🤖 machine learning

Channel Adaptation for EEG Foundation Models: A Systematic Benchmark Across Architectures, Tasks, and Training Regimes

This paper provides the first systematic benchmark of channel adaptation methods for EEG foundation models, revealing that the optimal approach depends on the specific model architecture and task, and demonstrating that compact, EEG-specific models can outperform significantly larger counterparts.

Original authors: Kuntal Kokate, Bruno Aristimunha, Dung Truong, Arnaud Delorme

Published 2026-04-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kuntal Kokate, Bruno Aristimunha, Dung Truong, Arnaud Delorme

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 you have a group of world-class musicians who have spent years mastering a very specific instrument—let’s say, a vintage 20-key piano. They are absolute geniuses at playing it.

Now, imagine you want to hire these musicians to play on a modern 88-key grand piano, or perhaps a strange, custom-built synthesizer with a completely different layout. If you just tell them, "Go play," they’ll be lost. They don't know where the notes are anymore.

This paper is about solving that exact problem, but for AI "brains" that listen to EEG (brainwave) signals.

The Problem: The "Language" of Brainwaves

Scientists are building "Foundation Models" for the brain—massive AI systems that have "listened" to thousands of hours of brain activity. They are incredibly smart, but they have a major flaw: they are picky eaters.

Most of these AI models were trained on a specific "montage" (a specific map of where the sensors are placed on the head). If one hospital uses 20 sensors and another uses 64, the AI gets confused. It’s like trying to plug a USB-C cable into an old iPhone charger; the "shape" of the data doesn't match.

The Experiment: The "Universal Adapters"

The researchers tested five different AI models and four different "adapters" (ways to translate the brainwaves so the AI can understand them).

Think of these adapters like different types of translation tools:

The "Quick Sketch" (Conv1d): This is like a fast, rough translation. It doesn't care where the sensors are; it just tries to mathematically "squish" the new signals into the old shape.
The "Map Maker" (SSI): This uses physics and geometry. It looks at where the sensors are on a 3D model of a head and "guesses" what the signal would have been if the sensors were in the "correct" spots.
The "Smooth Operator" (Riemannian): This focuses on the vibe or the statistical "texture" of the brainwaves to make sure the new data feels familiar to the AI.
The "Global View" (OmnEEG): Instead of looking at individual sensors, this turns the brainwaves into a "global map" of activity, ignoring the specific sensor locations entirely.

The Big Discoveries

1. The "Rigid" vs. "Flexible" Musicians
The researchers found that AI models fall into two camps:

The Rigid Musicians (BENDR, Neuro-GPT): These models are like the vintage piano players. They cannot function without an adapter. If you don't give them a translation tool, they simply fail.
The Flexible Musicians (EEGPT, LUNA, CBraMod): These are the modern synthesizers. They are built to handle any number of sensors natively. They are much more "plug-and-play."

2. The "Too Many Cooks" Problem (Negative Transfer)
This was a surprising finding. If you take a "Flexible" AI and try to force an external adapter on it while you are teaching it a new task (fine-tuning), the AI actually gets dumber.

It’s like trying to give a polyglot (someone who speaks many languages) a translation dictionary while they are trying to learn a new dialect. The dictionary actually confuses them and makes them stumble. The researchers call this "negative transfer."

3. Size Isn't Everything
In the AI world, "bigger is usually better." But this study found that a tiny, compact model called CBraMod (only 5 million parameters) actually beat models that were 31 times larger! It turns out that if you build a specialized, efficient "brain" for EEG, you don't need a massive, lumbering giant to get the job done.

Why does this matter?

In the future, we want to use AI to detect depression, monitor epilepsy, or help people control prosthetic limbs. For that to work, the AI needs to work in any hospital, with any equipment.

This paper provides the instruction manual for engineers, telling them: "If your AI is rigid, use this adapter; if it's flexible, leave it alone; and if you want to detect depression, use the global map approach."

