TNF-α Increase the Risk of Bleeding in Patients after CAR-T Cell Therapy: A Bleeding Model Based on a Real-World Study of Chinese Car-T Working Party

Backgroud: Chimeric antigen receptor (CAR) T-cell therapy has shown excellent clinical ecacy in patients with hematologic malignancies. However, severe bleeding after this treatment is a life-threatening complication for most patients. Objectives: This study evaluated the risk factors associated with bleeding in CAR-T treatment and developed a predictive model for this complication. Methods: We conducted a real-world study incorporating data from 400 patients with hematologic malignancies treated with CAR-T between November 1, 2015 and September 1, 2019. Also 39 patients from another hospital were selected for external validation. Results: Patients with severe bleeding (hazard ratio [HR] 13.04, 95% condence interval [CI] 5.82-29.18; p<0.001) had a higher risk of death after CAR-T. Stage III and IV cytokine release syndrome (CRS) (odds ratio [OR] 6.07, 95% CI 2.35-16.76; p<0.001) and higher tumor necrosis factor-α (TNF-α) levels (OR 4.00, 95% CI 1.53-11.35; p<0.001) were independent factors of bleeding in patients after CAR-T treatment. The predictive model developed by Lasso regression, which selected factors such as CRS period, transfusion volume, platelet percentage, platelet count, thrombinogen time, interleukin 6 and TNF-α levels, and showed Nomogram, yielded excellent agreement (C-statistics=0.905) with the calibration curve, which improved clinical benet with respect to established bleeding scores such as outpatient bleeding risk index (MOBRI). External validation was performed using 39 patients from another hospital with an AUC of 0.700. Conclusions: Patients with severe bleeding after Car-T therapy had increased the risk of death. A cross-validated bleeding risk score based on CRS stages and TNF-α level show signicant prognostic value in patients undergoing Car-T treatment.


Introduction
Advances in chimeric antigen receptor (CAR) T-cell therapy have revolutionized the treatment of hematologic malignancies, especially in relapsed and refractory patients [1]. Anti-cell differentiation 19 (CD19)-directed CAR T-cell therapy is effective and safe in non-Hodgkin lymphoma (NHL) [2] and is also useful in patients with relapsed and refractory acute lymphoblastic leukemia (ALL) [3,4]. CAR T-cell therapy has also shown e cacy in patients with acute myeloid leukemia (AML) [5] and multiple myeloma (MM) [6]. However, associated toxicity, particularly cytokine release syndrome (CRS) and hemorrhagic events, limit the prognosis of patients receiving this treatment [7]. Hemorrhagic diseases including partial thromboplastin time (PTT) [3,4] and prolonged thromboplastin time (PT) (PT) [3], disseminated intravascular coagulation (DIC) [8], and brinogen reduction often occur during treatment [3,6]. Bleeding events associated with CAR T-cell therapy can lead to patient death [9], although these events have a multifactorial cause and the contribution of CAR T-cell infusion is unknown.
CRS is an in ammatory process that occurs as a result of high levels of immune activation [10], and the level of immune activation required for modern immunotherapy to mediate clinical bene t exceeds that which occurs in more natural settings [11], In ammation-related hemostasis occurs in many diseases [12].
In ammation and severe infection inevitably cause hemostatic abnormalities, ranging from inconspicuous laboratory changes to severe DIC [13]. In ammation and coagulation as a response to acute infection manifests itself in the most extreme forms of multiple organ failure and DIC, the correlation of which is becoming increasingly clear [14,15]. Cytokines, especially tumor necrosis factor-α (TNF-α) and interleukin-6 (IL-6) are increased during CRS [16], which may be associated with in ammation-related hemostasis. However, the association between CRS, TNF-α and bleeding during CAR T-cell therapy is unclear.
To determine the role of bleeding in patients receiving CAR T-cell therapy and to elucidate risk factors for this complication, we conducted a single-center real-world study in a Chinese population. The prognostic value of bleeding was determined using multivariate adjusted survival analysis. Subsequently, we obtained a multivariate scoring model to predict the prognostic value of CAR T-cell therapy-related bleeding, which was validated using a logistic regression model. In addition, the prognostic value of the bleeding risk score was also con rmed. The validation cohort included a separate series of 39 patients from another hospital.

