Seizure forecasting results for three canines with naturally occurring epilepsy implanted with the NeuroVista Seizure Advisory System.

A logistic regression classifier was trained and used to predict iEEG data (1 minute blocks) as either pre-ictal or inter-ictal, using power-in-band (PIB) features. Each classifier used 10 of 96 available PIB features, chosen during the training process via forward selection. Forecasts and subsequent statistical analyses were based on different target proportions of total time in warning, ranging from 0.1 to 0.5. A 10-fold cross-validation scheme was used for all phases of feature selection, classifier training, seizure forecasting, and statistical analysis. For each dog a range of values for time in warning (TIW) are considered. The Sensitivity (Sn) and p-value (p) are reported separately for all seizures and for lead seizures only (Sn-lead, pn-lead) and have been adjusted to account for the performance of the chance prediction algorithm.