Continental generalization of an AI system for clinical seizure recognition

TL;DR

This study demonstrates a continental generalization of an AI system for clinical seizure detection, achieving high accuracy and efficiency across diverse EEG data, with performance comparable to human experts.

Contribution

It introduces a robust, generalizable AI model for seizure detection validated on large, diverse datasets, significantly reducing annotation time while maintaining expert-level accuracy.

Findings

Achieved 76.68% sensitivity with ~56 false alarms per 24 hours.

Reduced review time from 90 to 7.62 minutes per session.

Confirmed AI performance comparable to human experts at 92.19% accuracy.

Abstract

Electroencephalogram (EEG) monitoring and objective seizure identification is an essential clinical investigation for some patients with epilepsy. Accurate annotation is done through a time-consuming process by EEG specialists. Computer-assisted systems for seizure detection currently lack extensive clinical utility due to retrospective, patient-specific, and/or irreproducible studies that result in low sensitivity or high false positives in clinical tests. We aim to significantly reduce the time and resources on data annotation by demonstrating a continental generalization of seizure detection that balances sensitivity and specificity. This is a prospective inference test of artificial intelligence on nearly 14,590 hours of adult EEG data from patients with epilepsy between 2011 and 2019 in a hospital in Sydney, Australia. The inference set includes patients with different types and frequencies of seizures across a wide range of ages and EEG recording hours. We validated our inference model in an AI-assisted mode with a human expert arbiter and a result review panel of expert neurologists and EEG specialists on 66 sessions to demonstrate achievement of the same performance with over an order-of-magnitude reduction in time. Our inference on 1,006 EEG recording sessions on the Australian dataset achieved 76.68% with nearly 56 [0, 115] false alarms per 24 hours on average, against legacy ground-truth annotations by human experts, conducted independently over nine years. Our pilot test of 66 sessions with a human arbiter, and reviewed ground truth by a panel of experts, confirmed an identical human performance of 92.19% with an AI-assisted system, while the time requirements reduce significantly from 90 to 7.62 minutes on average.