An EEG spike detection algorithm using artificial neural network with multi-channel correlation

An automatic spike detection algorithm for classification of multi-channel EEG signals based on artificial neural network is presented. Radial basis function (RBF) neural network was chosen for single channel recognition, with model optimization using receiver operating characteristics analysis. Waveform simplification was employed for high noise immunity. Feature extraction with as few as three parameters was used as preparation for the inputs to the neural network. Identification of multi-channel geometric correlation was performed to further lower the false-positive rate by using an incidence matrix. Threshold value for spike classification was chosen for simultaneous maximization of detection sensitivity and selectivity. Evaluation with visual analysis in this preliminary study showed a 83% sensitivity using 16-channel continuous EEG records of four patients, while a high false positive rate was found, which was believed to arise from the extensive and exhaustive visual analysis process. The computation time required for spike detection was significantly less than that needed for online display of the signals on the monitor. We believe that the algorithm proposed in this study is robust and that the simple structure of RBF neural network yields high potential for real-time implementation.