We prospectively included adult patients with relapsed/refractory hematologic malignancies who were admitted to our center to receive CAR-T cell immunotherapy between June 2018 and October 2024. Our approach was a prospective, observational cohort study. Eligible patients received one of the following CAR-T-cell products available during the recruitment period: varnimcabtagene autoleucel, axicabtagen ciloleucel, lisocabtagene maraleucel, brexucabtagene autoleucel (all targeting CD19+ malignancies), cesnicabtagene autoleucel (ARI0002h, an academic BCMA-CAR T-cell product for multiple myeloma); and ARI0007 (an academic CD7-CAR T-cell product administered to one patient with acute myeloblastic leukemia). Prior to CAR T-cell administration, all patients received lymphodepleting chemotherapy with fludarabine and cyclophosphamide, as specified by each product. Under our non-randomized approach CAR-T products were allocated based on clinical criteria, product availability, and disease indication.

Clinical data were extracted from electronic medical records. They included the following variables: age, sex, significant comorbidities, underlying hematologic malignancy, previous treatments and related complications, disease status at admission, diagnosis of sepsis during hospitalization, and occurrence, timing, and severity grading of CAR-T cell-related toxicities. Information on the management of toxicities, including response to therapy, need for additional interventions, and intensive care unit (ICU) admission, was also recorded.

Patients with active infections due to Human Immunodeficiency Virus, hepatitis C virus, or hepatitis B virus were excluded.

Blood samples were collected at predefined time points throughout the immunotherapy course: before CAR-T cell infusion (time point A); 24–48h post-infusion (B); at the time of CRS suspicion if compatible signs and symptoms were present (C); 72–96h after CRS treatment (D); at the time of ICANS suspicion if compatible signs and symptoms were present (E); and 72–96h after ICANS treatment (F). If patients did not present any toxicity, a control sample was collected 96h after CAR T-cell infusion (G). Plasma was obtained in 3.2% citrated tubes and separated by centrifugation at 3000 g within 4 hours after extraction, aliquoted, and stored at −40 °C until use. Serum was obtained in serum-separating tubes, centrifuged at 3000g, and processed on the day as part of routine examinations.

In a previously published study (11), 62 of the included CAR T-cell patients underwent biomarker analyses performed in plasma samples. Markers of endothelial dysfunction, measured by ELISA, were the soluble vascular cell adhesion molecule 1 (sVCAM-1) (Sigma-Aldrich, USA); soluble TNF receptor 1 (sTNFRI) (Biomatik, Delaware, USA); soluble suppression of tumorigenesis-2 factor (sST2); thrombomodulin (TM), and angiopoietin-2 (Ang-2) (R&D Systems, Minnesota, USA). Biomarkers of innate immunity and complement activation included Neutrophil extracellular traps (NETs), using the Quant-iT PicoGreen dsDNA Assay Kit (Invitrogen, Thermo Fisher, Massachusetts, USA) by fluorimetry (Fluoroskan Ascent FL; Thermolab Systems, Massachusetts, USA); the soluble terminal fraction of the complement system membrane attack complex (sC5b-9), determined by ELISA (Quidel, California, USA). The hemostasis and fibrinolysis imbalance biomarkers measured included plasma levels of circulating von Willebrand Factor antigen (VWF: Ag) (vWF Ag, Siemens Healthineers, Germany) and α2-Antiplasmin activity (α2-AP), by Berichrom α2-Antiplasmin Kit, in the Atellica 360 COAG coagulometer (Siemens Healthineers, Germany); ADAMTS-13 activity (A13), by fluorescence resonance energy transfer (FRET-VWF73); plasminogen activator inhibitor-1 antigen (PAI-1 Ag), by ELISA (Imubind, Toronto, Canada). Each assay was performed in one run, including all samples in the same batch, and carried out as blinded operator assays. The levels of the different biomarkers were correlated with EASIX and m-EASIX scores.

Laboratory parameters used to calculate EASIX (LDH (U/L), creatinine (mg/dL), platelet count (10⁹ cells/L)), and m-EASIX (replacing creatinine with CRP (mg/dL)) were determined within the routine tests in serum samples (Advia 2400, Siemens Healthineers, Erlangen, Germany) and retrospectively collected at the same time points. When multiple values were available on the same day, the highest values for LDH, creatinine, and CRP, and the lowest platelet count were used. As platelet transfusion are systematically registered in the electronic medical records, the pre-transfusional value was selected). Log2-transformed values of EASIX and m-EASIX were also calculated for subsequent analyses.

