Predicting Lymphopenia Based on Dosiomics in Post-operative Breast Cancer Radiotherapy

Objective This study aimed to investigate the relationship between severe radiation-induced lymphopenia (sRIL) and dosiomics in post-operative radiotherapy for breast cancer, construct a machine learning model integrating dosiomic features and clinical factors to predict the risk of sRIL.Methods We retrospectively enrolled breast cancer patients who received adjuvant radiotherapy after surgery, and whose peripheral blood lymphocyte count (PLC) before radiotherapy were essentially normal (> 0.8 × 10⁹/L). Patients were categorized into sRIL (PLC < 0.6 × 10⁹/L) and non-sRIL groups (PLC ≥ 0.6 × 10⁹/L) according to their lowest PLC during radiotherapy. They were randomly split 7:3 into training and testing sets. Clinical data including patient characteristics, tumor parameters, treatment factors, and radiotherapy plan details (including DVH parameters) were collected. Organs at risk (heart, lungs, liver, spleen, thoracic bones) were contoured on planning CT scans. And then dosiomic features were extracted from 3D radiotherapy dose distribution maps based on planning CT images. The T-test and Least Absolute Shrinkage and Selection Operator (LASSO) regression were used to select dosiomic features significantly associated with sRIL occurrence (p < 0.05), and a dosiomics score was subsequently calculated. Univariate analysis, LASSO, and multivariate logistic regression analyses were performed to identify significant clinical predictors. The identified clinical predictors were then combined with the dosiomic score to train a Random Forest prediction model. Model performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC), Precision-Recall (PR) curves, and Decision Curve Analysis (DCA).Results Twelve dosiomic features were selected from the training set for model construction. Furthermore, baseline PLC level, Bone V50, receipt of neoadjuvant chemotherapy, and myelosuppression status were identified as significant clinical predictors. The combined model (integrating clinical factors and dosiomic features) achieved AUC values of 0.937 and 0.851 on the training and testing sets, respectively, outperforming the clinical-only model (AUCs of 0.884 and 0.802). DCA further demonstrated that the combined model yielded the highest net clinical benefit across the vast majority of threshold probabilities and exhibited good stability.Conclusion This study demonstrates that integrating clinical parameters with dosiomic features into a machine learning model can significantly enhance the prediction of sRIL risk in breast cancer patients undergoing post-operative radiotherapy. This represents a promising approach to guide individualized radiotherapy planning.