Search for a command to run...
Objective: To evaluate the performance of the ENCEVIS AI-based seizure detection algorithm in long-term EEG recordings and to identify seizure characteristics - specifically localization, rhythmicity, and duration - that influence detection sensitivity. Methods: This prospective study included 267 prolonged EEG recordings (>3 hours) collected from 2019 to 2023 at Khechinashvili University Hospital. Seizures were visually evaluated by experts, and onset zone, rhythmicity, and duration were defined. ENCEVIS detections were considered true positives if they occurred within 30 seconds before to 60 seconds after expert-defined seizure boundaries. Detection sensitivity was calculated and analyzed using multivariate logistic regression with interaction terms for rhythmicity × duration and localization × rhythmicity. Results: A total of 114 seizures were identified, of which 65 were correctly detected by ENCEVIS (overall sensitivity: 57.0%). Temporal lobe seizures were detected with the highest sensitivity (71.0%, p < 0.05), while generalized tonic seizures showed the lowest (35.1%). Rhythmic seizures had significantly better detection than arrhythmic seizures (71.4% vs. 22.7%, p < 0.0001). Peak sensitivity was observed for seizures lasting 38–68 seconds. In logistic regression, rhythmicity was the strongest independent predictor of detection (OR = 2.13), with a positive interaction observed between rhythmicity and seizure duration (OR = 1.61). Duration alone and localization alone showed limited predictive value. Significance: Seizure detection by ENCEVIS is primarily driven by rhythmicity and moderately long duration, particularly in temporal lobe seizures. Arrhythmic, short, and generalized tonic seizures remain underdetected, underscoring the need for algorithmic refinement. Enhancing detection performance for underrepresented seizure types will be essential for AI seizure detection model implementation in clinical practice.