Clinical outcome prediction with an automated EEG trend, Brain State of the Newborn, after perinatal asphyxia

(cid:1) Brain State of the New Born (BSN) serves as a reliable bedside EEG assessment tool for measuring clinical outcomes. (cid:1) Clear temporal dynamics in BSN levels are evident in outcome predictions during the ﬁrst two days of life. (cid:1) BSN provides an objective and high-resolution assessment of EEG background and enables accurate outcome prediction.

An automated analysis algorithm offers an alternative solution for bedside EEG review by providing the clinician with an immediate EEG interpretation.Many machine learning -based algorithms have been recently developed to give categorical classification of EEG background (Dereymaeker et al., 2019;Montazeri et al., 2021;Raurale et al., 2021) or other characteristics like sleep (Ansari et al., 2022;Montazeri et al., 2022b) and seizures (Ansari et al., 2019, Pavel et al., 2020, Stevenson et al., 2019, Tapani et al., 2022).They show generally good performance compared to experts, with deep learning -based (DL) classifiers outperforming the classifiers built on heuristic feature engineering.As an alternative to discrete background categories, our recent study used DL methods to develop a continuous measure of EEG background, Brain State of the Newborn (BSN) (Montazeri et al., 2022a).A continuous (a) EEG background measure provides two key advantages.First, it overcomes the inherent ambiguity related to discrete categorization, which is clearly reflected in considerable inter-rater disagreement within (a) EEG background assessments (Bourgoin et al., 2020a;Montazeri et al., 2021;Wusthoff et al., 2017).Second, it facilitates a smooth comparison between EEG recordings and clinical information, such as treatment effects or outcome prognostication (Kota et al., 2024).
Here, we aimed to test the idea that early prediction of clinical outcomes could be achieved by monitoring newborn cerebral recovery using BSN, a fully automated and objective DL-based trend measure of cortical activity.Our overarching hypothesis was that BSN trends would be associated with a favorable outcome, development of CP with or without epilepsy, or death.We also assessed the temporal dynamics in outcome prediction, i.e. how the outcome prediction changes over time at different BSN levels, to support future clinical management and trials.

Overview
The overall study design is presented in Fig. 1.In brief, all longterm EEG data was analyzed from a cohort of 80 infants monitored due to moderate or severe HIE.The BSN algorithm was previously trained and tested on different datasets (see Montazeri et al., 2022a).Here, we used the BSN algorithm to automatically compute BSN levels from the current dataset.The BSN values were used first to re-validate the algorithm by comparing them to expert's visual (a)EEG assessment.The full BSN time courses were then compared between patients with different 4-year clinical outcomes or neonatal neuroimaging outcomes.Both positive and negative predictions by BSN levels for each outcome were computed; finally, the dynamic relationship between BSN levels, outcome predictions and postnatal age were assessed.

Study cohort
A population-based cohort of 92 newborn infants (Nevalainen et al., 2020) (mean gestational age 39Á6 weeks, range 36Á1-42Á3 weeks) were collected after being admitted to the Helsinki University Hospital between 2011 and 2016 for treatment of moderate or severe HIE.An (a)EEG monitoring was commenced as per institutional routines for the first days of life as standard of care (see 2.3).A total of 80 infants were eligible for the present study after exclusions due to a lost (a)EEG file (N = 1); monitoring onset later than 12 hours of life (N = 3); missing perinatal details due to being born outside of the hospital district (N = 3); and missing follow-up information (N = 5).A total of (N = 69) infants were treated with therapeutic hypothermia.Antiseizure medication was given if necessary, according to hospital protocols.The Institutional Research Review Board at HUS diagnostic center approved the study, including a waiver of consent due to the observational nature of the study.

(a)EEG recordings
The (a)EEG recordings were started as soon as possible after birth (median 3Á4 hours, interquartile range, IQR 1Á1 hours, range 0Á7-11Á3 hours), and they continued for several days (median postnatal age of 90 hours, IQR 34 hours, range 27-237 hours; technical breaks median total 2Á75 hours per infant, range 0-19Á9 hours).The EEG was recorded using NicoletOne system (Cardinal Healthcare/ Natus, USA), with a sampling rate of 250 or 256 Hz.All analyses were done using three bipolar derivations: F3-P3, F4-P4, and P3-P4.

