TDMFS: Tucker decomposition multimodal fusion model for pan-cancer survival prediction

Jinchao Chen, Pei Liu, Chen Chen, Ying Su, Enguang Zuo, Min Li, Jiajia Wang, Ziwei Yan, Xinya Chen, Cheng Chen, Xiaoyi Lv

Abstract: Highlights•We propose TDMFS, a multimodal fusion pan-cancer survival-prediction model that leverages Tucker decomposition and interaction attention for deep, lightweight fusion of different modalities while supplementing single-modal features for excellent survival prediction with reduced complexity.•The IATDF strategy performs fusion modeling twice. Each fusion first applies Tucker decomposition to constrain the complex parameter tensor and reduce computational complexity, then employs an interaction attention strategy to obtain a deep fusion representation for co-learning multimodal information.•The BSC module uses bilinear pooling to decompose and fuse modality-specific information, followed by a signal modulation mechanism that produces self-attentive, modality-specific representations.•The effectiveness of TDMFS for cancer survival prediction was validated on TCGA's dataset of 33 cancer types.