EEG-SeeGraph: Interpreting functional connectivity disruptions in dementias via sparse-explanatory dynamic EEG-graph learning

Technical Deep Dive: EEG-SeeGraph

Overview

EEG-SeeGraph is a sparse-explanatory dynamic EEG-graph network for robust and interpretable diagnosis of neurodegenerative dementias like Alzheimer's disease (AD) and frontotemporal dementia (FTD). It models time-varying brain networks, capturing both regional oscillations and inter-regional functional connectivity. The key innovations are:

• A dual-trajectory temporal encoder that models node-level spectral features and edge-level amplitude correlations
• A topology-aware positional encoder that injects graph-spectral Laplacian coordinates into node embeddings
• A node-guided sparse explanatory edge mask that identifies a compact, interpretable subnetwork driving the diagnostic decision
• A gated graph predictor that operates on the sparsified connectivity for robust classification

Methodology

• EEG signals are segmented into overlapping time windows, from which two complementary feature streams are extracted:
  • Node trajectory: FFT-based spectral features capture regional oscillatory activity
  • Edge trajectory: Pearson correlation of amplitude time series captures inter-regional functional coupling
• These node and edge trajectories are independently encoded using multi-head self-attention to capture long-range temporal dependencies.
• Laplacian positional encodings derived from the connectivity matrix are appended to the node embeddings to inject topological information.
• A node-guided sparse mask is learned to identify a compact, explanatory subnetwork, with a sparsity-inducing regularizer to prevent mask degeneration.
• The gated graph predictor performs graph attention on the sparsified connectivity to produce the final diagnosis.

Data & Experimental Setup

The model was evaluated on two EEG datasets:

• AHEPA: A public resting-state EEG cohort with 88 participants (36 AD, 23 FTD, 29 HC), recorded with a 19-channel 10–20 system at 500 Hz.
• SZPH: A clinical cohort from Shenzhen People's Hospital with 50 participants (30 AD, 20 HC), recorded with a 64-channel 10–20 system.

Experiments used an 80/20 subject-independent train-test split. Robustness was assessed by injecting Gaussian noise (zero mean, std. dev. 0.3) into the input signals.

Results

• SeeGraph achieved state-of-the-art performance on both the AHEPA and SZPH datasets, outperforming baselines like EvolveGCN-O, DCRNN, STGCN, BIOT, CNN-LSTM, and MAtt.
• Under noisy conditions, SeeGraph exhibited only minor performance degradation, demonstrating robust generalization.
• Ablation studies confirmed the complementary contributions of the dual-trajectory encoding, Laplacian positional encoding, and sparse explanatory mask.
• Band-wise analyses showed that alpha, beta, and delta bands carry the strongest diagnostic signal, consistent with clinical evidence of dementia-related oscillatory changes.

Interpretation

• The node-guided sparse mask identified disease-relevant functional connectivity disruptions, with AD exhibiting temporal-lobe-centered ipsilateral clustering and healthy controls retaining more interhemispheric and bilateral integration.
• These localized subnetworks align with clinical findings on functional connectivity alterations in neurodegenerative dementias, providing transparent cues for neurological evaluation.

Limitations & Uncertainties

• The source text does not discuss potential limitations or uncertainties of the SeeGraph approach.

What Comes Next

The source text does not speculate on future work or next steps. The focus is on presenting the core SeeGraph methodology and validating its performance and interpretability.