Authors :
Presenting Author: Michelle Armenta Salas, PhD – Ceribell, Inc.
Parshaw Dorriz, MD – Department of Neurology – Providence Mission Medical Center; Kapil Gururangan, MD – Department of Neurology – David Geffen School of Medicine at UCLA; Richard Kozak, MD – Department of Emergency Medicine – Providence Mission Medical Center; Matthew Kaplan, MD – Department of Emergency Medicine – Providence Mission Medical Center
Rationale:
In 2018, a point-of-care electroencephalogram (POC EEG) was incorporated into acute care workflows at our community hospital. Since then, a machine learning algorithm (Clarity) has been released that provides seizure burden (SzB) monitoring and alerts for suspected status epilepticus (SE). In this study, we evaluated the real-world concordance between POC EEG interpretations by neurologists and Clarity SzB outputs.Methods:
We analyzed retrospective data from one year of POC EEG use at our hospital (N = 317). The on-site EEG-trained neurologist performed a post-hoc review of the neurology report and categorized the cases as either: Normal/Slow, Highly Epileptiform Patterns (non-seizure) (HEP), Seizure, or Status Epilepticus (SE). Clarity was run post-hoc on each of the POC EEG recordings to obtain a maximum SzB per file. We identified cases in which the neurologist’s categorization was SE and cases in which Clarity would provide an SE alert (maximum SzB ≥90%) to determine the frequency of true positive, false negative, and false positive cases. We also identified discordant cases and sent these POC EEGs for review by an epileptologist blinded to the initial neurologist’s categorization and Clarity SzB output.Results: Seven cases were categorized by an EEG-trained neurologist as SE, and 10 cases had a Clarity maximum SzB of ≥90% that would trigger an SE alert. Six cases categorized by the neurologist as SE had a maximum SzB of 100%, and one had a maximum SzB of 83.3%. On secondary review, the additional neurologist categorized this case as Seizure rather than SE.
All 4 cases that triggered a Clarity SE alert (but were not categorized by the neurologist as SE) were categorized by the neurologist as Seizure or HEP, indicating abnormal EEGs with abundant epileptiform discharges. The real-world diagnostic accuracy of Clarity SE alerts with respect to the neurologist categorization of SE was 98.4%, with 85.7% sensitivity, 98.7% specificity, and 99.7% negative predictive value.
Conclusions: We found that Clarity had a high level of concordance with interpretations by an EEG-trained neurologist. These results suggest that Clarity can accurately monitor for suspected SE in real-world acute care practice, providing initial guidance to the bedside team to either escalate care or rule-out ongoing seizures. It can also be valuable as a preliminary “second opinion” and screening tool to assist neurologists reviewing POC EEGs in making their final diagnoses.
Funding: Study funded by Ceribell, Inc.