Rule-enhanced Clustering in the Detection of Neonatal Seizures

This paper describes a three-stage system for the detection of neonatal seizures. The first stage detects 5-s seizure segments using signal processing and pattern recognition techniques. In the second stage, the seizure segments overlapping with artifactual segments are marked for post-processing using rules. Rules add intelligence to the spatio-temporal clustering in the third stage, by incorporating knowledge of known seizure characteristics, spatial context, and occurrence of artifacts, in order to reduce false detections. The false detection rate has been reduced without significantly lowering the sensitivity of the seizure detection process. In 21 subjects (11 with seizures and 10 without seizures), the false detection rate was less than 1/hr while sensitivity of seizure detection was 85%. The focus of this paper is the second and third stages of the system.