Abstract
Accurate survival prediction in breast cancer remains a key challenge in oncology, requiring models that can integrate diverse clinical, molecular, and imaging data sources to guide breast cancer management. While recent deep learning models have explored multimodal integration for cancer survival prediction, their generalizability to unseen data remains limited. In this study, we developed and optimized unimodal and multimodal models for breast cancer survival prediction, systematically assessing our optimized early and late integration strategies and their impact on out-of-sample generalization performance. We integrated clinical variables, somatic mutations, RNA expression, copy number variation, miRNA expression, and histopathology images from The Cancer Genome Atlas breast cancer dataset. Across all modality combinations, late fusion models consistently outperformed early fusion approaches and late and intermediate benchmark methods, with the combination of omics and clinical data yielding the highest test-set concordance indices. Explainability analyses showed that our models captured biologically relevant features associated with patient survival. These findings highlight the value of late-fusion multimodal deep learning frameworks for robust and explainable survival prediction in breast cancer.