Predictive models for hematologic toxicity after CAR-T therapy: A systematic review of current risk scores and unmet AI needs

Background: Hematologic toxicity, including early and prolonged neutropenia, anemia, and thrombocytopenia, is among the most common and clinically significant complications after CD19, CD20, CD22 and BCMA-directed CAR-T therapy. Several clinical scoresmost notably CAR-HEMATOTOX (CAR-HT) and the refined ALL-HThave been developed to predict post-infusion cytopenias, but their comparative performance across diseases has not been systematically evaluated. No prior review has summarized existing models or assessed the need for AI-based predictive approaches. Methods: We conducted a systematic review of studies that developed or validated predictive models or risk scores for hematologic toxicity after CAR-T therapy in hematologic malignancies. Eligible studies included clinical, cytokine-based, or composite predictors and reported discrimination metrics such as area under the curve (AUC). Extracted variables included CAR-T product, target antigen, disease type, predictors used, model type, validation strategy, and performance statistics. Results were synthesized qualitatively due to heterogeneity in model structure and outcome definitions. Results: Studies on the CAR-HT score in B-cell lymphomas and multiple myelomas reported AUC values of 0.89 and 0.82, respectively, for predicting severe or long-term neutropenia after CD19 and BCMA-directed CAR-T therapy. Studies on the ALL-HT score (a modified/refined version of CAR-HT replacing ferritin with bone-marrow burden to improve discrimination in B-ALL) reported AUC 0.84-0.90 for predicting prolonged neutropenia and worse overall survivability (OS) after CD19/CD22 CAR-T. A multicenter study including aggressive/indolent NHL, MM/PCL, and ALL patients treated with CD19, CD20, or BCMA CAR-T reported AUC 0.87 (eIPM-Pre) and 0.88 (eIPM-Post) for predicting ICA-HT. A CD19 CAR-T study in B-ALL and LBCL reported AUC 0.85 for an early hematotoxicity nomogram using TNF- and CRP. Conclusion: Existing prediction tools for hematologic toxicity after CAR-T therapy demonstrate strong but disease-specific performance. Yet inconsistent endpoints and heterogeneous predictors limit clinical generalizability. No current model incorporates contemporary AI or machine-learning methods, underscoring a significant unmet need. Advanced predictive frameworks may allow earlier identification of patients at risk for severe or prolonged cytopenias and guide more individualized supportive care, with the potential to reduce downstream complications.