💻 computer science

Temporal Feature Extractors in EEG Foundation Models: A Controlled Comparison Including a Pretrained Time-Series Model

This paper conducts a controlled comparison of linear, convolutional, and frozen pretrained time-series (MOMENT) feature extractors within EEG foundation models, finding that while simple temporal representations suffice for motor imagery tasks, richer modeling benefits emotion recognition and demonstrating that general-purpose pretrained time-series models can be effectively transferred as frozen extractors for EEG applications.

Original authors: Ayşe Betül Yüce, Chris Joey Leffler, Sarun Varghese, Myra Spiliopoulou, Sebastian Stober

Published 2026-06-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ayşe Betül Yüce, Chris Joey Leffler, Sarun Varghese, Myra Spiliopoulou, Sebastian Stober

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your brain is a busy city, and EEG (Electroencephalography) is a massive network of microphones placed all over the city square, recording the hum, buzz, and chatter of the crowd. These recordings are messy, noisy, and change constantly.

To make sense of this noise, scientists are building "Foundation Models." Think of these models as super-intelligent translators that listen to the brain's chatter and try to understand what the city is actually doing (like imagining moving a hand or feeling happy).

This paper asks a very specific question about how these translators are built: How should they listen to the timing of the sounds?

The Three Listening Styles

The researchers set up a controlled experiment to compare three different ways of "listening" to the brain's time-based signals before passing them to the main translator (the AI brain). They kept everything else exactly the same so they could see which listening style was best.

The "Simple Ear" (Linear Projection):
- The Analogy: Imagine a translator who just takes a quick, flat snapshot of the sound wave. They don't look for patterns or rhythms; they just take the raw numbers and pass them along. It's the simplest, most basic approach.
- The Paper's Claim: This simple method worked surprisingly well when the task was Motor Imagery (imagining moving a hand). It was competitive with much more complex methods.
The "Pattern Detective" (Convolutional Encoder):
- The Analogy: This translator is like a detective who looks for specific shapes and rhythms in the sound. They scan the audio for local patterns, like a specific drumbeat or a rising pitch, using "filters" to spot these features. This is the standard way most current EEG models work.
- The Paper's Claim: This method performed very similarly to the "Simple Ear" on the motor task but shined a bit brighter on the emotion task.
The "World-Traveling Expert" (Pretrained TSFM - MOMENT):
- The Analogy: This is the star of the show. Imagine a translator who has spent years listening to everything in the world—stock market trends, weather patterns, heartbeats, and even some brain waves. They are a "Time-Series Foundation Model" (TSFM). They are so smart they know how time works in general.
- The Twist: The researchers didn't teach this expert anything new about brains. They just "froze" the expert (kept their knowledge fixed) and asked them to listen to the brain data.
- The Paper's Claim: Even though this expert was trained on general data (not specifically brains), it worked very well. On the Emotion Recognition task, this "World-Traveling Expert" actually performed better than the simple methods, suggesting that general time-sense is useful for understanding brain emotions.

The Results: It Depends on the Job

The paper found that there is no single "best" listener for every job.

For Motor Imagery (Moving hands in your head): The Simple Ear was just as good as the complex detectives. You don't need a fancy time-analyzer to understand the rhythm of imagining movement.
For Emotion Recognition (Feeling happy or sad): The World-Traveling Expert and the Pattern Detective did better than the Simple Ear. Emotions seem to have richer, more complex time patterns that benefit from a deeper understanding of how time flows.

The Big Takeaway

The most exciting finding is that you can take a general-purpose time-series expert (trained on stock markets and weather) and use it as a "plug-in" for brain models without retraining it.

What the paper says: It works. A frozen, general-purpose model can act as a powerful feature extractor for EEG.
What the paper does not say: It does not claim this will immediately cure diseases, replace doctors, or be used in commercial brain-computer interfaces right now. It simply proves that this specific technical approach is viable and worth exploring further.

Summary

Think of it like building a car. The researchers tested three different engines (Simple, Pattern-Finder, and General-Expert) to see which one drove best. They found that for a short, straight track (Motor Imagery), a simple engine is fine. But for a winding, scenic road full of curves (Emotions), the General-Expert engine handled the turns better. Most importantly, they proved that you can use a General-Expert engine (trained for other roads) in a brain-car, and it still drives well.

Technical Summary: Temporal Feature Extractors in EEG Foundation Models

Problem Statement

Electroencephalography (EEG) foundation models aim to learn generalizable representations from large-scale brain recordings to address challenges such as high dimensionality, non-stationarity, low signal-to-noise ratios, and substantial inter-subject variability. While these models typically employ an embedding module to transform raw EEG data before refining it with a transformer backbone, the specific role of temporal feature extractors remains underexplored.

