Assessing a reduced-channel algorithm for end-to-end seizure detection on multiday EEG using inter-rater agreement...

Area of Science:

• Neurology
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Traditional electroencephalography (EEG) systems are cumbersome, limiting monitoring duration and accessibility.
• Wearable EEG devices offer discreet, reduced-channel monitoring for daily life use.
• Need for reliable seizure detection algorithms compatible with reduced-channel wearable EEG.

Purpose of the Study:

• To evaluate the performance of REMI Vigilenz AI for Event Detection (VED), a novel reduced-channel automated seizure detection algorithm.
• To compare VED's performance against expert epileptologists and a state-of-the-art high-channel algorithm (Persyst 15).
• To employ an inter-rater evaluation paradigm to assess algorithm-expert agreement in electrographic seizure detection.

Main Methods:

• Sixty standard-of-care wired EEG records (19+ channels, mean 67 hours) were annotated by three epileptologists and two algorithms (VED and Persyst 15).
• VED processed four differential EEG channels, simulating wearable device data.
• Experts and Persyst 15 reviewed full 19+ channel recordings; performance metrics included relative sensitivity, precision, and false positives per day (FPs/day).

Main Results:

• Experts exhibited high inter-rater agreement. Relative inter-rater sensitivity among experts ranged from 68.4% to 88.3%.
• VED achieved an average relative sensitivity of 77.0% with 4.79 FPs/day compared to experts.
• Persyst 15 achieved a relative sensitivity of 65.4% with 1.30 FPs/day. VED demonstrated non-inferiority in sensitivity at the Low confidence level.
• At the Moderate confidence level, VED showed statistical non-inferiority in FPs/day (p < 0.05) while retaining at least 75% of expert sensitivity (p < 0.01).

Conclusions:

• The VED algorithm demonstrates inter-rater sensitivity comparable to epileptologists and a full-channel algorithm, despite using only four EEG channels.
• VED's performance is influenced by expert agreement levels, suggesting the utility of inter-rater evaluation methods.
• Future VED versions may need to account for polyspike and spike-wave activity, identified in discordant records.
• This study provides a foundation for future validation studies on wearable EEG data in real-world settings.