Abstracts

AI-Based Detection of Seizure and Nonconvulsive Status Epilepticus in a Community Hospital

Abstract number : 2.167
Submission category : 3. Neurophysiology / 3B. ICU EEG
Year : 2025
Submission ID : 1240
Source : www.aesnet.org
Presentation date : 12/7/2025 12:00:00 AM
Published date :

Authors :
Presenting Author: Veeresh Kumar N. Shivamurthy, MD – Trinity Health Of New England

​Salvatore Criscuolo, ACNP-BC – Trinity Health Of New England Corporation, Inc.
Adrienne Clements, ACNP-BC – Trinity Health Of New England
ANYA OSORIO-OTERO, R. EEGT – St. Francis Hospital & Medical Center Trinity Health of New England
Damian Moskal, MD – Trinity Health Of New England

Rationale:

Electroencephalograph (POC EEG) with AI-based seizure detection was introduced at our hospital in 2020. To support timely clinical decision-making, we developed a treatment protocol based on the device’s 5-minute seizure burden (SzB) estimate, which has been shown to be highly accurate in detecting nonconvulsive status epilepticus (NCSE; Kamousi et al., 2021, Neurocrit Care). According to our protocol, SzB estimates ≥90% in the setting of high clinical suspicion warrant treatment and escalation to neurology for review. In all other cases, the protocol recommends either STAT or routine review of EEG results (Fig. 1). The goal of this study was to assess the accuracy of the POC EEG algorithm for NCSE compared with formal neurology interpretation.



Methods: Chart reviews for patients monitored with POC EEG between Apr-Jun 2023 were conducted. EEGs were categorized as “no seizure,” “seizure,” or “NCSE” based on original EEG reports, and compared to AI-based max SzB output.

Results:

Chart reviews for patients monitored with POC EEG between Apr-Jun 2023 were conducted. EEGs were categorized as “no seizure,” “seizure,” or “NCSE” based on original EEG reports, and compared to AI-based max SzB output.

Results: Of 166 patients, the EEG-based diagnoses were: NCSE (n = 1), seizure (n = 9), and no seizure (n = 156). The AI algorithm correctly identified the single case of NCSE (max SzB = 100%); IV medication was initiated within 30 minutes of the SzB exceeding 90%.
Max SzB exceeded 10% in 8 of 9 seizure cases; in the remaining case, artifact and impedance (coinciding with electroclinical seizure) prevented algorithm output, which is designed to detect only electrographic seizures.
Among the five EEGs with SzB estimates ≥90% but no confirmed NCSE, interpretations included: subclinical seizures, IIC, GPDs > 2 to 2.5 hz, sharp waves, BIPDs, and diffuse slowing. Cases with status, seizures, IIC, GPDs > 2 to 2.5 hz were treated based on the workflow and neurology input (Fig 1).



Conclusions:

Consistent with prior work (Kamousi et al., 2021), AI-based NCSE detection demonstrated high sensitivity. The moderate false-positive rate was mitigated by incorporating clinical judgment and STAT neurology review into our protocol.
In 160 cases, the algorithm correctly ruled out NCSE, potentially helping physicians avoid unnecessary escalation of medications and intubation.
The strength of this conclusion is limited by the small sample size and low incidence of confirmed NCSE.



Funding: This study received funding from Ceribell Inc.

Neurophysiology