263 A Regularity-Based Seizure Detection Algorithm for Neonates

Abstract

Background: Continuous EEG recordings in neonates are valuable both for monitoring of seizures and to measure response to therapy. Impediments to robust seizure detection include large datasets, limited availability of expert neurologist interpretation, the presence of low amplitude seizures and short seizures. Methods used to date in automatic seizure detection include Harmonie with Sensa software (Stellate Inc., Canada) based on Gotman et al (1997) algorithm.Objective: We evaluate a new approach to automated seizure detection (BrainZ Instruments Ltd., New Zealand) based predominantly on the assessment of relatively prolonged regularity in EEG waves.Design/Methods: The algorithm consists of the decomposition of EEG into adjacent waves and analysis of their regularity. Interval, amplitude and shape information were used in this regularity analysis. Conservative and liberal assessments (similar to integral-overlap and any-overlap, Wilson et al, 2003) of algorithm sensitivity and positive predictive value (PPV) were based on detected seizure duration. The algorithm performance was assessed on 52 multi channel EEG recordings of neonates totalling 20 hours, selected for the purpose of challenging the algorithm. The dataset included 14 EEG recordings with 81 seizures, 22 EEG recordings of normal patients with artefacts and 16 recordings of abnormal EEG.Results: For the BrainZ algorithm the conservative and liberal assessment of performance showed sensitivity of 83 and 95% and PPV of 52 and 75%, respectively. There were 1.8 false positive detections per hour. In comparison, the Sensa software (version 5.4) had sensitivity 37 and 85% and PPV 45 and 48%, respectively; with 11.7 false positives per hour.Conclusions: The regularity-based algorithm performed well and provides a basis for major improvements in neonatal seizure detection.