Multimodal fusion using sparse CCA for breast cancer survival prediction

Abstract

Effective understanding of diseases like cancer necessitates integrating diverse information across various physical scales through multimodal data. In this study, we introduce a novel feature embedding module based on canonical correlation analysis (CCA) to capture both intra-modality and inter-modality correlations. Our approach leverages CCA to develop multi-dimensional embeddings that align well across different data sources. We validated our method using both simulated and real datasets, demonstrating its capability to generate well-correlated embeddings. When applied to the one-year survival classification of breast cancer patients from the TCGA-BRCA dataset, our embeddings achieved competitive performance, with average F1 scores reaching up to 58.69% in 5-fold cross-validation.