John M. O' Toole

Aim: To assess the inter-rater agreement and accuracy of human experts’ estimate of EEG maturational age (EMA) and a computer algorithm’s estimate of EMA over the first days after birth in a cohort of normally developing preterm infants. In addition, we explore the influence of post-natal age (PNA) on EMA estimates. Methods: Analysis was performed on EEG records from newborns determined appropriate for gestational age (GA) with favorable neurodevelopment at 2 years of age and without significant neurological compromise at time of EEG monitoring. Three 1h epochs of EEG were selected from 29 newborns with GA ranging from 23 to r) and bias and standard deviation of error (SDE). Intra-newborn agreements for the EMA estimates were assessed using standard deviation. Linear mixed-effects models were used to quantify the effect of PNA on EMA estimates. Results: The algorithm provides a more accurate estimate of GA using 1 h EEG epochs for correlation and bias: algorithm r = 0.83 vs. experts r = 0.60 and 0.66, p < 0.05 for n = 29; algorithm bias = −0.8 days vs. experts’ bias = 3.6 and 7.0 days, p < 0.01 for n = 29. SDE of 8.7 days for the algorithm was not significantly lower compared to the experts’ SDE = 12.4 and 13.2 days, p > 0.05. The algorithm has higher intra-newborn agreement compared to the experts: algorithm SDE = 4.9 days vs. experts SDE = 7.4 and 7.4 days, p = 0.027. For the two experts, increasing PNA is associated with an increase in EMA estimates of 6.6 days/days and 3.7 days/days. The assessment of full, long-duration EEG recordings improved the experts’ estimate of EMA (r = 0.82; SDE = 9.2 days). Conclusions: Automated analysis outperforms visual interpretation of the EEG at estimating EMA for short-duration EEG recordings. PNA is an important factor in EMA estimates.

Neonatal seizures require urgent treatment, but often go undetected without expert EEG monitoring. We have developed and validated a seizure detection model using retrospective EEG data from 332 neonates. A convolutional neural network was trained and tested on over 50,000 hours (n = 202) of annotated single-channel EEG containing 12,402 seizure events. This model was then validated on two independent multi-reviewer datasets (n = 51 and n = 79). Increasing data and model size improved performance: Matthews correlation coefficient (MCC) and Pearson’s correlation (r) increased by up to 50% (15%) with data (model) scaling. The largest model (21m parameters) achieved state-of-the-art on an open-access dataset (MCC = 0.764, r = 0.824, and AUC = 0.982). This model also attained expert-level performance on both validation sets, a first in this field, with no significant difference in inter-rater agreement when the model replaces an expert (∣Δκ∣ < 0.094, p > 0.05).

Objective: To test the potential utility of applying machine learning methods to regional cerebral (rcSO₂) and peripheral oxygen saturation (SpO₂) signals to detect brain injury in extremely preterm infants. Study design: A subset of infants enrolled in the Management of Hypotension in Preterm infants (HIP) trial were analysed (n = 46). All eligible infants were 2 measurements performed over the first 72 h and cranial ultrasounds performed during the first week after birth. SpO₂ data were available for 32 infants. The rcSO₂ and SpO₂ signals were preprocessed, and prolonged relative desaturations (PRDs; data-driven desaturation in the 2-to-15-min range) were extracted. Numerous quantitative features were extracted from the biosignals before and after the exclusion of the PRDs within the signals. PRDs were also evaluated as a stand-alone feature. A machine learning model was used to detect brain injury (intraventricular haemorrhage-IVH grade II–IV) using a leave-one-out cross-validation approach. Results: The area under the receiver operating characteristic curve (AUC) for the PRD rcSO₂ was 0.846 (95% CI: 0.720–0.948), outperforming the rcSO₂ threshold approach (AUC 0.593 95% CI 0.399–0.775). Neither the clinical model nor any of the SpO₂ models were significantly associated with brain injury. Conclusion: There was a significant association between the data-driven definition of PRDs in rcSO₂ and brain injury. Automated analysis of PRDs of the cerebral NIRS signal in extremely preterm infants may aid in better prediction of IVH compared with a threshold-based approach. Further investigation of the definition of the extracted PRDs and an understanding of the physiology underlying these events are required.

Aim: To examine the impact of parent‐led massage on the sleep electroencephalogram (EEG) features of typically developing term‐born infants at 4 months.Method: Infants recruited at birth were randomized to intervention (routine parent‐led massage) and control groups. Infants had a daytime sleep EEG at 4 months and were assessed using the Griffiths Scales of Child Development, Third Edition at 4 and 18 months. Comparative analysis between groups and subgroup analysis between regularly massaged and never‐massaged infants were performed. Groups were compared for sleep stage, sleep spindles, quantitative EEG (primary analysis), and Griffiths using the Mann–Whitney U test.Results: In total, 179 out of 182 infants (intervention: 83 out of 84; control: 96 out of 98) had a normal sleep EEG. Median (interquartile range) sleep duration was 49.8 minutes (39.1–71.4) (n = 156). A complete first sleep cycle was seen in 67 out of 83 (81%) and 72 out of 96 (75%) in the intervention and control groups respectively. Groups did not differ in sleep stage durations, latencies to sleep and to rapid eye movement sleep. Sleep spindle spectral power was greater in the intervention group in main and subgroup analyses. The intervention group showed greater EEG magnitudes, and lower interhemispherical coherence on subgroup analyses. Griffiths assessments at 4 months (n = 179) and 18 months (n = 173) showed no group differences in the main and subgroup analyses.Interpretation: Routine massage is associated with distinct functional brain changes at 4 months.What this paper adds: Routine massage of infants is associated with differences in sleep electroencephalogram biomarkers at 4 months. Massaged infants had higher sleep spindle spectral power, greater sleep EEG magnitudes, and lower interhemispherical coherence. No differences between groups were observed in total nap duration or first cycle macrostructure.

This report describes a set of neonatal electroencephalogram (EEG) recordings graded according to the severity of abnormalities in the background pattern. The dataset consists of 169 hours of multichannel EEG from 53 neonates recorded in a neonatal intensive care unit. All neonates received a diagnosis of hypoxic-ischaemic encephalopathy (HIE), the most common cause of brain injury in full term infants. For each neonate, multiple 1-hour epochs of good quality EEG were selected and then graded for background abnormalities. The grading system assesses EEG attributes such as amplitude, continuity, sleep–wake cycling, symmetry and synchrony, and abnormal waveforms. Background severity was then categorised into 4 grades: normal or mildly abnormal EEG, moderately abnormal EEG, majorly abnormal EEG, and inactive EEG. The data can be used as a reference set of multi-channel EEG for neonates with HIE, for EEG training purposes, or for developing and evaluating automated grading algorithms.