Identification and Prediction of Severe Hematologic Toxicity after CAR T-Cell Therapy Using Machine Learning-Based Time-Series Clustering

• Published in: Transplantation and cellular therapy Vol. 30; no. 2; pp. S180 - S181
• Main Authors: Liang, Emily C., Albittar, Aya, Portuguese, Andrew J., Huang, Jennifer J., Wuliji, Natalie, Wu, Qian, Reyes, Joseph De Los, Pin, Nikki, Torkelson, Aiko, Kirchmeier, Delaney R., Chutnik, Abigail, Pender, Barbara S., Hill, Joshua A., Banerjee, Rahul, Cowan, Andrew J., Green, Damian J., Gopal, Ajay K, Poh, Christina, Shadman, Mazyar, Hirayama, Alexandre V, Till, Brian G, Kimble, Erik L., Iovino, Lorenzo, Chapuis, Aude G., Otegbeye, Folashade, Cassaday, Ryan D., Milano, Filippo, Turtle, Cameron J., Maloney, David G., Gauthier, Jordan
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.02.2024
• ISSN: 2666-6367; 2666-6367
• DOI: 10.1016/j.jtct.2023.12.234

Severe hematologic toxicity is a significant complication associated with CAR T-cell therapy, leading to infections, transfusion dependency, and mortality. Using data from >400 CAR T-cell patients (pts), we hypothesized a time-series clustering-based approach could: i) automate the identification of pts with impaired absolute neutrophil count (ANC) recovery, ii) enable the identification of factors associated with ANC recovery, and iii) assess predictive models of hematologic toxicity after CAR T-cell therapy. Adults who received their first infusion of CAR T cells for hematologic malignancies with commercial or investigational products at our center (2013-2023) were included. ANC trajectories were clustered using non-supervised longitudinal k-means based on Euclidean distances. Sensitivity and specificity were computed based on the Youden criteria. 403 pts were included. The most common disease types were aggressive NHL (n = 161; 40%), indolent NHL (n = 82; 20%), ALL (n = 74; 18%), and MM/PCL (n = 44; 11%). The most common CAR T-cell products were investigational CD19 or CD20 CAR T-cell products (n = 174; 43%), axi-cel (n = 101; 25%), and liso-cel (n = 46; 11%). The longitudinal ANC data clustered into 4 distinct trajectories: 1) very good recovery (high nadir followed by rapid recovery), n = 294 (73%); 2) good recovery (low nadir followed by rapid recovery), n = 87 (22%); 3) poor recovery (low nadir followed by intermittent recovery), n = 13 (3%); 4) very poor recovery (aplastic phenotype), n = 9 (2%) (Figure 1). Pts with poor/very poor ANC recovery had significantly shorter overall survival than those with good/very good recovery (median 3 vs. 19 months, p < 0.001; Figure 2). 100-day mortality in pts with in very good/good vs. poor/very poor recovery was 62% vs. 11%, respectively (p < 0.001). In univariate logistic regression, predictors of poor/very poor recovery included disease type and inflammatory biomarkers (Table 1). Next, we assessed the ability of the CAR-HEMATOTOX score (high vs. low risk; Rejeski, Blood 2021) to predict poor/very poor ANC recovery. Due to missing data, day +0 rather than pre-lymphodepletion (LD) values were used for CRP and ferritin. The specificity and sensitivity of the CAR-HEMATOTOX were 31% and 100%, respectively. A logistic regression model using restricted cubic splines and including pre-LD ANC, peak CRP, and peak ferritin showed high discrimination (C-index: 0.91) with 74% specificity and 95% sensitivity. We introduce an automated and scalable framework that successfully identifies pts with the most severe hematologic toxicity after CAR T-cell therapy, and specifically those displaying an “aplastic” trajectory. We identified key factors associated with poor ANC recovery. A new model including peak inflammatory biomarkers showed improved performance compared to the CAR-HEMATOTOX score.