Existing approaches vary widely: some use lightweight 1D convolutional encoders (often combined with frequency-domain features), while others rely solely on linear projection layers. However, differences in embedding modules are often entangled with variations in overall architecture and training paradigms, preventing a systematic study of the direct effect of temporal feature extraction. Furthermore, it is unclear whether pretrained time-series foundation models (TSFMs), which learn general-purpose temporal representations from diverse datasets, can be effectively transferred to the EEG domain as frozen temporal feature extractors.

Methodology

The authors conduct a controlled comparison of three temporal feature extraction strategies within a unified EEG foundation model framework. To isolate the effect of the extractor, all other components—including input representation, patching strategy, masking procedure, spherical positional encoding, and the transformer backbone—are held constant.

1. Input Representation and Patching

Input: Multichannel EEG recordings ( $X \in \mathbb{R}^{C \times T}$ ).
Patching: Following Vision Transformer practices, recordings are divided into 2-second patches.
Normalization:
- For the MOMENT (pretrained TSFM) setup, the model's native preprocessing (Reversible Instance Normalization) is used.
- For Linear and Convolutional baselines, absolute maximum normalization is applied to preserve zero-crossings and signal polarity.

2. Temporal Feature Extractor Strategies

The study compares three distinct approaches:

Linear Projection: A minimal baseline where each EEG patch is projected into the model embedding space via a single linear layer.
Convolutional Temporal Encoder: A depthwise separable convolutional encoder applying shared temporal filters to each channel-patch token. It uses kernels of varying sizes along the temporal dimension, followed by pointwise convolution and flattening.
Pretrained TSFM (MOMENT): The frozen MOMENT-small model is used as a feature extractor. Unlike forecasting-oriented TSFMs, MOMENT is pretrained with a masked reconstruction objective designed for representation learning.
- Adaptation: Since MOMENT expects sequences of length 512 and the EEG patches contain 200 time points, left zero-padding and input masking are applied.
- Aggregation: Mean reduction is applied over internal patch representations to produce a fixed-dimensional embedding.
- Freezing: The MOMENT encoder is kept frozen during both pretraining and downstream training; embeddings are computed offline.

3. Model Architecture and Training

Positional Encoding: Spherical Positional Encoding (SPE) is used for spatial information (electrode locations), and standard sinusoidal encoding is used for temporal patch indices.
Backbone: A transformer encoder with self-attention and feed-forward layers.
Objective: Masked reconstruction ( $L = \|\hat{X}_M - \tilde{X}_M\|_2^2$ ) over masked patches.
Datasets:
- Pretraining: Healthy Brain Network EEG (HBN-EEG) with ~3,000 participants.
- Downstream Tasks: PhysioNet Motor Imagery (MI) and FACED (Emotion Recognition).

Key Contributions

Systematic Analysis: The paper provides a controlled analysis of the role of temporal feature extractors in EEG foundation models across two distinct downstream tasks (motor imagery and emotion recognition).
TSFM Transfer Evaluation: It evaluates a pretrained, domain-general TSFM (MOMENT) as a frozen temporal feature extractor within an EEG foundation model, comparing it against standard linear and convolutional embedding strategies.

Results

The models were evaluated using linear probing (frozen backbone) and fine-tuning protocols.

Motor Imagery (PhysioNet-MI):
- In the linear probing setting, simple linear and convolutional models performed competitively, while the MOMENT model showed lower average performance.
- In the fine-tuning setting, the linear model achieved the highest average performance, followed by MOMENT.
Emotion Recognition (FACED):
- Both convolutional and MOMENT models outperformed the linear baseline under both protocols.
- The emotion dataset benefited more from richer temporal modeling (convolutional and TSFM) compared to the motor imagery dataset.
Statistical Significance: Friedman tests indicated significant differences across models for most settings. However, post-hoc comparisons with Holm-Bonferroni correction did not reveal significant pairwise differences, leading the authors to focus on consistent trends in mean performance.

Significance and Claims

The paper concludes that no single temporal feature extraction strategy consistently outperforms the others across all benchmarks.

Task Dependency: Simple linear projections are competitive for motor imagery tasks, whereas emotion recognition benefits from richer temporal modeling (convolutional or TSFM-based).
TSFM Viability: Although not specifically adapted to EEG, the pretrained TSFM (MOMENT) serves as an effective temporal feature extractor. This suggests that general-purpose time-series representations can be transferred as frozen extractors within EEG foundation models.
Limitations: The authors note that the frozen transfer setting may constrain the model's ability to fully capture the specific temporal dynamics of EEG signals, and broader conclusions are limited by the evaluation scope. Future work is proposed to investigate adapting TSFMs during pretraining and evaluating them across a wider range of tasks.