Abstracts

Rationale: Majority of neonatal seizures occurring in the neonatal ICU (NICU) are subclinical, being detected only by EEG monitoring. Expertise for recording and interpretation of EEG is not available around the clock in a NICU, resulting in a need for a reliable, automated EEG-based seizure detection method. Aim: To validate an improved automated neonatal seizure detection algorithm, in a large and independent data set of neonatal seizures recorded during continuous EEG monitoring and to evaluate its performance in relation to EEG background. Methods: We classified EEG background into eight grades based on evolution of discontinuity over 24 hours and presence of sleep-wake cycles. Patients were further sub-classified into two; gpI: those with mild to moderate (grades 1-5) and gpII: severe (grades 6-8) abnormalities of EEG background. Seizures were categorized as definite and dubious. Seizure characteristics were compared between the two EEG groups. The neonatal seizure detection algorithm (Deburchgraeve W. et al., Clin Neurophysiol 2008;119:2447-54)was developed jointly by the department of Electrical Engineering at the Katholieke Universiteit, Leuven and the department of Clinical Neurophysiology at the Erasmus MC, university medical center in Rotterdam. It has been further improved in artifact rejection. The algorithm was tested offline on a large new ( unseen ) EEG data set of 756 hours of monitoring from 24 consecutive neonates (median 25h per patient) with encephalopathy and recorded seizures. No selection was made regarding quality of EEG or presence of artifacts. Results: Seizure amplitude showed a significant negative correlation with worsening EEG background grade (Spearman s rho= -.475, p=0.019). In patients in EEG gpII, the total number of seizures expressed was significantly increased while the amplitude of the seizures was significantly decreased (Table 1 & Fig 1). There was a tendency for patients in this group to express more seizures with higher firing frequency, increased percentage of arrhythmic seizures and shorter duration of seizures (Fig 1), though the results were not statistically significant. After excluding 4 patients with persistent, severely abnormal EEG background, and predominantly (>90%) dubious seizures characterized by low amplitude arrhythmic discharges, the algorithm showed a median sensitivity per patient of 86.9% (1263/1538 seizures detected, total sensitivity 82.1%), positive predictive value (PPV) of 89.5% and false positive rate of 0.28/h. Sensitivity tended to be better for patients in gpI. Conclusions: The algorithm detects neonatal seizures well, has a good PPV, and is suited for long-term EEG monitoring. The changes in electrographic characteristics like amplitude, duration and rhythmicity in relation to worsening EEG background tend to negatively affect the performance of automated seizure detection.