Term newborns (≥36 weeks’ gestation) at Parkland Hospital (Dallas, TX) who met the following criteria were recruited for this prospective cohort study between 2017 and 2019: (1) a history of an acute perinatal event (e.g., placental abruption, cord prolapse, decreased fetal heart rate), (2) umbilical cord arterial pH or arterial blood gas pH of ≤7.0 or base deficit ≥15 mmol/L at <1 h postnatal age, and (3) signs of encephalopathy. Newborns were excluded from the study if they had any genetic or congenital condition, birthweight < 1800 g, and/or head circumference <30 cm, as these factors can interfere with the primary outcome. The study was approved by the Institutional Review Board at University of Texas Southwestern Medical Center, and written informed consent was obtained from a parent of each newborn prior to enrollment.

The newborns were evaluated within 6 h after birth using a modified Sarnat exam by trained clinicians to determine the severity of encephalopathy that included (1) level of consciousness, (2) spontaneous activity, (3) posture, (4) tone, (5) primitive reflexes (suck, moro), and (6) autonomic system (pupils, heart rate, respirations), with scores of normal (0), mild (1), moderate (2), or severe (3). The TSS was determined by adding the scores for each of the six categories, which ranges from 0 to 18, where 0 represents normal in all six categories, and 18 represents severe encephalopathy in all six categories.5 The clinical grade of encephalopathy was determined by the number of Sarnat abnormalities, with classifications ranging from mild to moderate or severe. In cases where equal numbers of abnormalities were observed, the grade was determined by the degree of reduced level of consciousness. The TSS scores were obtained from electronics medical records.

Whole-body TH was initiated within 6 h after birth for newborns with moderate and severe encephalopathy, following the National Institute of Child Health and Human Development (NICHD) protocol.6 A servo-controlled blanket (Blanketrol II, Cincinnati Sub-Zero Products LLC, OH) was used to maintain a core body temperature of 33.5 °C for 72 h, followed by rewarming at a rate of 0.5 °C per 1–2 h for the next 6 h. Newborns with mild encephalopathy received normothermia as per the standard of care. TH was initiated in accordance with the NICHD late hypothermia protocol if they progressed to more severe encephalopathy or experienced seizures within the first day of life.9

Neurodevelopmental impairment at age of 2 years was death or disability defined by a cognition, language, or motor score <85 on the Bayley Scales of Infant Toddler Development, third edition (BSID-III),5,18,19 which was performed by certified professionals in the follow up clinic at 18–24 months of age.

Continuous EEG acquisition was initiated at a sampling rate of 256 Hz as soon as newborns were admitted, following parental consent. EEG (Nihon Kohden America Inc., Irvine, CA) data were acquired from eight scalp electrodes (Fz, C3, Cz, C4, P3, P4, O1, O2) that were referenced to the Pz, placed according to the modified 10-20 montage for newborns.20 The Component Neuromonitoring System (CNS) Monitor (Moberg Research, Inc., Ambler, PA) was used as the bedside interface for EEG (Nihon Kohden America Inc., Irvine) and other physiological signals, including Near-Infrared Spectroscopy (NIRS) from the INVOS™ 4100–5100 oximeter (Somanetics, Troy, MI), the Blanketrol cooling device, and the Philips IntelliVue MP70 for electrocardiography, mean arterial pressure, and peripheral capillary oxygen saturation (SpO2). The EEG data were obtained from CNS monitor and processed offline using MATLAB (MathWorks Inc., Natick, MA).

Central and parietal electrodes were chosen for analysis because they reflect watershed injury patterns on MRI in HIE.21,22,23 Inter- (C3-C4, P3-P4) and intra- (C3-P3, C4-P4) hemispheric bipolar EEG data were obtained by taking a difference between pair of electrodes. The bipolar EEG data were high pass filtered at 0.3 Hz using a Butterworth filter of order 4.

The control cohort was collected retrospectively from the clinical archives in the Department of Clinical Neurophysiology, New Children’s Hospital, University of Helsinki. The EEG recordings used for the present purpose were part of a larger re-assessment of EEG recordings acquired and reviewed between 2011 and 2016. Here, we included all infants that were born at term age and the EEG was recorded for clinical indications (e.g., Exclusion of seizure activity), but the EEG report was fully normal and there were no neurological consequences found in the comprehensive review of the patient reports. All EEG data was recorded using a NicoletOne system (Cardinal Healthcare/Natus) at sampling rate of 256 Hz from four need electrodes (F3, F4, P3, and P4). All the bipolar signals (F3-P3, F4-P4, F3-F4, and P3-P4) were obtained for further analysis. This data collection was approved by the Institutional Research Review Board at Helsinki and Uusimaa Hospital district approved the study (HUS/244/2021) including waiver of consent due to the retrospective collection of data acquired as part of standard of care.

