Development and Validation of a Seizure Prediction Model in Neonates Following Cardiac Surgery

ABSTRACT BACKGROUND Electroencephalographic seizures (ES) following neonatal cardiac surgery are often subclinical and have been associated with poor outcomes. An accurate ES prediction model could allow targeted continuous electroencephalographic monitoring (CEEG) for high-risk neonates. METHODS Development and validation of ES prediction models in a multi-center prospective cohort where all postoperative neonates with cardiopulmonary bypass (CPB) underwent CEEG. RESULTS ES occurred in 7.4% of neonates (78 of 1053). Model predictors included gestational age, head circumference, single ventricle defect, DHCA duration, cardiac arrest, nitric oxide, ECMO, and delayed sternal closure. The model performed well in the derivation cohort (c-statistic 0.77, Hosmer-Lemeshow p=0.56), with a net benefit (NB) over monitoring all and none over a threshold probability of 2% in decision curve analysis (DCA). The model had good calibration in the validation cohort (Hosmer-Lemeshow, p=0.60); however, discrimination was poor (c-statistic 0.61) and in DCA there was no NB of the prediction model between the threshold probabilities of 8% and 18%. Using a cut-point that emphasized negative predictive value (NPV) in the derivation cohort, 32% (236 of 737) of neonates would not undergo CEEG, including 3.5% (2 of 58) with ES (NPV 99%, sensitivity 97%). CONCLUSIONS In this large prospective cohort, a prediction model of ES in neonates following CPB had good performance in the derivation cohort with a NB in DCA. However, performance in the validation cohort was weak with poor discrimination, calibration, and no NB in DCA. These findings support CEEG monitoring of all neonates following CPB.