Fig 1.
Our framework has three main steps: 1. Filtering, 2. Feature Extraction and 3. Classification. The output of each step is fed to the next step.
Table 1.
Specification of datasets used in this paper.
Table 2.
The accuracy of classifiers for synchronous BCI operation for all subjects.
For each subject the accuracy on the test data is shown. For each classification algorithm the first column shows the results of BP features and the second column shows the results of Morlet features.
Table 3.
The AUC of classifiers for self-paced subjects.
For each classification algorithm the first column shows the results of BP features and the second column shows the results of morlet features.
Table 4.
Average Rankings of the classification algorithms for both synchronous and self-paced datasets.
The number in the parenthesis corresponds to the average rank of the algorithm among different subjects. For each feature extraction method the classifiers typed in bold are the recommended ones. The recommended classifiers are selected based on the results of the statistical tests.
Table 5.
P-values corresponding to pairwise comparison of different classifiers.
α is chosen to be 0.1. For settings 1 and 2 all hypothesis with p-value less than 0.0333 are rejected. For setting 3 and 4 all hypothesis with p-value less than 0.05 are rejected. The results are rounded up to 4 decimal places.
Table 6.
Average Rankings of the classification algorithms for binary and multi-class classification in synchronous datasets.
The number of subjects in binary task was 12 and the number of subjects in multi-task BCIs was 9. The number in the parenthesis corresponds to the average rank of the algorithm among different subjects. For each feature extraction method the classifiers typed in bold are the recommended ones. The recommended classifiers are selected based on the results of the statistical tests.
Table 7.
P-values corresponding to pairwise comparison of different classifiers.
α is chosen to be 0.1. For binary task BCIs with BP features all hypothesis with p-value less than 0.02 are rejected. For multi-task BCIs with BP features all hypothesis with p-value less than 0.0333 are rejected. For binary task BCIs with Morlet features all hypothesis with p-value less than 0.1 are rejected. For multi-task BCIs with Morlet features all hypothesis with p-value less than 0.025 are rejected. The results are rounded up to 4 decimal places.
Table 8.
List of classifier parameters tuned in the training phase.