When a positive microbiological culture occurred in the hospitalization period after CAR T-cell infusion and was deemed sufficient to explain the clinical presentation (e.g., fever, hypotension, or hypoxia), the episode was classified as sepsis rather than CRS. Although overlap between syndromes is possible, this dichotomization was applied for analytical purposes. Patients who experienced both a toxicity episode and, later, a clearly separated sepsis episode during the same hospitalization, were included in both the toxicity and sepsis subanalyses. For each analysis, the parameters used to calculate the scores corresponded to the onset of the respective syndrome.

Treatment-related mortality (TRM) was defined as death occurring within 100 days after CAR T-cell infusion. Response to CAR-T cell treatment was assessed one month after therapy in all patients.

Finally, data from CAR-T cell recipients were compared with an independent historical cohort of 129 ICU patients (86 of whom had hematologic malignancies) admitted between 2014 and 2024 for sepsis or septic shock, following sepsis-3–2016 criteria (27). In these patients, the scores EASIX and m-EASIX were calculated retrospectively only at the time of their admission to the ICU.

Descriptive data are presented as frequencies and percentages for categorical variables, and as mean ± standard deviation (SD) or median ± interquartile range (IQR) for continuous variables, depending on the distribution. Normality was assessed using appropriate tests before selecting parametric or non-parametric methods. Univariate analyses were performed accordingly.

To evaluate the temporal evolution of EASIX and m-EASIX scores within patients across different time points during immunotherapy, paired comparisons were conducted using the Wilcoxon signed-rank test.

The association of EASIX, m-EASIX, and other clinical or laboratory variables with relevant outcomes, namely the development of severe toxicities, treatment-related mortality, and ICU admission, was assessed using logistic regression models. Binary logistic regression was used for dichotomous outcomes, while linear regression was applied for continuous dependent variables. Significant predictors were reported with corresponding odds ratios (ORs) and 95% confidence intervals (CIs). To assess the individual contribution of each component included in the EASIX and m-EASIX formulas, separate univariate logistic regressions were performed, never including the index and its component parameters in the same model to avoid multicollinearity. The Variance Inflation Factor (VIF) was calculated for CRP in models that included m-EASIX. VIFs below 5 were considered acceptable.

Correlations between EASIX, m-EASIX and the different biomarkers of endotheliopathy were analyzed with Pearson coefficients.

Receiver Operating Characteristic (ROC) curves were generated, and areas under the curve (AUCs) were calculated to evaluate the discriminatory capacity of EASIX and m-EASIX for different outcomes, including severe toxicity, ICU admission, and sepsis. Bootstrap corrections of AUCs based on 100 replicates were applied for each analysis. Brier scores, Hosmer-Lemeshow test and deviations between predicted and observed probabilities were used to complement ROC analysis of the main predictive models.

For survival analysis, overall survival (OS) was defined as time between CAR-T infusion and death for any cause. Univariate Cox proportional hazards analyses were first performed for all candidate variables impacting overall survival (OS). Those variables reaching statistical significance were included in multivariable Cox models, for which the proportional hazards assumption was verified using Schoenfeld residuals. Kaplan–Meier (KM) curves and log-rank tests were used to compare survival between categorical subgroups, excluding groups with very small sample sizes. Moreover, KM curves were also applied after patients were dichotomized into “high” and “low” EASIX or m-EASIX groups based on optimal cutoff points determined using log-rank statistics with the maxstat package in R. Statistical significance was defined as a two-tailed p-value <0.05.

All analyses and graphics were performed using R version 4.3.2. The following R packages were used: dplyr, survival, maxstat, car, ggplot2, pROC, boot, rmda, and rms. Statistical code is available from the corresponding author upon reasonable request. No formal power calculation was conducted; the sample size was determined based on all eligible patients treated at our center during the study period. Missing data were handled using the pairwise deletion method. No data imputation was applied. Outliers were retained unless linked to laboratory error.

This study was approved by the ethics committee of Hospital Clinic Barcelona (Registry HCB/2021/0608). All patients provided informed consent prior to their inclusion in the study.

De-identified data supporting the findings of this study are available from the corresponding author upon reasonable request. No public data repository was used due to privacy restrictions.

