Abstracts

Rationale:
Electroencephalographic seizures (ES) are common in encephalopathic critically ill children but identification requires extensive resources for continuous EEG monitoring (CEEG). In a previous study, we developed a clinical prediction rule to stratify patients at highest risk for ES for whom CEEG might be essential. The model incorporated age (greater or less than 1 year old), acute encephalopathy category (acute non-structural, acute structural, or epilepsy-related), clinically-evident seizure(s) prior to CEEG initiation (absent or present), EEG background category (normal/sleep, slow-disorganized, discontinuous, burst-suppression, or attenuated-featureless), and epileptiform discharges (absent or present). In the current study, we aimed to validate the ES prediction model using an independent cohort.
Method:
We performed a prospective observational study of 314 consecutive critically ill children from 2/19-12/19 treated in the Pediatric Intensive Care Unit of a quaternary care referral hospital with acute encephalopathy undergoing clinically-indicated CEEG to screen for ES based on a guideline-adherent institutional pathway. We used the multi-variate logistic model from the generation cohort to calculate the predicted probabilities of ES in the validation cohort, termed model scores. For clinical application, patients with model scores above a selected cutpoint would undergo CEEG. We evaluated the same model score cutpoints as studied in the generation cohort. We evaluated goodness-of-fit, discrimination, and calibration. We calculated test characteristics for the model in the validation cohort emphasizing high sensitivity.
Results:
We enrolled 314 consecutive critically ill children in the validation cohort. The generation and validation cohorts were alike in most clinical and EEG characteristics. The incidence of ES in the validation cohort was 22%. The Hosmer-Lemeshow test for goodness-of-fit (statistic 3.59; p=0.89) indicated that the observed event rates matched the expected event rates in subgroups of the model population. The model yielded an AUROC of 0.80 in the generation cohort, and it yielded an AUROC of 0.77 in the validation cohort. A 0.1 model score cutpoint selected to emphasize sensitivity yielded sensitivity of 90% and specificity of 37%, positive predictive value of 28%, and negative predictive value of 93%. If applied, the model would limit 31% of patients from undergoing CEEG while failing to identify 10% of patients with ES. The Lowess fit for the predicted versus observed probabilities indicated a slope of 0.954, indicative of good calibration.
Conclusion:
A model employing readily available clinical and EEG variables performed well when validated in a new consecutive cohort. Implementation would substantially reduce CEEG utilization although some patients with ES would not be identified. This model may serve a critical role in targeting limited CEEG resources to critically ill children at highest risk for ES.
Funding:
:NIH K02NS096058, NIH U54-HD086984, and Wolfson Family Foundation.