Abstracts

Rationale: An increasing number of critically ill children undergo resource-intense continuous EEG monitoring (CEEG) to identify electroencephalographic seizures (ES). Models are needed to identify children at risk for ES to optimally target limited CEEG resources. We aimed to identify clinical and EEG risk factors for ES, use these variables to develop ES prediction models, and determine whether a model incorporating data from a brief screening EEG had better performance characteristics than a model using only clinical variables. Methods: We performed a prospective observational study of consecutive critically ill children with acute encephalopathy. Predictor variables were chosen to be readily available to clinicians prior to CEEG onset and readily available to electroencephalographers using a 30-minute screening EEG. Variables that were associated with seizures on univariate analyses were considered in the multiple logistic regression to predict seizures. We evaluated the model performance with the areas under the receiver operating characteristic curve (AUROC), validated the optimal model which had the highest AUROC using a 5-fold cross-validation approach, and calculated test characteristics emphasizing sensitivity (to decrease the probability of failing to perform CEEG in a child who would experience ES). Results: The final model included age (less than or greater than 1 year), acute encephalopathy category (acute non-structural, acute structural, epilepsy related), seizures prior to CEEG, EEG background category (normal-sleep, slow-disorganized, discontinuous, burst-suppression, attenuated-featureless), and epileptiform discharges. Adding information from a screening EEG (AUROC 0.80) to clinically available information along (AUROC 0.70) led to significantly higher model performance (p<0.01) (Figure). At a 0.10 cutpoint to emphasize sensitivity, the optimal model had an AUROC of 0.80, sensitivity of 92%, and specificity of 37%. Given the seizure incidence of 26%, the test diagnostics included positive predictive value of 34% and negative predictive value of 93%. If applied, the model would designate that 29% of patients would not undergo CEEG but would fail to identify 8% of patients with ES. Varied cutpoints could be chosen to optimize sensitivity or specificity depending on available CEEG resources (Table). Conclusions: A model developed using a small number of readily available clinical and EEG variables could guide the use of limited CEEG resources. Data obtained from a brief screening EEG significantly improved model performance compared to clinical data alone. At a cutpoint selected to avoid failing to identify children experiencing ES, model implementation would lead to a substantial reduction in CEEG requirements but miss rare children experiencing ES. Funding: NINDS K02NS096058