Multistate survival analysis for seizure prediction on continuous EEG
Struck A, Osman G, Rampal N, Biswal S, Legros B, Hirsch L, Westover MB, and Gaspard N. Multistate survival analysis for seizure prediction on continuous EEG. Neurology 2017; 88(16 Suppl).
Objective: Determine risk factors and time dependence of seizure risk on continuous EEG. Background: Prior analyses of risk factors for continuous EEG prediction are retrospective and uncontrolled for the effects of censoring/subject drop out. To correct this shortcoming, we used a multistate survival analysis on 665 consecutive continuous EEG recordings. Design/Methods: Retrospective analysis of a prospectively acquired database of 665 consecutive continuous EEG sessions (>24hours) with associated clinical factors and EEG data including time to event. Elasticnet logistic regression was used to determine predictive risk factors of time independent variables, subsequently used in Cox proportional hazard model. Time dependent variables were used to create a multistate survival model with three states (entry, risk state, and seizure). The risk state was defined by emergence of epileptiform patterns: lateralized periodic discharges (LPDs), bilateral independent periodic discharges (BIPDs), brief rhythmic discharges (BRDs), lateralized rhythmic delta activity (LRDA), and/or sporadic epileptiform discharges (SED). Bootstrapping was used to generate 95% confidence intervals. Results: Electrographic seizures occurred in 23 % of the cEEG monitoring sessions. Time independent variables of greatest predictive value were coma (31% had seizures on EEG; O.R 1.8 p