-Author name in bold denotes the presenting author
-Asterisk * with author name denotes a Non-ASH member
Clinically Relevant Abstract denotes an abstract that is clinically relevant.

PhD Trainee denotes that this is a recommended PHD Trainee Session.

Ticketed Session denotes that this is a ticketed session.

3493 Deriving Predictive Features of Severe CRS from Pre-Infusion Clinical Data in CAR T-Cell Therapies

Program: Oral and Poster Abstracts
Session: 803. Emerging Tools, Techniques and Artificial Intelligence in Hematology: Poster II
Hematology Disease Topics & Pathways:
Research, Biological therapies, artificial intelligence (AI), Clinical Research, Chimeric Antigen Receptor (CAR)-T Cell Therapies, Therapies, Adverse Events, emerging technologies, Technology and Procedures, machine learning
Sunday, December 11, 2022, 6:00 PM-8:00 PM

Vibhu Agarwal1*, Jacob Aptekar, MD, PhD2*, David C Fajgenbaum, MD, MBA, MSc3,4 and Penelope Lafeuille, MS5*

1Medidata Solutions, New York, NY
2Medidata Acorn AI, a Dassault Systèmes Company, New York
3Center for Cytokine Storm Treatment & Laboratory, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA
4Castleman Disease Collaborative Network, Philadelphia, PA
5Medidata Solutions, NY


Cytokine release syndrome (CRS) is a life-threatening toxicity of chimeric antigen receptor (CAR) T-cell therapy that limits the widespread use of this lifesaving therapy. While only a portion of CAR T-cell therapy patients experience CRS, there is a limited understanding of its risk factors (Fajgenbaum et al. NEJM 2020). Known markers of severe CRS lack specificity or require central lab facilities, making them unsuitable for safety surveillance during trials or real-time clinical decision making. Earlier work by the authors suggests that patterns in longitudinal measurements of common laboratory markers distinguish patients who develop severe CRS after CAR-T infusion from others(Strait et al. JCO 40, 2022). Using pre-infusion clinical data from anti-CD19 CAR-T trials from the Medidata Enterprise Data Store (MEDS), we derived features that capture longitudinal patterns in common laboratory markers and vitals before and after the start of lymphodepleting chemotherapy (LDC). We fit penalized logistic regression models to these derived features using the occurrence of severe CRS (grade 3 and higher) as a binary response variable. Below we describe these novel features and their importance in predicting severe CRS following CAR-T therapy.


Subjects in the anti-CD19 CAR-T trials obtained from MEDS had relapsed or refractory disease in one of the following: B Acute Lymphoblastic Leukemia (B-ALL), Mantle Cell Lymphoma (MCL), Non-Hodgkin Lymphoma (NHL), or Diffuse Large B-cell Lymphoma (DLBCL). Laboratory markers with repeated measurements between screening and the start of LDC as well as between LDC start and CAR-T infusion that were present in all studies were used to compute the pre and post LDC features respectively. Only post LDC features were derived for heart rate, systolic blood pressure, weight and temperature (vitals) measured between LDC start and CAR-T infusion. Age, weight and ECOG at enrollment, race, number of prior therapies, response to last therapy and prior allogeneic or autologous hematopoietic stem cell transplant were included as baseline features. Since the prevalence of grade 3+ CRS varies between the studies, we split the data into training (80%) and test (20%) partitions by stratifying on grade 3+ CRS per each study. Penalized logistic regression models were fit on the training partition and evaluated on the test partition.


The dataset consisted of 361 patients: 19.1% B-ALL, 39.9% DLBCL, 21.6% MCL and 19.4% comprising patients with Transformed Follicular Lymphoma, Primary Mediastinal B-cell Lymphoma and High Grade B-cell Lymphoma; mean age 55.8 (SD=14.1) years, 69.5% males. 25.2% had CRS 3+ with a median time-to-event of 4 days. At enrollment, all patients had an ECOG score ≤ 1 with 47.7% patients having an ECOG = 0. During their last therapy prior to CAR-T cell therapy, 51.5% patients had progressive disease and 23.3% patients had a response followed by relapse.

Feature importance values in the pre and post LDC intervals are shown in Figure 1. Pre-LDC, the levels of leukocyte count, hemoglobin, aspartate aminotransferase and the change in alanine aminotransferase, albumin, and platelets were the strongest predictors of developing grade 3+ CRS. Post-LDC but pre-CAR-T therapy, the levels of alanine aminotransferase, alkaline phosphatase, hemoglobin, platelets and albumin and the change in platelets, albumin, and creatinine were the strongest predictors. High levels of Chloride, both pre and post LDC correlate with lower risk of grade 3+ CRS. Among vitals, heart rate, temperature and systolic blood pressure, as well as the rate of change in weight post LDC are the strongest predictors of grade 3+ CRS. While partial response or disease progression in prior therapy is correlated with grade 3+ CRS, the model with baseline features alone performed worst on the test partition. Adding predictors based on laboratory markers and vital signs, results in about 10% improvement in the AUROC over the baseline model (Table 1).


Absolute levels as well as the rate of change in laboratory markers and vitals may indicate immune pathway activity that correlates with a future grade 3+ CRS event. Modeling the pre and post LDC kinetics can yield novel predictors for grade 3+ CRS.The association between pre-infusion temporal patterns in clinical variables and grade 3+ CRS implies that dynamic tracking of patient status will be key to effective mitigation of grade 3+ CRS.

Disclosures: Fajgenbaum: EUSA Pharma: Consultancy, Research Funding.

*signifies non-member of ASH