Time-to-Event Machine Learning Model Incorporating MRI-Derived Intratumoral Heterogeneity Score for Predicting Invasive Breast Cancer Recurrence: A Dual-Center Study.
IntroductionPostoperative recurrence remains a major clinical challenge in invasive breast cancer (IBC), and conventional clinicopathologic factors and routine imaging assessment may not fully capture intratumoral spatial heterogeneity or nonlinear recurrence patterns. This study aimed to evaluate the prognostic value of a magnetic resonance imaging (MRI)-derived intratumoral heterogeneity (ITH) score incorporated into time-to-event machine learning models for predicting postoperative recurrence in patients with IBC.MethodsThis retrospective dual-center prognostic prediction-model study included 428 consecutive patients with IBC from two centers. Patients were randomly assigned to a training cohort (n = 256), validation cohort (n = 86), and independent testing cohort (n = 86). ITH scores were calculated from texture features and pixel intensity distributions on contrast-enhanced T1-weighted MRI using k-means-based intratumoral subregion quantification. Multiple time-to-event machine learning (ML) models were developed to predict recurrence-free survival (RFS), and model performance was assessed using Harrell's concordance index (C-index), time-dependent receiver operating characteristic curves, calibration curves, and SHapley Additive exPlanations (SHAP).ResultsAmong the evaluated models, the random survival forest (RSF) achieved the best predictive performance, yielding a C-index of 0.826 (95% confidence interval [CI], 0.735-0.894) in the validation cohort and 0.814 (95% CI, 0.722-0.905) in the testing cohort. SHAP identified the ITH score as the most critical predictor. Higher ITH scores were significantly associated with shorter RFS and upregulation of tumor proliferation-related pathways.ConclusionsIntegrating an MRI-derived ITH score with an RSF model provides a noninvasive and interpretable framework for recurrence prediction in IBC. This approach may support individualized risk stratification, postoperative surveillance planning, and tailored adjuvant management while requiring further prospective validation before routine clinical implementation.