Deep explainable multimodal graph pooling network with knowledge learning for Alzheimer's disease diagnosis

Medical imaging and gene sequencing techniques provide large amounts of multimodal data for Alzheimer's disease (AD) diagnosis. Brain imaging genetics mainly explores potential relations between multimodal imaging and genetic data. However, how to effectively fuse the multimodal data and extract complementary information from them is a challenge. Additionally, there is currently a shortage of explainable deep learning models to perform imaging genetics analysis. In this work, an explainable model named deep multimodal graph pooling network with knowledge learning (DMGPK) is proposed to predict disease status and identify candidate markers. First, the categorical-formed single nucleotide polymorphism (SNP) data are preprocessed into numerical-formed data by the subject-specific gene mapping matrix. To effectively fuse multimodal data, heterogeneous factors association graphs for each subject are constructed based on brain regions and risk genes. Second, we developed a deep graph pooling network, which utilizes edge-weighted graph attention convolutional layers for node feature updates and employs self-attention graph pooling layers to highlight significant nodes. Meanwhile, the deep graph pooling network with two explainable loss terms can identify individual and group-level biomarkers for AD diagnosis. Finally, to further use the clinical demographic data of subjects, a novel knowledge learning module is embedded after the pooling layer to enrich the subject-level clinical prior. Experiments on AD neuroimaging initiative datasets prove that DMGPK achieves higher classification accuracy and outperforms advanced methods. Biological experiments also suggest that several top-ranked candidate markers are biologically plausible and related to AD-relevant processes.