Authors :
Presenting Author: Zoltan Nadasdy, PhD – University of Texas at Austin
Adam Fogarty, BS – Stanford University
Christopher Primiani, MD – Stanford University
Johannes Koren, MD, PhD – Clinic Hietzing, Vienna, Austria
Clemens Lang, MD – Clinic Hietzing, Vienna, Austria
Christina Duarte, MD – Clinic Hietzing, Vienna, Austria
Christoph Baumgartner, MD – Clinic Hietzing, Vienna, Austria
Jason Shen, MD – Globus Medical Inc.
Hannes Stegmann, PhD – AIT Austrian Institute of Technology GmbH
Tilmann Kluge, PhD – AIT Austrian Institute of Technology GmbH
Manfred Hartmann, PhD – AIT Austrian Institute of Technology GmbH
Rationale:
Status epilepticus occurs in approximately 10-50% of critically ill patients, with more than 80% of these cases identifiable only through EEG. Early detection of nonconvulsive status epilepticus (NCSE) is critical for timely intervention to prevent irreversible brain damage, cognitive decline, and mortality. However, there is a lack of rapid-setup, full-montage EEG systems that combine ease of use with AI-powered NCSE detection and seizure burden (SB) tracking, specifically designed for acute care patients. This study evaluates an automatic NCSE detection algorithm integrated into the Zeto full-montage EEG device for use in acute care.
Methods:
A de-identified dataset comprising full-montage EEG recordings from 81 acute care patients was retrospectively reviewed by six expert neurologists who annotated seizures within the data. Using the ACNS Standardized Critical Care EEG terminology, we determined whether the aggregate seizure activity, based on expert versus algorithm annotations, met the criteria for electrographic status epilepticus (ESE). Automatically defined detections of ESE based on the AI model of encevis were compared between the AI-annotated and expert-annotated seizure datasets by computing the Positive Percent Agreement (PPA) and Negative Percent Agreement (NPA). Short-term seizure burden (STSB) and hourly seizure burden (HSB) based on expert and algorithmic seizure detections were calculated and compared at a threshold of 10% for HSB and threshold levels of 10% and 50% for STSB.
Results:
Encevis’ AI-powered algorithm for ESE detection showed a PPA of 82.6% and an NPA of 91.4% relative to expert annotations. The comparison of HSB between algorithm-detected and expert-detected seizures revealed a PPA of 86.8% and an NPA of 87.7%. Similarly, STSB calculations based on algorithmic detections relative to expert-detected seizures produced a PPA of 91.3%, and an NPA of 85.5% with a threshold of 10%, and a PPA of 88.6% and an NPA of 95.1% with a threshold of 50% applied.
Conclusions:
Timely recognition of ESE is crucial in managing NCSE, where missing or overtreating seizures pose significant risks. The algorithm presented in this study demonstrates high concordance with expert evaluations in detecting ESE, making it a valuable tool for use alongside physician supervision. An AI-powered seizure burden monitor that utilizes a robust seizure detection algorithm can significantly enhance the accuracy of ESE detection. This technology supports informed decision-making by physicians, enhances patient outcomes, and minimizes the risk of long-term cognitive impairment or disability.
Funding:
Funding for the study was provided by Zeto, Inc.