Abstracts

Rationale: The use of automated seizure detection is becoming increasingly widespread, but data proving its reliability is limited. This study is a retrospective analysis evaluating the reliability of Persyst 13, the only approved quantitative EEG software available on the market. Methods: 85 continuous critical care video EEGs (cEEGs) were analyzed. 23 cEEGs which documented discrete seizures in their initial report by an epileptologist were randomly selected, annotated by a primary EEG reader, and then independently reviewed by an epilepsy attending. Using a power analysis from a large study demonstrating that 27% of inpatient cEEG records at a tertiary care center contained seizures (Westover MB, Shafi MM, Bianchi MT, et al. The probability of seizures during EEG monitoring in critically ill adults Clin Neurophysiol. 2014;126(3):463–471.), 62 additional cEEGs without seizures were also similarly selected and reviewed. Individually marked seizures were reviewed in a subgroup analysis investigating the effects of EEG background and the level of artifact. All 85 cEEGs were then analyzed by Persyst. Statistical analysis was performed R Project for Statistical Computing. Results: Ultimately, 229 discrete seizures were marked by two independent EEG readers. Results were notable for Persyst having an overall seizure sensitivity of 48% and positive predictive value of 62%. Subgroup analysis was remarkable for a significant difference in seizure sensitivity in low voltage cEEGs, with a p value of 0.000625. When evaluating if Persyst could determine if an EEG record contained any seizures, Persyst had a sensitivity of 78% and had an overall negative predictive value of 88% on a cEEG level. If low voltage cEEGs were eliminated, Persyst had a 100% sensitivity and 100% negative predictive value on a cEEG level. Of 62 cEEGs without seizures, Persyst correctly identified that 35 cEEGs were seizure free for an overall specificity of 56%. Conclusions: Overall, Persyst had poor sensitivity and positive predictive value on an individual seizure level, but was fairly sensitive for detecting if there were any seizures present during a cEEG record. Persyst’s utility lies in the result that if it did not detect any seizures in cEEG, it was very unlikely that the record had seizures as it only missed seizures entirely in cEEGs with low voltage backgrounds. This study established a third party baseline assessment of the only FDA cleared seizure detection system in order to compare with future quantitative EEG analysis software. It is, to our knowledge, also the most likely to represent inpatient EEG records as it features the largest amount of carefully reviewed critical care inpatient EEG data to evaluate automated seizure detection software. Further corroborative studies may allow EEG readers to use Persyst to streamline the amount of cEEG that needs review. This study identifies Persyst as a useful tool in triaging and EEG data reduction in specific conditions if low voltage EEGs are not considered, otherwise our study demonstrates that Persyst fails to meet a clinically relevant standard for seizure detection. Funding: No funding