End-to-end electroencephalogram (EEG) motor imagery classification with Long Short-Term

Abstract

In this paper we generate an end-to-end model for electroencephalogram (EEG) motor imagery brain waves classification. EEG waves are considered a time series, however most of the literature focus on changing the representation of the waves or working on the data set as a whole. One of the goals of the investigation is reaching a conceptually simplified model so it can be generalized for the new approaches at EEG data acquisition (such as the novel EEG buds). First we mention some of the experiments and approaches that didn't obtain good metrics and then we show the results for MLP and LSTM neural networks. LSTM networks are slower to reach a higher accuracy compared to the MLP networks with less seconds of training, however they are better at reaching stable levels of accuracy when given enough data. Normalization plays an important role on the process, showing that the best and most consistent results are obtained when it is done locally at a sequence level, from where we can infer that the patterns are arguably most affected by the values (originally measured in micro volts and then normalized) in the local context of the sequence.