IG-GRD: A Model Based on Disentangled Graph Representation Learning for Imaging Genetic Data Fusion

Integrating imaging and genetic data provides a comprehensive approach to analyze brain disorders from different perspectives, which has important implications for the early diagnosis of Alzheimer’s Disease (AD) and the exploration of its underlying mechanisms. Current fusion methods focus primarily on the correlation between modalities or rely on decision-level fusion. However, due to the heterogeneity of imaging and genetic data, as well as the necessity to simultaneously consider their correlation and independence, current methods often face challenges in adequately integrating and fully learning from multimodal information. Therefore, in this paper, we propose a novel multimodal data fusion method, named IG-GRD, based on graph representation learning for imaging and genetic data. Firstly, we construct imaging graphs and genetic graphs based on the characteristics of fMRI and SNP data, mapping the data from these two modalities into a unified representation space. Subsequently, we use a disentangled representation learning method on multimodal graphs that considers structural information and complex relationships between nodes to capture common and private graph representations. Finally, the disentangled feature graphs are fused at the graph level to synthesize the collaborative and individual effects of imaging and genetic information on the disease. Experimental results demonstrate that IG-GRD excels not only in recognizing mild cognitive impairment (MCI), but also in identifying brain regions and genes closely associated with AD and cognition. This work offers a novel methodology for the fusion of imaging and genetic data and provides new directions for the early diagnosis of AD and the investigation of its pathogenesis.