Accuracy of Clarity Machine-Learning Algorithm in Detecting Status Epilepticus on Point of Care EEG
Abstract number :
1.133
Submission category :
3. Neurophysiology / 3B. ICU EEG
Year :
2023
Submission ID :
866
Source :
www.aesnet.org
Presentation date :
12/2/2023 12:00:00 AM
Published date :
Authors :
Presenting Author: Omar Hussein, MD – University of New Mexico
Masoom Desai, MD – Assistant Professor, Department of Neurology, University of New Mexico; Mariel Kalkach Aparicio, MD – Research Specialist, Department of Neurology, University of Wisconsin-Madison; Aaron Struck, MD – Assistant Professor, Department of Neurology, University of Wisconsin-Madison
Rationale:
Point-of-care EEG (POC EEG) systems enable quick triage of patients at risk for seizures in acute care settings. Machine learning algorithms can enhance the benefit of POC EEG for triaging patients at bedside by who are not EEG-specialists. It also can enhance care by providing continuous real-time monitoring of the brain with instant alerts when prolonged seizure activity is detected for clinical teams. In this study, we evaluated the accuracy of the Clarity algorithm (Ceribell Inc.) to detect status epilepticus in POC EEGs when compared to EEG interpretation by three independent experienced EEG specialists.
Methods:
Using data from the SAFER-EEG trial, we retrospectively analyzed an unbiased dataset of 264 POC EEG recordings across three academic hospitals. All EEGs were read post-hoc by three independent readers (trained epileptologists/neurophysiologists) and classified into four categories: 1) status epilepticus (SE), 2) seizure, 3) highly epileptiform patterns (HEP), and 4) non-HEP slow, or normal. Data from Clarity performance was collected and those with maximum seizure burden (max. SzB) threshold of ≥90% were categorized as SE. The concordance between EEG interpretations by experienced EEG readers and Clarity performance for SE was sought.
Results:
Among cases with 100% agreement of the three epileptologists (i.e., cases in which all three readers agreed on the category, N = 164), there were 8 cases that were categorized as SE. The sensitivity of Clarity performance to detect SE was 87%, while specificity was 98%, PPV was 70% and NPV was 99%.
Conclusions:
This study aligns with previous studies on this topic, indicating a high level of concordance for the detection or rule out of status epilepticus between the Clarity algorithm and human EEG reader reviews. The high sensitivity and NPV provide confidence for the use of this algorithm as a critical care triage tool.
Funding: Study funding provided by Ceribell, Inc.
Neurophysiology