Abstracts

Rationale: Accurately and rapidly identifying seizures in 24/7 continuous EEG data is a major challenge. A rapid method of Seizure Detection using Audio Transformation (SDAT) based on listening to EEG data at high speed and visually marking seizure onsets and durations is described. Methods: SDAT is validated by comparison with visual-only (VO) seizure scoring and measuring SDAT inter-scorer agreement. Three-channel EEG recordings (24-72 hr per data file) from two rat models of epilepsy (pilocarpine and tetanus toxin) were acquired at a sampling rate of 400Hz and stored as EDF files. Notch filters were applied to each EEG channel to minimize recording noise (e.g., 60 Hz) and the EDF files were converted into WAV files. The sampling rate in the WAV file header was adjusted to alter the playback speed. This made the EEG frequencies more audible to the human ear. For example, a 400 Hz EDF file was converted to a 32 kHz WAV file with 80s of EEG played in 1s and 24h of data played in 18min. WAV files were reviewed using audio editor software. Reviewers used aural and visual cues to mark the onset and duration of events of interest. These events included seizures, interictal epileptiform discharges, movement artifacts, and other artifacts. A team of nine reviewers scored a group of EEG files using SDAT. A subset of these files was scored using VO for comparison. Events scored by each reviewer were stored in a database and the results were analyzed to measure SDAT inter-scorer agreement and compare SDAT with VO. Results: Scoring a 24h file using SDAT took 25-40 min, compared to 3-5 hr using VO. Over 25,000 hours of EEG data from 275 rats were analyzed using SDAT resulting in 35,566 scored seizures. A preliminary inter-scorer comparison between two expert SDAT scorers based on 247 hr of EEG showed a total of 174 seizures identified with only one disagreement. Additional inter-scorer comparisons will be made as more EEG files are scored by multiple reviewers. A preliminary comparison between SDAT and VO scoring showed dramatic differences. In 121 hr of EEG, more seizures were scored by VO (199) than SDAT (19). Inspection of these differences suggests that many of the VO events were artifacts and not seizures. In 217 hr of EEG, more events were identified by SDAT (136) than VO (68). Inspection of these differences suggests that many seizures were missed by VO because of their short duration and low amplitude. In 97 hr of EEG with high amplitude seizures and few artifacts, agreement between SDAT (17) and VO (16) is excellent. Conclusions: SDAT was developed to improve the speed and accuracy of EEG seizure scoring. Our preliminary evaluation of SDAT suggests that it is faster than VO and may allow greater accuracy in files with considerable artifacts, such as rat EEG. (Supported by NINDS 5U01NS064049)