MLSP Data Analysis Competition 2006: Denoising of Magnetoencephelographic Data
Stephen James Redmond, Conor Heneghan · 2006
The study of human cognition, and preoperative functional brain mapping, are facilitated through the use of magnetoencephalography (MEG). However, the noise present in such recordings is significant relative to the signals of interest. To circumvent this issue multiple trials are performed for the same task and an ensemble average is performed to increase the signal-to-interference/noise ratio (SNIR). Unfortunately, perhaps 100-500 trials are required to achieve a sufficiently large SNIR. This paper describes a simple technique for improving the SNIR value given only 10 trials. The 10 trials from each of the 274 channels are first averaged. The 274 averaged channel estimates are then temporally low passed filtered, adjusting for the phase shift introduced by the filtering process. Finally, a spatial average is performed, where each channel is assigned the average value of itself and the three most similar channels. The three most similar channels are identified using correlation. The described method achieves an SNIR of 0.6 dB.