A Multi-Model Unbiased Algorithm for Reliable Detection of Seizures

Epilepsy is one of the most common neurological disorders and is characterized by recurrent spontaneous seizures, often monitored using electroencephalography (EEG) or intracranial electrocorticography (iEEG).We propose an unbiased and reliable system for detection of seizures, designed to replace manual inspection of iEEG recordings and to be potentially coupled with automatically triggered treatments. Cortical activity was continuously recorded in three different mouse models of epilepsy (genetic, status epilepticus-induced and albumin-induced), through two epidural electrodes connected to an implanted telemetry
transmitter. The acquired signals were filtered, segmented
and underwent extraction of features, to be classified by an
S58 J Mol Neurosci (2011) 45 (Suppl 1):S1–S137 artificial neural network (ANN). For classifier training, a dataset of seizure and non-seizure recordings was comprised and represented by 22 extracted features. Forward selection analysis led to the identification of a 5 feature subset, allowing optimal tradeoff between robust ANN classification and reduced computational time. Classification output was post-processed using sliding-window thresholding, applying a persistence rule for positive detection. A graphical user interface was created for simple execution of data analysis and seizure detection. System performance was assessed by analyzing over 2,800 hours of raw iEEG recordings from 15 animals (12 epileptic and 3 controls). Performance evaluation revealed overall sensitivity and positive predictive value above 98% in unedited signals containing noise, artifacts and interictal discharges. The system also successfully detected seizures in an IEEG recording of an epilepsy patient, suggesting the human applicability of the proposed approach.