Reconstructing signal during brain stimulation with Stim-BERT: a self-supervised learning model trained on millions of iEEG files.

Abstract

Brain stimulation has become a widely accepted treatment for neurological disorders such as epilepsy and Parkinson's disease. These devices not only deliver therapeutic stimulation but also record brain activity, offering valuable insights into neural dynamics. However, brain recordings during stimulation are often blanked or contaminated by artifact, posing significant challenges for analyzing the acute effects of stimulation. To address these challenges, we propose a transformer-based model, Stim-BERT, trained on a large intracranial EEG (iEEG) dataset to reconstruct brain activity lost during stimulation blanking. To train the Stim-BERT model, 4,653,720 iEEG channels from 380 RNS system patients were tokenized into 3 (or 4) frequency band bins using 1 s non-overlapping windows resulting in a total vocabulary size of 1,000 (or 10,000). Stim-BERT leverages self-supervised learning with masked tokens, inspired by BERT's success in natural language processing, and shows significant improvements over traditional interpolation methods, especially for longer blanking periods. These findings highlight the potential of transformer models for filling in missing time-series neural data, advancing neural signal processing and our efforts to understand the acute effects of brain stimulation.