Study design and participants
We conducted a real-world study incorporating data from 400 patients with hematologic malignancies who underwent continuous CAR T treatment at the First A liated Hospital of Suzhou University (Jiangsu, China) between November 1, 2015 and January 1, 2019. The end date of the follow-up visit is September 1, 2019. The validation cohort consisted of an independent series of 39 patients from Hongci Hematology Hospital. Patient data were collected in accordance with the standards for registration tracking (NCT03919240). Patients who were on antithrombotic therapy, had severe renal or hepatic insu ciency or hemorrhagic disease, lacked information on bleeding complications or had no follow-up information were excluded. A owchart of the patient enrollment process is shown ( Figure 1). Electronic medical records are used to obtain demographic variables. Gender; sex; age; diagnosis; percentage of vesicle cells in bone marrow; type of CAR T cell therapy; application of IL-6 knockout; hypertension and diabetes status; CAR T cell count; CRS staging; use of hematopoietic stem cell transplantation (HSCT); e cacy (complete remission); transfusion application; DIC incidence; platelet count before and after CAR T cell therapy; white blood cell count; hemoglobin levels. Active partial thrombinogen time (APTT); thrombinogen time (PT); thrombinogen time (TT); IL-2, IL-4, IL-6, IL-6, IL-10, IL-17A, TNF-α, and interferon gamma (IFN-γ) levels. Bleeding risk was estimated from modi ed outpatient bleeding risk index (MOBRI) and levels of hypertension, abnormal renal/hepatic function, stroke, history or susceptibility to bleeding, IL-2-4, IL-6, IL-10, TNF-17A, TNF-α and IFN-γ; as well as bleeding site and bleeding grade. Details of the above variants are given in Part III of the supplementary information. The mortality data are derived from the hospital death registers, and the time of death is con rmed through access to electronic medical records or telephone follow-up visits. The median follow-up time was 9.2 months. The creation of the database was carried out by an independent researcher who was not involved in the care of the patients.
Informed consent was obtained from all patients or their immediate family members. All research programmes are in line with the guidelines of the Ethics Committee of Soochow University and follow the Declaration of Helsinki.

CRS grading system
The CRS stage was evaluated by the CAR T-cell therapy-associated TOXicity Working Group [17]. Grade 1 organ toxicities were: temperature ≥ 38.0°C, systolic blood pressure ≥ 90 mmHg, and arterial oxygen saturation > 90%. Grade  with neurologic signs and symptoms, hemorrhage associated with hemodynamic instability (hypotension, systolic or diastolic blood pressure changes >30 mmHg), and fatal hemorrhage of any origin. Bleeding levels 0 to 2 were classi ed as low bleeding levels, while levels 3 and 4 were classi ed as high bleeding levels [18].
Assays for plasma cytokines Peripheral blood specimens from patients treated with CAR T cells were collected and all specimens were treated with EDTA as an anticoagulant. The peripheral blood specimens were then centrifuged at 800×g for 10 min at room temperature to collect plasma. Plasma levels of IL-2, IL-4, IL-6, IL-10, IL-17A, TNF-α and IFN-γ were determined by ow cytometry using the corresponding antibodies (BD Biosciences, Franklin Lakes, NJ, USA) according to manufacturer's instructions. Details of this antibody are given in the supplementary material. Flow cytometry was performed using the BD FACSCalibur system (BD Biosciences).

Statistical analysis
Sample size assessment was performed using NCSS-PASS software (https://www.ncss.com/software/pass/) version 11.0. Power was set as 0.90 and alpha was 0.5. Mortality rates (0.40 and 0.029) for the severe and non-severe bleeding groups from previous data were entered into PASS. the actual hazard ratio was set to 10. sample size was then calculated using PASS and the minimum sample size was found to be 233 (control group = 203 and experimental group = 30).
Our sample size of 400 (363 in the control group and 37 in the experimental group) is more appropriate. The sample size assessment report (Supplementary Material Part II) was also presented. Missing data were estimated and a random forest algorithm using the mouse package in RStudio (R version 3.6.1).
Continuous variables with skewed and normal distributions are expressed as median to interval mean and mean ± standard deviation. The Mann-Whitney U test and unpaired t test were used for intergroup comparisons. Categorical variables were expressed as percentages and compared using the κ² test. Cumulative mortality was shown using Kaplan-Meier curves and analyzed using the log-rank test.
Univariate and multivariate survival analyses for total survival (OS) were assessed using Cox regression models. The signi cance of covariates on prognosis was visually analyzed using forest plots. Restricted stereo spline analysis was performed using Harrell's Regression Modelling Strategies (rms) package.
To build a bleeding risk model, Lasso regression was used to identify factors associated with bleeding. The contribution of each covariate was quanti ed and presented in the form of a Nomogram plot with 1000 self-directed internal validations. The consistency of the model created was assessed using a calibration assay. The net clinical bene t of the model compared with traditional bleeding scores was evaluated using decision curve analysis (DCAs). The coordination of each model was visualized using scatterplots and analyzed using 1000 bootstrapping. The association between bleeding grade and survival endpoint was analyzed using Kaplan-Meier curve and log-rank test. Statistical analyses were performed using RStudio (R version 3.6.1) with rms, risk regression, ggplot2, PredictABLE and surminer packages.

