Abstract
BACKGROUND Severe hematotoxicity has been reported in a subset of patients (pts) after chimeric antigen receptor (CAR) T-cell therapy, leading to severe infections, transfusion dependency, and worse outcomes. Early identification of pts at risk for severe hematotoxicity could inform the consideration of an autologous stem cell boost, allogeneic HCT, and/or additional infectious prophylaxis. While distinct archetypal trajectories of blood count recovery have been proposed (e.g., quick vs. intermittent vs. aplastic), manual trajectory classification is subjective and labor-intensive. To address these limitations, we automated the identification of distinct trajectories of hematotoxicity by applying time-series clustering to longitudinal absolute neutrophil count (ANC) data from >400 pts treated with CAR T-cell therapy. Next, we sought to identify factors associated with the identified trajectories and to assess the predictive ability of the CAR-HEMATOTOX score (Rejeski, Blood 2021). STUDY DESIGN AND METHODS Adults ≥18 years of age who received their first infusion of CAR T cells for hematologic malignancies with commercial or investigational products at the Fred Hutchinson Cancer Center (2013-2023) were included. ANC trajectories were clustered using non-supervised longitudinal k-means based on Euclidean distances using the latrend and kml packages in R 4.1.3. We applied logistic regression to pre-lymphodepletion (pre-LD) variables to predict poor/very poor ANC recovery. Sensitivity and specificity were computed using a probability threshold based on the Youden criteria. RESULTS A total of 509 pts were identified; 106 were excluded due to insufficient data, with 403 pts included for the analysis. The most common disease types were aggressive NHL (n = 161; 40%), indolent NHL (n = 82; 20%), and ALL (n = 74; 18%). CAR T-cell products were axi-cel, n= 101 (25%); brexu-cel, n = 24 (6%); cilta-cel, n = 21 (5%); liso-cel, n = 46 (11%); ide-cel, n = 25 (6%); tisa-cel, 12 (3%); and investigational CD19 or CD20 CAR T-cell products, n = 174 (43%). As shown in Figure 1A, the ANC longitudinal data clustered into four distinct trajectories as follows: 1) high nadir followed by rapid recovery (“very good recovery”), n = 294 (73%); 2) low nadir followed by rapid recovery (“good recovery”), n = 87 (22%); 3) low nadir followed by intermittent recovery (“poor recovery”), n = 13 (3%); 4) aplastic phenotype (“very poor recovery”), n = 9 (2%). In univariate logistic regression, ALL (reference: aggressive NHL; OR = 5.43, 95% CI, 1.71-20.6, p = 0.006), pre-LD lower ANC (OR = 3.33 per log 10 , 95% CI, 1.82-6.25, p < 0.001), lower pre-LD platelet count (OR = 10 per log 10, 95% CI, 3.70-33.3), p < 0.001), and higher pre-LD disease burden as measured by lactate dehydrogenase (LDH) were associated with poor/very poor ANC recovery (OR = 16.6 per log 10, 95% CI, 4.79-59.3, p < 0.001). Next, we assessed the predictive ability of CAR-HEMATOTOX to identify poor/very poor ANC recovery in n = 356 pts with a calculable score. The CAR-HEMATOTOX score high/low showed modest discrimination (C-index: 0.65), which improved when used as a continuous variable (C-index: 0.85). As previously reported, these models had high sensitivity (>90%) but low specificity (high/low: 31%, continuous: 58%). To improve risk prediction of poor/very poor ANC recovery, we built a logistic regression model using restricted cubic splines - allowing for non-linear effects - including pre-LD ANC, platelet, Hb, LDH, CRP, and ferritin, which showed higher discrimination (C-index: 0.91), high sensitivity (88%), and higher specificity (79%) ( Figure 1B). CONCLUSION We introduce an automated, scalable, and disease-agnostic framework to analyze hematologic toxicity after CAR T-cell therapy. K-means clustering of longitudinal ANC data categorized pts into archetypal trajectories of hematologic recovery. A logistic regression model using pre-LD ANC, platelet, Hb, LDH, CRP, and ferritin showed improved discrimination and sensitivity compared to CAR-HEMATOTOX in our training set, but both models had low specificity (poor ability to “rule in” pts at risk of severe hematotoxicity) and thus low clinical utility at this time. We plan to further improve our predictive models by incorporating post-infusion biomarkers of systemic inflammation that we have previously shown to be associated with delayed count recovery after CAR T-cell therapy (Juluri, Blood Advances 2022).