Abstracts

Rationale: We assessed inter-rater reliability (IRR) of EEG interpretation among 6 board-certified clinical neurophysiologists (IRR reported separately). Exploratory analysis of those data revealed that the level of agreement among raters varied significantly with the electrographic category being assessed. We chose to pursue this finding by using latent class analysis (LCA) to identify 'classes' of EEG whose characteristics, whatever they may be, allow raters to score or interpret those EEGs similarly.Methods: 200 EEGs from patients ? 1 year old were scored by 6 board certified EEGers (A-F). The 6 raters were divided into 20 groups of 3 (ABC, ABD DEF). Each group interpreted 10 EEGs and each EEG was interpreted by 3 raters. Raters were aware of patient age and medications, but were blinded to clinical history, reason for EEG request, and each other s scores. Raters were asked to assign probabilities to one or more of 7 EEG diagnostic categories (Table 1, legend), with the stipulation that one category had to have the highest probability. This category was deemed to be the rater's best assessment of the true category, and was used as input for the analysis. LCA was used to identify classes of EEGs, such that members of each class exhibit similar rating patterns.Results: In the context of this study, LCA is based on the speculation that there may be different classes or groupings of EEG readings that have specific characteristics. If this is true, the first question is: How many classes are there? For the data obtained in this study, the right answer seems to be 3 (Table 1). Class 1 readings (estimated to comprise 30% of the sampled population) are usually classified by all readers as Normal. However, Reader A, and to a lesser extent Reader B, are less likely to do this than the others, often interpreting these studies as Slow. At the other extreme, Reader E nearly always interprets Class 1 studies as Normal. Class 2 readings (44% of sampled population) are generally agreed to be Slow, though Readers D & E often say Normal, and Reader B often sees epileptiform discharges as well. Class C readings (26% of sampled population) are typically agreed to contain epileptiform discharges and independent slowing though there is considerable disagreement about the epileptiform component.Conclusions: Based on these data obtained from 6 readers and 200 EEGs, the LCA resulted in 3 EEG classes characterized by considerable agreement among raters. This finding is consistent with the concept of superior EEG features formulated by Abend NS et al, which relates inter-rater agreement to the characteristics of the EEG category being assessed. The EEG diagnostic categories for which there was relatively little agreement among the raters should be a priority for continuing education of experienced EEGers.