Prediction of seizure spread network via sparse representations of overcomplete dictionaries

Abstract

Epilepsy is one of the most common brain disorders and affect people of all ages. Resective surgery is currently the most effective overall treatment for patients whose seizures cannot be controlled by medications. Seizure spread network with secondary epileptogenesis are thought to be responsible for a substantial portion of surgical failures. However, there is still considerable risk of surgical failures for lacking of priori knowledge. Cortico-cortical evoked potentials (CCEP) offer the possibility of understanding connectivity within seizure spread networks to know how seizure evolves in the brain as it measures directly the intracranial electric signals. This study is one of the first works to investigate effective seizure spread network modeling using CCEP signals. The previous unsupervised brain network connectivity problem was converted into a classical supervised sparse representation problem for the first time. In particular, we developed an effective network modeling framework using sparse representation of over-determined features extracted from extensively designed experiments to predict real seizure spread network for each individual patient. The experimental results on five patients achieved prediction accuracy of about 70%, which indicates that it is possible to predict seizure spread network from stimulated CCEP networks. The developed CCEP signal analysis and network modeling approaches are promising to understand network mechanisms of epileptogenesis and have a potential to render clinicians better epilepsy surgical decisions in the future.