Computer recognition of generalized spike-wave discharges.

Abstract

Twelve 12h daytime telemetered EEGs were read for generalized spike-wave (S-W) bursts by a computer and independently by three experienced electroencephalographers. The computer system was a hybrid of analog devices (tape-recorder, voltage summator, and high-pass filters) and a multipurpose laboratory digital device (PDP-12). The computer was programmed to identify only the location on the record and the duration of the generalized S-W paroxysms. From 1,462 discharges identified by both the readers and the computer, a consensus list of 609 was derived by selecting only those bursts identified by all three readers. The computer recognized 516 (85%) of those bursts on the consensus list; the computer's accuracy, as compared with the consensus, rose to 92% when portions of the EEG containing sleep were discounted. The computer also recognized as generalized S-W paroxysms 15 high-voltage transients, which none of the three readers selected. Ten of these, however, occurred during sleep. When the computer results were added to those of the three readers to permit a four-way inter-reader comparison, the computer was found to have had about as many isolated decisions as any of the three readers. In the number of bursts recognized, the computer tended to be about as conservative as the reader under whose guidance it was originally programmed. The computer's accuracy in measuring burst duration was not evaluated statistically.