Visual (a)EEG background assessment
For validating the BSN output, the (a)EEG background was scored at 1-hour epochs visually by a neonatal EEG expert (PN) blinded to clinical information according to a previously published scheme (Nevalainen et al., 2020).EEG background was classified in hourly epochs into four grades ranging from normal to inactive (see Supplemental File page 2).
In addition, we estimated the time when the (a)EEG background had improved from an initial inactive trace to widely used (Nyman et al., 2022) recovery milestones: Recovery to burst suppression, to continuous EEG, and to continuous EEG with SWC.These three timings were identified by two experts, and they were compared to the automatically computed level of BSN.

Computation of BSN trends
BSN is a continuous scale of EEG background activity produced by the output of a DL-based background classifier (Montazeri et al., 2022a).Previously published classifiers give discrete background predictions, whereas BSN gives a continuous background scale with higher resolution.It also offers a higher temporal resolution, at the minute level, compared to the conventional hourly or multihourly rating.The value of BSN ranges from 0 to 100: A BSN value of 0 corresponds to an inactive EEG, while values in the range up to 100 represent increasing levels of improvement, from burstsuppression and declining degrees of discontinuity to a fully continuous background pattern typically observed during active sleep or wake states.

Analysis pipeline through the computational cloud
The BSN pipeline is freely available via our cloud service, https:// www.babacloud.fi.No special equipment or scientific support is required; the users only need to upload the EEG files in the standard EDF format (available in most clinical EEG systems) using their preferred bipolar montage, and the results can be downloaded within a few minutes.In the BSN pipeline, the fully automated algorithm reads the unprocessed EEG file.The pipeline includes an automated protocol for identifying potential artifacts and seizures in each channel (Stevenson et al., 2019, Webb et al., 2021) that are used in the postprocessing phase.

Postprocessing
The BSN measures were computed for each EEG signal using our automated analysis pipeline (Montazeri et al., 2022a) which produces an estimate of EEG background for each 1-minute EEG segment.The temporal BSN course was aligned with expert's scoring by computing hourly averages of all the 6783 hours of (a)EEG data.
To achieve this, we initially configured the algorithm to reject a 1-minute segment of the given channel if it included >30 s artifacts or >6 s seizure detections.Subsequently, the BSN score of the 1minute epoch was not computed if there were at least two channels labeled as artifacts (30.8%, 2089 hours) or seizures (7.5%, 509 hours), respectively.The median BSN level was then used to summarize scores across EEG channels in each segment.For the BSN trend output, 1-hour epochs were rejected if >50% of their 1-minute segments had been rejected, yielding a total of 5427 hours.The details of the proportions of rejected data in each outcome group are presented in Supplemental File Table S1.For full transparency, we display the periods of rejected EEG segments in gray in the BSN visualizations (see Fig. 3A).

Clinical outcomes
The children's medical records were reviewed retrospectively.An experienced pediatric neurologist made the CP diagnosis based on repeated neurological assessments and repeated assessments by an occupational therapist and a physiotherapist.Observations made by the guardians/parents were also considered.
At the time of the review, the children ranged in age from 4 to 10 years old.Following the criteria explained in the prior study, the pediatric neurologist or pediatric appointment closest to age 4 was identified to determine outcome (Nyman et al., 2022).Based on the outcome at the age 4 years, the children were categorized into four groups: deceased (N = 11), epilepsy (N = 6, all also developed CP), CP without epilepsy (N = 7) and favorable outcome meaning no CP or epilepsy diagnosis (N = 56).Most of the deceased neonates died during the neonatal period, except for one that died in early infancy (before 3 months of age).Of those who developed epilepsy, five were diagnosed with infantile spasms syndrome and one with focal onset epilepsy.All children with a diagnosis of epilepsy at age 4, had been diagnosed before age 1, and all but one of those with a CP diagnosis at age 4 had been diagnosed before age 1.
All neonates in the cohort received anti-epileptic drugs (AEDs), except 23 who had favorable outcome (41.1% of this outcome group).

Neuroimaging outcomes
Based on the imaging protocol described in (Nyman et al., 2022), all 80 neonates underwent brain MRI with a 1.5 T scanner (Philips Intera Achieva, Philips Medical Systems, Best, The Netherlands) or a 3 T scanner (Siemens Magnetom Skyra, Siemens Healthcare GmbH, Erlangen, Germany) between 1 and 16 days of age (median 5 days).We used previously published MRI classification scores (Nyman et al., 2022, Shankaran et al., 2012), in brief: 0 = normal, score 1A = minimal cerebral lesions, score 1B = more extensive cerebral lesions alone (no involvement of basal ganglia, thalamus or anterior or posterior limb of the internal capsule, and no area of watershed infarction), score 2A = any involvement of the basal ganglia, thalamus, anterior or posterior limb of the internal capsule or watershed infarction (no other cerebral lesions), score 2B = 2A + addition cerebral lesions, and score 3 = cerebral hemispheric devastation.

