Abstracts

Rationale: Approximately 90% of seizures in critically ill patients are unrecognized by clinical observations and are only diagnosable by continuous EEG (cEEG) monitoring. Electrographic seizures are potentially harmful; particularly the longer diagnosis and treatment are delayed. Although there are published criteria defining electrographic seizures and periodic discharges (PDs), the inter-rater agreement (IRA) among experts applying these criteria is not known, especially for cases with equivocal patterns or complex background activities. Since accurate interpretation of these clinically significant patterns directly impacts the value of EEG as a diagnostic tool, it is important to assess IRA using methods that closely resemble actual clinical practice. The purpose of this study was to investigate IRA among EEG experts for identification of electrographic seizures and PDs in prolonged EEG recordings of critically ill patients. Methods: Eight board-certified clinical neurophysiologists were recruited to identify seizures and PDs in 30 one-hour EEG segments. The study used a web-based EEG review system (EEGnet) that allowed reviewers to classify and mark the beginning and end of each event. Kappa statistic was calculated to assess IRA for each pair of experts. The analysis for kappa statistic adopted a method developed for dependent serial responses. The mean kappa values were compared between seizure and PD detections, as well as among rater groups that have passed the Critical Care EEG Monitoring Research Consortium (CCEMRC) ICU EEG Certification Test versus those who have not. We further investigated how events were determined differently among raters, in terms of occurrence and event durations. A new event-based IRA statistic was developed based on this investigation. Results: Both kappa and event-based IRA statistics showed higher mean values for identification of electrographic seizures compared to PDs (p< 0.001): The mean kappa statistic for identification of seizures was 0.58, suggesting an overall "moderate" agreement among raters. The kappa statistic for documentation of PDs was 0.38, suggesting "fair" agreement. The mean of the event-based IRA statistic was not significantly different from the mean kappa statistic in regards to seizure or PD identification, and the two IRA statistics were highly correlated (correlation coefficients = 0.98 and 0.88). Further, the group of rater pairs that passed the CCEMRC Certification Test had the highest mean IRA statistic, whereas the group of rater pairs without this certification had the lowest mean IRA statistic. Difference between these groups was most significant for PD identification. Conclusions: Inter-rater agreement among experts is significantly higher in identification of electrographic seizures compared to PDs. However, with additional training the IRA for identification of PDs could be greatly enhanced. Improving IRA can not only further the quality of research in the field, but also increase the diagnostic value of cEEG monitoring in intensive care patients.