Dg-Hgcn: a Multimodal Survival Prediction Model Based on Cross-Modal Dynamic Graph Construction
Aiping Qu, Mao Tang, Weirong Wang, Hao Ming Peng · 2025
In recent years, with the rapid development of deep learning, multimodal techniques have significantly improved their application in predicting the prognostic survival of cancer patients. In cancer treatment, a single data source may fail to comprehensively reflect a patient's condition. By integrating multiple types of information, multimodal learning can more effectively capture the multidimensional characteristics of cancer patients, thereby enabling accurate survival predictions. However, since medical modalities are often heterogeneous, existing methods lack sufficient intramodal and inter-modal interactions. To address this issue, we propose the enhanced the Hybrid Graph Convolutional Network model based on Cross-Modal Dynamic Graph Construction (DG-HGCN). This model first builds crossmodal feature graphs between the two modalities by computing cross-modal attention between pathology, genomic, and clinical, enabling improved intra-modal and inter-modal interactions. Our method achieved results superior to existing approaches on six cancer datasets from TCGA. Our code are available in: https://github.com/airleaya/DG-HGCN