Automatic seizure onset detection in drug-resistant epilepsy: A Bayesian optimal solution

Automatic seizure onset detection (ASOD) from intracranial EEG recordings (iEEG) in drug-resistant epilepsy has recently gained large interest. Effective ASOD can shorten the time for offline review of the iEEG recordings, help with the development of unsupervised online monitoring systems, and contribute to responsive neurostimulation. Depending on the application, the performance goals may vary and a tradeoff between detection delay and probability of false positives must be solved. We recently developed a cost-based framework for ASOD. Applied to iEEG recordings, this framework (i) constructs a statistic to distinguish seizure and non-seizure states; (ii) models the dynamics of this statistic in each state and the state transitions; and (iii) develops a Bayesian 'quickest detection' strategy to estimate online the transition into seizure from sequential iEEG measurements. A cost function of the detection delay and probability of false positives is explicitly minimized, and the solution is a threshold which decreases over time non-monotonically to avoid responding to rare events that would normally trigger false positives. Applied to iEEG data from two drug-resistant epileptic subjects (202h, 11 seizures), our framework achieved low false positive rates (0.07 false positives/h) and decreased the average detection delay by 30% when compared with existing detection policies.