Automated neonatal EEG analysis: multi-center validation of NeoNaid pipeline (FBA and sleep staging)

• NeoNaid is a fully automated software tool for neonatal EEG analysis based on functional brain age (FBA) estimation and sleep staging, with quality control for artifacts, out-of-distribution inputs, and uncertain predictions.¹
• Multi-center validation on internal and external datasets showed median absolute FBA errors of 0.50 and 0.55 weeks and Cohen’s Kappa ~0.89 for sleep staging.¹
• Strong fit for neural signal processing and EEG-based tools in clinical neurophysiology (Frontiers in Neuroscience).¹
• Architecture: multi-task deep learning model with a shared encoder, multiple output heads, and a channel-agnostic design with shared weights processing each EEG channel independently; an attention-based mechanism aggregates per-channel predictions into a global output.¹
• Training data comprised ~1,326 h of EEG from 124 recordings for FBA; 565 h from 132 recordings for sleep staging; 44 h from 73 recordings for artifact annotations — using non-overlapping 30-second segments at 64 Hz.¹
• External validation dataset (Oxford, John Radcliffe Hospital): 38 EEG recordings from 24 neonates (PMA 29.4–41.4 weeks); internal dataset (KU Leuven / University Hospitals Leuven): 33 recordings from 17 neonates (PMA 27.3–47 weeks), with results maintained across two different EEG acquisition systems and electrode configurations.¹
• FBA results: mean absolute error 0.60 weeks (internal) / 0.69 weeks (external); 79% of internal recordings and 74% of external recordings estimated within 1 week of true postmenstrual age; all internal recordings within 2 weeks.¹
• Sleep staging results: per-recording Cohen’s kappa median of 0.89 (internal) and 0.87 (external); overall pooled kappa 0.874 (internal, 95% CI: 0.865–0.884) and 0.831 (external); quality control provided modest further improvement.¹
• Quality control pipeline uses three criteria: artifact content (>50% flagged segments excluded), novelty detection (isolation forest on nine spectral features), and uncertainty estimation (attention-weight threshold from calibration set) — reducing extreme errors without degrading median performance.¹
• Corresponding author Tim Hermans (KU Leuven, ESAT/STADIUS); collaborative team spans KU Leuven (NICU, Child Neurology) and University of Oxford (Department of Paediatrics); the channel-agnostic design enables use with limited-channel setups comparable to aEEG monitoring.¹
• Clinical deployment pathway: NeoNaid’s FBA and sleep outputs provide quantitative, expert-consistent measures of neonatal neurodevelopment scalable to large datasets and multi-site NICU settings — addressing a major bottleneck in neonatal neurophysiology research.¹

Footnotes

1. https://www.frontiersin.org/articles/10.3389/fnins.2026.1750045 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