Identifying neonatal seizures with a machine-learning algorithm

August 31, 2020
Miranda Hester

Ms. Hester is Content Specialist with Contemporary OB/GYN and Contemporary Pediatrics.

Diagnosing neonatal seizures can be complicated and difficult. Could an algorithm help? An investigation examines.

Technology has helped rapidly evolve infant care in recent years, but even with improvements, there are still some difficulties. Continuous conventional electroencephalography (cEEG) hasn’t been able to make an accurate diagnoses of neonatal seizures easier. However, an automated seizure detection algorithm named Algorithm for Neonatal Seizure Recognition (ANSeR) could provide help. An investigation in the Lancet Child and Adolescent Health sought to find some answers.1

Investigators conducted a multicenter, randomized, 2-arm, parallel, controlled study. It was run in 8 neonatal centers in Ireland, the Netherlands, Sweden, and the United Kingdom. The study ran from February 2015 to February 2017. Newborn children with a corrected gestational age between 36 and 44 weeks who either had seizures requiring electroencephalography monitoring or were at significant risk of them were randomized to either receiving cEEG monitoring plus ANSeR or cEEG monitoring alone.

The study recruited 264 newborn children who were evenly divided into the 2 treatment groups, but 6 were subsequently excluded: 4 from the cEEG plus ANSeR group and 2 from the cEEG only group. Electrographic seizures were found in 32 of 128 newborns in who received both cEEG and ANSeR and 38 of the 130 newborns who had cEEG monitoring alone. The sensitivity for recognizing newborns with electrographic seizures was 81.3% (95%, 66.7 – 93.3) in the algorithm group and 89.5% (95% CI, 78.4 – 97.5) in the cEEG only group. The specificity was 84.4% (95% CI, 76.9 – 91.0) in the algorithm group and 89.1% (95% CI, 82.5 – 94.7) in the cEEG group. False detection was 36.6% in the cEEG and ANSeR group and 22.7% in the cEEG only group.

Investigators found 659 hours where a seizure occurred. In the algorithm group they found 268 hours and 391 hours in the cEEG only group. Correct identification of the seizure hours was higher in the cEEG plus ANSeR group than the cEEG only group (177 [66.0%; 95% CI, 53.8–77.3] of 268 hours vs 177 [95% CI, 45.3%; 34.5–58.3] of 391 hours; difference 20.8% [3.6–37·.]). There was no significant difference found in the percentage of newborns with seizures being given at least 1 antiseizure medication didn’t work (37.5% [95% CI, 25.0 to 56.3] vs 31.6% [95% CI, 21.1 to 47.4]; difference 5.9% [–14.0 to 26.3]).

The researchers concluded that ANSeR was safe and could accurately detect neonatal seizures. The algorithm didn’t improve the identification of seizures beyond what cEEG could do. However, the algorithm did improve the recognition of seizure hours. They did note that further study should be done.


1. Pavel A, Rennie J, de Vries L et al. A machine-learning algorithm for neonatal seizure recognition: a multicentre, randomised, controlled trial. Lancet Child Adolesc Health. August 27, 2020. Epub ahead of print. doi:10.1016/s2352-4642(20)30239-x