CBraMod

Config name: CBraMod

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

Reference: Wang et al., CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding, 2025

CBraMod is a criss-cross transformer for EEG that decomposes attention into separate spatial (S-Attention across channels) and temporal (T-Attention across patches) streams. The pretrained checkpoint was trained on the Temple University Hospital EEG corpus (TUEG) at 200 Hz with a patch size of 200 samples (1 s) and 19 standard 10-20 channels. CBraMod accepts any number of input channels without a fixed channel set or channel projection.

Pretraining data overlap

Warning

CBraMod was pretrained on the full TUEG corpus. The TUEG corpus overlaps with the NeuralBench pathology task (TUH Abnormal EEG Corpus), artifact task (TUH EEG Artifact Corpus), and clinical_event task (TUH EEG Events Corpus), all of which are subsets of TUEG. Results on these tasks may therefore be inflated.

Known limitations

• n_times must be exactly divisible by patch_size (200). Tasks with non-integer-second durations at 200 Hz will fail unless the window duration is adjusted.

• The spectral projection path contains a hardcoded reshape that assumes patch_size=200. Using a different patch size would require changes to the braindecode implementation.