Assessing BSN -based prediction
First, we re-validated BSN for use in this dataset by comparing BSN distributions to the visually defined background (a)EEG scores as well as to the time points where clinical experts indicated EEG recovery to burst-suppression, to continuous activity, or the onset of sleep-wake cycles (SWC) (Nyman et al., 2022).
Second, we studied cortical recovery of BSN time courses from each infant, grouped into clinical outcome categories.The summarized data for BSN at six specific timepoints (6,12,24,36,48, and 72 hours) ± 1 hour were computed for both clinical and MRI outcome categories.Pair-wise comparisons between groups and the associated significance levels were determined using Generalized Estimating Equations (GEE).GEE is a statistical method suitable for handling the correlated nature of longitudinal data (repeated measures) and adjusting for multiple comparisons.We utilized a freely available toolbox accessible at (https://www.jstatsoft.org/article/view/v025i14).Detailed results can be found in Supplemental File page 5. Additionally, p-values from two-tailed t-test analyses are presented in Fig. 3B and C to illustrate group comparisons per each timepoint.This aims to show that the group difference persists over time.
Third, BSN-based prediction of clinical outcomes was assessed.The ROCs and AUCs were estimated for each outcome at the first three specified timepoints (6, 12, and 24 hours) ± 1 hour of postnatal age.All ROCs and AUCs were computed using the binary one-vsrest scheme.The ROCs and AUCs can be found in Supplemental File, page 7. To further explore the utility of BSN-based prediction, PPV, NPV, ROCs, and AUCs were computed and reported for two extreme outcome groups: 1) Favorable outcomes only, and 2) The two poorest outcomes, including deceased or CP with epilepsy.PPV and NPV allow for the assessment of prediction utility while considering both time and BSN on a continuous scale.For the ROCs and AUCs, the same six specified time points (6, 12, 24, 36, 48 and 72 hours) ± 1 hour were used.
Fourth, by utilizing two fixed BSN thresholds selected from ROC curves, we depicted the changes in PPV and NPV over time.Additionally, for these two fixed values, we provided accuracy, recall, and F1 scores for specified times in Supplemental File page 8.

Results
A total of 80 infants were included in the analysis.Their demographic data and clinical features are summarized in Table 1.The BSN trends were generated from the hourly average BSN values after automated rejection of those 1-hour epochs that were found to have too much artifact or seizures, as identified by the automated artifact and seizure detections, respectively.As a results, we retained a total of 5427 hours (80%) of (a)EEG recording data.There were no infants with entire recording rejected, however the proportions of rejected data varied somewhat between the outcome groups (see Supplemental File, Table S1 for details).

Comparison of BSN with expert scores and (a)EEG recovery
The BSN levels were significantly different between all visually assigned (a)EEG background categories (Fig. 2; p < 1Á1e-7, the Kruskal Wallis test).Post hoc pairwise comparison using a Wilcoxon test for each pair were conducted, with a Bonferroni correction for multiple hypothesis testing (correction threshold at p < 0Á05/3 for *, p < 0Á01/3 for **, and p < 0Á001/3 for ***).The hourly BSN level during an inactive background showed median BSN of 14Á2 (IQR 7Á8-16Á6); epochs with burst suppression showed median BSN of 30Á9 (IQR 28Á2-44Á1), and continuous (a)EEG showed a range of BSN levels 70-90 (85Á0; IQR 75Á3-90Á0) with generally higher levels if the SWC was present in the epoch.Only negligible overlap was seen between inactive and burst-suppression, while hourly epochs assigned as burst suppression overlapped somewhat with hourly epochs assigned as continuous EEG without SWC, a finding that likely reflects the genuine visual ambiguity in these EEG background categories as the EEG evolves over category boundaries (Montazeri et al., 2021).A clearer overlap in BSN levels was between categories assigned as continuous EEG without SWC vs. with SWC, reflecting their highly comparable appearance at higher temporal resolution, such as the 1-minute epochs used in the BSN.
We then examined BSN levels at times where clinician experts had indicated the key EEG recovery milestones at a range of postnatal ages (1 to 181 hours).Onset of burst-suppression corresponded to BSN levels (median 23Á6, IQR 18Á1-37Á1) far apart from the BSN levels related to the onset of continuous EEG (median 70Á8, IQR 52Á7-83Á4), which in turn showed some overlap with the BSN levels when visual assessment indicated emergence of SWC (median 82Á1, IQR 70Á6-89Á3).
Taken together, the BSN levels compare well with the visual background categories, and they link to the key landmarks in the visually observed (a)EEG monitoring.