An automated cloud service tool (https://babacloud.fi/) was employed to calculate BSN from aEEG from bipolar combinations, and identify seizures and artifacts for every 2 s segments of data.17 The cloud service is fully automated and it needs no prior experience from the user, apart from uploading the EEG file, followed by downloading the analysis results for each EEG file. The analysis algorithm has been previously explained in full detail,17 including its rationale, technical design, training, and external validations. Montazeri et al.17 collected visually classified EEGs from newborns recovering from birth asphyxia or stroke.24,25 Unsupervised learning methods were used to explore latent EEG characteristics using a training dataset of 2561 h of multi-channel EEG recordings from 39 term newborns from previously published clinical cohorts.24,25 These insights guided the supervised training of a deep learning-based classifier that achieved an accuracy comparable to human inter-rater agreement. The BSN was calculated by combining the novel EEG background classifier and sleep state trend. The algorithm’s performance was validated using an external dataset26 consisting of 105 h of multi-channel EEG recordings from 31 newborns with HIE. Residual deep neural networks27 identified artifacts like device interference, electromyograph (EMG), movement, electrode pops, non-cortical rhythms like electrocardiogram (ECG), high amplitudes, and zeros. Hybrid neonatal EEG seizure detection algorithms were used to calculate seizure probability.28 Seizure detection probability greater than 0.5 considered as presence of seizure.

All raw analysis outputs from Babacloud are given as numerical vectors for each 2 s sample of the given EEG recording. All further analysis was done offline, including compression by averaging over 3-h epochs, and other statistical analyses. To maximize real-world utility and validity of the pipeline, the full EEG data was always analyzed, and the signal quality automatically assessed using the integrated artifact detector/classifier algorithm. This approach aimed to remove subjective components and other potential biases in the data selection. BSN is a continuous score that ranges from 0 (inactive EEG) to 100 (fully active), allowing a straightforward comparison to other clinical information, as well as avoiding the inherent ambiguity related to classification with discrete EEG scores.17 Artifact rejection was performed on every 1-min nonoverlapping segment of the BSN if they contained more than 30 s of artifacts or more than 6 s of seizure detections. Additionally, if at least two of the four electrodes in a 1-min segment were labeled as artifacts or seizures, the BSN value for the entire 1-min segment were removed from subsequent analysis. The mean BSN value was calculated from artifact-free segments within a 3-h segment for statistical analysis.

Demographic and clinical characteristics of newborns, stratified by non-mild HIE (including mild-moderate, moderate, and severe) and mild HIE, were summarized with descriptive statistics, where continuous variables were presented as means with standard deviations or medians with interquartile ranges (IQRs), and categorical variables were presented with counts with percentages. The non-mild and mild HIE groups were compared using Student’s t test or Wilcoxon rank-sum test for continuous variables, while \(}}}}}^\) test or Fisher’s exact test for comparisons among categorical variables. Univariate linear regression analysis was performed to determine the association between BSN and TSS. The assumptions for linear regression model, including normality of residuals and homoscedasticity, were evaluated using Shapiro-Wilk test and Breusch-Pagan test. Univariate logistic regression models were used to assess the relationships of BSN on Bayley-III neurodevelopmental outcomes, including 2-year Bayley, cognitive, language, and motor outcomes. To evaluate the prediction ability of BSN on Bayley-III neurodevelopmental outcomes, we conducted the receiver operating characteristic (ROC) curve, with the area under the ROC curve (AUC). The optimal cut-off values of BSN were obtained with the Youden method to distinguish between normal and abnormal neurodevelopmental outcomes in infants. Results were reported as odds ratios (ORs) with 95% confidence intervals in logistic regression models, and regression coefficients with 95% confidence intervals in linear regression models. A 2-tailed p value less than 0.05 was considered the threshold for statistical significance. No adjustments were made for multiple comparisons, thus our findings for secondary outcomes (including the severity of encephalopathy and TSS) and subsequent analyses (such as exploring different post-birth time windows both prior to and beyond 6–9 h) should be interpreted as exploratory. All statistical analyses were performed using R version 4.2.2.