Accuracy of Seizure Burden Monitoring and Detection of Status Epilepticus with the Clarity AI Algorithm
Josef Parvizi1, Archit Gupta2, Michelle Armenta Salas2, Suganya Karunakaran2, Tanaya Puranik2, Baharan Kamousi2
1Neurology, Stanford, 2Ceribell
Objective:

To evaluate the performance of Clarityv8 artificial intelligence (AI) algorithm (Ceribell, Inc.) in detecting status epilepticus (SE) and seizures in a large retrospective adult dataset. 

Background:
Point-of-care electroencephalography (POC EEG) systems enhanced with AI algorithms may bridge gaps in timely EEG access and seizure triage.
Design/Methods:

We analyzed 1340 adult POC EEGs acquired with the Ceribell EEG system.  Two or more experts marked epochs of seizure and other pathological activity; from these an EEG label was assigned per file per reviewer.  EEGs were labeled as SE if the seizures’ cumulative duration over any 5-min window was ≥4.5min or if seizures were present for ≥20% of any 1-hour window.  The Clarityv8 AI maximum estimated seizure burden (SzB) was extracted per file.  Detection of SE was evaluated at SzB ≥50% and ≥90%, including if any 1-hour window had seizures for ≥12-min. We also evaluated detection of seizures/SE at ≥1% and ≥10% SzB.

Results:

Of 1340 EEGs, expert consensus identified 25 EEGs meeting criteria for SE and 51 with non-SE seizures. For SE detection, Clarityv8 AI at SzB ≥90% correctly identified 24 cases for a 96.0% sensitivity, with 94.8% specificity and 99.9% negative predictive value (NPV). For SzB ≥50%, the sensitivity, specificity, and NPV were 96%, 93.3%, and 99.9%, respectively.  In the missed SE case, Clarity’s SzB was 10%. For seizures/SE detection, with SzB ≥10%, the sensitivity, specificity, and NPV were 88.2%, 86.1%, and 99.2% respectively. And with SzB ≥1%, the sensitivity was 96.1% with a specificity of 74.0% and a NPV of 99.7%.

Conclusions:

In a large real-world dataset, Clarityv8 AI demonstrated high performance for SE and seizure detection. Its high NPV offers a clinically useful tool for immediate rule out of seizures and SE during neuroemergencies, underscoring the value of AI algorithms as flexible tools to triage patients and potentially reducing empiric treatments.

10.1212/WNL.0000000000216518
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.