Removing TMS Artifacts from EEG Recordings Using a Deep Gated Recurrent Unit

The combination of transcranial magnetic stimulation (TMS) and electroencephalography (EEG) provides a direct means of assessing focal and distributed cortical behavior such as excitation/inhibition, intrinsic oscillatory activity and connectivity. However, TMS-EEG poses a number of technical challenges, foremost of which is removal of stimulation-induced artifacts that are several orders of magnitude larger than the neural signals of interest and typically obscure critical early neural responses to the stimulation. Here we describe a non-linear non-causal neural network predictor, built using a Gated Recurrent Unit architecture, and demonstrate its use to remove TMS artifacts from EEG recorded on a phantom, as well as real EEG synthetically contaminated by artifacts from the phantom experiment. Our results indicate that this artifact removal algorithm may decontaminate EEG signals as early as 6ms following stimulation. Given this result we discuss the future development of neural network based predictors for TMS artifact rejection.