Prolonged status epilepticus: Early recognition and prediction of full recovery in a 12-year cohort

Early identification of patients who are at risk of prolonged status epilepticus (SE) and patients with high chances of full recovery despite prolonged SE may urge clinicians to intensify treatment rather than to withdraw care. We aimed to develop prediction models based on readily available clinical parameters to predict prolonged SE at seizure onset and to identify patients with high chances for full recovery.; From 2005 to 2016, all adult SE patients treated at the University Hospital Basel, a Swiss medical care center, were included. Multivariable Poisson regression was performed to identify predictors of prolonged SE (defined as SE for >12, >24, and >48 hours) and return to baseline from prolonged SE. The area under the receiver-operating characteristic curves (AUROC) for prediction models was calculated.; Of 467 patients, the median age was 66.7 years and mortality was 12%. Relative risk (RR) for death was 1.06 (P < 0.0001) with every SE day. In multivariable analysis, nonconvulsive SE with coma, SE severity score ≥3, and acute brain lesions at SE onset independently predicted prolonged SE with an AUROC of 0.68 for >12, 0.67 for >24, and 0.72 for >48 hours of SE. Absence of nonconvulsive SE with coma and a decreasing Charlson comorbidity index were independent predictors for return to baseline in prolonged SE with an AUROC of 0.82 and 0.76 following cross-validation. Both associations remained significant despite adjustments for determinants of adverse outcome, such as anesthetics and vasopressors (nonconvulsive SE with coma RR = 0.24, 95% confidence interval [CI] 0.07-0.86; comorbidity index RR = 0.87, 95% CI 0.76-0.99).; Although our data indicate that identification of prolonged SE at seizure onset is unreliable, timely recognition of patients with high chances of good outcome despite prolonged SE is promising on the basis of comorbidities, type of SE, and level of consciousness. Further external validation of this prediction model is needed.