Intelligent immunotherapy: AI-driven optimization of CAR-T therapy in hematologic malignancies

Chimeric antigen receptor (CAR)-T cell therapy has transformed outcomes in selected hematologic malignancies, yet its delivery remains constrained by variable efficacy, acute and prolonged toxicities, and complex manufacturing logistics. Artificial intelligence (AI) is being applied across the CAR-T continuum at three distinct levels of evidentiary maturity. Externally validated parsimonious models built on routine clinical and laboratory data including CAR-HEMATOTOX, endothelial activation and stress index (EASIX) and m-EASIX, MyCARe, and gradient-boosting cytokine release syndrome predictors provide deployable risk stratification today. Methodologically sophisticated tools using deep-learning positron emission tomography (PET) segmentation, attention-based digital-pathology analysis, and multimodal fusion of imaging with laboratory and circulating tumor DNA features have demonstrated technical promise but generally lack the multi-center external validation required for clinical-grade deployment. Forward-looking applications, including digital-twin frameworks, large language models, and closed-loop adaptive therapy, remain at the prototype stage and have not yet been validated for CAR-T. This review synthesizes these three tiers, summarizes 18 published AI and statistical-modeling tools in a comparative performance table, and compares data characteristics and AI modeling paradigms across large B cell lymphoma, B cell acute lymphoblastic leukemia, and multiple myeloma. We also identify the implementation realities—absent regulatory clearance, demographic and racial bias in training data, and post-deployment model drift—that condition the field's evolution. The contribution of AI to CAR-T therapy, as the evidence currently stands, is the calibrated quantification of established risk factors rather than the discovery of novel biology, but the methodological and data infrastructure foundations are now in place for the broader transformation that the next decade may enable.