Fig 1.
An overview of MI-BCI classification using machine learning vs. deep learning approaches.
In ML approach, EEG signals are first pre-processed and relevant features are extracted before applying a classifier. In DL approach, raw signals are directly fed into the model.
Fig 2.
The time course of each trial in the BCI task.
(a) shows the calibration run and (b) the feedback runs. In all trials, participants saw a fixation cross and thereafter an arrow pointing to either left or right, which indicated the corresponding hand for the MI task in the trial. In feedback runs, the blue bar indicated the direction and certainty of the classifier’s prediction in order to feedback to the participants. The grey area indicates the time course of the MI task.
Fig 3.
CNN architecture.
Table 1.
Comparison between training and test accuracies of CNN and CSP+LDA models.
Fig 4.
Mean difference between accuracies of CNN and CSP+LDA models (AccuCNN−AccuCSP+LDA) for Low Performer and High Performer groups.
Low Performers showed significantly higher improvement in MI-BCI accuracy after using a CNN classifier.
Fig 5.
Improvement in the accuracy rate of the subjects using CNN model against CSP+LDA in percent points (i.e., absolute difference between the two accuracies; AccuCNN–AccuCSP+LDA).
Table 2.
Average F-score obtained by the CNN and CSP+LDA models for each MI class.