109 Rapid Response Electroencephalograph With Artificial Intelligence for Diagnosing Seizures and Highly Epileptiform Patterns in Emergency Medicine

Abstract

To test the performance of a recently FDA-cleared machine learning method (Claritγ) that generates bedside alerts for possible status epilepticus and measures in real time the burden of seizure activity. We designed a retrospective study of electroencephalographs from adult patients (n=353) who underwent monitoring with Rapid Response EEG system (Rapid-EEG, Ceribell Inc.) for evaluation of possible seizures between January 2018 and April 2019. We developed a machine learning method for automated detection of seizure activity and seizure burden throughout each recording. We measured sensitivity and specificity of possible status epilepticus (seizures>5 minutes) and real-time seizure burden measurements generated by Claritγ compared to the majority consensus of at least 2 expert neurologists reviewing the same EEGs. Various thresholds of seizure burden were tested (≥10% indicating at least 30 seconds seizure activity in the last 5 minutes, ≥50% indicating at least 2.5 minutes of seizure activity, and ≥90% indicating at least 4.5 minutes of seizure activity and triggering an alert for status epilepticus). Majority consensus of neurologists labeled the 353 EEGs as normal or slow activity (n=249), highly epileptiform patterns (HEP, n=87), or seizures (N=17, nine longer than five minutes and eight shorter than five minutes). The sensitivity and specificity of various thresholds for seizure burden during EEG recordings for detecting patients with seizures was 100% and 82% for 50% seizure burden and 88% and 60% for 10% seizure burden. The algorithm generated a status epilepticus alert (≥90% seizure burden) with 100% sensitivity and 93% specificity. Of the 179 EEG recordings in which the algorithm detected no seizures, seizures were identified by the expert reviewers in only 2 cases, indicating a negative predictive value of 99%. Combination of high sensitivity for status epilepticus events combined with a high negative predictive value for negative cases makes our novel algorithm a useful tool for triaging EEGs in emergency care settings. Confirming cases of non-convulsive subclinical status epilepticus cases within minutes of their arrival to emergency department and independently from neurologists and or EEG technicians will expedite the triage of these critically ill patients and will expedite their treatment. In addition, ruling out seizures accurately in cases being suspected to have seizures can help prevent unnecessary or aggressive over-treatment of such patients.