David M. Groppe 1, Gerard O’Leary 2, Victoria Barkley 1, Richard Wennberg 3, Esther Bui 3, Peter Tai 3, Aylin Reid 3, Danielle Andrade 3, Martin Del Campo 3, Roman Genov 2, Taufik A. Valiante 1,3
1 Computational Neuroscience, Krembil Research Institute
2 Electrical & Computer Engineering, University of Toronto
3 Toronto Western Hospital
Surgically removing seizure generating brain regions is the most effective way of preventing seizures in patients with focal, drug resistant epilepsy. Identifying these brain regions typically entails acquiring a vast amount of different kinds of data (e.g., 10-30 days of continuous intracranial EEG recordings with simultaneous video; brain MRI and CT scans). Interpreting these data is done manually, is labor intensive, and requires unique clinical expertise. The scale and diversity of data make epilepsy surgery an ideal target for AI in many ways. For example, computer vision can be used to analyze video of patients and brain scans, supervised and unsupervised learning of time series data can identify brain states/regions of interest from the EEG data, and reinforcement learning can be used to stimulate the brain to prevent seizures. Our lab has been leveraging the power of AI to develop novel anti-seizure devices and assisted diagnosis software. Here we present a summary of (1) the clinical data we have archived, (2) software we have developed for data visualization/AI-aided expert annotation, (3) an implantable machine learning-accelerated chip for brain state classification and control that we have created, and (4) our ongoing clinical trial of the chip for seizure control at Toronto Western Hospital. Analysis of archived data suggests that our chip should significantly outperform the current clinically approved responsive brain stimulator for seizure control.