April 16: Piotr Mirowski, CIMS
Towards seizure prediction from EEG
Could we predict seizures? Clinical case studies reported that some
epileptic patients have premonitory symptoms minutes to hours before the
seizure onset, and recent research has measured corresponding
electrophysiological changes in the brain activity.
In this talk, I will present our patent-pending algorithm for epileptic
seizure prediction, developed in collaboration with Yann LeCun and the NYU Medical Center, and bridging a long-standing neurophysiological research problem
with machine learning algorithms. Our software extracts patterns of
synchronization between intracranial EEG, and feeds them into an
L1-regularized convolutional network. Evaluated on datasets of patients
with medically-intractable epilepsy, our method significantly outperformed
previous research in the field.
There is still room for improvement before we could translate our
methodology to human neuro-prosthetics. I will therefore briefly mention
my PhD research on time series modeling, using state-space models and
gradient-based approximations to the EM algorithm, and explain how I am
working on improving our current seizure prediction system.