Comprehensive Summary
Kartik et al. (2025) sought to explore whether the childhood sleep cycle and sleep spindle progression can serve as markers of brain maturation and neurodevelopmental health by predicting Functional Brain Age (FBA) via overnight polysomnography and electroencephalography (EEG). To examine the usefulness of overnight FBAs, the relationship between brain age estimations with quantitative EEG features, neural network explainability techniques, and sleep architecture was analyzed. Supervised machine learning (ML) using deep neural networks was applied to further enhance FBA estimation from EEG and sleep stage labeling, enabling visualization of cortical development related to childhood sleep architecture. FBA was determined across wake, NREM (N1-N3), and REM sleep stages in 814 children with clinically normal sleep studies. The neural network models (Res-NN and Res-NN-Seq) were trained on EEG data with corresponding sleep stage labels to predict age and assess how well predicted FBA matched the children's age. The hybrid Res-NN-Seq model outperformed other models with a mean absolute error (MAE) of about 0.96 years across all age ranges. Prediction accuracy varied across age groups, with MAES of 0.78 years in age groups below 2 years old, 0.87 years in groups between 2-12 years old, and 1.55 years in age groups between 12-18 years old. FBAs were also determined to fall within ±25 % of the real age for over 95% of participants. FBA performance varied across sleep stages. The N2, N3, and REM sleep stages provided the most accurate FBAs with well-characterized EEG features such as sleep spindles. Analyses conducted on the explainability of the neural networks (e.g., Grad-CAM) demonstrated the biological plausibility of FBA, as internal network activations tended to correspond with known quantitative EEG features. The authors emphasized that sleep architecture through development can track underlying cortical and thalamocortical maturation. This supports the use of sleep EEG as a scalable, non-invasive device for monitoring brain maturation, a marker of neurodevelopmental health.
Outcomes and Implications
This research supports the use of a potentially scalable, non-invasive biomarker to track neurodevelopment in children. This technique can work along with, or occasionally substitute for, neuroimaging or behavioral assessments, particularly in early developmental stages, where traditional evaluations are more limited in scope. Using FBA predictions from sleep EEG can help track neuromaturation in individuals during childhood and screen for whether development is delayed or accelerated, signaling potential neurodevelopmental conditions or cognitive impairments. Due to the presence of age bias (e.g., a higher proportion of children between 4-9 years old versus other age groups), further validation is likely needed before widespread clinical use via long-term longitudinal studies to determine associations between FBA deviations and neurodevelopmental outcomes.