Abstracts

Rationale: Prompt identification of seizures allows rapid intervention that reduces seizure-related morbidity and mortality. Seizure detection in real-time includes multiple approaches including signal processing, machine learning, and training non-expert readers for seizure identification. Translating this into clinical practice is challenging due to highly variable patient population and EEG artifacts. To circumvent this, we propose automated detection of spectral signatures that represent seizure-like activity, and subsequent notification of non-expert readers (Figure 1). We aim to evaluate its effectiveness in identifying seizure-like activity, and characterize artifactsMethods: Patients aged >17 years admitted to the intensive care units (ICU), floor or epilepsy monitoring unit (EMU) were prospectively identified over 28 days. Demographic and clinical data including EEG recordings alongside videos and reports, and ""events"" such as electrographic seizures, epileptiform discharges (EDs), generalized and periodic lateralized EDs (GPEDs, PLEDs), and nonconvulsive status epilepticus (NCSE) were recorded daily. Commercially available seizure detection software (Persyst version 12) was used to identify events by compressing raw EEG signals to one hour scale. Visual EEG analysis was considered gold standard. Performance of the software in detection of events was quantified by calculation- true and false positives (TP, FP), and false negatives (FN). Artifact was classified as physiological (movement, chewing); mechanical or electrical intereference, and electrode contact-related (faulty connections, breech rhythm).Results: A total of 308 EEG studies totaling 6164 hours were performed. Average age was 54.3 years (range 18-91). 42 studies were positive for seizures and status epilepticus (SE), and 29 for EDs, GPEDs, and/or PLEDs. A total of 1580 events were detected over 28 days with 50.9% of events classified as true positives (TP), and 49.1% false positives (FP). No events were detected in 11 studies (false negative) (fig.1). The TP rates were 50.2%, 55.2%, and 49% whereas the FP rates were 49.8%, 44.8%, and 51% for patients admitted to ICU, hospital floor and EMU respectively. These results are not significantly different (p=0.20). The variability of TP and FP rates over 28 days is highlighted in Figure 1. Patterns of artifact leading to false positives were evaluated in 90 studies. Artifact due to muscle activity was most frequent and present in 28/90 studies, followed by artifacts due to faulty electrode contacts (11/90), and miscellaneous (8/90 – electrocardiogram, mechanical events). No artifacts were detected in 43/90 studies.Conclusions: Although the software can detect majority of events concerning for seizure-like activity, there is a high false-positive rate caused by different artifacts, the most common of which is muscle activity. By identifying these artifacts, the software can be potentially modified in order to increase the specificity. We believe that this approach would enable non-expert readers to readily identify EEG patterns in real-time that warrant intervention.