BIOT

Config name: biot

Pretrained checkpoint: braindecode/biot-pretrained-six-datasets-18chs (HuggingFace Hub)

Reference: Yang et al., BIOT: Biosignal Transformer for Cross-data Learning in the Wild, NeurIPS 2023

BIOT is a transformer-based foundation model for biosignal classification. The pretrained checkpoint was trained on six EEG datasets (PREST, SHHS, CHB-MIT, IIIC Seizure, TUAB, TUEV) at 200 Hz using 18 bipolar channels derived from the Temporal Central Parasagittal (TCP) montage.

In NeuralBench, a Conv1d channel projection maps from the recording’s arbitrary channel count to the 18 expected channels, allowing BIOT to be evaluated on datasets with diverse montages without manual channel selection.

Pretraining data overlap

Warning

The biot-pretrained-six-datasets-18chs checkpoint was pretrained on six EEG datasets including TUAB and TUEV (Section 3.6 of the paper). These are used both in pretraining and as NeuralBench evaluation datasets, so results on the affected tasks may be inflated.

Pretraining dataset	NeuralBench task	NeuralBench study
TUAB	pathology	Lopez2017
TUEV	clinical_event	Harati2015

Known limitations

• Fixed spatial embeddings – The pretrained positional embeddings are tied to the 18 TCP channels. The Conv1d channel projection learns a linear mapping into this space, but there is no guarantee that the resulting channels align with the original TCP bipolar semantics.