C W Ko1, H W Chung. 1. Department of Electrical Engineering, National Taiwan University, Room 238, Taipei, Taiwan.
Abstract
OBJECTIVE: Automatic detection of epileptic EEG spikes via an artificial neural network has been reported to be feasible using raw EEG data as input. This study re-investigated its suitability by further exploring the effects of data preparation on classification performance testing. METHODS: Six hundred EEG files (300 spikes and 300 non-spikes) taken from 20 patients were included in this study. Raw EEG data were sent to the neural network using the architecture reported to give best performance (30 input-layer and 6 hidden-layer neurons). RESULTS: Significantly larger weighting of the 10th input-layer neuron was found after training with prepared raw EEG data. The classification process was thus dominated by the peak location. Subsequent analysis showed that online spike detection with an erroneously trained network yielded an area less than 0.5 under the receiver-operating-characteristic curve, and hence performed inferiorly to random assignments. Networks trained and tested using the same unprepared EEG data achieved no better than about 87% true classification rate at equal sensitivity and specificity. CONCLUSIONS: The high true classification rate reported previously is believed to be an artifact arising from erroneous data preparation and off-line validation. Spike detection using raw EEG data as input is unlikely to be feasible under current computer technology.
OBJECTIVE: Automatic detection of epileptic EEG spikes via an artificial neural network has been reported to be feasible using raw EEG data as input. This study re-investigated its suitability by further exploring the effects of data preparation on classification performance testing. METHODS: Six hundred EEG files (300 spikes and 300 non-spikes) taken from 20 patients were included in this study. Raw EEG data were sent to the neural network using the architecture reported to give best performance (30 input-layer and 6 hidden-layer neurons). RESULTS: Significantly larger weighting of the 10th input-layer neuron was found after training with prepared raw EEG data. The classification process was thus dominated by the peak location. Subsequent analysis showed that online spike detection with an erroneously trained network yielded an area less than 0.5 under the receiver-operating-characteristic curve, and hence performed inferiorly to random assignments. Networks trained and tested using the same unprepared EEG data achieved no better than about 87% true classification rate at equal sensitivity and specificity. CONCLUSIONS: The high true classification rate reported previously is believed to be an artifact arising from erroneous data preparation and off-line validation. Spike detection using raw EEG data as input is unlikely to be feasible under current computer technology.