A total of 119 patients who received CAR T-cell therapy were included in the analysis. Baseline clinical characteristics at the time of inclusion, the specific CAR-T construct administered, and the incidence of associated toxicities are summarized in Table 1.

CRS of any grade was observed in 76 patients, while 23 patients developed ICANS. All patients who experienced ICANS had received the full dose of the CAR-T product in a single aliquot, and in 21 of the 23 cases, the construct administered was axicabtagene ciloleucel. Confirmed sepsis, defined as having clinical criteria of sepsis (organ dysfunction with a change in SOFA score ≥2 and infection with a clinically significant positive result in a microbiological culture) was diagnosed in 10 patients in the CAR T-cell cohort.

We assessed the correlation between EASIX and m-EASIX scores and circulating biomarkers of endotheliopathy, hemostasis imbalance, fibrinolysis, and innate immune activation across multiple time points during the immunotherapy course (Figure 1). Overall, both EASIX and m-EASIX presented a positive correlation with most biomarkers, except for ADAMTS-13 and sC5b9, which demonstrated negative correlations with the scores at certain time points. Importantly, the relationship between the scores and biomarkers varied over time, reflecting the dynamic nature of endothelial and immune activation during CAR T-cell therapy. The most relevant changes were observed at the onset of ICANS, where both scores displayed a strong negative correlation with sVCAM-1 and sC5b-9 (Figure 1E). These inverse correlations were partially attenuated following ICANS-directed treatment (Figure 1F).

Baseline EASIX and m-EASIX values varied according to patients past medical history. Patients with lymphoid neoplasms (n= 95) exhibited significantly higher EASIX and m-EASIX scores compared to those with plasma-cell dyscrasias (n=23) (mean and IQR: EASIX 1.39 ± 1.29 vs. 0.8 ± 0.56, respectively, p=0.008); m-EASIX 1.59 ± 3.99 vs. 0.63 ± 1.75, respectively, p=0.026).

Patients with only a partial response or progressive disease before CAR T-cell infusion (n=90) had significantly higher m-EASIX scores compared to those who were in complete remission or very good partial response (n=21) (1.84 ± 4.25 vs. 0.41 ± 1.48, respectively; p=0.00032). No significant differences were observed in baseline EASIX between groups.

Similarly, patients who had previously undergone autologous hematopoietic cell transplantation (auto-HCT) (n=40) exhibited lower m-EASIX values than those who had not (n=79) (1.12 ± 1.9 vs. 1.65 ± 5.67, p=0.032), whereas patients with a history of acute graft-versus-host disease (GVHD) following allogeneic HCT (n=13) had significantly lower EASIX scores than those without GVHD (n=12) (0.99 ± 1.16 vs. 2.43 ± 3.12, p= 0.026). This was not observed for m-EASIX. No differences in these scores were observed between patients with (n=3) or without previous chronic GVHD, although the small sample size limits interpretation.

There were no significant differences in baseline EASIX or m-EASIX scores with respect to the number of previous treatment lines or previous exposure to inotuzumab-ozogamicin (n=19) in patients with ALL.

No significant differences were observed in EASIX or m-EASIX values between baseline (time point A) and the immediate post-infusion period (time point B), in the overall cohort.

However, among patients who developed CRS, a significant increase in m-EASIX values was observed from time point A to time point B (1.61 ± 3.34 vs. 2.56 ± 6.1, respectively; p = 0.003247). Also, both EASIX and m-EASIX significantly increased at the time of CRS onset (point C) compared to the post-infusion values (point B) (EASIX: 1.82 ± 2.13 vs. 1.38 ± 1.45; p=4.1·10^-6; m-EASIX: 11.83 ± 22.64, p=1.14 ·10^-9.

Following the administration of CRS-specific therapy, a significant decrease in m-EASIX, but not in EASIX, was observed (m-EASIX: 11.83 ± 22.64 at point C vs. 3.16 ± 7.73 at point D; p), p=5.4 ·10^-7; EASIX: 1.82 ± 2.13 at point C vs. 2.02 ± 4.33 at point D; p=0.256) (Figure 2A). In patients who developed ICANS, neither EASIX nor m-EASIX showed significant changes at ICANS onset (point E) compared to either post-infusion (point B) or CRS onset values (point C). Nevertheless, m-EASIX, and again not EASIX, significantly decreased following corticosteroid treatment (m-EASIX: 7.1 ± 12.5 at point E vs. 1.03 ± 1.43 at point F, p=0.00038; EASIX: 1.57 ± 1.42 at point E vs. 1.47 ± 1.54 at point F, p=0.527) (Figure 2B).

