CSBrain

EEG foundation models have converged on a recipe imported wholesale from language and vision: chop the signal into uniform patches, embed them, and let dense self-attention sort out the rest. CSBrain argues that this scale-agnostic treatment discards something essential about neural activity. EEG phenomena live at very different scales — a spike or evoked response lasts tens of milliseconds while a sleep rhythm unfolds over seconds, and a motor-imagery pattern is anatomically localized while an emotional response is distributed across cortex. A single fixed tokenization resolution serves one of these regimes well and the others poorly.

The model addresses this with two coupled components. Cross-scale Spatiotemporal Tokenization (CST) aggregates features over multiple temporal window lengths and multiple anatomical groupings simultaneously, producing compact tokens that carry an explicit notion of scale. Structured Sparse Attention (SSA) then relates those tokens across windows and across regions with linear rather than quadratic cost, which both keeps long-context modelling tractable and prunes the spurious correlations that dense attention tends to pick up in noisy multichannel recordings. CST and SSA are stacked alternately so that scale information is integrated progressively through depth.

CSBrain was developed by Yuchen Zhou, Jiamin Wu, and colleagues at Shanghai AI Laboratory with Sun Yat-sen University, the Chinese University of Hong Kong, and the Karlsruhe Institute of Technology, posted in June 2025 and accepted as a Spotlight at NeurIPS 2025. It positions itself directly against LaBraM and CBraMod, the two EEG foundation models it most closely resembles in training recipe.

#Key Features

• Cross-scale tokenization: Multi-scale temporal and spatial kernels (K = 3 scales, with exponentially decaying embedding dimension across scales) fold short bursts and slow rhythms into the same token vocabulary.
• Linear-cost structured attention: Inter-window and inter-region attention runs in O(N·k) rather than O(N²), making long recordings and high channel counts affordable.
• Single backbone, eleven tasks: One pretrained model fine-tunes to motor imagery, emotion recognition, seizure detection, sleep staging, imagined speech, vigilance estimation, stress detection, mental disorder diagnosis, event-type classification, abnormality detection, and slowing-event classification.
• Montage-agnostic evaluation: The 16 benchmark datasets differ in electrode layout, sampling rate, and duration, so results reflect transfer across recording setups rather than within one lab's protocol.
• Released code and weights: Pretraining and fine-tuning code is public, with pretrained and per-task checkpoints distributed through Google Drive.

#Technical Details

Pretraining used masked autoencoding over the Temple University Hospital EEG Corpus (TUEG) — 1,109,545 EEG segments totalling more than 9,000 hours — with a 50% masking ratio, run for 40 epochs on four NVIDIA A100 GPUs in roughly 101 hours. The architecture is a stack of L paired CST and SSA layers; the paper does not report a total parameter count. Across 11 tasks and 16 public datasets (BCIC-IV-2a, PhysioNet-MI, SHU-MI, FACED, SEED-V, CHB-MIT, Siena, ISRUC, HMC, BCIC2020-3, SEED-VIG, MentalArithmetic, Mumtaz2016, TUEV, TUAB, TUSL), CSBrain reaches a macro-average of 0.7095, ahead of CBraMod at 0.6760, LaBraM at 0.6697, and BIOT at 0.6322. Individual gains are largest on the harder decoding problems: 56.57% balanced accuracy on BCIC-IV-2a motor imagery against CBraMod's 51.38%, 60.04% on BCIC2020-3 imagined speech against 53.73%, and 85.71% on TUSL slowing-event classification against LaBraM's 76.25%.

#Applications

The model is intended as a general backbone for EEG analysis rather than a finished clinical tool. Brain-computer interface groups can fine-tune it for motor imagery or imagined-speech decoding with limited labelled data; sleep and epilepsy researchers get a starting point for staging, seizure detection, and abnormality screening; and affective computing and driver-vigilance work can reuse the same encoder. The practical draw is that a single pretrained model covers task families that previously each required a bespoke architecture, which matters most for labs with small, heterogeneous EEG collections.

#Impact

CSBrain's argument — that cross-scale structure is a necessary inductive bias for neural signals, not an optional refinement — is the kind of claim that shapes how the next generation of biosignal models is designed, and the consistent margins over CBraMod and LaBraM across 16 datasets give it empirical weight. The NeurIPS 2025 Spotlight designation reflects peer assessment of that contribution. Two limitations are worth stating plainly: pretraining draws on a single clinical corpus (TUEG), so transfer to consumer-grade or dry-electrode hardware is untested, and the repository carries no licence file, which leaves the terms of reuse for both code and the Google Drive checkpoints formally undefined.