Results
A total of 400 patients receiving CAR T-cell therapy were enrolled in the study. The median age was 44 years (26- patients achieved complete remission after receiving CAR T-cell therapy (Table 1).
Severe bleeding as a fatal event after CAR T-cell therapy has shown signi cant prognostic value for multiple outcomes. Successful prediction may help prevent bleeding and suggests an improved prognosis after CAR T-cell therapy. Therefore, we modeled the association of multivariate risk scores with severe bleeding. Univariate analysis showed that stage III and IV CRS, hemoglobin < 100 g/L, platelets < 30 × 109, APTT(s) ≥ 31.4, PT(s) ≥ 12.2, CRP(mg/L) ≥ 23.325, IL-2(pg/mL) ≥ 4.875, IL-4(pg/mL) ≥ 2.42, IL-6(pg/mL) ≥ 15.09, IL-10(pg/mL) ≥ 5.91, TNF-α(pg/mL) ≥ 2.12 were risk factors for severe bleeding (Table 3). When adjusting for model-1 (age and sex) and model-2 (age, sex, HSCT, IL-6 knockout, BLAST percentage, CAR-T type, diagnosis, CAR-T cell count and complete remission), TNF-α and CRS stages were also risk factors for CAR-T treatment-related bleeding ( Table 2). Multivariate regression was statistically signi cant (P < 0.05) for all variants, showing that stage III and IV CRS (OR 6.07, 95% CI 2.35-16.76; P < 0.001) and high TNF-α levels (OR 4.00, 95% CI 1.53-11.35; P < 0.001) were independent factors of bleeding in patients after CAR T treatment (Table 3). Without assuming a linear relationship, a Logistic regression model was constructed using restrictive cubic splines to determine whether TNF-α levels affect the risk of severe bleeding and to show age-and sex-adjusted OR for severe bleeding, and TNF-α levels were found to be linearly correlated with increased risk over the entire range of TNF-α level values (Fig. 3). Next, we selected CRS stage, transfusion volume, platelet percentage, platelet count, PT, IL-6, and TNF-α levels to build predictive models by multivariate Lasso regression and Logistic regression ( Figures S1A,  1B). To quantify the contribution of each covariate to the severe bleeding grade, a logistic model of the nomogram was generated, as shown in Fig. 4A. The novel bleeding scores derived from this model showed good correction (C-statistic = 0.905) (Fig. 4B), and the model conveyed a greater clinical bene t compared to established bleeding scores such as MOBRI ( Figure S2A) and HAS-BLED (19). In addition, clinical impact curves for patients with hematologic malignancies after CAR T-cell therapy also showed higher accuracy of our new bleeding model compared to the HAS-BLED and MOBRI models ( Figure S3A-C).

External validation of nomogram for Car-T associated bleeding
Nomogram was externally validated by the calibration plots in Fig. 5 and by calculating bootstrap C statistics in an independent validation cohort of 39 patients. In the external validation step, Nomogram predicted a C index of 0.700 for Car-T-related bleeding (Fig. 5A), indicating that the model is well discriminated. The calibration curve shows that the nomogram was well calibrated; the 5-year OS shows optimal agreement between the actual observations and the nomogram predicted values (Fig. 5B).

