Authors :
Presenting Author: Zubeda Sheikh, MD, MSCTS, FACNS, FAES – Virginia Commonwealth University
Jennifer Haynes, MD – Virginia Commonwealth University
Robert Culley, DO – Virginia Commonwealth University Health System
Christine Baca, MD, MSHS – Virginia Commonwealth University
Rationale:
Continuous electroencephalography (cEEG) is essential for diagnosing non-convulsive seizures, but is limited mostly to large academic medical centers due to limited availability of neurophysiology resources. Rapid-response EEG (RR-EEG) systems provide potential EEG access in such resource-limited settings allowing for potential appropriate triage and transfer of high-risk patients to medical centers with CEEG resources. Several of these systems (Ceribell) have an associated automated seizure detection algorithm (Clarity Pro). Its accuracy has been tested in a large real world study using data from a tertiary care hospital and two affiliate hospitals1. Our study aims to assess the algorithm’s accuracy and clinical utility in a community hospital setting with the long-term goal to improve appropriateness of transfers of high-risk patients for higher level of CEEG care.
Methods:
All consecutive rapid-response EEG studies performed at VCU (Virginia Commonwealth University) Tappahannock hospital (a community-based hospital without in-house neurology or neurophysiology) were included. Two ABPN certified clinical neurophysiologists at the VCU tertiary care medical center in Richmond served as reviewers to set the reference standard, against which the algorithm’s performance was compared. ACNS ICU EEG terminology was used for classification of EEG patterns.
Results:
Eighty-three RR-EEGs were reviewed. At least 39 patients (47%) were awake at baseline mental status at the time of EEG. Six patients had seizures, 3 electrographic (Esz or ESE) and 3 electroclinical (ECSz or ECSE) and 3 of these were classified as status epilepticus (5 min or >90% worst 5 min burden). The measures of sensitivity and negative predictive value were limited by the low number of patients with seizures in this cohort and observations of discordant patient and event level data (Fig 1 ) Specificity was low at lower seizure burden thresholds and progressively improved with use of higher cutoffs with 97% specificity at >90% cutoff (Fig 1). False positive detections at >90% threshold were seen with higher amplitude theta or delta background (Fig 2A/2B). Additional false positive detections at thresholds between 20-90% were seen with artifact from rapid eye blinks, rhythmic EMG artifact from chewing and other electrode artifacts (Fig 2C/2D).
Conclusions:
Low seizure detection likely reflects a selection of patients at low risk for ongoing ESE (as noted half of the cohort was at their baseline mental status), which also generates a high negative predictive value. High false positive detections especially those related to artifacts ( are also likely related to inclusion of awake and interactive patients (low risk). Lack of neurologist supervision in using the device and the algorithm appropriately is a limitation at community hospital settings, which likely affects the reliability of the algorithm. Future work is ongoing: (1) to modify the electronic order for RREEG to help refine and guide utilization for appropriate high-risk patients, in addition to providing (2) interactive teaching and feedback sessions with ordering providers at VCU Tapp.
Funding: No funding was received for this work