Forecasting Seizure Risk over Days

For persons with epilepsy, much suffering stems from the apparent unpredictability of seizures. Historically, efforts to predict seizures have sought to detect changes in brain activity in the seconds to minutes preceding seizures (pre-ictal period), a timeframe that limits preventative interventions. Recently, converging evidence from studies using chronic intracranial electroencephalography revealed that brain activity in epilepsy has a robust cyclical structure over hours (circadian) and days (multidien). These cycles organize pro-ictal states, hours- to days-long periods of heightened seizure risk, raising the possibility of forecasting seizures over horizons longer than the pre-ictal period. Here, using cEEG from 18 subjects, we developed point-process generalized linear models incorporating cyclical variables at multiple time-scales to show that seizure risk can be forecasted accurately over days in most subjects. Personalized risk-stratification days in advance of seizures is unprecedented and may enable novel preventative strategies.