Abstracts

Interictal short oscillations between 100-500 Hz have been reported in EEG acquired with implanted microelectrodes in epileptic patients. These oscillations last 10 to 50ms and it is very time consuming to analyze and mark them visually because the EEG must be simultaneously displayed with different filters and with very low time scales, such as 0.5s/screen. It can take several hours to mark one channel for a 20 min recording. We report the performance of an automatic high frequency oscillation detection method., EEGs from 4 epileptic patients were evaluated. They were recorded using depth macroelectrodes with 500Hz filtering and 2000Hz sampling. 5-minute EEG segments with 5 channels in each were used for training and similar segments were used for validation. In each segment, an expert selected a 2s section as baseline and marked the locations of high frequency transients, which were classified as follows: 1) oscillations superimposed on spike (OWS), 2) oscillations with no associated spike (ONS), 3) very sharp spikes (VSS) which upon high-pass filtering resemble oscillatory events. The information about locations and numbers of events were not used in the software.
Using the training data, the optimal parameters were computed for use with validation data. Events detected by the software and classified by the expert as ONS or OWS were treated as True Positives. Events detected by the software and classified as VSS were treated as False Positives. Events classified as ONS or OWS by the expert and not detected were treated as Missed Events. Events detected by the software and not classified by the expert were false positives. Although we did not analyze these events, it was apparent that many were oscillations not marked by the expert.
The software computes the spectral power as a function of time and detects changes significantly different from the baseline in different frequency bands. Events that occurred in multiple frequency bands were excluded in order to avoid the detection of oscillations caused by sharp transients., Details are in the Table. The method detected an average of 73% of the marked events in the validation set. Its specificity was low (18%), but a full evaluation of the false detections is required to determine their nature. The automated detection took [sim]1 min per 20 min of EEG per channel., The automatic method has a good sensitivity and can save a considerable amount of time for the analysis of high frequency oscillations over long periods and many channels in intracerebral human EEGs. False detections need to be visually eliminated to ensure the validity of the final results.[table1], (Supported by Can. Inst. of Health Research.)