Detection of movement-related desynchronization of the EEG using neural networks

We aimed to detect the non-averaged MRD of the mu rhythm from the background activity of the EEG. The MRD was produced in the contralateral sensorimotor cortex by means of a movement of the right hand. Thirteen volunteers took part in the experiments. They were all right-handed and had ages between 22 and 37 years old. The volunteers were asked to do one of two tasks: (a) just one movement by trial; (b) repetition of movements. The EEG signal was taken in the scalp near to the C3 lead. The signal was amplified, sampled and digitally filtered off-line. The time response of the mu band was computed with the FFT. Windows of the time response was used as predicates in the training and testing of a neural network classifier. Representative samples of MRD and background activity was taken for each class. The neural networks used were LVQ and they were trained by session, by task and by volunteer. The performance was measured by the geometric mean of the sensibility and specificity indexes, calculated for every training epoch over test data. The best detection performance was 88% for the single movement task and 78% for repetition of movements, with an average of 66% for both tasks. For the LVQ, only 3 subclasses for every class were sufficient, and a constant learning rate of 0.01 with the number of training epochs calculated by the relation of 250×(total number of subclasses) was enough to get the best results