554 - The Efficacy of an AI-based Seizure-detection Algorithm for Term vs. Preterm Neonates

Poster Number: 554
Publication Number: 554.1342

Background: The development and clinical application of EEG devices equipped with seizure-detection algorithms using artificial intelligence in Neonatal Intensive Care Units has grown. Although some of these algorithms, trained using term seizure data, have reported sensitivities of approximately 0.5, their adaptability to seizures observed in preterm infants remains uncertain.

Objective: This study examined the validity of automatic seizure detection in term neonates using a seizure-detection algorithm in our cohort and assessed its validity for seizures in preterm infants.

Design/Methods: Between 2019 and 2020, EEG recordings from two institutions that had been recorded with a nine-channel bipolar montage, and included seizure patterns, were collected. The EEG records were classified into those before 37 weeks postmenstrual age (preterm EEGs) and those from 37 weeks onward (term EEGs). EEG records without seizure patterns were also collected. Using the Nihon Kohden seizure-detection program (model QL-162A), these recordings were retrospectively re-analyzed to assess the seizure-detection capability. Experienced investigators evaluated all EEGs for the presence or absence of a seizure pattern. The performance of the program at detecting seizures was investigated.

Results: Of the 48 term EEGs, 26 contained 496 seizure episodes. Of the 68 preterm EEGs, 17 included 276 seizure episodes. The average gestational age at recording was 39.6 weeks for the term EEGs and 33.9 weeks for the preterm EEGs. Other characteristics are shown in Table1. The seizure-detection program accurately identified seizures in 25 of the 26 term infants with seizures, for a sensitivity of 0.96 and specificity of 0.41. At the individual seizure level, the sensitivity averaged 0.48, with a false-positive rate of 0.33 (/hour). For preterm infants, 14 of 17 were correctly identified with a sensitivity of 0.82 and specificity of 0.49. The seizure detection sensitivity was 0.33 and the false-positive rate was 0.33 (/hour) (Figure 1). Notable characteristics of undetected seizures in preterm infants included rhythmic delta activity and brief seizure episodes. Events often misclassified as seizures encompassed delta activity specific to preterm infants and various artifacts.

Conclusion(s): Compared to term infants, the seizure-detection sensitivity was lower in preterm infants. There is a need to develop an artificial intelligence algorithm specifically trained on the unique background EEG patterns and seizure waveforms of preterm infants.