Enhancing the detection of seizures with a clustering algorithm

Automated detection algorithms of EEG seizures or similar clinical events typically analyze a finite epoch a given channel at a time, producing a probability or a weight estimating how likely it is for the event to resemble a clinical pattern. Epochs are normally shorter than the duration of a seizure, which may spread to more than one electrode. This may result in a weak correspondence between the seizure pattern in the record and its calculated detector event counterpart. As a result, such algorithms suffer from a high rate of false detections. We show that the weights/probabilities of a generic detector can be described as a weight function embedded in a directed graph (digraph). Extended objects such as seizures therefore correspond to the connected components of the digraph. We introduce a clustering algorithm that accounts for the shortcomings of a generic detector of the type described above. By correlating detector results with respect to both time and channel, we effectively extend the detection to an unlimited number of electrodes over an indefinite time. The algorithm is fast (linear - O(m)) and may be implemented in real time. We argue that the algorithm enhances the detection of seizure onset and lowers the rate of false detections. Preliminary results demonstrate a strong correlation between the seizure and the cluster's boundaries and over 50% reduction of false detection rate.