ISSC 2005, Dublin. Sept 1–2 A Novel Automatic Neonatal Seizure Detection System Stephen Faul †∓ , Geraldine Boylan ‡ , Sean Connolly ∗ , William Marnane † , and Gordon Lightbody † † Department of Electrical ‡ Department of Paediatrics ∗ Department of Clinical and Electronic Engineering, and Child Health, Neurophysiology, University College Cork, University Hospital, St. Vincents Hospital, Cork, Cork, Dublin, Ireland Ireland Ireland E-mail: ∓ stephenf@rennes.ucc.ie Abstract — A novel neonatal seizure detection system is proposed including work in the areas of Independent Component Analysis, feature extraction, probability and classification networks. The sys- tem comprises of a preprocessing stage to reduce the effect of EEG artifacts and incorporate multichannel analysis and data reduction. A feature extraction stage examines the EEG using techniques from various signal processing approaches. A Probability Estimator com- pares current and past features to emphasise changes in the state of the EEG. Finally, a classification stage uses the results from the probability estimator to make a decision as to whether the EEG is non-seizure or seizure. Results show promising performance, detect- ing 45 of 46 seizures in the test data with low false detection rates. Keywords: neonatal seizure detection, feature extraction, classifi- cation, EEG analysis, EEG modelling I Introduction A PPROXIMATELY 1 in every 200 newborns experience seizures [1], and in premature and low birth weight infants, this figure rises to up- ward of 1 in 4 [2]. Controversy exists as to whether seizures damage the brain. Although the healthy immature brain does not appear to incur injury from prolonged seizures, in an immature brain that has suffered some injury, seizures can cause brain damage or death [2]. Quick detection of seizures presents the best window of opportunity for treat- ment. Clinical signs of seizure in newborns can be very subtle or non-existent [3] and constant supervision of the EEG signal by a trained EEG specialist is needed to detect neonatal seizures; an unrealistic prospect. An automatic seizure detection system would overcome the problem of constant supervi- sion and make seizure detection and analysis faster and more reliable. Of course, such a system would have to have a high detection rate, but also it is imperative that a low false detection rate is main- tained for the system to be successful. Previous work on neonatal seizure detection has focused on frequency analysis [4, 5]. Work has also been carried out in modelling and complexity analysis [6]. Analysis of this work has shown that the variety of the characteristics related to seizure events makes it difficult for these approaches to perform reliable neonatal seizure detection [7, 8]. Furthermore, the effect of EEG artifacts on the false detection rate and the use of simple threshold- ing for classification have been shown to greatly in- hibit the achievable performance of these systems. The neonatal seizure detection system proposed in this paper aims to overcome these problems. This paper is organised as follows. Section II details the data used for the development of this system. Section III will briefly describe the pro- posed neonatal seizure detection system. Section IV introduces the preprocessing section. Section V explains the features extracted from the EEG. Sec- tion VI introduces the probability estimator and Section VII explains the classifying routines. Fi- nally, results will be given in Section VIII and con- clusions will be drawn up in Section IX. II Data Acquisition All EEG data was collected from newborn babies with seizures in the neonatal intensive care units of Kings College Hospital in London, UK and Cork University Maternity Hospital, Ireland. Written consent was obtained from the parents of each