Spectral Considerations on Unsupervised Time-domain EEG Representation Learning with Deep Networks
Authors: David Aquilué-Llorens, Abel Puigseslloses, Angel Blanco, Claire Braboszcz, Aureli Soria-Frisch
Organizations: Starlab Barcelona S.L., Neuroscience Business Unit, Barcelona, Spain · Center for Brain and Cognition, Computational Neuroscience Group, Universitat Pompeu Fabra, Barcelona, Spain
Electroencephalography (EEG) is a non-invasive, scalable source of quantitative biomarkers for the early diagnosis of brain disorders, and deep representation learning has become a powerful paradigm for extracting information-rich latent features from raw signals. However, neural networks tend to under-represent high-frequency information, a concern that is especially acute for EEG given its oscillatory nature and characteristic 1/fχ power decay, which risks masking disease-relevant high-frequency signatures. In this study, we systematically investigate the spectral constraints of unsupervised time-domain feature extraction across three autoencoder architectures (LSTMVAE, VAEEG, Transformer) and a state-of-the-art EEG foundation model. Using synthetic experiments and empirical validation, we show that while autoencoders capture macro-scale shifts such as alpha slowing, they remain biased against higher frequencies due to a magnitude-driven masking effect induced by the Mean Squared Error loss under 1/f-decaying spectra. Increasing latent dimensionality mitigates this attenuation but sacrifices latent compactness and generative stability. We argue that, as the field advances toward large-scale foundation-level feature extractors, it is imperative to move beyond simple time-domain reconstruction and to integrate spectral-domain loss terms and gradient-balancing techniques, so that the subtle high-frequency transients critical for diagnosing complex brain disorders are preserved in the latent representation.
Figures & tables
Figure 1: Performance Evaluation on Synthetic EEG Datasets. (a) Characterization of Synthetic Datasets: The upper (middle) panels display the group-specific Power Spectral Density (PSD), averaged across channels and subjects within the validation partition for the α ( β/γ ) datasets. The lower panels show representative 2-second EEG segments for the three simulated midline channels across both datasets. (b) Reconstruction Fidelity across Latent Dimensions: Model performance is reported as 1−MSE (where 1 indicates perfect reconstruction) on the held-out test partition. Results are shown for the LSTMVAE (upper), VAEEG (middle), and Transformer (lower) architectures as a function of the latent space dimensionality ( NLS ) and dataset type. (c) Latent Space Discriminative Power: Classification accuracy on the held-out test partition of a non-linear MLP trained to predict diagnostic group (HC, SCD, MCI, AD) from the latent embeddings. Accuracy is reported for the LSTMVAE, VAEEG, Transformer, and REVE (top to bottom), across varying latent dimensions. (d) provides a detail-view of REVE’s high accuracy. Theoretical chance accuracy (25%) is denoted by the horizontal red dashed line. Error bars indicate 95% confidence interval.
Figure 2: UMAP Projections of Latent Space Embeddings ( NLS=50 ). Two-dimensional UMAP visualizations of the latent manifolds learned by the LSTMVAE (left), VAEEG (middle), and Transformer (right) architectures. The upper row displays projections for the α slowing dataset, while the lower row represents the β/γ slowing dataset. Each point corresponds to a multi-channel EEG segment from the held-out test partition, color-coded by the synthetic diagnostic group (HC, SCD, MCI, AD) to illustrate the degree of class-wise clustering and separability within the learned latent space.
Figure 3: Spectral Learning Order Synthetic experiment adapted from Rahaman et al. [ 44 ] illustrating the evolution of frequency learning as a function of training progress. The heatmaps represent spectral fidelity, defined as the ratio between the reconstructed and target amplitudes for each frequency component ( f∈{5,10,…,50} Hz). A ratio of 1 (yellow) indicates successful reconstruction (hence encoding), while a ratio of 0 (blue) indicates the component has not yet been captured. Two scenarios are compared across the three architectures (LSTMVAE, VAEEG, and Transformer) with NLS=50 : the equal-amplitude condition ( Ai=1 , upper panels), isolating architectural bias; and the decaying-amplitude condition ( Ai∝1/fi0.75 , lower panels), simulating the power distribution of empirical EEG to observe magnitude-driven masking effects.
Figure 4: Spectral Organization of the LSTMVAE Latent Space for the held-out test partition of the BBHI Dataset. (a) UMAP projections of the latent embeddings color-coded by the grand-average (average over all channels) relative power within standard EEG frequency bands: theta ( 4−8 Hz), alpha ( 8−13 Hz), and beta ( 13−30 Hz). (b) Topographically resolved UMAP projections of alpha relative band power for the LSTMVAE (upper) and REVE (lower) models. Color-coding alpha band power is the result of averaging alpha band power over different sets of channels: Frontal (F7, F3, Fz, F4, F8), Temporal (T7, T8), and Central (C3, Cz, C4). These projections aim to illustrate the emergence of spectral and spatial representational structure within the unsupervised latent manifold.
Figure 5: Frequency-resolved spectral decay across reconstruction autoencoders as a function of latent space dimension ( NLS ), evaluated on the held-out test partition of the BBHI dataset. Power spectral densities (PSDs) of target 2-second EEG windows from the BBHI dataset (black) are compared against the PSDs of reconstructed outputs for the LSTMVAE (top, red gradient), VAEEG (middle, blue gradient), and Transformer (bottom, green gradient) models. Three latent space dimensions are shown, NLS={50,250,500} , with color intensity increasing from lighter to darker shades as NLS increases. Higher NLS improves high-frequency reconstruction fidelity, particularly for the Transformer architecture.
