (not shown).

The three single EASIX parameters were analyzed individually in separate explorative models ( Supplementary Figures 2A, B and Supplementary Tables 1, 2 ). Whereas there was a correlation between LDH-pre and grade ≥3 CRS/ICANS in both cohorts, and also measurable, but inconsistent effects of creatinine and platelets determined before lymphodepletion, CRP levels at lymphodepletion had no impact.

Prognostic values of EASIX derivatives before chemotherapy were tested in the training cohort by multivariate logistic regression including age, gender, diagnosis, and disease status as confounding variables ( Supplementary Tables 3, 4 ). Both derivatives associated with grade ≥3 CRS/ICANS after CART infusion with an odds ratio of 1.63, p-value = 0.004 (sEASIX) and 1.22, p-value = 0.015 (mEASIX). Validation of univariate simplified (s)EASIX and modified (m)EASIX models calculated in the training cohort showed that both EASIX derivatives did not compare favorably to EASIX-pre in ROC analysis ( Figure 3 ). Validation of multivariable models revealed that mEASIX showed lower prediction errors [Brier score model with confounders (no EASIX) 0.184, model with EASIX 0.173, with sEASIX 0.173, mEASIX 0.155]. Confining the analysis to Axi-cel-treated patients, we calculated Brier scores in the validation cohort (n = 45) for the model with confounders 0.211, model with EASIX 0.176, with sEASIX 0.179, mEASIX 0.180.

The maximally selected Chi-Squared statistic revealed three potential univariate cutoff values for EASIX-pre in the training cohort (3.09, 4.67, and 5.73). The conditional inference tree with EASIX as the only splitting variable splits the dataset exactly once at the cutoff 4.67. Graphical validation based on the partial dependence plot of EASIX-pre in a multivariate conditional inference forest (including all other confounders as potential splitting variables as well) supported the cutoff 4.67 as the optimal value in the training cohort.

Comparing the prediction performance of a multivariate model including the derived EASIX cut point, a model with only the confounders, and a reference model with the intercept only, prediction errors were smallest in the model with EASIX cutoff 4.67 (Brier score 0.162 vs. 0.166 with confounders only, vs. 0.173 reference model).

When testing the derived cutoff in a newly fitted multivariate model, EASIX-4.67 constitutes a prognostic factor for severe complications with an OR of 4.3 (1.5–12.7, p = 0.006) ( Table 3 ), which can be interpreted as further statistical validation of the prognostic effect of EASIX-pre cut 4.67 for severe CRS/ICANS.

Looking at the two evolving subgroups in an unadjusted manner, 11/94 (11.7%) patients had severe complications in the subgroup with EASIX<4.67, compared to 10/27 (37%) patients in the high EASIX subgroup (p = 0.0071). For patients treated with Axi-cel, 7/35 (20%) developed severe CRS/ICANS in the low EASIX subgroup, whereas 6/10 (60%) developed such events in the high EASIX subgroup (p = 0.022).

Finally, patients in the EASIX-pre low-risk group (<4.67) who did not develop severe CRS/ICANS until day+7 had a very low risk of developing these complications subsequently (3/83, 3.5% in the validation cohort).

Serum markers of endothelial distress were measured in 47 patients of the validation cohort before and after CAR-T cell infusion. ST2 and sCD141 showed a weak, ANG2, and CXCL8 no correlation with EASIX-pre when assessed before lymphodepletion, but all four markers moderately correlated with EASIX-pre when assessed on day+7 ( Supplementary Table 5 ). Univariable logistic regression analyses revealed an association with severe complications for ST2-pre [OR per log2 1.49 (0.98–2.26), p = 0.061], and no association for the other markers (ANG2-pre p = 0.986, sCD141-pre p = 0.341; IL-8-pre p = 0.676). In contrast, marker values measured on day+7 were associated with severe CRS/ICANS occurring before or after day 7 for all four serum markers (ORs for doubling of marker values): ST2-d7 OR 2.9 (1.7–5.2), p < 0.001; ANG2-d7 OR 2.9 (1.5–5.2), p = 0.001; sCD141-d7 OR 13.4 (3.0–59.8), p = 0.001; IL-8-d7 OR 1.9 (1.2–2.9), p = 0.004 ( Figure 4 ).