ENCEVIS: Evaluating the feasibility of an AI-based algorithm as an assistant to neurophysiologists in clinical...

This study evaluated the performance of the ENCEVIS artificial intelligence (AI)-based algorithm as a screening tool to predict the presence of ictal and/or interictal epileptiform discharges (IEDs) in electroencephalography (EEG) recordings.

This prospective study was conducted from 2019 to 2023 at Khechinashvili University Hospital, Tbilisi. EEG recordings over 3 h were included; standard EEGs and recordings with EEG-negative seizures were excluded. Two independent EEG experts performed blinded visual analyses. In case of disagreement, a third neurophysiologist was consulted, and the final consensus served as the reference standard. ENCEVIS annotations were compared to this reference.

A total of 267 EEG recordings were analyzed. Clinical events occurred in 54 patients (20.2%): 43 had epileptic seizures, 11 had nonepileptic events, and 2 had both. A total of 114 seizures were captured, of which ENCEVIS correctly detected 65 (sensitivity 57.0%, p > 0.05). Detection sensitivity varied by seizure type, FB-TCS bilateral and GTCS 100%, focal seizures with impaired consciousness 66.7% (median 50 s), for focal seizures with preserved consciousness 36.4% (median 28 s), and for tonic seizures 23.1% (median 11 s). Longer seizure duration was associated with higher detection rates. False positive seizure detection rate was 0.27/h. ENCEVIS detected at least one seizure in 42 of 43 seizure-positive recordings (97.7%). Specificity was 48.2%, a positive predictive value (PPV) of 26.6%, and a negative predictive value (NPV) of 99.1%. Performance for interictal detection demonstrated a sensitivity of 97.4%, a specificity of 40.2%, a PPV of 69.3%, and an NPV of 91.8%.

The ENCEVIS algorithm demonstrates high sensitivity in detecting EEG recordings with ictal and interictal epileptiform activity. However, its limited specificity necessitates neurophysiological review to validate positive findings. Its high NPV highlights ENCEVIS's potential as a prescreening tool for identifying EEG recordings without ictal or interictal abnormalities, thereby reducing the workload on neurophysiologists.

This study evaluated the ENCEVIS artificial intelligence (AI) algorithm as a tool to support electroencephalography (EEG) analysis in epilepsy care. Researchers analyzed 267 long-term EEG recordings and compared ENCEVIS results to expert neurophysiologists' evaluations. The algorithm showed high accuracy in detecting normal EEGs and most seizures, especially longer ones. It also performed well in identifying interictal epileptiform activity. However, ENCEVIS sometimes incorrectly annotated normal recordings as abnormal. While it cannot replace expert review, ENCEVIS may serve as a helpful screening tool to reduce the time experts spend reviewing EEGs without epileptic activity.