From Prediction to Intervention: Artificial Intelligence for Adaptive Response and Toxicity Modeling in Cellular...

Hematologic malignancies, including acute myeloid leukemia, myelodysplastic syndromes, lymphoma, and multiple myeloma, are characterized by profound biological heterogeneity and highly dynamic treatment trajectories that conventional, static prognostic systems incompletely capture. Cellular therapies, such as chimeric antigen receptor T-cell therapy and hematopoietic stem cell transplantation, offer potentially curative options for relapsed or refractory disease, yet outcomes remain highly variable, and management decisions regarding conditioning intensity, lymphodepletion, immunosuppression, and toxicity surveillance continue to be largely protocol-driven rather than individually adapted. Artificial intelligence (AI) and machine learning (ML) have demonstrated substantial promise in diagnostic support, prognostic stratification, and multimodal data integration across hematologic malignancies, but existing models remain predominantly static and related to pre-treatment in orientation, limiting their utility for real-time clinical guidance. This review summarizes current AI applications in hematologic oncology; critically compares the strengths, limitations, and clinical applicability of major AI model classes, including traditional machine learning, deep learning, multimodal integrative frameworks, reinforcement learning, digital twins, and emerging foundation models and large language models; and proposes an adaptive, multimodal paradigm. We examine key enabling technologies and address the clinical, regulatory, ethical, and implementation challenges that must be resolved before these systems can be deployed at the bedside. We argue that the central challenge facing the field is no longer whether AI can predict outcomes, but whether it can actively guide real-time therapeutic decisions, and that achieving this transition will require interdisciplinary collaboration, prospective validation, and governance frameworks capable of ensuring interpretability, equity, and clinical trustworthiness.