Abstracts

Rationale: Nonconvulsive seizures and status epilepticus (SE) are common neurological emergencies in pediatric critically ill patients (Abend, et al. 2011). Timely management of these seizures is key to avoiding negative neurological outcomes (Payne, et al. 2014). Point-of-care electroencephalography (POC EEG) has shown promise in reducing time to EEG in pediatric settings (Rajan, et al. 2025). Use of artificial intelligence (AI) algorithms to aid in the monitoring and detection of seizures may further facilitate rapid management in this at-risk population. Recently, the FDA cleared a newly developed version of the Clarity seizure detection algorithm (Ceribell, Inc.) specifically designed and optimized to expand the intended patient population to pediatric patients 1 year and older. Here, we validated the performance of Clarity for detecting suspected SE in a large real-world dataset of pediatric patients.

Methods: A retrospective dataset of 644 pediatric POC EEGs was annotated by two or more pediatric neurologists. Reviewers marked epochs of pathological and normal EEG activity in each study, and the most severe epoch reported in a study was used to create a per-reviewer label. We subsequently selected the majority label from all reviewers. Seizures lasting 5 minutes or more were labeled as suspected status epilepticus. Pediatric Clarity was run post-hoc on the EEGs. The algorithm calculates seizure burden as the percentage of seizure activity detected within a 5-minute moving window. Then, the peak 5-min seizure burden was obtained for each EEG (SzB). Performance was evaluated using a 90% SzB threshold for detecting suspected SE across the entire pediatric cohort and in age subgroups.

Results: Pediatric Clarity correctly identified 17 out of 18 suspected SE cases, demonstrating a sensitivity of 94.4% across the entire cohort, with a specificity of 93.1%. In the 1-7 y cohort (N = 242), Clarity detected 9/10 suspected SE cases for a sensitivity of 90% and a specificity of 87.9%. The missed SE case had a peak SzB of 87%. In the 8-17 y cohort (N = 402), the algorithm detected 8/8 suspected SE cases for 100% sensitivity and 96.2% specificity. In the full cohort, the negative predictive value for ruling out suspected SE was 99.8% with a positive predictive value of 28.3%. In 43 false positives for suspected SE, 12 had short seizures, 16 had highly epileptiform patterns (non-seizure), 7 had rhythmic patterns or other abnormalities, and 8 were normal. Using 0% SzB for ruling out seizures in the full cohort, the algorithm demonstrated a NPV of 99.0%, with a sensitivity of 91.5% and a specificity of 71.3%.

Conclusions: The latest version of Clarity, an AI seizure-detection algorithm, detects suspected status epilepticus in pediatric patients with high sensitivity and specificity, and maintains high performance across age subgroups. The algorithm also has high negative predictive value to rule out seizures and status epilepticus at 0% SzB, which makes it a valuable tool for continuous monitoring in critical care environments.

Funding: Funded by Ceribell, Inc.