Advancing Preterm EEG Analysis: A Stochastic-Gradient-Descent-Enhanced Adaptive Algorithm for Interburst Interval Detection

Background and aims: Previous studies have highlighted the importance of discontinuous activity in preterm EEG, characterized by bursts of activity1,2. Traditional burst detection has been manual, time-consuming, and prone to variability. The aim of this study was to optimize a dynamic, adaptive threshold algorithm for enhanced burst detection in preterm neonates and evaluate it on clinical real-world data.
Methods: We developed the AT-IBI (Adaptive Threshold Interburst-interval) algorithm and applied it to 30 EEG recordings from 15 preterm infants, born between 23 and 25 weeks of gestation. Two expert reviewers manually annotated 5-minute EEG epochs for burst activity. Our algorithm, AT-IBI, works with adaptive thresholds which are based on the mean peak-to-peak amplitude. Its inner parameters were fine-tuned through training a neural network that simulates the algorithm's internal logic. Performance was compared to two previously published methods based on line-length and the non-linear-energy-operator3–5.
Results: Interrater reliability between human experts was significant, with a kappa statistic of 0.73. Following optimization, AT-IBI improved agreement with the consensus human rating by 19.7%, increasing the kappa value from 0.66 to 0.79. The algorithm successfully identified 65% of interburst intervals (IBIs) and 60% of bursts with at least 80% overlap, surpassing the other algorithms in performance.
Conclusions: The AT-IBI algorithm exhibited strong capability in identifying burst patterns within preterm infants' clinical EEG data, highlighting its applicability in real-world settings. It achieved performance levels comparable to those of human raters, underscoring its effectiveness as an automated solution for detecting bursts in preterm infant EEG.