BSN trends in different clinical outcomes
The hourly BSN time courses in different outcome categories during the first three days of life is shown in Fig. 3A.Summary distributions at six time points (6, 12, 24, 36, 48 and 72 hours) show substantial difference in BSN levels between outcome categories (Fig. 3B & Table S2 in the Supplemental File).Most of the infants with favorable outcomes show a very rapid BSN recovery to normal levels, however some individuals in the favorable group might exhibit a more gradual BSN recovery over several days of life.In contrast, infants with later CP show an incomplete and clearly slower BSN recovery, which is accentuated in infants who develop additional epilepsy.The poorest BSN recovery was found in infants who died.There were two apparent outliers: One infant with a favorable outcome presented with a very slow BSN recovery (the lowest BSN trend line, Fig. 3A), which was compatible with a severe, clinical grade 3 HIE followed by an unexpectedly favorable recovery.Another infant with a death outcome presented initially high BSN level followed by a drop by about 20 hours of age (purple BSN trend in the corresponding Fig. 3A), and the retrospective evaluation was suggestive of an antenatal asphyxia with frequent seizures, which may have biased the early BSN estimate in this infant.

Comparison of the BSN trends to MRI imaging
The BSN trends over the first five postnatal days are largely linked to the severity of brain damage defined later from the MRI scores (Fig. 3C & Table S3 in the Supplemental File).Expectedly, infants with the worst structural lesions (MRI score 3) had low BSN levels without apparent recovery.The BSN trends in infants with the second worst lesions (MRI score 2A and 2B) show a generally modest recovery, particularly in infants with more extensive lesions (MRI score 2B).More superficial MRI lesions (MRI score 1A and 1B) were associated with a generally faster BSN recovery.A normal MRI was generally associated with a rapid BSN recovery, however a few infants presented with slower BSN recovery that overlapped with other MRI categories (MRI score 1 and 2).During the first two days, there were widespread group differences between non-neighbor MRI scores, however the difference disappears gradually over time.

BSN -based prediction of clinical outcomes
The BSN-based prediction of clinical outcomes was initially assessed by calculating the ROC curves and AUC values for predicting the outcomes at three time points (please refer to the Supplemental File, page 7).The prediction accuracy was notably high from six hours to 24 hours of life for predicting favorable outcomes, with AUC levels ranging from 96Á1% to 97Á2%, CP with epilepsy outcomes, with AUC levels ranging from 88Á1% to 92Á7%, death outcomes, with AUC levels ranging from 96Á7% to 98Á9%, and moderately accurate for CP outcomes, with AUC levels ranging from 78Á3% to 67Á3%.
The individual BSN time courses in Fig. 3A suggest that outcome prediction by EEG background varies over time.Therefore, we evaluated systematically how PPV and NPV change over time across all BSN threshold levels.The optimal BSN threshold for maximizing PPV increased from about 60 to 80 until 36 hours of life when comparing favorable to non-favorable outcomes, whereas the optimal BSN threshold for maximizing NPV increased from 15 to about 40 (Fig. 4A1&2).In contrast, for predicting patients with the poorest outcomes (death or CP with epilepsy), the optimal BSN threshold for maximizing PPV increased steadily from about 20 to about 30 during the first 48 hours, whereas optimal BSN threshold for maximizing NPV was approximately 60 during the 6 to 12 hours period (Fig. 4B1&2).Subsequently, the threshold decreased to 40 by 24 hours, and remained steady until 48 hours.
These findings together suggest that thresholds for optimal outcome prediction change rapidly during the first days of life, which is a challenge when translating to clinical work-up.
We also inspected temporal changes in PPV and NPV at two fixed BSN thresholds (35 and 80) selected as optimal operating points at 12 hours in the AUC (Fig. 4A3&B3).Expectedly, the PPV and NPV levels evolve over time and differently for each combination of outcome categories (Fig. 5).Moreover, prediction of favorable outcomes is relatively stable, while prediction of unfavorable outcomes shows much clearer change during the first days of life.

