HSCFA: Hierarchical sparse and collaborative fusion attention with large foundation models for diagnosing Alzheimer’s disease

Abstract

Integrating macro-level neuroimaging data with micro-level genetic data offers mechanistic understanding into Alzheimer’s Disease (AD). However, existing methods fail to fully exploit multi-level features and their collaborative patterns. To address this limitation, this paper proposes a unified framework incorporating large foundation models and attention mechanisms to construct, extract, and fuse hierarchical features. We first construct a Hierarchical Sparse and Collaborative Fusion Attention (HSCFA) model to characterize AD pathogenesis, where two sparse attention mechanisms are utilized to extract hierarchical features and co-attention is used to achieve feature fusion. Subsequently, we implement an HSCFA algorithm based on the model, leveraging biomedical large foundation models to construct high-quality features and applying attention mechanisms to capture characteristic AD-specific association patterns. Finally, experiments on public datasets validate the superiority of HSCFA in sample classification and pathogeny extraction, achieving 3-class classification accuracy of 88.41 %. This work provides an effective algorithm for the early diagnosis of AD and identifies AD-related risk genes and abnormal brain regions, offering novel insight for the pathological research of AD. The code of HSCFA can be accessed at the following link: https://github.com/fmri123456/HSCFA.