Technical Summary: Channel Adaptation for EEG Foundation Models

1. Problem Statement

The deployment of EEG foundation models in real-world scenarios is hindered by montage heterogeneity. Most pretrained EEG models are trained on specific electrode configurations (e.g., a fixed 19-channel or 22-channel system). However, downstream clinical or consumer datasets frequently use varying channel counts (e.g., 24, 64, or 128 channels) and different spatial layouts. There is currently no systematic understanding of how to best map these arbitrary source montages to the fixed input dimensions required by pretrained models, nor how different architectural designs (rigid vs. flexible) interact with various adaptation strategies.

2. Methodology

The authors conducted a large-scale, controlled systematic benchmark involving over 200 experimental conditions.

Models Evaluated: Five EEG foundation models ranging from 5M to 157M parameters, representing different architectures:
- Rigid-montage models: BENDR (Conv encoder) and Neuro-GPT (Conv + GPT-2).
- Flexible-montage models: EEGPT (Patch Transformer), LUNA (Cross-attention), and CBraMod (Criss-cross Transformer).
Adaptation Methods: Four primary families of channel mapping:
1. Learned Projection (Conv1d): A trainable $1 \times 1$ convolution mapping source to target channels.
2. Spatial Interpolation (SSI): A physics-based reconstruction using 3D electrode positions and spherical splines.
3. Source-Space Decomposition (OmnEEG): Projecting signals into a topology-agnostic representation of 25 spherical harmonic coefficients.
4. Geometric Domain Adaptation (Riemannian Re-centering): Per-subject whitening on the SPD manifold to align distributions.
Training Regimes:
- Linear Probe: The encoder is frozen; only the classification head is trained.
- Supervised Fine-Tuning (SFT): All model parameters are unfrozen and trained end-to-end.
Datasets: Five diverse datasets covering four task domains: Motor Imagery (BCIC2A, PhysioNet), Clinical Event Detection (TUEV), Emotion Recognition (FACED), and Depression Screening (MDD).

3. Key Contributions

The Rigid/Flexible Dichotomy: The study identifies a fundamental split in model design. Rigid models require external adaptation to function, whereas flexible models can handle arbitrary channels natively through internal architectural mechanisms.
The Probe-SFT Asymmetry: The authors discover that external adaptation can be detrimental during fine-tuning. While external methods help frozen models (Probe), they can cause negative transfer during SFT in flexible models by creating conflicting optimization targets between the adapter and the internal channel-handling layers.
Architecture-Dependent Optimization: The study proves there is no "one-size-fits-all" adapter; the optimal method depends entirely on the backbone's inductive bias.
Scale vs. Efficiency: The benchmark provides evidence that compact, EEG-specific architectures (like the 5M-parameter CBraMod) can outperform models up to 31 times larger.

4. Key Results

Performance by Model Type:
- BENDR/Neuro-GPT: Require external mapping. Conv1d is best for BENDR, while SSI/Riemannian is superior for Neuro-GPT.
- EEGPT/CBraMod: Native handling is highly effective. For CBraMod, native SFT achieved the study's highest accuracy (87.6% on MDD).
- LUNA: Uniquely benefits from external adaptation (SSI/Riemannian) in both Probe and SFT regimes due to its cross-attention bottleneck.
Task-Specific Success: Source-space decomposition (OmnEEG) was uniquely effective for depression detection (MDD), as it preserves global hemispheric patterns necessary for clinical biomarkers while discarding focal spatial noise.
Task Difficulty: Emotion recognition (FACED) proved to be the most challenging task, with all models significantly underperforming compared to specialized methods.

5. Significance and Practical Guidelines

This paper provides a decision framework for researchers and clinicians deploying EEG models:

If using a frozen encoder: Use external adaptation (Conv1d for convolutional models; Riemannian for spatial-heavy models).
If fine-tuning a rigid model: Use Conv1d or SSI/Riemannian depending on the architecture.
If fine-tuning a flexible model: Use native handling. Adding an external adapter during SFT risks "catastrophic" performance collapse due to gradient conflict.

The study shifts the focus of EEG foundation model research from simply "scaling up parameters" to "optimizing architectural flexibility and adaptation compatibility."