Variational Mixture of Graph Neural Experts for Alzheimer's Disease Biomarker Recognition in EEG Brain Networks

TL;DR

The paper addresses the challenge of differentiating Alzheimer's disease and frontotemporal dementia from EEG signals by introducing VMoGE, a variational mixture of graph neural experts that leverages frequency-specific biomarkers. It combines a multi-granularity transformer-based node feature extractor with band-specific GMRF priors and a gating-based mixture of experts to dynamically fuse information across delta, theta, alpha, and beta bands. The approach demonstrates consistent improvements in AUC and accuracy over state-of-the-art baselines on two dementia datasets, while offering interpretable biomarkers through gating weights and spatial activation maps that align with clinical indicators and neuropathology. These results suggest substantial potential for frequency-specific EEG biomarker discovery and robust dementia diagnosis and progression monitoring in practical clinical settings.

Abstract

Dementia disorders such as Alzheimer's disease (AD) and frontotemporal dementia (FTD) exhibit overlapping electrophysiological signatures in EEG that challenge accurate diagnosis. Existing EEG-based methods are limited by full-band frequency analysis that hinders precise differentiation of dementia subtypes and severity stages. We propose a variational mixture of graph neural experts (VMoGE) that integrates frequency-specific biomarker identification with structured variational inference for enhanced dementia diagnosis and staging. VMoGE employs a multi-granularity transformer to extract multi-scale temporal patterns across four frequency bands, followed by a variational graph convolutional encoder using Gaussian Markov Random Field priors. Through structured variational inference and adaptive gating, VMoGE links neural specialization to physiologically meaningful EEG frequency bands. Evaluated on two diverse datasets for both subtype classification and severity staging, VMoGE achieves superior performance with AUC improvements of +4% to +10% over state-of-the-art methods. Moreover, VMoGE provides interpretable insights through expert weights that correlate with clinical indicators and spatial patterns aligned with neuropathological signatures, facilitating EEG biomarker discovery for comprehensive dementia diagnosis and monitoring.

Figures (7)

• Figure 1: Diagram of MGT-NFE for node feature extraction, incorporating multi-granularity hierarchical feature extraction and spatial positional encoding at different granularities.
• Figure 2: Overview of the VMoGE framework for AD biomarker identification and prediction. VMoGE framework first extracts node features from multi-channel EEG signals (channels C1–C19) by integrating spatial and frequency band features through a 1D-CNN and FFT-based MGT-NFE module. The prior graph structure is constructed using a GMRF, and a variational router models the latent distribution to capture structural correlations across multiple frequency bands. Finally, the framework performs a weighted summation of four expert networks based on k-th gating probabilities to obtain the final classification result.
• Figure 3: Ablation study of MGT-NFE module using different granularity components for EEG feature extraction.
• Figure 4: Box plots showing the distribution of mixture weights across different bands under three subtyping comparison conditions. The weights represent the contribution of different components in distinguishing between HC, FTD, and AD in the Open AD dataset in Fig. (a), and Session-based AD staging based on CDR=0, CDR=1, and CDR=2 in Fig.(b).
• Figure 5: The expert gating weight analysis for different dementia indicator analysis. The top row shows age-based comparisons, and the bottom row shows MMSE-based comparisons. Each column represents different pairwise comparisons between HC and FTD, HC and AD, and FTD and AD.