A Neural Minimum Input Model to Reconstruct the Electrical Cortical Activity

In recent years, technology has allowed the progressive increase in the number of channels for EEG recording. The scientific rationale is the demand for an increase of the spatial resolution of the recording to better locate the sources of the underlying cortical activity. Despite some papers confirm the improvement of the spatial resolution by using 256 channels we wonder if in fact this density of electrodes on the scalp does not constitute an useless spatial oversampling. Thus we set out to determine whether the amount information derived from a standard 19 channel EEG recording was obtainable with a smaller number of electrodes, in particular with a mounting to 8 channels. Were used and compared the performance of a Perceptron, a Feed-Forward and a Recurrent neural networks, after supervised training by the back-propagation algorithm. The target was to reconstruct the signals of all the 19 channels starting from only 8 input channels. The data-set was built by using multi-subjects 19 channels recordings containing examples of normal, generalized and focal abnormal EEG activity. All the types of network have been able to reconstruct the missing channels with an error lower than 1%. From this pilot study seems to conclude that the information content of this 8- channel EEG is equivalent to that obtainable with a number of channels more than double. Further developments will check the optimal ratio between the number of recorded and reconstructed channels and the applicability of the approach in reallife contexts.