A Computational Framework to Support the Automated Analysis of Routine Electroencephalographic Data

Epilepsy is a condition in which a patient has multiple unprovoked seizures which are not precipitated by another medical condition. It is a common neurological disorder that afflicts 1% of the population of the US, and is sometimes hard to diagnose if seizures are infrequent. Routine Electroencephalography (rEEG), where the electrical potentials of the brain are recorded on the scalp of a patient, is one of the main tools for diagnosing because rEEG can reveal indicators of epilepsy when patients are in a non-seizure state. Interpretation of rEEG is difficult and studies have shown that 20-30% of patients at specialized epilepsy centers are misdiagnosed [18, 73]. An improved ability to interpret rEEG could decrease the misdiagnosis rate of epilepsy. The difficulty in diagnosing epilepsy from rEEG stems from the large quantity, low signal to noise ratio (SNR), and variability of the data. A usual point of error for a clinician interpreting rEEG data is the misinterpretation of PEEs (paroxysmal EEG events) – short bursts of electrical activity of high amplitude relative to the surrounding signals that have a duration of approximately .1 to 2 seconds [4]. Clinical interpretation of PEEs could be improved with the development of an automated system to detect and classify PEE activity in an rEEG dataset. Systems that have attempted to automatically classify PEEs in the past have had varying degrees of success [47]. These efforts have been hampered to a large extent by the absence of a “gold standard” data set that EEG researchers could use. In this work we present a distributed, web-based collaborative system for collecting and creating a “gold standard” dataset for the purpose of evaluating spike detection software. We hope to advance spike detection research by creating a performance standard that facilitates comparisons between approaches of disparate research groups. Further, this work endeavors to create a new, high performance parallel implementation of ICA (independent component analysis), a potential preprocessing step for PEE classification. We also demonstrate tools for visualization and analysis