Neural Noise- Stim-BERT reconstructs iEEG signals during brain stimulation by masking and predicting stimulation artifacts with a BERT-style transformer.1
- This enables closed-loop recording during stimulation, which is critical for adaptive neurostimulation BCIs.1 1
Weekly enrichment (2026-07-20)
- The peer-reviewed version appeared in Frontiers in Artificial Intelligence (2025;8:1502504, DOI 10.3389/frai.2025.1502504), authored by K. Menon, T. Tcheng, C. Seale, D. Greene, M. Morrell, and S. A. Desai (NeuroPace/Stanford).2 3
- Stim-BERT was trained on a very large iEEG corpus: 4,653,720 iEEG channels (over 4 million ~90-second signals) drawn from 380 NeuroPace RNS System patients being treated for epilepsy.2 3
- Recordings were converted into tokens by binning spectral power into 3 or 4 frequency bands over 1-second non-overlapping windows, yielding a discrete vocabulary of 1,000 or 10,000 tokens, comparable in size to BERT’s language vocabulary.2
- Training used masked-token self-supervised learning in the style of BERT, with about 18% of tokens masked, so the bidirectional transformer learns to predict the missing (blanked) segments from surrounding context.2
- The model captures both short- and long-term temporal context and significantly outperformed traditional interpolation methods, with its advantage growing as the duration of the stimulation blanking period increased.2 3
- Because RNS and similar devices blank or corrupt recordings during stimulation, recovering that lost signal is a prerequisite for analyzing the acute effects of stimulation and for closed-loop, adaptive neurostimulation BCIs.2
- The authors frame the contribution as two findings: neural time series can be effectively tokenized from spectral information, and a bidirectional transformer can recover neural spectral data lost to stimulation artifact.3
- Reconstruction was performed in the spectral/token domain rather than as raw voltage waveforms; exact error-metric values were not captured in the sources reviewed here.2