Dynamic multi-hypergraph structure learning for disease diagnosis on multimodal data

With its superior capability in complex data modeling, hypergraph computation is a powerful tool for many applications. In this work, we propose using hypergraph computation for disease prediction. Hypergraphs allow for the representation of higher-order relations, called hyperedges, spanning possibly more than two nodes to capture complex correlations within multimodal medical data and patients' characteristics. We propose a dynamic bi-clustering approach to learn a multi-hypergraph structure based on node embedding to model high-order multimodal patient interaction. We have conducted experiments on benchmark real-world datasets for Alzheimer's Disease and Autism Spectrum Disorder prediction. Experimental results demonstrate that the proposed Hypergraph Neural Network method outperforms state-of-the-art methods.