Abstracts

Rationale: Seizures are a common neurological emergency in the neonatal period. They are associated with significant morbidity and mortality. Neonatal seizures typically occur within 72 hours of birth and are commonly caused by acute brain injuries such as hypoxic-ischemic encephalopathy and intracranial haemorrhage, with around 10-15% of cases early indicators of epilepsy [1]. Despite their clinical importance, EEG characteristics of seizures over the first days of life remain largely unexplored due to resource intensive annotation requirements. AI seizure-detection models provide new tools to describe the seizure characteristics on a large-scale dataset of prolonged EEG recordings.


Methods: An anonymised, retrospective dataset of continuous, multi-channel EEG from 427 neonates was used in this study. The EEG data was recorded from the Cork University Maternity Hospital, Ireland (see Table 1). Over 25k hours of EEG was collected using 9 electrodes over the frontal, central, temporal, and occipital regions, with a reference and ground.



We previously developed a deep-learning model to detect seizures on a per-channel basis [2]. The model obtained state-of-the-art performance on an open-access benchmark, and more importantly, achieved expert-level equivalence on 2 held-out validation sets. The latter represents a milestone achievement in the field of neonatal seizures and opens the possibility of using an automated approach to uncover hereto hidden information in large datasets. We run the entire dataset through our seizure detection model and generate binary masks for each channel for all EEG records.



To quantify seizure characteristics, we assess attributes of seizure events and relate them to the total seizure burden (TSB). TSB is known to be correlated with clinical outcome such as poor neurodevelopmental outcomes.


Results: Of the 427 neonates we find 162 with a TSB > 60 seconds. For this set, median (interquartile range, IQR) TSB was 25.8 (4.0 to 88.0) minutes. Within the first hour after seizure onset, neonates with localised seizures (maximum of 2 channels) have a median (IQR) TSB of 4.7 (2.2 to 22.2) minutes vs 50.5 (21.3 to 122.1) minutes for babies with > 2 channels (p< 0.001) (Figure 1). The duration of seizure period is also shorter for the localised seizure group with median (IQR) of 4.7 (0.1 to 20.1) hours vs 19.7 (7.1 to 46.9) hours (p< 0.001).


Conclusions: The new era of expert-level AI systems for neonatal EEG interpretation presents a promising opportunity to study previously unexplored facets of EEG. We show that this technique can uncover significant insights into seizure evolution in neonates, with just 1 hour of AI-reviewed EEG showing significant predictive power of total seizure burden and duration of seizure period.


References

[1] Ziobro J, Pilon B, Wusthoff CJ et al. Neonatal Seizures: New Evidence, Classification, and Guidelines. Epilepsy Currents 2024;0(0).

[2] Hogan R, Mathieson SR, Luca A et al. Scaling convolutional neural networks achieves expert-level seizure detection in neonatal EEG, arXiv:2405.09911




Funding: NA