Abstracts

Rationale: The seemingly random occurrence of seizures creates constant uncertainty, a source of suffering for patients with epilepsy. A pioneering prospective study predicting the preictal state, minutes ahead of impending seizures, demonstrated the feasibility of seizure prediction in some patients1. Recently, using chronic EEG recordings (NeuroPace RNS® System), we showed that proictal states of heightened seizure risk recur cyclically at longer time-scales. Indeed, circadian and multidien (multi-day) rhythms of interictal epileptiform activity (IEA) strongly correlate with the likelihood of seizure occurrence2, suggesting that seizure forecasting may be achievable at unprecedented horizon lengths. In this study, we hypothesized that the phase of IEA rhythms at different time-scales (hours – days) can inform seizure likelihood modeled as a point-process. Methods: Using recordings of hourly counts of IEA and seizures over several months in 17 patients, we trained point process generalized linear models that included circadian and multidien rhythms as covariates. Optimal history length was first selected using the validation dataset. Results: Model performance at horizons of one and 24 hours was evaluated on a held-out test dataset (40% of data) using the area under the sensitivity versus proportion of time in warning curve. In both cases, the phase of the multidien rhythms significantly improved the prediction scores compared to when only the recent history of seizures and IEA was used. In a subset of patients, it was possible to forecast days of heightened seizure risk 24 hours in advance, and sensitivity greater than 80% was achievable with time in warning under 30%. Conclusions: These results show that the inclusion of IEA rhythms at multidien time-scales can improve seizure prediction algorithms. Forecasting proictal states with a 24-hour horizon represents a paradigm shift in the field of seizure prediction that has so far mostly focused on the preictal state. Funding: No funding