MMBCSurv: Multimodal Breast Cancer Survival Prediction Integrating Histopathology Images and Multi-Omics Data via Dynamic Fusion

Accurate survival prediction in breast cancer remains challenging because of tumor heterogeneity and the limitations of existing unimodal approaches. This study introduces a multimodal framework designed to improve overall survival prediction by combining advanced computational pathology with multi-omics data analysis. The framework features two main innovations: (1) a dynamic fusion method that adaptively combines pathological imaging with multi-omics data, and (2) the use of graph attention networks (GAT) to examine transcriptomic pathway structures for enhanced representation learning. Whole slide images were analyzed using attention-based multiple instance learning and transformer encoders. mRNA expression was modeled through graph construction and graph-based embeddings, whereas somatic mutations and copy number variations utilized RFE for feature extraction. The performance on the TCGA-BRCA and CPTAC-BRCA datasets showed that MMBCSurv outperforms current leading methods, achieving higher concordance index (c-index) scores. Integrated gradient analysis further improves interpretability by assessing the contributions of each modality, with histopathology images providing the most significant predictive value. As a result, MMBCSurv is an effective and interpretable tool for predicting survival outcomes, thus advancing precision oncology. Future studies should focus on enhancing the model’s generalizability to other organs.