Patients who did not develop CRS or ICANS did not exhibit significant changes in EASIX or m-EASIX values throughout immunotherapy (median values of EASIX were 1.06 and m-EASIX were 0.87 at all time points). Different sub-analyses were applied to assess the kinetics of the scores during immunotherapy with the CAR T-cell constructs with more casuistry. The same tendency was observed with all CAR T-cell products, as in the entire cohort, although not all changes were significant due to the limited sample size in some subgroups (Supplementary Figures S1-S4).

The predictive value of EASIX and m-EASIX was analyzed at baseline (time point A) and after infusion (time point B). For the prediction of severe toxicities, 77 patients were included, and for the ICU admission subanalysis, 118 patients were analyzed, representing those with sufficient data available for each analysis.

At baseline, patients who developed severe (grade ≥3) toxicities had significantly higher m-EASIX values compared to those who did not (4.65 ± 6.3 vs. 1.51 ± 3.1, p=0.0254). In contrast, no significant difference was observed in EASIX values (1.4 ± 1.5 vs.1.3 ± 1.1, p=0.56).

ROC curve analysis, corrected by bootstrap based on 100 replicates, showed that baseline m-EASIX outperformed EASIX and their log₂-transformed counterparts in predicting severe toxicities, with AUCs of 71% for m-EASIX (CI 95% 0.48-0.84), 68% for log₂ m-EASIX (CI 95% 0.46-0.82), 57% for EASIX (CI 95% 0.41-0.73), and 57% for log₂ EASIX (CI 95% 0.43-0.73) (Figure 3A). A baseline m-EASIX value > 3.68 had a sensitivity of 73% and specificity of 74% for predicting grade ≥3 CRS or ICANS (Figure 3A). For the prediction of ICU admission, both scores showed similar performance at time point A: AUC of m-EASIX 71% (CI 95% 0.62-0.80), AUC of EASIX 70% (CI 95% 0.57-0.82), AUC of log₂ m-EASIX 65% (CI 95% 0.41-0.77), and AUC of log₂ EASIX 0.70 (CI 95% 0.54-0.81). The m-EASIX threshold >0.69 yielded a 100% sensitivity and 38% specificity for this outcome (Figure 3B).

Univariate logistic regression at baseline revealed that CRP was the only individual score component significantly related to the development of severe toxicities (OR 1.3, 95% CI 1.01-1.6, p = 0.042). In contrast, a low platelet count was the only variable significantly associated with ICU admission (OR: 0.99, 95% CI: 0.98-0.99, p = 0.03).

At the post-infusion time point (B), m-EASIX remained significantly higher in patients who later developed grade ≥3 toxicities compared to those who did not (7.4 ± 13.4 vs. 2.3 ± 5.4, p = 0.01, while EASIX values did not differ significantly (1.4 ± 3.4 vs.1.4 ± 1.4, p=0.38). Again, and after bootstrap analysis correction, m-EASIX demonstrated superior predictive accuracy for severe toxicity compared to the other scores at point B: AUC of m-EASIX 73% (CI 95% 0.56-0.88), AUC of EASIX 57% (CI 95% 0.41-0.74), AUC of log₂ m-EASIX 70% (CI 95% 0.51-0.87), and AUC of log₂ EASIX 60% (CI 95% 0.45-0.77) (Figure 3C). After conducting performance assessments, the predictive model based on m-EASIX at time point B for ≥3 grade toxicities displayed no global miscalibration (Brier score 0.122 and Hosmer-Lemeshow test p=0.18). In addition, the average deviation error was 0.11 while the maximum error was 0.65, suggesting that the average differences of the predicted and observed probabilities were small, though a local miscalibration was observed due to the small sample size.

Moreover, m-EASIX also showed the best predictive power for ICU admission a time point B: AUC of m-EASIX 77% (CI 95% 0.66-0.87), AUC of EASIX 73% (CI: 0.61-0.84), AUC of log₂ m-EASIX 73% (CI 95% 0.53-0.85), and AUC of log₂ EASIX 72% (CI 95% 0.60-0.83). A m-EASIX cutoff >5.94 provided a 62% sensitivity and 84% specificity for the prediction of need for ICU admission at this time point (Figure 3D). After applying model refinement analyses, we did not observe global miscalibration for m-EASIX at time point B for the prediction of ICU admission (Brier score 0.14, Hosmer-Lemeshow test p=0.28). The average calibration deviation was 0.08 and the maximum deviation was 0.29, therefore local miscalibration was smaller for the ICU model compared to the ≥3 grade toxicities model.

