Peer-reviewed article
The Use of Artificial Intelligence in Neonatal Seizure Detection: An Artificial Intelligence-Assisted Systematic Review.
Nai Ming Lai, Kee Thai Yeo, Juin Yee Kong, Menghao Duan, Shaun Wen Huey Lee et al. ·
Summary
Written by AI, not yet AICS's view · claude-haiku-4-5-20251001This systematic review examined how artificial intelligence can detect seizures in newborn babies. The review included 26 studies: one randomised trial of 258 newborns and 25 smaller validation studies. The trial found that an AI algorithm did not reduce seizure duration or the number of babies with seizures, although it did detect more seizure activity. Overall, the evidence is too limited and uncertain to guide clinical practice.
Key points
- One RCT (n=258) of automated algorithm ANSeR showed no difference in seizure burden (MD: -14.5 minutes) or number of babies with seizures (RR 0.86)
- The ANSeR algorithm did detect higher percentages of seizure-hour activity (difference: 20.8%)
- Twenty-five validation studies reported machine learning algorithms with widely varying sensitivity (32%-98.6%) and specificity (70.1%-100%)
- Most studies had small samples, unrepresentative populations, and unclear reference standards
- Current evidence is insufficient to recommend AI-assisted seizure detection in clinical practice
Limitations
The review found only one RCT and 25 smaller validation studies with significant methodological limitations including small sample sizes and unrepresentative populations. The review cannot determine whether AI-assisted seizure detection improves long-term outcomes for newborns or is ready for widespread clinical use.
Abstract
BACKGROUND: Artificial intelligence (AI) is increasingly used in health care. We systematically reviewed evidence on the accuracy of AI in detecting neonatal seizures. METHODS: We searched PubMed and CENTRAL (covering Medline, EMBASE, CINAHL and trial registries) (29 December 2025) for studies that evaluated the use of all AI applications for detecting and managing neonatal seizures, screened articles using ASReview (a MLA-based platform) and assessed risk-of-bias using Cochrane and QUADAS 2 tools. As data precluded meta-analysis, we performed narrative synthesis. RESULTS: We included 26 studies out of 2051 articles. One RCT (n = 258) evaluated an automated algorithm (ANSeR) and showed no differences in seizure burden (minutes) (mean difference (MD): -14.5, 95% CI -37.8, 15.9), number with seizures (risk ratio (RR) 0.86, 95% CI 0.57, 1.28) and diagnostic accuracy (%) (sensitivity: -8.2, 95% CI -25, 7.7; specificity: -4.8, 95% CI -14.1, 4.6), although higher seizure-hour (%) were detected with ANSeR (difference: 20.8, 95% CI 3.6, 37.1) (low-certainty evidence). Twenty-five validation studies (n = 1021) reported MLAs with varying sensitivity (32%-98.6%) and specificity (70.1%-100%) in seizure detection. DISCUSSION: Evidence from one RCT shows no clear improvements in seizure-related outcomes, while preliminary validation studies show promising but varied MLA performances. However, most studies had small samples and significant methodological concerns including unrepresentative population and unclear reference standards, and lack of clinical validation. The current evidence is therefore insufficient to inform clinical practice. Future studies should validate the findings by enrolling representative high-risk neonates and assess long-term patient-important outcomes. TRIAL REGISTRATION: PROSPERO: CRD42021243455.