Improving the ICAHT grading criteria using time-series clustering.

7018 Background: Hematotoxicity is a major cause of morbidity after CAR T-cell therapy. We compared the performance of unsupervised time-series clustering versus the EHA/EBMT immune effector cell-associated hematotoxicity (ICAHT) grading system in describing ANC recovery after CAR T-cell therapy. Methods: Adults who underwent CAR T-cell therapy for hematologic malignancies with commercial or investigational products at our center between 2013 and 2023 were included (n = 602). Log10-transformed ANC trajectories were clustered using non-supervised longitudinal k-means based on Euclidean distances (latrend and kml packages in R). Overall survival (OS) was modeled using Cox regression. Results: The most common disease types were aggressive NHL (n = 293; 49%), indolent NHL (n = 150; 25%), and ALL (n = 94; 16%). The most common CAR T-cell products were the investigational CD19 CAR T-cell product JCAR014 (n = 197; 33%), axi-cel (n = 129; 21%), and liso-cel (n = 73; 12%). Median follow-up was 31 months (IQR, 11-60). The longitudinal ANC data clustered into 4 distinct trajectories (“clusters”): 1) very good (high nadir followed by rapid recovery), n = 460 (76%); 2) good (low nadir followed by rapid recovery), n = 109 (18%); 3) poor (low nadir followed by intermittent recovery), n = 20 (3%); 4) very poor (aplastic phenotype), n = 13 (2%). Grade 1, 2, 3, and 4 ICAHT occurred in 319 (53%), 96 (16%), 60 (10%), and 35 (6%) patients, respectively. Clusters were more strongly associated with OS (poor vs. very good, HR = 3.30, 95% CI, 1.88-5.80, p < 0.0001; very poor vs. very good, HR = 11.5, 95% CI, 6.45-20.4, p <0.001), compared to ICAHT categories (grade 3 vs. 0-1, HR = 2.00, 95% CI, 1.43-2.81, p < 0.0001; grade 4 vs. 0-1, HR = 4.64, 95% CI, 3.10-6.96, p < 0.0001). When evaluating concordance of clusters versus ICAHT grades, we found that 65% of grade 3-4 ICAHT patients were in the very good (17%) or good (48%) clusters; they exhibited distinct and more favorable patterns of ANC recovery (e.g., stable ANC recovery by day +15) compared to poor/very poor clusters. Among grade 3-4 ICAHT patients, those in the good/very good clusters had longer median OS than those in the poor/very poor clusters (9 vs. 3 months, p < 0.001). This led us to modify the ICAHT criteria (mICAHT). When patients with grade 3-4 ICAHT and good/very good recovery were re-classified as grade 2 mICAHT, we observed greater distinction and improved model fit in median OS between mICAHT grades compared to original ICAHT grades. The median OS of grade 3-4, 2, and 1 mICAHT patients was 3, 19, and 35 months, respectively, compared to 7, 35, and 35 months using the original ICAHT criteria. Conclusions: Unsupervised time-series clustering identified patterns of ANC recovery not captured by the ICAHT grading system and more strongly associated with OS than ICAHT grades. Incorporating these distinct patterns of recovery into a modified ICAHT grading system improved predictions of OS compared to the original ICAHT grades.