Inter-Ictal Seizure Onset Zone Localization Using Unsupervised Clustering and Bayesian Filtering

Surgical removal of seizure-generating brain tissue can cure epilepsy in patients who do not respond to medications. However, identifying seizure-generating regions is difficult and fails in many cases. In this paper, we report a fully unsupervised and automated approach to seizure focus localization using a Bayesian filter. This method uses a spectral domain feature, Power in Bands (PIB). PIB is extracted from inter-ictal (non-seizure) intracranial EEG recordings of patients with focal epilepsy to differentiate normal and abnormal brain regions. This study was carried out using data collected from 34 patients with focal epilepsy at the Mayo Clinic. Experiments show that using a Bayesian filter for capturing temporal properties of the iEEGs recorded from epileptic brains remarkably improves localization accuracy (AUC: 0.63 -> 0.72). Our study also reaffirms that high-frequency oscillations and inter-ictal spikes are useful inter-ictal biomarkers of the epileptic brain, and PIB, which could be implemented with relatively low computational burden, performs as well as the standard bio-markers when used in this setting. We conclude that the technique of extracting spectral features from inter-ictal iEEGs and capturing their temporal properties via a Bayesian filter markedly improves our ability to localize seizure onset zones.