Early prediction of severe ICANS after standard-of-care CD19 CAR T-cell therapy using gradient-boosted classification trees.

7034 Background: Severe immune effector cell-associated neurotoxicity syndrome (ICANS) has arisen as a major complication leading to morbidity, mortality, and increased resource utilization. There is an urgent need to accurately predict the development of severe ICANS after CAR T-cell therapy. Here, we demonstrate the high performance of the XGBoost (“eXtreme Gradient Boosting”) machine learning algorithm in predicting the development of severe ICANS using commonly available laboratory and vital sign data from patients with B-cell lymphomas undergoing standard-of-care CD19 CAR T-cell therapy. Methods: We included patients who received axicabtagene ciloleucel (axi-cel) or brexucabtagene autoleucel (brexu-cel) per standard of care. ICANS was graded by CTCAE 4.03 criteria (pre-12/2018) or ASTCT Consensus Grading System (post-1/2019). Predictions were generated using XGBoost, a widely used supervised machine learning algorithm that trains ensembles of decision trees to learn iteratively from prior trees and allows for very flexible modeling (e.g. non-linear effects, complex high-order interaction effects). K-fold cross-validation was used to assess calibration and overfitting. Variables assessed included age, ferritin, CRP, LDH, IL-6, fibrinogen, platelet count, and temperature at pre-infusion, day 0, and day +3 post-infusion timepoints. Results: A total of 175 patients were included with 39 receiving brexu-cel (22%) and 136 receiving axi-cel (78%). Grade ≥3 ICANS occurred in 40 patients (23.3%). XGBoost modeling of factors demonstrated that serum ferritin level on day +3 was the most important variable predictive of grade ≥3 ICANS (accuracy gain: 0.28). Other influential factors included day 0 and +3 platelet count, patient age, day +3 IL-6 level, day 0 fibrinogen level, and day 0 and +3 CRP level, in descending order of importance. The XGBoost model maintained high discrimination (ability to distinguish high-risk from low-risk patients; mean AUROC 0.74, sensitivity 0.95, specificity 0.90). Longitudinal LOESS smoothing confirmed associations between severe ICANS and elevation in serum ferritin at day 0 (OR 3.01, 95% CI 1.41 – 6.73, p = 0.005) and day +3 (OR 5.50, 95% CI 2.31-14.5, p <0.001). Conclusions: A supervised machine learning model using XGBoost incorporating age, temperature and commonly accessible laboratory data including ferritin, CRP, LDH, IL-6, fibrinogen, and platelet count predicted severe ICANS prior to the development of symptoms with high discrimination (AUROC 0.74) in patients undergoing standard of care CAR T-cell therapy. This allows for the early identification of patients at highest risk for developing severe ICANS who may benefit from prophylactic interventions. Once externally validated, we plan to develop a user-friendly web application to generate individualized predictions from our model.