Authors :
Presenting Author: Suma Anand, PhD – Ceribell
Archit Gupta, PhD – Ceribell
Leonardo Tozzi, MD, PhD – Ceribell
Tanaya Puranik, MS – Ceribell
Courtney Juliano, MD – Icahn School of Medicine at Mount Sinai
Maite La Vega-Talbott, MD – Icahn School of Medicine at Mount Sinai
Rana Said, MD – The University of Texas Southwestern Medical Center
Avantika Singh, MD – Medical College of Wisconsin
Joseph Sullivan, MD – University of California San Francisco School of Medicine
Nicholas Abend, MD – Children's Hospital of Philadelphia and University of Pennsylvania
Cecil Hahn, MD – The Hospital for Sick Children, and Department of Paediatrics University of Toronto
Baharan Kamousi, PhD – Ceribell
Rationale:
The risk of seizures is particularly high in the first year of life (Annegers et al., 1995). In neonates, seizures are especially common and are associated with higher mortality and risk of long-term neurological disability (Kim et al., 2022). However, prompt detection and treatment of seizures requires access to both continuous electroencephalogram (EEG) monitoring and interpretation, which varies within and across institutions (Boylan et al., 2010).
Clarity for Neonates and Infants (“Clarity”, Ceribell, Inc.) is a new algorithm for automated seizure detection from EEG in neonates and infants, which could facilitate continuous EEG interpretation and timely treatment of seizures. Here, we validate the performance of Clarity in an independent real-world dataset of pre-term neonates, term neonates and infants.
Methods:
We obtained 539 EEGs recorded in three hospitals from patients of postnatal age 0-1 years and gestational age 22-42 weeks. Two or more board-certified pediatric neurologists annotated seizures according to established criteria (Tsuchida et al., 2013). In case of disagreement, we used the majority opinion.
Clarity was then run on the EEGs. The algorithm detects seizures in non-overlapping 10 second EEG segments. It continuously estimates seizure burden as the cumulative duration of seizure activity within the prior 5 minutes (from 1 or more seizures). It generates an alert when seizure burden is equal or longer than a threshold. Users can set thresholds based on the minimum seizure burden deemed clinically significant.
To evaluate the algorithm, we first calculated the area under the curve (AUC) for seizure detection in 10 second EEG segments. We then evaluated the algorithm’s ability to detect whether the maximum seizure burden in a recording exceeded two possible thresholds: ≥75 seconds and ≥150 seconds (i.e. ≥25% and ≥50% of the 5-minute window). For these two thresholds, we calculated sensitivity, specificity, and negative predictive value (NPV) on a per-recording basis.
We evaluated algorithm performance in the whole cohort, and separately in pre-term neonates (post-menstrual age (PMA)< 37 weeks, N=127), term neonates (PMA 37-44 weeks, N=235) and infants (PMA≥44 weeks, N=177).
Results:
The 539 EEGs comprised 3849 hours, with 13.78 hours of seizure activity in 74 EEGs (Table 1).
The algorithm achieved AUC=0.96 for seizure detection in term neonates and infants, and AUC=0.91 in pre-term neonates. For a maximum seizure burden threshold ≥75 seconds the algorithm had sensitivity=91% and specificity=84%. When ruling out seizures, the algorithm was correct in 99% of recordings (NPV). For a maximum seizure burden threshold ≥150 seconds, sensitivity was 87%, specificity was 88%, and NPV was 99%.
Conclusions:
In this real-world multi-center validation cohort, Clarity demonstrated high performance in detecting and ruling out seizures in preterm neonates, term neonates, and infants. Clarity may help to identify neonates and infants with a high seizure burden, assist in ruling out seizures and enable more timely seizure management in the first year of life.Funding: Funded by Ceribell, Inc.