Foreseeing Electrical Activity of the Brain – Generative Deep Learning Models for EEG Time Series Forward Prediction

Abstract

Recent findings suggest that efficacy of transcranial magnetic stimulation (TMS) can be substantially improved with brain-state dependent stimulation. This can be done with a brain-computer interface (BCI) that triggers the stimulation based on real-time measured EEG. However, this is challenging, as algorithmic decision making takes time and brain states are known to change rapidly. One solution here is to forward predict the EEG time series – this enables the BCI to anticipate the occurence of brain states that are suitable for stimulation.

In this thesis we propose two convolutional neural network models for forecasting EEG time series. The first one is an adaptation of the WaveNet model developed for processing audio signals. The second one in turn is a multivariate adaptation of the first one.

We found that our univariate model is better at estimating instantaneous phase of an EEG signal compared to an autoregressive forward prediction model that has been previously used for brain-state dependent TMS. In addition, our multivariate model was not able to achieve more accurate predictions than our univariate model, but it did show slightly improved phase estimation accuracy.

In conclusion, results reported here indicate that deep learning is a feasible approach for EEG time series forward prediction.