Improving severity grading of chemotherapy-induced myelosuppression in AML via data-driven and model-based deep learning

Chemotherapy-induced myelosuppression in acute myeloid leukemia (AML) frequently leads to life-threatening complications, yet current assessment standards lack the specificity required for personalized risk prediction. We present MM-AI-AML, a two-stage framework merging mechanistic mathematical modeling (MM) with artificial intelligence (AI) to predict myelosuppression severity using pre-treatment clinical data. Initially, a dynamic model simulating post-chemotherapy kinetics across four blood cell lineages was developed to derive a quantitative severity indicator, providing objective labels for 479 AML patients and 900 virtual cases. Subsequently, a TabNet deep learning classifier was trained on 51 clinical features to predict risk. MM-AI-AML demonstrated robust performance, achieving AUCs of 0.85 and 0.78 in internal and external validation cohorts, respectively, significantly outperforming traditional classifiers. Key predictive features included serum albumin, A/G ratio, and lactate dehydrogenase. High-risk stratification by the model was significantly associated with reduced in-hospital survival. By bridging mechanistic insights with interpretable machine learning, MM-AI-AML enables precise, personalized clinical decision-making for managing chemotherapy-related complications in AML.