BreastSurPro: A Multimodal Deep Survival Model for Breast Cancer

Accurate survival prediction for breast cancer is challenged by pronounced tumor heterogeneity and the presence of censored data. While multimodal deep learning offers promise, performance estimates are often optimistic due to inconsistent validation frameworks. To develop and internally validate BreastSurPro, a multimodal deep survival model integrating clinical, genomic, and imaging data for breast cancer prognosis under a rigorous, censoring-aware evaluation paradigm. This study utilized The Cancer Genome Atlas Breast Invasive Carcinoma (TCGA-BRCA) cohort. Clinical and genomic data (n=952) were integrated with MRI-derived radiomic features from a subset (n=92). The model employed a modular neural network with Cox partial-likelihood loss. Performance was evaluated using stratified 5-fold cross-validation, reporting the concordance index (C-index), time-dependent AUC, Kaplan-Meier analysis, and calibration plots. Benchmarks included Cox proportional hazards and DeepSurv models. BreastSurPro achieved a mean C-index of 0.73 for overall survival and 0.72 for disease-free survival, significantly outperforming unimodal baselines (p<0.05). In the radiogenomic subset, the tri-modal model (C-index=0.71) modestly surpassed the dual-modality version. Ablation studies confirmed complementary prognostic information from all modalities. The model demonstrated effective risk stratification (log-rank p<0.001) and good calibration. SHAP analysis identified tumor size, nodal status, and driver genes (TP53, PIK3CA) as top features. BreastSurPro provides a reproducible, interpretable framework for multimodal survival prediction, demonstrating that integrating clinical, genomic, and imaging data yields consistent, incremental improvements in prognostic accuracy.