Interobserver variability in the classification of neonatal seizures based on medical record data

This population-based, retrospective cohort study of neonatal seizures included all 16,428 neonates born to residents of Fayette County, Kentucky, from 1985 to 1989. Eighty potential cases were ascertained by computer search of hospital-based medical record systems, birth certificate data files, and multiple-cause-of-death mortality data files. Medical records for potential cases were abstracted, and relevant portions were reviewed independently by three neurologists using prospectively determined criteria. Both unweighted and weighted κ statistics were used to measure agreement between each pair of observers in the classification of potential cases as seizures, possible seizures, or not seizures, adjusting for the proportion of agreement expected by chance. Agreement in the classification of potential cases was excellent (κ = 0.72–0.79, average = 0.76; weighted κ = 0.85–0.88, average = 0.87). The κ extension statistic of Kraemer was used to assess agreement in the classification of seizure types by a simplification of the classification scheme of Volpe. This documented excellent agreement between raters in the classification of seizure types ( κ e = 0.72). Experienced raters can reliably classify potential cases of neonatal seizures using seizure descriptions transcribed from medical records.