Abstracts

Rationale: The ability to accurately predict seizures in high-risk neonates would have important management implications. We aimed to understand how data extracted from standardized EEG templates can be used to generate seizure prediction models to guide the care of neonates, particularly those with hypoxic-ischemic encephalopathy (HIE).

Methods: In April 2017, our center implemented a novel EEG reporting system based on common data elements (CDE) in the electronic medical record (EMR) which incorporated standardized terminology from the American Clinical Neurophysiology Society (ACNS). Neonates with HIE were selected from the EMR using the diagnosis codes for “hypoxic-ischemic encephalopathy” and “therapeutic hypothermia.” The list of patients obtained was then cross-validated against an independently curated list of individuals with HIE, with manual chart review of selected patients. The diagnoses and basic demographic data extracted from the EMR were combined with the EEG CDEs and exported using Clarity, a SQL database. The data were analyzed in R Studio using feature correlation and clustering analyses. Seizure prediction models were developed and tested using both logistic regression and decision tree analyses.

Results: Implementation of a standardized EEG reporting system directing the documentation of discrete and standardized EEG variables results in >94% of reports incorporating the recommended terminology. From April 2017 – March 2021, 889 neonates, including 126 neonates with HIE, had continuous EEG data reported using the CDE-based templates and were included in this study. The median age of EEG initiation was 3 days for all neonates and 0 days for neonates with HIE. The median duration of monitoring was 3 days (1, 18) for all neonates and 5 days (1, 10) for neonates with HIE. Several EEG features were highly correlated, and patients could be clustered based on the presence or absence of specific EEG features. Logistic regression analysis revealed that the only feature significantly predictive of seizures on subsequent recording days was the presence or absence of seizures on day 1 (adjusted OR 3.81, 95% CI 1.19-12.19, p=0.024). However, decision tree models incorporating background features had improved performance with accuracies >90% and were able to segment groups at very low and very high risk of seizure. Furthermore, the prediction accuracy improved over time with increasing data accrual.

Conclusions: Adopting a standardized EEG reporting system yielded >94% implementation of ACNS-recommended terminology, enabling analysis of EEG-guided management. Using data extracted from the standardized EEG report on the first day of recording, we could predict the presence or absence of neonatal seizures on subsequent days with classification accuracies of >90%. Furthermore, decision tree analyses identified groups of patients with very low and very high probability of future seizures. This information, incorporated into routine care, could be used to guide decisions about the necessity of long-term EEG monitoring and the allocation of limited EEG resources.

Funding: Please list any funding that was received in support of this abstract.: Children’s Hospital of Philadelphia, The Hartwell Foundation, NINDS.