Authors :
Jennifer Keene, MD, MBA – University of Utah
Nicholas Abend, M.D., PhD – Children's Hospital of Philadelphia
Giulia Benedetti, MD – University of Michigan
Zach Fullmer, BS – N/A
Maire Keene, BS – N/A
Saeed Montazeri, MD – University of Helsinki
Craig Press, MD – Children's Hospital of Philadelphia
Nathan Stevenson, PhD – QIMR Berghofer Medical Research Institute
Sampsa Vanhatalo, MD – University of Helsinki
Presenting Author: Amanda Sandoval Karamian, MD – University of Utah
Rationale:
Neonatal seizures occur in 1-3.5/1000 term births and up to 40% of infants with neonatal encephalopathy (NE). High seizure burden is associated with unfavorable neurodevelopmental outcomes. Early treatment of neonatal seizures increases the odds of successful seizure cessation and decreases total seizure burden; rapid and effective identification and treatment of neonatal seizures is critical. However, neonatal seizures are often electrographic-only with identification dependent upon electroencephalogram (EEG) interpretation. Optimal EEG monitoring is limited by availability of technologists and neurophysiologists experienced in neonatal EEG, and recognition of neonatal seizures is often delayed by several hours. Advances in artificial intelligence (AI) have enabled the development of AI-derived automated neonatal seizure detectors (NSDs), which have the potential to increase timely and accurate neonatal seizure diagnosis, particularly in resource limited settings. We sought to compare the accuracy and false detection (FD) rates of three AI-derived automated NSDs using an annotated neonatal EEG dataset. It is critical to understand the accuracy of these detectors in artifact rich neonatal intensive care units (NICUs) prior to broad implementation.
Methods:
The test dataset comprised 34 EEGs of term neonates at risk for seizure due to NE with a median duration of 100 hours. Two independent pediatric epileptologists agreed upon the presence or absence of seizures in all EEGs and one epileptologist marked 1247 individual seizures. We compared the accuracy of three NSDs; one that uses a hybrid feature-based/convolution neural network approach (NSD1) and two that utilize deep learning of raw EEG (NSD2 and Persyst-15). NSD identification of individual seizures and FDs was compared to seizure identification using an any-overlap method. Results:
A total of 3402 hours of EEG from 34 neonates (65% male, median gestational age 39 weeks interquartile range (IQR) 37-40 weeks) was evaluated. Seizures occurred in 30 subjects (88%). All three NSDs detected at least 90% of neonates with seizures. All NSDs had at least one FD per neonate without seizures. NSDs varied in accuracy of individual seizure detections, ranging from 33% (NSD2, IQR 15%-62%) to 59% (Persyst-15, IQR 33-79%) and 69% (NSD1 algorithm, IQR 51-90%) (Table 1). Median false detections per day (FD/day) were significantly different between NSD1 at 87 (IQR 56-140), NSD2 at 5.2 (IQR 2.2-13), and Persyst-15 at 2.5 FD/day (IQR 1.1-5.4) (p < .0001). Analysis of an additional 144 neonates is ongoing.