Cross-session reproducibility of resting-state EEG metrics in high-risk very-preterm infants: A preliminary reliability study

Introduction

Resting-state EEG metrics—aperiodic 1/f slope, signal complexity, and spontaneous activity transients (SATs) with interburst intervals (IBIs)—are proposed as bedside markers of brain maturation and injury in preterm infants. Their use assumes that a single recording yields a stable readout, but this is rarely tested. We assessed the cross-session reproducibility of these metrics in high-risk very-preterm infants.

Methods

Continuous 8-channel resting-state EEG (F3/F4, C3/C4, P3/P4, T7/T8; 512 Hz; ~11.5 min; no stimulation) was recorded in 16 very-preterm infants (gestational age ≈ 25–33 weeks) with substantial morbidity (intraventricular hemorrhage, periventricular leukomalacia, necrotizing enterocolitis, bronchopulmonary dysplasia, retinopathy of prematurity). Each was recorded across 2–3 sessions (33 recordings). After band-pass filtering (0.5–45 Hz) and artifact rejection, we computed the aperiodic 1/f exponent (specparam), relative delta power, spectral entropy, Lempel–Ziv complexity, SAT rate, and median IBI. Test–retest reliability across the first two sessions used intraclass correlation (ICC) with leave-one-out analysis.

Results

All recordings showed structured, non-random neural activity (aperiodic exponent 2.09 ± 0.23; range 1.56–2.58; white noise ≈ 0). Reproducibility was strongly metric-dependent. Median IBI was most robust (ICC 0.56; leave-one-out 0.46–0.65; Spearman ρ = 0.55, p = 0.026). Spectral entropy and Lempel–Ziv complexity appeared reliable (ICC 0.75, 0.70) but were driven by one infant, collapsing under leave-one-out (0.18, 0.08). Aperiodic slope, relative delta power, and SAT rate were poorly reproducible (ICC 0.09–0.41, non-significant)—the state-sensitive metrics.

Conclusions

In high-risk very-preterm infants, resting-state EEG metrics are readily computable and clearly neural, but cross-session reproducibility varies markedly. Interburst interval is comparatively stable, whereas amplitude-, rate-, and slope-based metrics fluctuate between sessions, plausibly with behavioral state. The single-session resting-state biomarkers reported here warrant caution, requiring behavioral-state control or repeated sampling.