Deep convolutional recurrent neural network for short-interval EEG motor imagery classification

Abstract

In this paper, a high-performance short-interval motor imagery classifier is presented that has good potential for use in real-time EEG-based brain-computer interfaces (BCIs). A hybrid deep Convolutional Recurrent Neural Network with Temporal Attention (CRNN-TA) is described that achieves state-of-art performance in four-class classification (73% accuracy, 60% kappa, 3% higher than the winner of the BCI IV 2A competition). An adaptation of the guided grad-CAM method is proposed for decision visualization. A novel EEG data augmentation technique, shuffled-crossover, is introduced that leads to a 3% increase in classification accuracy (relative to a comparable baseline). Classification accuracies for different windows sizes and time intervals are evaluated. An attention mechanism is also proposed that could serve as a feedback loop during data capture for the rejection of bad trials (e.g., those in which participants were inattentive).

Tasks

Reproductions