Predicting pediatric cardiac arrest outcomes using early quantitative EEG

Abstract

Aim: Accuracy of neuroprognostication after pediatric cardiac arrest (CA) is critical for directing clinical care. Current limitations include imprecise neuroprognostication models, inability to discriminate between degrees of disability, and lack of modifiable post-CA biomarkers. Models including quantitative EEG (qEEG) characteristics may improve post-CA prognostic accuracy.

Methods: Retrospective multicenter cohort of children (3mo-18 yr) without return to neurologic baseline post-CA at two pediatric tertiary care hospitals (2010-2016) with ≥ 6-hours of EEG within 24-hours post-CA and baseline Pediatric Cerebral Performance Category (PCPC) 1-3. Primary outcome measure was 6-month PCPC dichotomized into favorable (1-3) and unfavorable (4-6 and Δ > 1). Training and validation sets were derived from clinical variables, qualitative EEG (qualEEG) features, and qEEG analysis using Persyst software.

Results: Among 221 subjects, 84 (38%) had favorable 6-month outcomes. All models including clinical features (AUC 0.73 [0.59-0.87]), qualEEG (0.90 [0.81-0.97]) and qEEG features (0.85 [0.74-0.94]) predict outcomes well. A parsimonious model incorporating clinical, qualEEG and qEEG variables had an AUC of 0.92 (0.85-0.97) for predicting outcome. Increased SR was associated with degree of disability and unfavorable outcomes. Machine learning models were not superior to the more transparent parsimonious model.

Conclusions: qEEG features measured with 24-h post-CA add to predictive outcome models and can be trended at the bedside. SR is an objective measure that may improve the precision of outcome prediction. qEEG features may be targetable dynamic brain injury biomarkers which could aid in future studies of neuroprotective interventions.

Keywords: Electroencephalography; Machine learning; Neurocritical care; Neuroprognostication; Prediction model.