Title: Integrating Convolutional Neural Networks and Probabilistic Graphical Models for Epileptic Seizure Detection and Localization

 

Abstract: Epilepsy affects 1% of the population worldwide, and roughly 30% of these patients do not respond to medication. If we can trace the seizures to a singe brain region, then the best course of action is to surgically resect this seizure onset zone (SOZ). Electroencephalography (EEG) is the first and foremost modality used in epilepsy management. However, due to the wide variation in epileptic pathologies, seizure detection and localization rely almost exclusively on expert visual inspection of the EEG signals. This process is time consuming and prone to human error. Moreover, the electrographic signatures can be difficult, if not impossible, to isolate in many patients due to motion artifacts, apparent multifocal onsets, and rapid spreading patterns.

This scenario provides an ideal opportunity for automated methods to simultaneously mine and integrate discriminative cues in the data that can augment expert review. In this talk, I will describe a new framework for seizure detection and localization from multichannel EEG data. The crux of our approach is that the propagation of seizure activity provides valuable information about its onset. Our unique modeling strategy combines the interpretability of probabilistic graphical models with the representational power of deep learning. Specifically, the latent variables in our PGM will capture the spread of seizure activity; they are complemented by a nonparametric likelihood based on convolutional neural networks. We demonstrate that our approach achieves better detection accuracy than competing baseline models, and that it can identify the SOZ across a heterogeneous patient cohort without any a priori information.