Development and Validation of an Interpretable Model Integrating Radiomics and Clinical Data for Predicting 70-gene Signature Risk in Breast Cancer: A Multicenter Study.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, The First People's Hospital of Foshan (Foshan Hospital Affiliated to Southern University of Science and Technology), School of Medicine, Southern University of Science and Technology, Foshan, China (Y.Y., Z.X.).
- The Second Clinical Medical College, Jinan University), Shenzhen, China (T.Z., H.H.); Division of Breast surgery, Department of General Surgery, Shenzhen People's Hospital (The First Affiliated Hospital, Southern University of Science and Technology), Shenzhen, China (T.Z., H.H.).
- Guangdong Provincial Key Laboratory of Malignant Tumor Epigenetics and Gene Regulation, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou, China (L.J.); Breast Tumor Center, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou, China (LJ).
- Department of Pathology, The First People's Hospital of Foshan (Foshan Hospital Affiliated to Southern University of Science and Technology), School of Medicine, Southern University of Science and Technology, Foshan, China (Q.L.). Electronic address: [email protected].
Abstract
The 70-gene signature (MammaPrint) guides risk assessment and treatment in the hormone receptor-positive/human epidermal growth factor receptor 2-negative (HR+/HER2-) early breast cancer but is limited by cost and laboratory requirements. This study aimed to develop and externally validate an interpretable MRI radiomics-clinical machine learning model to predict 70-gene signature status in patients with HR+/HER2- early-stage breast cancer. This study retrospectively enrolled patients with HR+/HER2- breast cancer from three medical centers, comprising a training cohort (n = 80 from center 1), external validation cohort 1 (n = 60 from center 2), and external validation cohort 2 (n = 62 from center 3). Radiomics features were derived from multiparametric MRI. A total of eight machine learning classifiers were assessed, with the best-performing model chosen for subsequent analysis. A radiomics-clinical model integrating radiomics and clinical variables was constructed and assessed using the area under the curve (AUC) and decision curve analysis (DCA), with interpretability evaluated by the Shapley Additive exPlanations (SHAP) method. Age and Ki-67 emerged as key clinical predictors. From the multiparametric MRI data, four radiomics features were ultimately selected for model construction. Among classifiers, the multilayer perceptron (MLP)-based radiomics model performed best. The combined model integrating radiomics and clinical variables achieved AUCs of 0.860, 0.764, and 0.792 in the training cohort and two external validation cohorts, respectively, surpassing both the clinical model (AUCs: 0.694, 0.655, 0.674) and the best radiomics-only model (AUCs: 0.773, 0.674, 0.681). DCA demonstrated that the combined model yielded better clinical net benefit, and SHAP analysis identified Ki-67 as the most influential predictor. The interpretable machine learning model integrating MRI radiomics with clinical data demonstrated promising performance in stratifying 70-gene signature risk in HR+/HER2- early-stage breast cancer, providing a potential noninvasive tool to support personalized risk assessment and guide treatment decisions.