@InProceedings{boubchir:embc:2015,
  author    = {Boubchir, Larbi and Touati, Youcef and Daachi, Boubaker and Ch{\'e}rif, Ali, Arab},
  title     = {EEG error potentials detection and classi- fication using time-frequency features for robot reinforcement learning},
  booktitle = {International Conference of the IEEE Engineering in Medicine and Biology Society},
  year      = {2015},
  address   = {Milan, Italy},
  month     = {August 25-August 29},
  url       = {https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=\&arnumber=7318719},
  keywords  = {Electroencephalography, Feature extraction, Robots, Time-frequency analysis, Accuracy, Training},
  doi       = {10.1109/EMBC.2015.7318719},
  abstract  = {In thought-based steering of robots, error potentials (ErrP) can appear when the action resulting from the brain-machine interface (BMI) classifier/controller does not correspond to the user\textquotesingle s thought. Using the Steady State Visual Evoked Potentials (SSVEP) techniques, ErrP, which appear when a classification error occurs, are not easily recognizable by only examining the temporal or frequency characteristics of EEG signals. A supplementary classification process is therefore needed to identify them in order to stop the course of the action and back up to a recovery state. This paper presents a set of time-frequency (t-f) features for the detection and classification of EEG ErrP in extra-brain activities due to misclassification observed by a user exploiting non-invasive BMI and robot control in the task space. The proposed features are able to characterize and detect ErrP activities in the t-f domain. These features are derived from the information embedded in the t-f representation of EEG signals, and include the Instantaneous Frequency (IF), t-f information complexity, SVD information, energy concentration and sub-bands\textquotesingle  energies. The experiment results on real EEG data show that the use of the proposed t-f features for detecting and classifying EEG ErrP achieved an overall classification accuracy up to 97\% for 50 EEG segments using 2-class SVM classifier.}
}