Discussion
This real-world study demonstrates the prognostic value of hemostatic disorders in patients treated with CAR T cells. A total of 11.0% of patients developed bleeding events after receiving CAR T-cell therapy.
Severe bleeding levels were associated with a signi cant increase in overall mortality. In addition, severe bleeding grade was independently associated with adverse outcomes and had signi cant authority in identifying high-risk patients. A multivariate bleeding risk model, including CRS stage, transfusion, platelet percentage, platelet count, PT, IL-6 and TNF-α levels, successfully predicted the risk of severe bleeding in patients receiving CAR T-cell therapy.
Bleeding events are associated with poorer prognosis in patients treated with CAR T cells. Despite preliminary ndings from cross-sectional studies and case reports, only a few studies have suggested a prognostic role for bleeding in patients with hematologic malignancies after CAR T-cell therapy. Maude et al [9] conducted a phase 2, 25-center, single cohort global study in 75 young adult and pediatric patients with CD19 + relapse or refractory B-cell ALL. In their study, patients were found to develop cerebral hemorrhage during CAR T-cell therapy and this hemorrhagic event resulted in death.Turtle et al [19] Another clinical trial of CD19 CAR T-cell therapy applied to 32 adult patients with relapsed and/or refractory B-cell NHL showed that 2 patients died of pelvic and gastrointestinal hemorrhage, respectively.Locke et al [20] showed that 1 out of 7 patients treated with CAR T-cells died of intracranial hemorrhage in the multicenter ZUMA-1 phase 1 study.
Our study included a large sample size and recruited only Chinese, which is an advantage because the ndings can be applied to a speci c ethnic group. Interestingly, the results of the Cox regression model showed that patients with severe bleeding had a poorer prognosis compared to those without this complication. Even though bleeding is not the direct cause of death in some patients, its malignant role in overall mortality clearly leads to a poor clinical prognosis. Considering the incidence and mortality of severe bleeding events in patients receiving CAR T-cell therapy, effective control of hemostatic disorders may be a critical step in improving prognosis. Although prophylactic infusion of platelets and supplemental clotting factors can improve bleeding risk, a subset of patients with severe bleeding are somewhat resistant to conventional therapy and are prone to an increased risk of death [21]. Thus, identifying risk factors for bleeding, in addition to thrombocytopenia and coagulation dysfunction, may shed new light on bleeding care and improved prognosis after CAR T cell therapy.
We used the machine learning method Lasso regression to screen for factors associated with severe bleeding, using a machine learning approach to control for collisionality [22], multivariate logistic regression of 400 patients showed that CRS and high TNF-α concentrations were independent risk factors for severe bleeding in patients with hematologic malignancies after CAR T-cell therapy.CRS is an in ammatory process. Acute in ammation and infection almost always result in hematologic abnormalities, ranging from inconspicuous laboratory changes to severe DIC [7], and it has long been known that in ammation can lead to activation of the coagulation system. Acute in ammation as a response to a serious infection or trauma can lead to systemic activation of the coagulation system and bleeding [23]. Endotoxins are lipopolysaccharide compounds of Gram-negative bacteria that induce sepsis syndrome and DIC [24]. In addition to the direct positive effect of endotoxins on tissue factor (TF) synthesis, synthesis of cytokines IL-1, TNF-α, and IL-6 all stimulate TF formation [25]. An in vitro study showed that TF mRNA production was rapidly induced in blood cells after injection of endotoxin into human volunteers [26] and in vivo, whole blood cells produced TF after incubation with endotoxin or IL-1, TNF-α and IL-6. in vitro, cultured human endothelial cells also induced TF after incubation with TNF-α and IL-1, but their role in DIC remains unknown. To date, the only direct evidence of endothelial cell involvement in TF generation is the presence of TF-expressing circulating endothelial cells in patients with sickle cell disease (a possible DIC-related disease) [27]. Studies in primates have revealed the molecular mechanisms of endotoxin-induced coagulation activation [28]. Rapid production and release of proin ammatory cytokines was observed following an intravenous endotoxin challenge.
TNF-α [29] and IL-6 [30] are important factors in inducing brinolysis and procoagulant changes in the blood. Our study highlights the role of CRS and proin ammatory cytokines during CAR T-cell therapy, and our cohort was signi cantly different from other studies on bleeding risk factors. Notably, our study not only incorporated a larger sample size, but also underwent cross-validation and internal validation using a bootstrap sampling method. In addition, we con rmed the results of Logistic regression using multivariate models and explored the clinical signi cance using DCA to assess the prognostic value of bleeding scores [31]. Future multicenter studies, including different ethnic groups, larger sample sizes, and various treatment options, are needed to identify additional risk factors associated with post-CAR T-cell therapy bleeding. Targeted control of these factors may provide new insights into the active control of bleeding and provide new information about the survival bene ts for these patients.
Bleeding as a fatal event after CAR T-cell therapy requires early prediction and intervention to reduce subsequent adverse events. To the best of our knowledge, no bleeding scores have been proposed for patients receiving CAR T-cell therapy. In contrast, bleeding scores, including MOBRI and HAS-BLED, have been established as a low-cost, convenient tool for cardiovascular care. The predictive role of these scores in CAR T-cell therapy-related bleeding remains unknown, considering the differences between patients receiving CAR T-cell therapy and those with cardiovascular disease. Through internal guidance, we developed a new bleeding score consisting of 7 factors associated with this therapy, which in our study has a superior estimated clinical bene t in predicting bleeding risk compared to other bleeding scores. The new score not only predicts bleeding, but also has a prognostic effect on OS in these patients. Our results highlight the need for a prognostic model for bleeding in patients receiving CAR T-cell therapy.

Conclusions
In conclusion, our results show that CAR T-cell therapy-related bleeding predicts worse OS in patients.
Bleeding grade provides these patients with additional predictive power beyond other factors. A crossvalidated multivariate score showed good agreement between bleeding grade and e cacy in predicting long-term prognosis in patients receiving CAR T-cell therapy.

Consent for publication
Written informed consent for publication was obtained from all participants.

Availability of data and materials
All the data and materials are available if necessary.