Abstract
Acute hematologic toxicity (HT) remains a critical dose-limiting complication in gynecologic cancer patients undergoing pelvic radiotherapy, particularly when combined with chemotherapy. Early prediction of severe HT could inform personalized management and minimize toxicity. We developed and validated a predictive model integrating clinical parameters and radiomic features, evaluating five machine learning approaches. Clinical data, dosimetric parameters, and pelvic bone marrow radiomic features extracted from MRI and CT images were analyzed. Feature selection was performed using LASSO and random forest algorithms, followed by comparison across multiple classification models. In the independent test set, the combined clinical and MRI-radiomics model showed superior predictive performance (AUC=0.927, accuracy=85.5%, sensitivity=92.3%, specificity=66.7%) compared to clinical-only (AUC=0.703), MRI-only (AUC=0.925, but low specificity of 38.1%), and CT-only models (AUC=0.54). The model performed notably better in patients receiving concurrent chemoradiotherapy. Key predictors included baseline hemoglobin, white blood cell count, bone marrow dosimetry, and MRI-derived texture and fat fraction features. Integrating clinical data with MRI-based radiomics provides a robust approach for predicting acute HT, potentially guiding personalized management strategies and improving safety during gynecologic cancer radiotherapy.