Artificial Intelligence-Based Neonatal Seizure Detection: A Pragmatic Evaluation of Clinical Performance Tradeoffs in Clinical EEG

Purpose: The aim of this study was to evaluate how three artificial intelligence (AI)-based neonatal seizure detectors (NSDs) perform in a large, clinically acquired data set and their potential to support timely seizure identification. Methods: We conducted a secondary analysis of continuous EEG (cEEG) recordings from 169 neonates born at ≥35 weeks gestation with neonatal encephalopathy. Seizure presence was determined by independent review by two or three pediatric encephalographers, and 2,455 individual seizures were annotated by a single expert as a pragmatic, clinically representative reference standard. Three NSDs were evaluated: a feature-based support vector machine detector (NSD1) and two deep-learning detectors (NSD2 and Persyst-15). Using an any-overlap approach, outcomes included neonate-level seizure identification, seizure-level detection, false detections per day (FD/day), and detection latency. Results: We analyzed 15,343 hours of cEEG (median 98 hours per neonate); 55 neonates (33%) had seizures. All NSDs identified ≥95% of neonates with seizures. Seizure-level detection varied: NSD2 detected 35% (IQR 17–53%), NSD1 63% (IQR 38–82%), and Persyst-15 60% (IQR 37–80%) of seizures. NSD1 generated significantly more FDs (46 FD/day, IQR 22–82) than NSD2 (3.2 IQR 1.2–9.3) and Persyst-15 (2.8 IQR 1.1–6.2; P < 0.0001). The median time to first true detection was <1 minute for NSD1 and Persyst-15, versus 26 minutes for NSD2. Conclusions: AI-based NSDs demonstrated high infant-level sensitivity but different tradeoffs between seizure-level detection and alarm burden. These findings highlight how AI tools may augment clinical EEG workflows and support earlier seizure recognition, while underscoring the need for evaluation under real-world conditions to guide safe and effective implementation.