Abstract
To study the temporal structure of EEG microstates, we trained recurrent neural networks (RNNs) consisting of long-short-term-memories (LSTMs) with microstate sequences of different lengths to 1) reconstruct the input microstate sequence and 2) predict the future trajectory of microstates. We tested the reconstruction and prediction accuracies on nonoverlapping subsets of resting state data preceding and following social stress within and between subjects and investigated the activation patterns of the neurons in the hidden layer. The results show that the microstates’ trajectory 1) can be learned successfully across sessions by RNNs, 2) is largely subject-invariant at shorter time scales, 3) is affected by stress. These findings suggest that the sequence of microstates is governed by different processes at different time scales.
