LaBraM

Config name: Labram

Pretrained checkpoint: braindecode/labram-pretrained (HuggingFace Hub)

Reference: Jiang et al., Large Brain Model for Learning Generic Representations with Tremendous EEG Data in BCI, ICLR 2024

LaBraM is a patch-based EEG transformer with learned spatial and temporal embeddings. The pretrained checkpoint was trained on 16 public EEG datasets totalling ~2534 hours at 200 Hz with a patch size of 200 samples (1 s) and up to 128 channels from the standard 10-20/10-10 system. Input channels are matched case-insensitively against the 128-channel LABRAM_CHANNEL_ORDER; unmatched channels are dropped.

Pretraining data overlap

Warning

LaBraM was pretrained on 16 public EEG datasets totalling ~2534 hours (Appendix D of the paper). The pretraining corpus includes PhysioNet Motor Imagery (Schalk et al., 2004) and TUAR (TUH EEG Artifact Corpus), both of which are also NeuralBench evaluation datasets. Results on the affected tasks may be inflated because the model has seen the same recordings (though not the same labels) during self-supervised pretraining.

The paper explicitly states that TUAB and TUEV recordings were excluded from pretraining, so pathology and clinical_event are not affected.

Pretraining dataset	NeuralBench task	NeuralBench study
PhysioNet Motor Imagery	motor_imagery	Schalk2009
TUAR	artifact	Hamid2020

Known limitations

• Zero-padding for non-divisible durations introduces artefactual silence at the end of the window, which may slightly reduce performance compared to natively aligned durations.