BLPM

An EEG foundation model that fits a fresh classification head per dataset — six outputs for clinical event typing, five for sleep staging, four for motor imagery — is one initialization wearing seven hats. The encoder transfers; the decision layer does not. BLPM (Brain Latent Predictive Model) removes that layer: each class label is written out in natural language and embedded with a pretrained text encoder, and the model predicts where in that text embedding space an EEG segment belongs. Classification becomes a nearest-neighbour lookup against whatever candidate labels the task supplies, so the candidate set is an input rather than a fixed architectural choice.

The mechanism has two halves. A Continuous EEG Latent Predictive (CELP) encoder is pretrained to predict the latent representations of masked channel-time regions from the visible context, following the joint-embedding predictive architecture recipe of an online encoder, an exponential-moving-average target encoder, and a Smooth L1 loss. Because the objective lives in embedding space rather than signal space, it does not reward fitting the amplitude noise and subject-specific idiosyncrasies that dominate raw EEG. A Multi-Query Semantic Decomposition (MQSD) module then cross-attends natural-language queries — temporal dynamics, spectral characteristics, spatial relationships, morphology — over the EEG tokens, so one segment can be read out along whichever axis the task cares about.

BLPM comes from Seong-Whan Lee's group at Korea University and sits against the dominant EEG-language approach: NeuroLM and LaBraM quantize EEG into a discrete codebook before handing it to a language model, while CBraMod and LUNA pretrain by reconstructing masked signal. BLPM does neither.

#Key Features

• Semantic matching in place of a classifier: inference is one forward pass producing an answer embedding, scored by dot product against the encoded candidate labels. Adding a task means writing its label texts, not training a head.
• Latent predictive pretraining: the CELP encoder predicts masked latent targets from visible context under structured temporal-span and channel-block masks, never reconstructing the underlying signal.
• Language-guided query decomposition: MQSD's queries are tokenized with the language model's own embedding layer, so the readout axes carry linguistic meaning rather than being generic learnable slots.
• Dual-branch patch embedding: each channel segment is encoded by 1D convolutions and by a linear projection of its FFT magnitude spectrum, fusing time- and frequency-domain structure.

#Technical Details

The CELP encoder is a 12-layer transformer with a hidden dimension of 384 and six attention heads, paired with a six-layer predictor of the same width. It is pretrained on the Temple University Hospital EEG Corpus (TUEG) — 27,062 hours across 69,652 recordings from 14,987 subjects, cut into 30-second segments. Semantic alignment then freezes the encoder and trains MQSD with a bidirectional multi-positive InfoNCE objective, treating samples that share a label text as positives rather than false negatives. Instruction tuning adapts a Llama 3.2-1B-Instruct backbone through LoRA (rank 8) on its query and value projections, with the attention mask switched from causal to bidirectional; label texts are encoded with EmbeddingGemma.

Evaluation follows the NeuralBench protocol. Balanced accuracy reaches 82.3% on TUAB, 57.2% on TUEV, 77.1% on HMC sleep staging, 69.1% on COG-BCI workload, 75.4% on Mental Arithmetic and 34.3% on nine-class FACED emotion recognition — best of the compared methods on six of the seven, with PhysioNet-MI motor imagery at 68.2% trailing REVE by 1.2 points. The baselines are each adapted separately to every dataset, while BLPM reports one multi-task checkpoint throughout. TUAB and TUEV are TUH-family corpora sharing provenance with the pretraining data; the other five are external. Ablations isolate the contributions — dropping semantic alignment costs 6.91 points on FACED, swapping embedding prediction for autoregressive token prediction costs 1.91 points on TUEV, and replacing latent prediction with masked reconstruction 2.55.

#Applications

The evaluation tasks map onto where EEG is actually read: abnormality screening and event typing in clinical neurophysiology, sleep-stage scoring, motor-imagery decoding for brain-computer interfaces, and workload, stress and emotion estimation. Because the label set is supplied at inference, a group with an in-house paradigm can describe its classes in text and score them against the same weights — useful where labelled recordings are too few to fit a reliable head. Per-channel patching lets 4- to 64-electrode recordings share one encoder.

#Impact

BLPM's contribution is an argument about interface rather than scale: that a continuous shared embedding space preserves discriminative structure vector quantization discards, and that a text-embedding target is a workable substitute for a learned classifier across heterogeneous protocols. Several limits bound the evidence. It is a preprint awaiting peer review, and no code or weights have been released, so results are not yet independently reproducible. Pretraining draws on a single clinical corpus, narrower in recording conditions than the multi-dataset corpora behind some baselines. And the model picks among supplied candidates rather than generating free-text interpretations — the trade for skipping autoregressive decoding.