Optimal time-frequency bases for EEG signal classification in BCI context
Résumé
We consider the problem of classifying multi-sensor signals, more precisely EEG signals in the context of Brain Computer Interfaces (BCI), by selection of time-frequency features. The features are determined among local cosine bases (MDCT) by a "best basis" type algorithm adapted to the classification context. In the BCI domain, or more generally in neuroscience, many classification algorithms are based upon automatic approaches (artificial neural networks, SVM, ...) which do not allow a simple interpretation of the features. The proposed approach allows such interpretation, since the features are determined in the form of time-frequency atoms, similarly to classic analyses of EEG signals which involve specific frequency bands and time intervals. The proposed algorithm generalizes the best discriminant basis algorithm by Saito, employing pairwise comparisons between the signals belonging to two classes of data. Results on artificial data show that the method is able to determine simulated differences between signals. Results on real data are competitive with state of the art classification algorithms and more easily interpretable.
Origine | Fichiers produits par l'(les) auteur(s) |
---|
Loading...