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

Background Chemotherapy for acute myeloid leukemia (AML) frequently induces myelosuppression, leading to treatment delays and life-threatening complications. Current assessment methods, based on WHO grading standards, are not AML-specific and overlook individual patient dynamics. Accurate pretreatment risk prediction is essential for personalized therapy. Methods We propose MM-AI-AML, a novel two-stage framework integrating mathematical modeling (MM) with artificial intelligence (AI) to predict the severity of chemotherapy-induced myelosuppression using pre-treatment data including 51 features from blood routine tests, clinical characteristics and laboratory tests. First, A dynamic model was constructed to simulate postchemotherapy changes in four key blood cell types, creating a quantitative indicator for accurate myelosuppression severity assessment. This indicator was then used to label patient cases as ’severe’ or ’mild’ myelosuppression, providing the ground truth for training the deep learning model. We augmented real-world data (479 AML patients) with 900 virtual patients generated via a radial basis function network. Second, using TabNet, a deep learning model tailored for tabular data, we built a predictive classifier and identified key contributing features. Model performance was validated across training, internal, and external cohorts, and compared with other classifiers. Results This study involved 1,379 patients, including 479 AML patients from Zhongnan Hospital of Wuhan University and 900 virtual patients. Our model achieved high predictive accuracy, outperforming traditional classifiers. Patients were stratified into severe and mild myelosuppression groups, showing distinct clinical outcomes. Key predictive features included SA, A/G ratio, and LDH. High-risk patients exhibited significantly reduced in-hospital survival times. Conclusions MM-AI-AML combines mechanistic modeling with interpretable AI to enhance risk prediction in AML. This approach enables more precise, personalized treatment planning and has potential applications in managing chemotherapy-related complications across various cancers.