The Effect of Vibrotactile Feedback on ErrP-based Adaptive Classification of Motor Imagery.

This work presents an implementation of Error-related Potential (ErrP) detection to produce progressive adaptation of a motor imagery task classifier. The main contribution is in the evaluation of the effect of vibrotactile feedback on both ErrP and motor imagery detection. Results confirm the potential of self-adaptive techniques to improve motor imagery classification, and support the design of vibratory and in general tactile feedback into Brain-Computer Interfaces to improve both static and adaptive performance.

Mutual information-based feature selection for low-cost BCIs based on motor imagery.

In the present study a feature selection algorithm based on mutual information (MI) was applied to electro-encephalographic (EEG) data acquired during three different motor imagery tasks from two dataset: Dataset I from BCI Competition IV including full scalp recordings from four subjects, and new data recorded from three subjects using the popular low-cost Emotiv EPOC EEG headset. The aim was to evaluate optimal channels and band-power (BP) features for motor imagery tasks discrimination, in order to assess the feasibility of a portable low-cost motor imagery based Brain-Computer Interface (BCI) system. The minimal sub set of features most relevant to task description and less redundant to each other was determined, and the corresponding classification accuracy was assessed offline employing linear support vector machine (SVM) in a 10-fold cross validation scheme. The analysis was performed: (a) on the original full Dataset I from BCI competition IV, (b) on a restricted channels set from Dataset I corresponding to available Emotiv EPOC electrodes locations, and (c) on data recorded with the EPOC system. Results from (a) showed that an offline classification accuracy above 80% can be reached using only 5 features. Limiting the analysis to EPOC channels caused a decrease of classification accuracy, although it still remained above chance level, both for data from (b) and (c). A top accuracy of 70% was achieved using 2 optimal features. These results encourage further research towards the development of portable low cost motor imagery-based BCI systems.