Real-time noise cancellation for EEG monitoring and brain-computer interfaces

Abstract

Raw electroencephalography (EEG) data suffer from a low signal-to-noise ratio (SNR). Effective real-time cleaning of EEG would improve the robustness of EEG monitoring, brain-computer interfaces, and brain-state-dependent stimulation applications in both laboratory and real-life environments. Here, we present a real-time source-estimate-utilizing noise-discarding algorithm (rtSOUND) for EEG noise suppression. We adapted the established offline SOUND algorithm for real-time use and implemented it into an open-source EEG-processing software. Instead of using conventional temporal filtering, SOUND operates the data with a spatial filter. This makes it particularly suitable for real-time applications, since each streamed data sample can be cleaned separately in near real-time. Unlike temporal filtering methods, SOUND does not make assumptions about the frequency content of noise and brain signals but assumes that the noise topographies differ from those generated by brain activity. With an EEG forward model, rtSOUND cross-validates the EEG time series and estimates the spatial distribution of noise, resulting in a data-cleaning linear spatial filter. Each streamed EEG sample is processed synchronously, and the filter is periodically updated in an asynchronous parallel process. The algorithm rapidly updates the spatial filter in response to changes in the data, ensuring that noise reduction is both immediate and accurately adapted to the current EEG state. rtSOUND was validated with transcranial magnetic stimulation (TMS)-EEG and resting-state EEG datasets measured from 25 healthy subjects. SNR of both TMS-evoked potentials and resting-state data consistently improved across various metrics. rtSOUND particularly benefitted channels with poor SNR, while retaining the waveforms in high-quality channels. The algorithm was fast enough to process data streamed at 5 kHz. rtSOUND improves the SNR of EEG in real time, providing a robust method to improve signal quality for brain-state-monitoring applications.