Prediction of Scalp EEG Waveforms from Forehead Electrodes Using Convolutional Neural Networks to Improve Signal-to-Noise Ratio

The electroencephalogram (EEG) is a non-invasive method for measuring brain activity, and event-related potentials (ERPs)—EEG responses observed to be time-locked to events—have been used for brain-computer interfaces (BCI) in real-world environments. An EEG is generally measured from electrodes placed on the scalp. However, it is not suitable for daily use because the preparation time is relatively long, and the electrodes are likely to cause discomfort to the user. EEG measurements from disposable electrodes placed on the forehead (forehead EEG) have been used to mitigate this disadvantage. However, because many ERP components used in BCI show the maximal voltage on the scalp, the signal-to-noise ratio (SNR) of ERPs obtained from a forehead EEG is low, which may affect the reliability of a BCI system. To address this shortcoming, we propose convolutional neural networks that predict the EEG signal measured from electrodes placed on the scalp (scalp EEG) from forehead EEG. In the study, we focused on predicting the mismatch negativity (MMN) responses, and single-trial scalp EEG at Fz was predicted from three forehead EEG measures (Fpz, horizontal, and vertical electrooculograms). Data were measured while nine subjects performed a passive auditory oddball task. To evaluate the proposed model, the mean squared error (MSE) between the observed single-trial EEG at Fz and the predicted single-trial EEG from three forehead EEG measures was calculated, as well as the MSE between the observed ERP difference wave (deviant – standard) at Fz and the difference wave predicted from three forehead EEG measures within the time window in which MMN was observed. The result showed that within the time window in which MMN was observed, the MSE between the ERP difference wave at Fz and the ERP difference wave predicted from three forehead EEG measures was significantly smaller than the MSE between the ERP difference wave at Fz and the ERP difference wave at the forehead (Fpz). This indicates that the proposed neural network improved the SNR of the forehead EEG for predicting ERP responses at the scalp and could lead to enhancing the usefulness of forehead EEG for BCI use in daily life.