Abstracts

Rationale: Recently a feature baseline correction technique (FBC) has been introduced for neonatal seizure detection in the EEG using a support vector machine (SVM). It has been shown that a SVM classifier developed for neonatal seizure detection can be used without alteration for the detection of seizures in adults [Faul, S., A. Temko and W. Marnane Age-independent seizure detection. Conf. Proc. IEEE Eng. Med. Biol. Soc. 5:5332553, 2009] . We examined whether FBC could also improve epileptic seizure detection in adult patients using a SVM classifier trained on neonatal EEG. Furthermore a classifier trained on adult EEG is evaluated on neonatal patients. Methods: The dataset consists of 57 neonatal and 21 adult standard EEG registrations (± 20 minutes). Using patient leave-one-out cross validation the performance of both neonatal and adult classifiers are tested. Performance is evaluated in terms of area under the ROC curve, percentage of seizures detected and amount of false detections per hour. Results: FBC improves seizure detection performance on both adult and neonatal detection for both classifiers. The neonatal classifier outperforms the adult classifier on both patient groups. With FBC mean ROC curve areas are increased from 0.87 to 0.95 and 0.81 to 0.91 for the adult and neonatal patients respectively. Approximately 90% of the seizures in the adult patients and 80% in the neonatal patients are detected at a cost of ~1.5 false detections per hour. Conclusions: Adult seizure detection performance significantly improves using FBC. Training with neonatal data results in higher performance compared to training with adult data. This might be due to the wider range of seizure morphologies in neonates. The lower amount of detected seizures also indicates that seizure detection is more difficult in neonates than in adults. We conclude that FBC is a useful alteration of the neonatal SVM classifier for detection of seizures in both adult and neonatal EEG.