Automatic detection of electrographic Status Epilepticus based on seizure burden estimation in critically ill patients

Background
In critically ill patients, electroencephalogram (EEG) monitoring is crucial for the diagnosis of non-convulsive status epilepticus. However, visual EEG interpretation requires a high degree of availability of trained experts. Recent advancements in artificial intelligence enable the automatic evaluation of EEG recordings for the assessment of seizure burden (SB) and the detection of electrographic status epilepticus (ESE). This could significantly reduce resource requirements, as only a few suspected cases recognized by the algorithm would need to be verified by experienced EEG experts.
Methods
An artificial neural network was trained to recognize electrographic seizures (ES) and their temporal extent (onset and offset), using EEGs recorded from the 10-20 electrode positions, plus optional frontotemporal (FT) or temporal-posterior temporal (TP) positions. Training data were obtained from 859 intensive care- and epilepsy monitoring patients with a total EEG duration of 21400 hours (approximately 2.4 years). Based on this data, hourly seizure burden (HSB) was continuously estimated within a moving one-hour window. According to the ACNS’s Standardized Critical Care EEG Terminology, ESE was detected in case of single seizures lasting for more than 10 minutes, or if the estimated HSB exceeded 20% (12 minutes). For validation we retrospectively collected EEG recordings from 81 patients at Clinic Hietzing in Vienna and at Stanford University in Palo Alto, California. Experienced, board-certified neurologists annotated begin and end of electrographic seizures according to the ACNS. Seizures in the Stanford dataset were independently annotated by two reviewers, seizures in the Hietzing dataset were annotated by one reviewer, a second reviewer independently reviewed for the presence or absence of seizures in each file. Sensitivity, specificity, positive- and negative predictive values were evaluated for the detection of ES and of ESE respectively.
Results
According to the consensus of reviewers, 37 patients (45%) had ES, of whom 21 patients (26%) had ESE. A total of 53 (65%) patients had no ESE, of whom 37 patients had no ES at all. In the remaining 7 patients (9%) no consensus was found among reviewers. The automatic algorithm correctly identified 18/21 patients (86%) with ESE and 36/37 patients (97%) with ES. This corresponds to sensitivities of 86% and 97%, specificities of 91% and 76%, positive prediction values of 78% and 80%, and negative predictive values of 94% and 97%, for automatic recognition of ESE and ES, respectively.
Conclusions
Our study assessed an artificial intelligence-based algorithm for the automatic recognition of electrographic seizures and status epilepticus in critically ill patients. The algorithm showed high sensitivity in identifying patients with electrographic seizures and status epilepticus when compared to expert annotations. These findings suggest that such automated systems have the potential to significantly aid clinical decision-making by detecting and quantifying electrographic seizure burden, reducing file review workload, and facilitating timely diagnosis and management of non-convulsive status epilepticus. If successfully implemented, these advances could enable standardized seizure monitoring of patients in the ICU and thus significantly improve patient care.