B[FM]²

Electroencephalography (EEG) foundation models aim to learn one backbone that transfers across clinical diagnosis and brain-computer interface (BCI) decoding. Almost all of them borrow the recipe that worked for text and images: cut the continuous multi-channel waveform into patches or codebook tokens, mask some of them, and train a transformer to reconstruct the rest — the strategy behind LaBraM, CBraMod, and NeuroLM. B[FM]² (Brain Foundation Model via Flow Matching) argues that this discretization works against the data, fragmenting continuous brain rhythms and obscuring fine-grained temporal dynamics.

The model instead pretrains directly on the raw signal with continuous-time flow matching: no patches, no tokenization, no masking. It learns a velocity field that transports noise to real EEG, making pretraining a generative process over the waveform itself rather than a reconstruction task over discrete symbols. Applying flow matching to EEG exposes an architectural mismatch, though. Time is densely sampled and highly autocorrelated across thousands of timepoints, while the electrode axis is short — tens of channels at fixed, physically meaningful scalp positions. A standard 2D velocity network treats the two axes alike and downsamples both.

B[FM]²'s answer is SplitUNet, a velocity network that factorizes each block into separate 1D temporal and 1D electrode convolutions and downsamples only along time, preserving electrode topology through the whole hierarchy. The work comes from researchers at MIT and KU Leuven, and is currently an arXiv preprint.

#Key Features

• Tokenizer-free pretraining: Optimal-transport conditional flow matching (OT-CFM) is applied to the raw 19-channel waveform, so no patching, codebook, or masking step stands between the model and the continuous signal.
• SplitUNet velocity network: Every 2D convolution is replaced by a factorized (1+1)D operator — a temporal Conv1D followed by an electrode Conv1D — with four encoder stages that halve only the time axis and a self-attention bottleneck for global mixing.
• Data-efficient pretraining: The released results use 36,895 thirty-second segments (roughly 307 hours, about 1% of the TUEG corpus hours), one to two orders of magnitude less pretraining data than comparable EEG foundation models.
• Broad downstream transfer: A single backbone sets the best reported results on 7 of 9 standard EEG classification tasks spanning seizure detection, sleep staging, event and abnormality classification, mental state, and emotion recognition.
• Realistic EEG synthesis: Because pretraining is generative, the model samples new recordings; two board-certified neurologists scoring interleaved real and synthetic 30-second windows agreed with the ground truth at below chance (Cohen's κ = −0.096).

#Technical Details

Pretraining draws on the Temple University Hospital EEG Corpus (TUEG), which holds 69,652 recordings from 14,987 subjects. After band-pass and notch filtering, segmentation into non-overlapping 30-second windows on the standard 19-channel 10–20 montage, amplitude thresholding, and normalization, the pool contains 1,109,545 segments; B[FM]² samples 36,895 of them (about 3.3%) for pretraining. Downstream evaluation fine-tunes the pretrained network end-to-end with cross-entropy, conditioning the backbone on a fixed timestep t = 0 and attaching a single linear head — not a frozen linear probe. The nine benchmarks are Mumtaz, MAT, Siena, ISRUC, HMC, TUEV, TUAB, BCIC-IV-2a, and SEED-V, scored against BIOT, LaBraM-Base, CBraMod, REVE, and CSBrain. B[FM]² leads on 7 of the 9; on the motor imagery task BCIC-IV-2a it places second, behind REVE.

#Applications

The pretrained backbone is a general-purpose starting point for laboratories with limited labeled EEG, covering clinical neurophysiology tasks such as seizure detection, sleep staging, and abnormal-recording screening as well as BCI decoding of motor imagery, mental state, and affect. Its generative half is equally usable: sampling realistic multi-channel EEG offers a route to data augmentation for rare events and to sharable surrogate recordings in a field where patient data is tightly restricted.

#Impact

B[FM]² makes a concrete architectural argument for biosignal foundation models — that continuous-time generative objectives on the raw waveform can beat masked token prediction, and that the asymmetry between a long time axis and a short, spatially structured electrode axis deserves to be built into the network rather than averaged away. The efficiency result is the sharper claim: matching or exceeding larger models from roughly 307 hours of pretraining suggests that inductive bias, not corpus size, is the binding constraint in EEG pretraining. Two caveats accompany the work. It remains an unrefereed preprint, and no code or model weights have been released — the project page hosts the paper and a blog post only. Pretraining also draws on a single clinical corpus recorded at one montage, so transfer to different electrode configurations or non-clinical populations needs empirical validation.