Classifying MEG 20Hz rhythmic signals of left, right index finger movement and resting state using cascaded radial basis function networks

A cascaded Radial-Basis Functions (RBF) network was devised to classify the magnetoencephalography (MEG) rhythmic signals of the left, right index finger movement and resting state. Four right-handed subjects were instructed to perform self-paced index finger lifting in a rate of every 8 sec. MEG epochs from 4000 ms pre-movement to 3000 ms post-movement were digitized. Each trial was decomposed by Principal Component Analysis (PCA) and the task-related components were selected to reconstruct data followed by band-passed filtering around 16-20 Hz. Every five filtered epochs were averaged and processed by the Hilbert transform to produce a beta-band envelope. All the envelopes around sensorimotor channels were normalized and down-sampled to 100 samples. A movement feature vector was defined by concatenating two mean envelopes in the left and right channels. We defined a resting feature vector by averaging five envelopes randomly selected in the left and right channels respectively during the resting state, and then concatenating two means. Two cascaded 3-layer RBF networks were constructed in which the first RBF network was to discriminate the movement from the resting state, while the second one was to distinguish the left and right index finger movement. The classification rates (rest, left, right) for four subjects achieve (100, 91, 66)%, (100, 80, 72)%, (100, 76, 73)% and, (100, 74, 94)%, respectively.