Time-series models can forecast long periods of human temporal EEG responses to randomly alternating visual stimuli

Visual stimuli with constant temporal frequency input is known to induce peaks in the driving frequency of the power spectrum of the electroencephalogram (EEG) over the visual cortex. While EEG responses with random temporal frequencies (m-sequences) have been studied, the underlying biophysical mechanisms that shape these responses are not fully understood. We analyze our new EEG data from a controlled experiment with m-sequence inputs and model the EEG using statistical time series models that mimic biophysical mechanics: an autoregressive (AR) model, adding exogenous input to AR (ARX), adding moving average terms (ARMAX), and finally adding a seasonality term (SARMAX). We implement computational methods to robustly handle model instabilities induced by this data, fitting these models with the Box-Jenkins methodology and assessing forecast performance for long periods of several seconds out-of-sample. We find in-sample fits are good in all models despite the complexities of the visual pathway, and that all models can forecast aspects of EEG: including the distribution of point-wise values in time, the point-wise Pearson’s correlation of EEG and model, and the frequency content. Surprisingly, we find little variation in the performance among these models, with the most sophisticated model (SARMAX) performing comparatively poorly in some instances. Our results suggest the simplest AR model is viable and can out perform more complicated models. Since these models are relatively simple and more transparent than contemporary models with numerous parameters, our study could inform future mechanistic studies of the temporal dynamics of human EEG responses to visual stimuli.