The statistics of a practical seizure warning system

Statistical methods for evaluating seizure prediction algorithms are controversial and a primary barrier to realizing clinical applications. Experts agree that these algorithms must, at a minimum, perform better than chance, but the proper method for comparing to chance is in debate. We derive a statistical framework for this comparison, the expected performance of a chance predictor according to a predefined scoring rule, which is in turn used as the control in a hypothesis test. We verify the expected performance of chance prediction using Monte Carlo simulations that generate random, simulated seizure warnings of variable duration. We propose a new test metric, the difference between algorithm and chance sensitivities given a constraint on proportion of time spent in warning, and use a simple spectral power-based measure to demonstrate the utility of the metric in four patients undergoing intracranial EEG monitoring during evaluation for epilepsy surgery. The methods are broadly applicable to other scoring rules. We present them as an advance in the statistical evaluation of a practical seizure advisory system.