Predictive Models for Toxicities after CAR T-cell Therapy: Challenges and Opportunities

Chimeric antigen receptor (CAR) T-cell therapy is increasingly utilized with expanding indications beyond hematologic malignancies. Here, we review existing models developed for predicting toxicities in the CAR T-cell setting and identify both strengths and challenges emerging with their application. Predictive modeling approaches offer potential to guide risk stratification and inform clinical decision-making, but small sample sizes, overfitting, and poor data quality have limited model reproducibility and widespread adoption. As utilization of CAR T-cell therapy broadens, identifying additional biomarkers, developing context-specific models, standardizing guidelines for emerging toxicities, and leveraging federated learning to promote collaborative data sharing will be critical.

Predictive models integrating biomarkers and clinical variables are increasingly used to forecast potential toxicities after CAR T-cell therapy. However, due to heterogeneity in patient populations and cellular therapy products, the rapidly evolving nature of the field, and continued advancements in management of inflammatory toxicities, modeling in CAR T-cell therapy faces significant challenges. This comprehensive review of existing/emerging models serves to delineate components of developing predictive models including discrimination, calibration, biomarker integration, validation, and mitigation of overfitting while highlighting strengths, opportunities for improvement, and future directions applicable to CAR T-cell therapy.