Window-Delay Analysis on EEGNet

Brain-Computer Interfaces (BCIs) have grown significantly in recent years. Several BCI competitions have been proposed for further research in this area. However, they evaluated the results based on kappa and accuracy. This has created a tendency in the literature to evaluate BCIs only on their performance. However, BCI is a multi-objective problem and performance is not enough to evaluate the models. In this work, we propose a new metric, Window-Delay score (WD-score), to evaluate the time that the model needs to reach its maximum performance. In several applications, it is necessary that the model can reach a good performance as quickly as possible. EEGNet was used to evaluate the WD-score with different parameters and window positions. The obtained results show that the WD-score is important in analyzing the relationship between the delay and the accuracy of the model. For the case of 2 classes, the analysis reduces the delay from 2 seconds to 0.5 seconds maintaining an accuracy above 80%. Moreover, this work points out that accuracy varies over time differently according to the window chosen for training.