Figure 6: UMAP Projections of Latent Space Embeddings ( NLS=250 ). Two-dimensional UMAP visualizations of the latent manifolds learned by the (left to right) LSTMVAE, VAEEG, Transformer, and REVE architectures. The upper row displays projections for the α slowing dataset, while the lower row represents the β/γ slowing dataset. Each point corresponds to a multi-channel EEG segment from the held-out test partition, color-coded by the synthetic diagnostic group (HC, SCD, MCI, AD) to illustrate the degree of class-wise clustering and separability within the learned latent space.
Figure 7: Spectral Learning Order - Reconstruction Examples Evolution of reconstructed segments in the synthetic experiment adapted from Rahaman et al. [ 44 ] illustrating the evolution of frequency learning as a function of training progress in the equal-amplitude condition ( Ai=1 ) across three architectures: LSTMVAE (upper panels), VAEEG (middle panels), and Transformer (lower panels), all with NLS=50
Figure 8: UMAP projection of LS embeddings for LSTMVAE NLS=50 , coloured by demographic information (sex, age, and cognitive performance measured with MMSE)
Figure 9: UMAP projection of LS embeddings for REVE Base NLS=512 , coloured by demographic information (sex, age, and cognitive performance measured with MMSE)
Figure 10: Spectral Learning Order Study on REVE . For Figure details, refer to Figure 3 in the manuscript.
Figure 11: Spectral Learning Order Study on LSTMVAE and Transformer reconstruction autoencoders with a modified masked autoencoding objective . For Figure details, refer to Figure 3 in the manuscript.
Figure 12: Topographical Mapping of Spectral Decay in the held-out test partition of the BBHI Dataset for LSTMVAE (upper left), VAEEG (upper right), and Transformer (lower). Regional distribution of band-specific spectral attenuation ( ΔPdB ) as a function of increasing latent space dimensionality ( NLS ). Results are shown for the delta, theta, alpha, beta, and gamma bands across the scalp. Lighter hues indicate minimized attenuation (higher reconstruction fidelity), whereas darker hues denote regions of significant power loss, highlighting the spatial and spectral dependencies of the LSTMVAE. reconstruction process.
Increasing EEG pretraining data scale or model capacity does not consistently improve downstream performance. We identify a persistent low-frequency bias in representations learned by diverse EEG foundation models, which remains across dataset scales, model capacities, and pretraining objectives. Our analysis links this bias to the interaction between EEG's 1/fα-like spectral structure and neural networks' tendency to preferentially learn low-frequency components. In masked autoencoders, the ℓ2 reconstruction objective further amplifies this imbalance: under comparable relative reconstruction errors, high-power low-frequency components contribute disproportionately to the loss. To address this issue, we introduce FAME, a frequency-balanced masked autoencoding framework that reconstructs time--frequency activity in predefined EEG bands from masked EEG inputs. FAME independently standardizes the reconstruction targets within each band and assigns equal weight to all band-specific losses, thereby balancing supervision across the EEG spectrum. Evaluated on 41 downstream tasks in OmniEEG-Bench, FAME learns more spectrally balanced representations and achieves state-of-the-art performance on 24 of them. These results underscore the importance of balanced spectral supervision for learning transferable EEG representations.
Junjie Yu, Zihan Deng, Jianyu Zhang +8
Department of Biomedical Engineering, Southern University of Science and Technology, Shenzhen, China · Omni-Intelligence, Shenzhen, China · College of Design and Engineering, National University of Singapore, Singapore +4
Electroencephalograph (EEG) timeseries signals are characterized by significant noise and coarse spatial resolution, which complicates the classification of neurodegenerative diseases. Even SOTA deep learning architectures struggle to distinguish between healthy controls and diseased subjects, or between different disease types, due to high intergroup similarity. In this paper, we show that a spectrally selective approach to feature construction enhances class separability. By isolating signal strengths within the primary brainwave bands, we transform high dimensional raw data into high value spectral features. Our results demonstrate that in small datasets a) features derived from frequency and time frequency domain allow traditional machine learning models to match or exceed the performance of SOTA deep learning models, b) Attention mechanism is unable to distill the stable feature signatures that characterize healthy neural activity in both resting and task EEGs, and c) the limitations of attention based models in finding relevant spectral features appear to be robust in that providing frequency selective time domain input do not appreciably improve their performance. We validate our methodology across three open source resting EEG datasets and one task EEG dataset, providing robust empirical evidence for our claims.
EEG foundation models, pre-trained on large-scale unlabelled EEG data, have emerged as a promising direction towards learning generalizable EEG representations. Despite showing positive results in data-rich regimes, they often fail to outperform significantly smaller supervised models in low-resource settings compared to fully supervised models. We provide a mechanistic account of this shortcoming, attributing it to a fundamental mismatch between reconstruction-based pretext tasks and the idiosyncratic spectral structure of EEG signals, which decompose into distinct high-power aperiodic and low-power oscillatory components. Using controlled, synthetically-generated EEG inputs, we demonstrate that EEG foundation model embeddings are biased to capture the aperiodic components of the EEG signal while under-representing oscillatory components, particularly at higher frequencies. Additionally, linear probe evaluations on real-world BCI datasets further reveal that embeddings encode subject identity more strongly than task-relevant information, thereby reinforcing the low-frequency and aperiodic component bias in foundation model embeddings trained primarily on reconstruction based objectives. Together, these findings elucidate a failure mode in reconstruction based EEG foundation models and motivate future work to incorporate auxiliary losses explicitly targeting high-frequency oscillatory structure as a path toward more capable and generalizable EEG representations.
Ming Hsieh Department of Electrical and Computer Engineering1 University of Southern California · Thomas Lord Department of Computer Science2[cs.LG] University of Southern California · Department of Mathematical Sciences3 Aalborg University