At time point B, CRP was again the only individual parameter significantly associated with severe toxicity (OR 1.4, 95% CI 1.01-1.8, p = 0.009). In contrast, elevated LDH and CRP levels and low platelet count were independently associated to ICU admission: elevated LDH had an OR of 1.004, (95% CI 1.0002-1.007, p=0.04), elevated CRP had an OR of 1.38 (95% CI 1.13-1.68, p=0.0012), and decreased platelet count had an OR of 0.99 (95% CI 0.98-0.99, p=0.011). In the multivariant regression, only elevated CRP and low platelet count remained independently associated with ICU admission. The Variance Inflation Factor (VIF) was calculated for CRP in models including m-EASIX, yielding values of 1.30 and 1.29 for the prediction of severe toxicity and ICU admission, respectively. These low VIF values indicate mild correlation but fall well below commonly accepted thresholds for problematic multicollinearity (VIF > 5). This finding supports the fact that m-EASIX’s predictive performance is not solely driven by CRP. While the role of CRP alone for the prediction of severe toxicities and need for ICU admission was acceptable (AUC of 0.71 and 0.72 for each outcome, respectively) the values were slightly lower than for m-EASIX, although not significantly.

The performance of the scores at predicting these outcomes at early time points depending on the main groups of CAR T-cell products is shown in Table 1. Supplementary Materials.

Patients with baseline or post-infusion EASIX or m-EASIX values above the optimal cutoff thresholds had significantly worse overall survival (OS) compared to those with lower scores. This association was evident for both scores, but more significant for m-EASIX at both time points: pre-infusion (Figures 4A, B, p = 0.0011 for m-EASIX) and after infusion (Figures 4C, D, p= 7.93·10–⁶ for m-EASIX at time point B). Moreover, m-EASIX demonstrated exemplary performance as a mortality risk stratifier when subanalyzing patients with lymphoid neoplasms and plasma cell dyscrasias (Supplementary Figure S5).

In multivariable Cox proportional hazards analyses, including variables that were significant in univariate models, higher m-EASIX remained an independent predictor of poorer OS (Table 3). When probability of survival was assessed according to CAR-T type (only including the three constructs with more casuistry), no significant differences were observed (Supplementary Figure S6).

In contrast, neither EASIX nor m-EASIX demonstrated significant predictive value for treatment-related mortality (TRM), with p-values of 0.31 and 0.57 at baseline, and 0.18 and 0.65 post-infusion, respectively. Similarly, neither scores were significantly associated with early treatment response (assessed at 1 month) (p=0.28 and 0.91 at baseline; p=0.38 and 0.48 post-infusion for EASIX and m-EASIX, respectively), nor with time to relapse among patients who eventually relapsed (p=0.74 and p=0.81 at baseline; p=0.17 and 0.13 post-infusion, for EASIX and m-EASIX, respectively).

ROC analyses were performed to evaluate the diagnostic performance of EASIX and m-EASIX in distinguishing between CRS and sepsis. Patients with CRS were compared with two groups: 1) patients with confirmed sepsis following CAR T-cell infusion (n=10), and 2) a historical cohort of ICU patients with sepsis (n=119), including 86 with hematologic malignancies. The scores were calculated at the onset of symptoms suggestive of either CRS or sepsis (fever, hypoxia, hypotension), or at ICU admission in the historical cohort. In Table 2 are displayed the baseline characteristics of the patients included in this subanalysis.

m-EASIX outperformed EASIX in discriminating sepsis from CRS, with an AUC of 81% vs. 78%, respectively, although the difference between the scores was not statistically significant (p=0.28) (Figure 5).

When the analysis was restricted to severely ill ICU patients (defined as those with an APACHE-II score greater than 12 points), EASIX showed the best performance (AUC 77%), outperforming m-EASIX (AUC 62%), p = 0.0072.

Among the score components, elevated CRP (OR 1.15, 95% CI 1.10-1.21) and elevated creatinine (OR 4.31, 95% CI 1.50-12.40) were independently associated with sepsis diagnosis.