Machine detection of spike-wave activity in the EEG and its accuracy compared with visual interpretation

Machine detection of epileptiform activity in the EEG is useful in seizure monitoring because of its inherent consistency and the rapid data reduction it can provide. Devices based on a few detection criteria have lacked reliability of detection and those with more complex algorithms have sacrificed operating speed and portability. This paper describes a largely analog device which detects irregular as well as classic spike and wave activity. It is portable and it can process the accelerated playback of 24 h tape recorders as well as real-time EEG. It recognizes spikes by their shape and waves by their frequency. It makes inter-channel comparisons to identify trains of bilateral synchronous spikes, generalized waves, and coincidence of spikes and waves and furnishes a limited description of each event in terms of these characteristics. The device was tested against the judgment of 3 experienced and certified electroencephalographers in 18 h of EEG containing 769 bursts of spike-wave activity from 6 patients. It detected 96.5% of the consensus spike and wave activity (i.e., activity identified by all 3 electroencephalographers). Only 0.56% of the machine's detections were false positives (i.e., activity identified by none of the electroencephalographers), though the false positive rate was higher in the presence of chewing artifact. It measured burst duration with an average error of 0.43 sec/burst. While reader-machine agreement varied somewhat by patient, in general, the machine disagreed with the consensus no more than the readers disagreed with each other. In a second reading session after 6 months, the amount of activity identified by the readers changed by an amount ranging from 2.4% to 57% while the machine was consistent within a few tenths of 1%. Hence, this paper demonstrates that by implementing a multi-criteria detection algorithm in special purpose circuitry, a cost-effective solution to the problem of reliable machine detection of spike and wave activity can be obtained.