EEG-X: Device-Agnostic and Noise-Robust Foundation Model for EEG

AI summary

Overview

• Research area: Self-supervised representation learning for electroencephalography (EEG) — specifically, building a "foundation model" for brain signals that transfers across recording devices and tasks.
• Technical level: Advanced. The paper assumes familiarity with masked autoencoding, student–teacher / EMA self-distillation, transformer tokenizers, independent component analysis (ICA), and convolutional feature extraction.
• Scope (one sentence): The paper proposes EEG-X, a pretrained EEG model built from location-based channel embeddings, dual latent-and-raw reconstruction of artifact-removed signals, and a dictionary-inspired convolutional transformation that replaces direct time-domain MSE, and evaluates it on seven datasets from four headsets.

What This Paper Is About

EEG foundation models lag behind those in vision, language, and speech for two reasons the authors identify: EEG datasets vary wildly in sampling frequency, electrode placement, channel count, and recording length across devices, and EEG has a low signal-to-noise ratio, with brain activity frequently buried under artifacts and non-brain sources. EEG-X addresses both by making the model spatially aware of where on the scalp each electrode sits (rather than relying on channel names that differ between devices) and by training it to reconstruct artifact-removed signals — in both raw and latent space — using a loss that is less dominated by amplitude than plain mean squared error.

Key Contributions

1. Location-based channel embedding. Each electrode is mapped to Cartesian coordinates on the scalp via a universal location mapping, then encoded with sinusoidal position encodings adapted from transformers. Nearby electrodes receive similar embeddings, so the model can handle varying channel numbers and combinations, and can fall back on adjacent electrodes when one is missing or too noisy.
2. Noise-aware reconstruction in both raw and latent spaces. Unlike prior work that masks and reconstructs raw noisy EEG, EEG-X reconstructs signals that have been passed through an artifact-removal pipeline (ICA with ICLabel), and in latent space predicts the representations of unmasked inputs from masked ones via a student–teacher setup.
3. Dictionary Convolution Transformation (DiCT). A lightweight random-convolution projection layer, inspired by Hydra, that maps both the artifact-removed input and the decoder output into a structured feature space before computing MSE, reducing noise sensitivity, balancing comparisons across frequency bands, and capturing shape similarity rather than pointwise amplitude matching.
4. Extensive cross-device evaluation. EEG-X is benchmarked against six state-of-the-art EEG representation-learning models on four in-domain and three cross-domain datasets spanning four headsets, including settings where the pretraining and downstream electrode layouts differ.

Main Findings

• In-domain: best average performance across all four tasks. Averaged over five runs, EEG-X reached 80.32 ± 1.21 accuracy / 88.37 ± 1.43 AUROC on STEW, 58.14 ± 1.35 / 57.64 ± 1.29 on DREAMER, 81.02 ± 1.48 / 88.97 ± 1.62 on TUAB, and 91.15 ± 1.41 / 96.18 ± 1.39 on TUEV_binary. For comparison, the strongest baseline on TUEV_binary was EEG2Rep at 89.83 ± 1.38 / 95.57 ± 1.47, and the best on STEW other than EEG-X was EEGPT at 77.65 ± 2.95 / 85.98 ± 3.01.
• Latent-space models beat raw-space models on noisy datasets. The authors report that latent reconstruction models such as EEG2Rep outperform raw-space reconstruction models such as MAEEG and EEGPT on the noisy TUAB and TUEV_binary datasets, and that EEG-X surpasses both by combining artifact removal with DiCT-enhanced reconstruction.
• Cross-domain: large gains as a frozen foundation model. When all models were pretrained only on TUAB and evaluated with a frozen encoder plus linear regression, EEG-X achieved 86.89 ± 1.28 accuracy / 93.51 ± 1.41 AUROC on Crowdsourced (best baseline EEGPT: 79.55 ± 1.36 / 87.79 ± 1.52), 37.58 ± 1.19 balanced accuracy / 35.32 ± 1.15 weighted F1 on BCICIV_2A (best baseline LaBraM: 32.44 ± 1.13 / 27.37 ± 1.21), and 57.17 ± 1.42 / 55.05 ± 1.33 on SSVEP (best baseline BioT: 42.12 ± 1.41 / 40.17 ± 1.36).
• Learnable channel embeddings barely help cross-domain; location-based ones do. In the ablation, the reported average cross-domain drop relative to EEG-X (60.55) was 8.81 without any channel embedding and 8.52 with a learnable channel embedding — nearly identical — while the location-based embedding is what EEG-X uses. In-domain, the drop was 1.16 without channel embedding and 0.43 with the learnable version, against EEG-X's average of 77.66.
• DiCT helps most in transfer. Removing DiCT cost 3.10 in-domain and 3.80 cross-domain on average; the authors describe the in-domain improvements as moderate and the cross-domain benefits as clearer, attributing them to structured feature alignment across recording setups.
• Artifact-removed reconstruction targets beat raw reconstruction. Replacing the artifact-removed reconstruction target with raw-signal reconstruction cost 3.88 in-domain and 6.85 cross-domain on average, and removing raw-space reconstruction entirely cost 2.42 in-domain and 2.93 cross-domain.
• ICA is better as a loss target than as a preprocessing step. Running EEG-X supervised for 100 epochs on raw versus ICA-preprocessed data gave STEW 75.12 vs 70.68, DREAMER 55.69 vs 56.24, TUAB 78.12 vs 78.25, and TUEV 88.24 vs 83.36, showing ICA preprocessing can remove useful brain components.
• Synthetic evidence for DiCT's frequency and shape awareness. On a composite signal of a 2 Hz component (amplitude 5.0), a 20 Hz component (amplitude 1.0), and a 100 Hz component (amplitude 0.5), direct MSE scored a low-frequency-only reconstruction at 0.58 while scoring a phase-shifted, more spectrally complete reconstruction at 13.88. With DiCT, the same three reconstructions scored 9.26, 6.74, and 7.25 respectively.
• ICA-relevant artifact magnitude. The paper notes artifacts such as muscle activity, eye blinks, cardiac signals, respiration, motion artifacts, and power line noise can be 5–50 times larger than typical brain signals.

