DGSurv: Dynamic Graph-Based Multimodal Learning for Interpretable Cancer Survival Prediction

Multimodal learning in cancer research offers transformative potential for enhancing medical care and guiding clinical decisions. Most analyses rely on unimodal inputs or employ simplistic multimodal fusion techniques, which do not optimally integrate the diverse data types. Additionally, there is a critical need for enhanced interpretative methods to fully exploit the depth of multimodal patient data. To address these issues, we propose DGSurv, a novel multimodal learning approach that utilizes a graph neural network (GNN) to dynamically map inter-modality relationships for cancer survival prediction. We demonstrate the utility of our proposed approach on cancer survival prediction, highlighting its potential to inform more accurate clinical decision-making. We perform empirical evaluations on four cancer datasets from The Cancer Genome Atlas Program (TCGA) and demonstrate that DGSurv outperforms existing fusion techniques. For interpretability, our study advances multimodal cancer analysis by effectively harnessing the full spectrum of multimodal data and significantly boosting its interpretability.