Estimation of Imagined Rhythms from EEG by Spatiotemporal Convolutional Neural Networks

The problem of estimating imagined music from electroencephalogram (EEG) is very challenging. In this paper, we focused on beats (pulse trains of single notes), one of the components of music, and attempted to estimate imagined beats from an EEG. First, we presented two types of beat patterns and asked 17 experimental participants to imagine them. Next, the imagined beat pulses were estimated from the EEG during the task based on spatiotemporal convolutional neural network models. We employed a CNN and an EEGNet to evaluate the model's performance with binary cross entropy and focal loss as AUC and F1-measure. Although AUCs between the CNN model and EEGNet are competitive, the number of parameters of the EEGNet is much smaller than that of the CNN. Moreover, we have observed the effect of the loss functions in the F1-measure. Overall, the EEGNet model with the focal loss efficiently performed in imagined beat identification.