P1401: identification of biomarkers and risk factors for immune effector cell-associated neurotoxicity syndrome (icans) in cd19-directed car t-cell therapy: a retrospective machine learning-based analysis.

Topic: 25. Gene therapy, cellular immunotherapy and vaccination - Clinical Background: CD19-directed chimeric antigen receptor (CAR) T-cell therapy has shown great potential as a treatment for patients with relapsed or refractory diffuse large B-cell lymphoma (DLBCL) and acute lymphoblastic leukemia (ALL). However, the occurrence of immune effector cell-associated neurotoxicity syndrome (ICANS) is a severe adverse event that can be life-threatening. Aims: This retrospective study aimed to develop a predictive model for the occurrence of ICANS in real-world settings and identify potential biomarkers and risk factors for ICANS. Methods: We conducted a retrospective analysis of 97 patients with DLBCL or ALL who received CAR-T cell therapy at our center (HGUGM). We collected clinical data, baseline characteristics, correlative biomarkers, and prior treatments. Out of a total of 50 parameters, we used a machine learning linear model (LASSO model - regularized linear binary classifier) to select parameters that predict the occurrence of ICANS. The features include demographic characteristics, clinical variables, prognosis scales, genetic features, type of CAR-T cell therapy, metabolic tumor burden before CAR-T cell therapy, and analytical variables that encompass cytokine profile, flow cytometry of lymphocyte subpopulations, complete blood count, and biochemistry results at baseline and after CAR-T cell infusion. To reduce the potential bias introduced by the data partition due to the reduced sample size and ensure generalizability, we trained the model with 100 random stratified train/test splits. From the average weights associated with each feature across all splits, we selected the 15 most significant variables to obtain a practical model that could be used in real clinical scenarios. Results: In our cohort, ICANS occurred in 33 (34%) patients, 11 (33%) of whom had severe ICANS (ASTCT grades 3-4). Our model demonstrated an accurate prediction of ICANS occurrence, with a ROC AUC of 0.7674 ± 0.0752. The model showed particular efficacy in predicting ICANS development in patients with elevated levels of IL-6 on days 0 and 3 after infusion, higher levels of D-dimer on day 0 before infusion, and those who had previously suffered from microangiopathy observed by neurological screening (MRI). Additionally, CAR-T cell therapy with axi-cel and higher SUV (max) values from PET-CT scans before infusion were identified as relevant predictors of ICANS. Summary/Conclusion: Our study demonstrates the potential of using machine learning algorithms to identify patients who are at higher risk of developing ICANS and to aid in the management of this serious adverse event. The identified biomarkers and risk factors could be used to develop a predictive tool for the occurrence of ICANS and improve patient outcomes.Keywords: Prediction, Machine learning, Toxicity, CAR-T