TL;DR
ReBrain is a novel framework that uses high-fidelity, invasive intracranial EEG (iEEG) to create a "neural prior" for cleaning up noisy, non-invasive scalp EEG. By re-expressing scalp signals through a discrete codebook learned from iEEG, it consistently boosts the performance of downstream seizure detection models across various architectures and hospitals without requiring paired data or fine-tuning.
What Problem Does This Paper Address?
Scalp electroencephalography (EEG) is the workhorse of clinical brain monitoring, but it has a significant "signal-to-noise" problem. Because the signal must travel through the skull and scalp, it becomes blurred (volume conduction) and heavily contaminated by non-neural interference like eye blinks, muscle movements, and environmental noise.
In contrast, intracranial EEG (iEEG)—recorded via electrodes implanted directly inside the skull—offers a much cleaner, high-fidelity view of neural activity. However, iEEG is invasive and scarce, making it impossible to use for routine monitoring. The challenge is that these two modalities are rarely recorded simultaneously in the same patient, they use different electrode layouts (channel spaces), and there is no "clean" ground truth for scalp EEG to train a standard denoiser. The authors ask: Can we use the scarce, clean iEEG to help us understand the abundant, noisy scalp EEG?
Methods
The authors introduce ReBrain, a framework based on the concept of cross-modal reprogramming. Instead of trying to translate one signal to another directly, they learn the "vocabulary" of clean brain activity from iEEG and use it as a filter for scalp EEG.
- Stage A (Fitting the Prior): They train an encoder-decoder model with a Hierarchical Vector-Quantized (VQ) bottleneck exclusively on iEEG data. This forces the model to represent brain signals as a sequence of discrete "atoms" drawn from a finite codebook of clean neural patterns. They use a multi-scale STFT discriminator to ensure the model captures sharp, high-frequency transients (like seizure spikes).
- Stage B (Freezing): Once trained, the model is frozen. It now acts as a fixed, channel-agnostic transformation.
- Stage C (Integration): This frozen prior is inserted as a "plug-and-play" front-end for any existing scalp EEG model. When a noisy scalp signal passes through ReBrain, it is re-expressed using the iEEG-learned vocabulary. Components of the signal that don't match this neural vocabulary (like noise) are naturally attenuated.
Crucially, ReBrain is channel-agnostic, meaning it processes each electrode's signal independently, allowing it to work across different clinical setups without modification.
Datasets
The study utilizes a diverse range of datasets to prove the generalizability of the approach:
- iEEG (Prior Training): The SWEC database (Bern University) and a separate Multi-Center (MC) iEEG corpus.
- Scalp EEG (Evaluation): Four major clinical datasets: CHB-MIT (pediatric), TUSZ (the largest public clinical corpus), Siena (adult focal epilepsy), and Beirut (adult epilepsy).
- Non-human iEEG: A canine epilepsy dataset was used for mechanistic analysis to see if neural priors translate across species.
Key Insights and Ideas
- The "Neural Vocabulary" Effect: The VQ bottleneck acts as a physical constraint. Because the codebook only contains patterns found in clean iEEG, the model literally lacks the "words" to describe scalp noise, effectively filtering it out during reconstruction.
- Denoiser vs. Amplifier: Through geometric analysis, the authors found that ReBrain doesn't just push seizure and non-seizure classes further apart; it significantly reduces the "scatter" (variance) within each class. It makes the neural patterns more consistent and recognizable for downstream models.
- Data, Not Architecture: The authors ran a control experiment where they trained the same architecture on scalp EEG instead of iEEG. This version failed to provide the same gains, proving that the benefit comes specifically from the high-fidelity structure inherent in iEEG.
- Cross-Species Transfer: Interestingly, a prior trained on canine iEEG also improved human scalp EEG models. This suggests that the fundamental "language" of mammalian brain rhythms is conserved enough to be useful across species.
Key Results
- Universal Improvement: ReBrain improved performance for every host model (CNN, RNN, Transformer, GNN) across every dataset tested.
- Performance Boost: On average, it provided a 0.08 increase in F1-score and a 0.04 increase in AUROC. The gains were most pronounced on the "hardest" datasets where raw signals were most difficult to classify.
- Edge-Ready Efficiency: The deployed model is lightweight (2.24M parameters). It can process one hour of 18-channel EEG recording in just 157 seconds on a single CPU thread—over 20 times faster than real-time. This makes it highly practical for deployment on low-power clinical devices like a Raspberry Pi.
Limitations and Future Work
- Task Specificity: The current prior was trained on iEEG from epilepsy patients, making it highly effective for seizure detection. Its impact on non-pathological tasks (like Brain-Computer Interfaces) was not yet statistically significant.
- Future Directions: The authors suggest that future work could involve building more general "foundation priors" by training on larger, more diverse iEEG corpora that include non-pathological brain states. This could potentially turn ReBrain into a universal enhancement layer for all types of brain-sensing technology.
Summary updated