文献类型：期刊文献

英文题名：LMFN: A Lightweight Multimodal Fusion Network for Subject-Independent EEG-Based Alzheimer's Disease Diagnosis

作者：Fang, Yaohui[1]; Guo, Weibin[1]

机构：[1] East China University of Science and Technology, School of Information Science and Engineering, Shanghai, China

年份：2025

起止页码：473

外文期刊名：Proceedings of 2025 5th International Symposium on Artificial Intelligence and Big Data, AIBDF 2025

收录：EI(收录号：20261720600373)

语种：英文

外文关键词：Biomedical signal processing - Brain mapping - Electroencephalography - Electrophysiology - Graph neural networks - Graph theory - Learning systems - Neurodegenerative diseases - Parallel architectures

摘要：Alzheimer's Disease (AD) has become an increasingly pressing global health concern, emphasizing the critical need for accessible, early-stage diagnostic solutions. Electroencephalography (EEG) emerges as a promising low-cost alternative to traditional neuroimaging techniques. However, current deep learning methods for EEG-based AD diagnosis exhibit notable limitations. They often struggle to effectively capture long-range temporal dependencies, fail to adequately represent the brain's complex functional topology (due to overly simplified homogeneous graph assumptions), and have difficulty generalizing to previously unseen subjects. To address these issues, this paper introduces the Lightweight Multimodal Fusion Network (LMFN). LMFN utilizes a parallel architecture to independently extract complementary features from the temporal, spectral, and spatial domains of EEG signals, minimizing inter-modal interference. A central innovation of LMFN is the integration of a heterogeneous graph attention mechanism within the spatial branch. This mechanism explicitly models three critical inter-electrode relationships: spatial proximity, functional connectivity, and inter-hemispheric symmetry, thereby overcoming the limitations of simplistic homogeneous graph models. Additionally, factorized spatio-temporal convolutions are employed to efficiently refine features while maintaining a compact model size. LMFN was rigorously tested using a subject-independent protocol - an essential step for ensuring real-world clinical applicability. Extensive experiments were conducted across three public EEG datasets (ADSZ, APAVA, ADFD). The results demonstrate that LMFN outperforms 19 state-of-the-art baseline models, achieving 95.54% accuracy on ADSZ and 56.33% accuracy on the multi-class ADFD task. These findings reinforce LMFN's potential as a robust, resource-efficient, and clinically useful tool for supporting AD diagnosis in resource-limited settings. ? 2025 IEEE.