Abstracts

Rationale: Infantile spasms (ISS) is a devastating epileptic syndrome that occurs in infants beyond the neonatal period characterized hypsarrhythmia (HYPS) on EEG. Variations on the classic hysparrhythmia pattern have been described including “modified” and “atypical” forms. A recent study by Hussain et al. found that even skilled electrophysiologists may interpret the EEG of children with infantile spasms differently.[1] Since EEG is a key factor in the diagnosis of ISS, misinterpretation could result in serious consequences including inappropriate treatment. Quick and effective treatment of infantile spasms is thought to alter the neurodevelopmental outcome for some children. Modified forms of HYPS continue to pose challenges, and their relationship to severity of seizures, response to treatment, and long-term outcome need to be studied. EEG spectral analysis, allows for objective measures of amplitude, frequency, and spike density to be captured to create a spectral definition of HYPS. We developed an algorithm to improve upon established methods of auto-detecting epileptiform discharges in pre-treatment EEG’s of infants with infantile spasms. 1. Hussain, S.A., et al., Hypsarrhythmia assessment exhibits poor interrater reliability: A threat to clinical trial validity. Epilepsia, 2015. 56(1): p. 77-81.Methods: Five EEG’s of infants 4 -9 months of age were used to develop this algorithm. Five-minute samples of awake EEG were used for each subject. Two electrodes were used to apply the algorithm, P4 and O2. The approach to identify the epileptiform discharges consisted of three stages: 1. Constructing the time frequency time-frequency domain (TFD) using matching pursuit TFD (MP-TFD), 2. Decomposing the TFD matrix into two submatrices using non-negative matrix factorization (NMF), 3. Using the decomposed spectral and temporal vectors to locate the epileptiform discharges. The spikes detected using this algorithm were compared to spikes manually counted by an epileptologist using conventional EEG and also to spikes detected using commercially available software (Persyst).Results: This algorithm successfully identified spikes with an average true positive detection rate of 86% among this cohort (figure 1).The average false negative rate was 14% and false positive was 53% when compared with the manual counting of spikes using conventional EEG.Comparatively, using commercially available Persyst software the average true positive rate was 4% and false negative rate was 96% (table 1).Conclusions: Detection of spikes manually in hypsarrhythmia is difficult and time consuming. The matching pursuit time-frequency domain algorithm successfully identified spikes with a true positive rate of 86%.Future work will focus on refining the methods to decrease the false positive spike detections.It is hypothesized that aspects of the pretreatment EEG in infantile spasms using components such as spike counts will correlate with neurodevelopmental outcomes including intractable epilepsy, developmental delay, and autism.