Wavelet preprocessing for automated neural network detection of EEG spikes

The purpose of this study was to investigate the feasibility of using a wavelet transforms (WT) as a preprocessor for an artificial neural network (ANN) based EEG spike detection system. The study aimed at decreasing the input size to the ANN detector, without decreasing the information content of the signal and degrading the detection performance. Since routine clinical EEG requires recordings from many channels (generally 32 or 64), input size becomes a critical design parameter for real-time multichannel spike detection systems. For a sliding window of 20 points, more than 600 input lines will be necessary for a 32-channel system, which is not easily manageable with current ANN technology. One approach to this problem is to use a single ANN module for each EEG channel and integrate the outputs across channel information with a second module. The authors have successfully developed a 16 channel prototype of such a system working in real-time. This system used a 20 point (100 ms) sliding time window and employed a floating point digital signal processor for real-time operation. Adding more channels to this system would be difficult for real-time operation. In addition, the authors´ recent studies showed that the best detection performance is attained with a 30 point (150 ms) time window, further increasing the computational load. Thus, the development of a preprocessor to reduce the input size without significantly reducing the information content would be very helpful in developing large multichannel EEG spike detection systems