Answers are grounded in this paper and cite their sources.
Objective The value of electroencephalography (EEG) relies on accurate and reliable interpretation. Prior research has demonstrated imperfect interrater reliability, but intrarater reliability remains poorly characterized. We performed a “noise audit” primarily to measure intrarater variability in EEG interpretation by epileptologists. Methods Experts independently rated the same 100 full‐length EEGs on two separate occasions ≥60 days apart, unaware the EEGs were the same. EEGs were recorded from adult (56%) and pediatric patients in outpatient (78%) and non‐critical care inpatient settings. There were 20 studies in each of five categories: normal, focal epileptiform (FE), generalized epileptiform (GE), focal nonepileptiform (FNE), and generalized nonepileptiform (GNE) abnormalities. EEGs were presented using our online platform (EEGHub) without clinical information except for age and sex. Experts rated each recording and indicated their level of confidence (LOC) for each EEG and overall. We analyzed percent as well as chance‐corrected intra‐ and interrater agreement. Results Fourteen experts (mean of 19 years EEG reading experience, range = 4–52 years) were included in the study. Percent self‐agreement was 96% for GE, 90% for FE, 88% for FNE, and 87% for GNE abnormalities; 90% for any epileptiform abnormality; and 82% for any nonepileptiform abnormality. Self‐agreement was strongly correlated with interrater agreement (mean κ ML ), which was almost perfect for GE (.90, 95% confidence interval [CI] = .87–.94) and FE (.82, 95% CI = .77–.87) and substantial for GNE (.75, 95% CI = .69–.81) and FNE (.74, 95% CI = .68–.80) abnormalities. LOC correlated with self‐agreement on the presence of nonepileptiform but not epileptiform abnormalities. Significance We found generally strong within‐expert agreement in EEG interpretation, especially in the identification of epileptiform abnormalities. Nevertheless, the number of flipped decisions was high enough to be clinically consequential. This undesired variability (“occasion noise”) may be addressed by interventions to (1) decrease this noise directly and (2) reduce case ambiguity with longer recordings and better EEG data whenever feasible.
No takes yet. Share an insight, caveat, or question.
Yuan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: