BrainWave (Brant-2)

BrainWave (developed under the codename Brant-2) is a foundation model for brain electrical signals built by the BrainNet group at Zhejiang University (Zhizhang Yuan and colleagues), first released as an arXiv preprint in February 2024. It is positioned as the first foundation model to span both invasive recordings — stereo-EEG (SEEG) and intracranial EEG (iEEG) — and non-invasive scalp EEG within a single pretrained backbone, where prior efforts typically targeted one modality.

The core problem BrainWave addresses is the extreme heterogeneity of neural recordings: sampling rates, channel counts, electrode montages, recording sites, and patient populations all vary widely, which has historically forced researchers to train bespoke models per dataset and per disorder. By pretraining a single channel-agnostic encoder on a very large, mixed corpus, BrainWave learns representations that transfer across recording conditions and across diseases, including in zero-shot and few-shot settings without task-specific fine-tuning.

BrainWave is the successor to the authors' earlier Brant model (which focused on intracranial signals). Brant-2/BrainWave broadens the scope to unify invasive and non-invasive modalities and to target clinical diagnosis of neurological disorders, making it a notable entry in the growing landscape of biosignal foundation models.

#Key Features

• Cross-modality coverage: A single backbone ingests both invasive (iEEG/SEEG, 48–238 channels) and non-invasive (EEG, 1–64 channels) recordings, rather than specializing in one signal type.
• Scale-alignment layer: Each 1-second signal patch is converted to a spectrogram with constant time–frequency resolution, letting the model handle diverse sampling rates (EEG ~100–1024 Hz; iEEG ~1000–4096 Hz) without resampling.
• Channel-agnostic attention: Channels are encoded independently, then a bidirectional self-attention stage models inter-channel correlations, so the model accepts variable channel counts and montages.
• Zero- and few-shot transfer: BrainWave supports cross-hospital and cross-subtype zero-shot transfer and strong 3-/8-shot classification, reducing reliance on large labeled clinical datasets.
• Broad clinical scope: Evaluated on epilepsy/seizure detection, Alzheimer's disease, major depressive disorder, schizophrenia, and ADHD.

#Technical Details

BrainWave is a RoBERTa-style transformer encoder with roughly 86 million parameters: 10 layers, 16 attention heads, hidden size 768, and a maximum sequence of 61 tokens (60 one-second signal patches plus a [CLS] token). It is pretrained by masked reconstruction of time–frequency representations over ~3.16 billion signal patches. The pretraining corpus totals 13.79 TB across 40,907 hours from roughly 16,000 individuals — 10.63 TB of iEEG (5,231 hours, 91 subjects) and 3.16 TB of EEG (35,675 hours, 15,906 subjects), drawing on public datasets such as TUEG, Sleep-EDF, CAP, HMC, Siena, SRM, and CCEP alongside private corpora. Across 13 downstream datasets and 20+ tasks, the authors report average gains of roughly 11.9% AUROC for cross-subject diagnosis over the next-best baseline, zero-shot cross-hospital seizure transfer around 93.8% AUROC, and 8-shot results such as ~91.9% AUROC on absence-seizure data and ~89.8% AUROC for major depressive disorder.

#Applications

BrainWave targets clinical neurophysiology workflows where labeled data is scarce and recording setups differ across sites. Use cases demonstrated in the paper include seizure detection and seizure-onset-zone localization, cross-hospital and cross-subtype seizure transfer, and screening or biomarker prediction for Alzheimer's disease, depression, schizophrenia, and ADHD. Because the same encoder serves both intracranial monitoring (e.g., presurgical epilepsy evaluation) and routine scalp EEG, it is of interest to epileptologists, sleep and psychiatric researchers, and machine-learning groups building diagnostic tools that must generalize across hospitals and patient populations.

#Impact

BrainWave is one of the first efforts to unify invasive and non-invasive brain recordings under a single foundation model, extending the authors' earlier Brant work and contributing to the rapid emergence of biosignal foundation models. Its emphasis on zero- and few-shot clinical transfer is significant for settings where collecting large labeled corpora is impractical. Important caveats remain: the strongest results are on internal benchmarks reported in a preprint (peer review of the latest version was ongoing), and as of this writing the authors state that code and model weights will be released upon publication, so independent reproduction and broad adoption are still pending.