Multimodal adaptive fusion deep analysis model for Alzheimer's disease exploration and diagnosis.

Abstract

Alzheimer's disease (AD) is a chronic progressive neurodegenerative disorder, the etiology and pathogenesis of which are currently unclear. Brain imaging genetics, which analyzes genetic factors and neuroimaging phenotypic data in association, is an effective technique for identifying AD-related biomarkers. With the rapid advancement of imaging and genetic sequencing technologies, the association between multimodal imaging data and genetic data has gradually received widespread attention. However, how to make full use of the complementary information between multimodal data is an urgent problem to be solved. Therefore, Multimodal Adaptive Fusion Deep Association Analysis Model (MAFDAA) is proposed to solve the above problems. Firstly, a novel multi-head attention graph convolutional model is proposed to extract and reconstruct complementary information among multimodal data, thus constructing a comprehensive representation of brain information. Secondly, the representation module statistically encodes genetic data to obtain genetic representations, while embedding demographic information as a supplement to the genetic representation. Subsequently, in the association analysis module, nonlinear correlation analysis is conducted between genetic representations and brain reconstruction data, yielding latent association vectors for subsequent research. Finally, the diagnostic module diagnoses the subjects and identifies AD-related biomarkers based on the association analysis results. The experimental results demonstrate that MAFDAA exhibits excellent diagnostic performance. Additionally, the identified biomarkers were analyzed from different perspectives, establishing a new approach for studying the complex genetic mechanisms of neurodegenerative diseases from a micro to macro scale.