Methodology in Plain English

EEG-X treats each EEG channel like a sentence. The recording is cut into overlapping fixed-length windows (overlap is typically one quarter of the window size), each window is converted to a frequency representation using the magnitude of a short-time Fourier transform, and a linear layer turns those into tokens. This tokenization handles recordings of different lengths.

To handle different devices, the model adds a spatial embedding instead of a channel-name embedding. Every electrode name is looked up in a dense scalp mesh of 348 positions drawn from the 10-05, 10-10, and 10-20 systems, giving Cartesian coordinates; sinusoidal encodings of those coordinates become the channel embedding. Because the encoding is smooth over space, electrodes that are physically close get similar vectors — the paper illustrates this with F4 being more similar to F2 and F6 than to P7 — so a channel that is absent or too noisy can effectively be covered by its neighbors.

Pretraining has two objectives running in parallel. In latent space, a 4-layer transformer student sees a masked version of the input while a 4-layer transformer teacher, updated as an exponential moving average of the student, sees the unmasked input; a 2-layer cross-attention predictor tries to match the teacher's representations of the masked tokens. Masking follows the Semantic Subsequence Preserving method, and the objective combines an L2 alignment loss with a variance/covariance regularization term to prevent collapse. In raw space, an encoder–decoder path reconstructs a cleaned version of the signal, where cleaning is done with ICA plus ICLabel, which classifies components into seven categories (brain activity, eye movements, muscle activity, heart signals, line noise, channel noise, and others

Authors’ abstract

Foundation models for EEG analysis are still in their infancy, limited by two key challenges: (1) variability across datasets caused by differences in recording devices and configurations, and (2) the low signal-to-noise ratio (SNR) of EEG, where brain signals are often buried under artifacts and non-brain sources. To address these challenges, we present EEG-X, a device-agnostic and noise-robust foundation model for EEG representation learning. EEG-X introduces a novel location-based channel embedding that encodes spatial information and improves generalization across domains and tasks by allowing the model to handle varying channel numbers, combinations, and recording lengths. To enhance robustness against noise, EEG-X employs a noise-aware masking and reconstruction strategy in both raw and latent spaces. Unlike previous models that mask and reconstruct raw noisy EEG signals, EEG-X is trained to reconstruct denoised signals obtained through an artifact removal process, ensuring that the learned representations focus on neural activity rather than noise. To further enhance reconstruction-based pretraining, EEG-X introduces a dictionary-inspired convolutional transformation (DiCT) layer that projects signals into a structured feature space before computing reconstruction (MSE) loss, reducing noise sensitivity and capturing frequency- and shape-aware similarities. Experiments on datasets collected from diverse devices show that EEG-X outperforms state-of-the-art methods across multiple downstream EEG tasks and excels in cross-domain settings where pre-trained and downstream datasets differ in electrode layouts. The models and code are available at: https://github.com/Emotiv/EEG-X