Machine learning-based predictive model for high- grade cytokine release syndrome in chimeric antigen receptor T-cell therapy

Introduction: The development of robust predictive models for high-grade cytokine release syndrome (CRS) in CAR-T recipients remains limited by sparse clinical trial data. Methods: We analyzed of 496 COVID-19 patients revealed that CRS plays a pivotal role in disease progression and serves as a valuable data source for understanding CRS progression. Building on this insight, we evaluated and compared the predictive performance of three machine learning models, with the ultimate goal of developing a predictive model for high-grade CRS in patients receiving CAR-T therapy. Results: Among evaluated algorithms (XGBoost, Random Forest, Logistic Regression), XGBoost demonstrated superior performance in high-grade CRS prediction. Feature importance analysis identified SpO2, D-dimer, diastolic blood pressure, and INR as key predictors, enabling development of a validated riskassessment algorithm. In an independent CAR-T cohort (n=45), the algorithm achieved impressive predictive performance for high-grade CRS prediction. Discussion: Using machine learning, we identified key clinical biomarkers strongly associated with high-grade CRS. This tool efficiently predicts progression to high-grade CRS post-onset and shows significant potential for clinical deployment in CAR-T therapy.