[Paper Review] Fusion of Spatio-Temporal and Multi-Scale Frequency Features for Dry Electrodes MI-EEG Decoding

Tianyi Gong, Can Han+2 more|arXiv (Cornell University)|Jan 26, 2026

EEG and Brain-Computer Interfaces0 citations

TL;DR

STGMFM introduces a tri-branch architecture that fuses dual ordered spatio-temporal graphs with a multi-scale frequency envelope mixer for dry-electrode MI-EEG decoding, achieving state-of-the-art cross-session/ cross-subject results.

ABSTRACT

Dry-electrode Motor Imagery Electroencephalography (MI-EEG) enables fast, comfortable, real-world Brain Computer Interface by eliminating gels and shortening setup for at-home and wearable use.However, dry recordings pose three main issues: lower Signal-to-Noise Ratio with more baseline drift and sudden transients; weaker and noisier data with poor phase alignment across trials; and bigger variances between sessions. These drawbacks lead to larger data distribution shift, making features less stable for MI-EEG tasks.To address these problems, we introduce STGMFM, a tri-branch framework tailored for dry-electrode MI-EEG, which models complementary spatio-temporal dependencies via dual graph orders, and captures robust envelope dynamics with a multi-scale frequency mixing branch, motivated by the observation that amplitude envelopes are less sensitive to contact variability than instantaneous waveforms. Physiologically meaningful connectivity priors guide learning, and decision-level fusion consolidates a noise-tolerant consensus. On our collected dry-electrode MI-EEG, STGMFM consistently surpasses competitive CNN/Transformer/graph baselines. Codes are available at https://github.com/Tianyi-325/STGMFM.

Motivation & Objective

• Motivate robust MI-EEG decoding with dry electrodes, addressing low SNR, baseline drift, and cross-subject/session variability.
• Propose a tri-branch network that combines spatio-temporal graph reasoning with envelope-based frequency features.
• Leverage physiologically meaningful priors (PLV) and simple decision-level fusion for robust generalization.
• Show superiority over CNN/Transformer/GCN baselines on a dry-electrode MI-EEG dataset.
• Provide insights via ablations on the contribution of each component.

Proposed method

• Construct a PLV-initialized channel graph (A) to encode functional connectivity as a prior.
• Implement two dual-order spatio-temporal branches: CCG → TSG and TSG → CCG to yield complementary evidence.
• Add a Multi-Scale Frequency Mixer (C) operating on amplitude envelopes to capture ERD/ERS-like dynamics across multiple scales.
• Fuse branch logits at the decision level with a lightweight linear head (no gating) to form a robust final prediction.
• Train with cross-entropy plus L1/L2 regularization on graph modules and cosine-annealing learning rates.
• Use windowed sliding segments to create per-window inputs and stabilize cross-trial comparisons.

Experimental results

Research questions

• RQ1 Can a PLV-informed, learnable adjacency improve dry-electrode MI-EEG decoding across sessions/subjects?
• RQ2 Do dual-ordered spatio-temporal graphs provide complementary robustness to dry-electrode noise paths?
• RQ3 Does a Multi-Scale Frequency Mixer capturing envelope dynamics enhance robustness to contact variability?
• RQ4 Is decision-level fusion sufficient to leverage multi-branch representations without overfitting in small EEG datasets?

Key findings

• STGMFM outperforms CNN/Transformer/GCN baselines across three evaluation protocols on a 23-channel dry-EEG MI-EEG dataset.
• Ablations show the full model (dual graph orders + MFM + PLV initialization with regularization) yields the best cross-subject performance (ACC 57.26%, kappa 0.3592, F1 56.52%).
• Dual graph orders provide complementary inductive biases, reducing noise propagation in dry EEG.
• The Multi-Scale Frequency Mixer effectively recovers ERD/ERS-like envelope patterns under dry-electrode conditions.
• Regularization and cosine-annealed training help prevent overfitting and promote generalization across subjects/sessions.