AUTOMATIC EEG SPIKE DETECTION - WHAT SHOULD THE COMPUTER IMITATE

We conducted a study to explore how electroencephalographers (EEGers) read EEGs and reach clinical impressions based upon them. Eight EEGers and a rule-based computerized ''spike'' detector marked epileptiform discharges (''EDs'') in 12 test records. Of all marked events, 18% were marked by all readers and 38% were marked by only one reader. Readers agreed on basic clinical features of the records, such as whether a record demonstrated EDs, the rank order of ED sources by location, and the ranking of test records in order of the number of EDs detected. Readers marked records in a consistent pattern that was independent of an objective measure of expertise and experience. Our computerized ED detector had lower sensitivity and selectivity than human readers, but either parameter could be adjusted to be comparable to human EEGers at the expense of the other. We propose that EEGers employ reproducible, quantitatively different styles of reading EEG tracings to reach qualitatively similar clinical impressions. In practice, EDs are not absolutely defined, but appear to represent a continuum of activity which lends itself better to description and rank ordering than to absolute quantitation. More than just counting EDs, a successful computerized ED detector should be adaptable to the style of individual readers in order to help them efficiently formulate their clinical impressions.