The work, led by Alban Gallard and colleagues, focused deliberately on the bursts of spontaneous neural activity that dominate the electroencephalogram during the third trimester of gestation equivalent. During this period the human brain undergoes rapid structural and functional transformation, with neural activity evolving from sporadic, discontinuous bursts toward more organized and sustained patterns as thalamocortical and cortico-cortical networks mature. Rather than feeding the entire EEG signal into the model, as previous machine-learning approaches have done, the researchers excluded discontinuous periods and targeted only the activity bursts, which they argue are tied to the underlying mechanisms driving evolving bouts of neural activity.

The cohort consisted of 165 preterm newborns, 53 percent of them female, born at a mean gestational age of 31.3 ± 1.9 weeks. EEG recordings took place a mean of 1.9 ± 1.7 weeks after birth, at 33.1 ± 2.0 weeks postmenstrual age on average. After preprocessing, the team detected bursts of spontaneous activity with a mean duration of 7.82 ± 1.61 seconds and extracted 17,221 concatenated 20-second burst epochs, an average of 104 ± 68 epochs per newborn. All signals were transformed to a custom 16-channel bipolar montage before entering the network.

The CNN architecture processed these 20-second burst windows through four blocks, each combining a convolution layer with a pooling layer, separated by normalization layers and ending in flatten, dropout and dense layers. Training used 10-fold cross-validation, with roughly 15,499 epochs per training fold and 1,722 per test fold. At the epoch level, 54.6 percent of predictions fell within one week and 85.2 percent within two weeks, with a mean predicted age difference of 1.10 ± 0.90 weeks. Averaging over epochs for each newborn improved accuracy considerably, yielding the headline population-level figure of 0.75 ± 0.60 weeks (R² = 0.77), which outperformed a predefined baseline of 1.61 ± 1.23 weeks corresponding to each participant’s distance from the mean population age of 33.1 weeks.

Accuracy was not uniform across the age range. Divided into three subgroups, the mean predicted age difference was 1.01 ± 0.62 weeks for infants aged 28–32 weeks, 0.51 ± 0.35 weeks for 32–34 weeks, and 0.81 ± 0.71 weeks for 34–38 weeks. Within one week, the model succeeded for 51.1 percent of the youngest subgroup, 90.6 percent of the middle group, and 74.1 percent of the oldest. An analysis of variance confirmed a significant effect of age group on prediction error, and post-hoc tests showed the middle group differed significantly from the other two, while the youngest and oldest groups did not differ from each other. The authors attribute this pattern partly to clinical and logistical constraints that left the postmenstrual age distribution uneven, with denser sampling around the middle of the third trimester.

Visualizations of the network’s internal representations showed how maturation information emerged layer by layer. Uniform manifold approximation and projection (UMAP) plots of epoch activations progressed from a random pattern in the first block to systematic organization by postmenstrual age in the final block. The Pearson correlation between pairwise distances among all 148,272,810 epoch pairs and their absolute difference in postmenstrual age climbed from r = 0.005 after the first block to r = 0.70 after the fourth, and the mean age difference among neighboring epochs fell from 15.3 ± 6.4 weeks in the first block to 6.4 ± 3.0 weeks in the fourth, with a highly significant main effect of block.

The study’s most revealing experiments involved deliberately corrupting the EEG in three ways and watching the model’s accuracy collapse or hold. Removing high-frequency content above 12 Hz by filtering to 0.5–12 Hz barely dented performance: the mean absolute error rose only slightly to 0.78 ± 0.67 weeks (R² = 0.74), with 73.3 percent of predictions still within one week. Removing low-frequency content below 4 Hz by filtering to 4–20 Hz was far more damaging, pushing the error to 1.58 ± 1.26 weeks, driving R² down to 0.008, and cutting the within-one-week rate to 38.2 percent. A third manipulation disrupted the phase-amplitude relationship between slow and fast oscillations while preserving power in both bands; this raised the error to 1.09 ± 0.87 weeks (R² = 0.53) and lowered the within-one-week rate to 56.4 percent.

Statistical analysis across subgroups sharpened the picture. Removing high-frequency content had no significant impact on prediction accuracy at the population level or in any age group, though the authors caution that the weak Bayes factors involved should be interpreted carefully. Removing low-frequency content significantly degraded accuracy across all three age groups, a result the authors describe as coherent with the established importance of the delta band at this developmental stage. Disrupting phase-amplitude coupling significantly impaired predictions only for infants below 34 weeks postmenstrual age, with the most pronounced effect in the 28–32 week group, a stage marked by the emergence of nested oscillatory patterns in which faster oscillations occur at specific phases of slow waves.

That developmental specificity fits known neonatal neurophysiology. Theta oscillations nested in slow waves first appear around 24 to 25 weeks postmenstrual age, and sigma-band oscillations nested in slow waves, the hallmark delta brushes of human neonatal EEG, first appear around 27 to 28 weeks. Beyond these classical neurobiomarkers, slow oscillations modulate the amplitude of faster activity during spontaneous bursts, and these cross-frequency interactions evolve systematically with cerebral maturation, likely reflecting the development of thalamocortical pathways and the maturation of local corticocortical and subplate-cortical circuits. Computational studies suggest that changes in thalamocortical and hippocampocortical synaptic strength can shape both slow-oscillation morphology and their phase-amplitude relationships. While the study cannot provide direct evidence about the underlying networks, the authors conclude that the model captures developmental EEG signatures generated by those circuits, and that it relies less on phase-amplitude relationships near term, when nested-oscillation biomarkers progressively disappear from the neonatal EEG.

The authors frame the work as a normative reference built exclusively on typically developing preterm infants, and they are explicit that clinical translation awaits further steps. Sleep-wake states were not considered, since bursts were detected across entire recordings, though incorporating vigilance-state information could reduce inter-participant variability and improve sensitivity; sedative and analgesic medications, which can influence neonatal EEG, should also be accounted for in future work. The cohort, though relatively large for neonatal research, remains modest against methodological recommendations favoring several hundred participants, and multi-center validation across recording systems, clinical environments and demographic variability is required before deployment. Future directions include longitudinal EEG monitoring to define individual developmental trajectories, the brain-age gap as a quantitative marker of neurodevelopmental divergence, and multimodal models combining EEG with MRI-based structural morphometry and biological markers. The code has been deposited publicly on GitHub, though the EEG data cannot be shared publicly because participants did not consent to release outside the research consortium.

Subject of Research: EEG-based deep learning estimation of cerebral age in preterm infants

Article Title: EEG-based deep learning estimates preterm infants’ cerebral age: Contribution of slow-wave and cross-frequency coupling

Article References: Gallard, A., Billiot, A., Bourel-Ponchel, É., Brunel, P., Chazal, C., Kongolo, G., Ghostine, G., Wallois, F., Routier, L., & Moghimi, S. (2026). EEG-based deep learning estimates preterm infants’ cerebral age: Contribution of slow-wave and cross-frequency coupling. iScience, 29(11), Article 118052. https://doi.org/10.1016/j.isci.2026.118052

Image Credits: AI Generated

DOI: 10.1016/j.isci.2026.118052

Keywords: EEG, deep learning, preterm infants, cerebral maturation, postmenstrual age, convolutional neural network, slow waves, phase-amplitude coupling, delta brushes, neonatal brain development, cross-validation, third trimester