Discussion
Our study shows that neonatal EEG background can be monitored automatically and reliably using the BSN trend, which may support early, accurate predictions of both favorable and unfavorable clinical outcomes.Moreover, the findings demonstrate clear temporal dynamics in the BSN-based outcome predictions; the optimal BSN levels for positive and negative predictions change during the first two days of life, in particular for the unfavorable outcomes.The findings are fully compatible with prior literature showing that improvement in the visually defined, discrete EEG background categories relates to later clinical outcome (Chandrasekaran et al., 2017, Csek} o et al., 2013, Hallberg et al., 2010, Massaro et al., 2012, Murray et al., 2009, Murray et al., 2016, Nyman et al., 2022, Sewell et al., 2018, Thoresen et al., 2010, Toet et al., 1999, Watanabe et al., 1999).The continuous BSN measure extends these findings by adding an objective, automated assessment with much higher resolution in time and severity; therefore, BSN can disclose the gradual changes in the cortical recovery (EEG background) that support individual level outcome predictions (Kota et al., 2024).The high resolutions of BSN in both time and depth of background changes surpass previous attempts in the field, providing direct additional support to personalized medical evaluation.
A particular bedside challenge is how to assess recovery of spontaneous cortical activity (Walsh et al., 2011) in an objective manner, and preferably using non-categorical measures.It has been clearly established that the temporal evolution of EEG background after perinatal asphyxia carries significant prognostic information about acute and long-term recovery (Chalak et al., 2021, Menache et al., 2002, Murray et al., 2009, Shellhaas et al., 2011, Walsh et al., 2011, Watanabe et al., 1999).Most importantly, the reviewing expert needs to define the delays from brain injury to recovery of the key EEG landmarks: burst suppression, continuous Data shown as n (%), mean [SD], or median {IQR}.There were no significant differences between the different outcome groups in sex, TH treatment (Chi square), GA, or birth weight (ANOVA).There were significant differences between the outcome groups in Apgar scores (Kruskall-Wallis test).The diseased and CP groups had lower scores than the favorable group at both 1-and 10-min tests.At 10 min the diseased group also had a lower score than the CP group.TH = therapeutic hypothermia, GA = gestational age.a = not available for two neonates, b = not available for four neonates.*p < 0Á05.cortical activity, or the emergence of sleep-wake cycling (SWC) (Chalak et al., 2021, Shellhaas et al., 2011).These are typically assessed visually using one of the many possible discrete EEG categories (Dilena et al., 2021, Walsh et al., 2011), which are all characterized by substantial levels of ambiguity (Bourgoin et al., 2020b;Montazeri et al., 2021;Wusthoff et al., 2017).The challenges in visual (a)EEG review can be partly tempered by smoothing the background assessments over multi-hour epochs (Csek} o et al., 2013, Massaro et al., 2012, Nash et al., 2011, Thoresen et al., 2010), however, visual assessment still imposes substantial variability to the definitions of recovery landmarks used in both clinical trials and bedside reviews.
The above considerations together imply that scaling up human resources for EEG interpretation cannot offer a sound solution for bedside brain monitoring.Instead, EEG review needs to be facilitated by automated and objective measures with clinically useful outputs and validated reliability.The BSN trend offers a potential solution for all these issues by providing a reliable, category-free measure with high temporal precisions.Yet, there is an apparent ceiling effect also in the true accuracy of BSN, both in severity and in time: The underlying classifier in BSN needs to be trained with the visual background scores that introduce their inherent ambiguity (Montazeri et al., 2021(Montazeri et al., , 2022a)), however, the practical performance of BSN may exceed visual reading by providing a widely scalable, objective and consistent EEG assessment.Temporally, BSN is initially computed for every minute of EEG, but excessively high temporal resolution does not likely add information about true changes in the EEG background that changes much slower in the HIE context.However, a higher temporal accuracy may be relevant in assessing faster cortical responses, such as drug effects or acute injury (El-Dib et al., 2022a, El-Dib et al., 2022b).
A direct quantitative comparison of our present findings to prior studies is challenged by the conceptual and practical differences between BSN trend and visual EEG categories.In general, earlier studies have shown that timing of EEG recovery from an inactive trace to burst suppression (Nyman et al., 2022), continuous (Csek} o et al., 2013, Massaro et al., 2012, Nyman et al., 2022, Sewell et al., 2018, Thoresen et al., 2010) or continuous plus sleep-wake cycles (Csek} o et al., 2013, Massaro et al., 2012, Nyman et al., 2022, Sewell et al., 2018, Thoresen et al., 2010) may be all predictive of later clinical or neurocognitive outcomes.Those studies have generally indicated that during the first day  S2 and Table S3 for group comparisons using Generalized Estimating Equations (GEE).BSN: Brain State of the Newborn.SWC: Sleep-wake cycling.MRI: Magnetic resonance imaging.
of life normal or near normal EEG background has high PPV (>90) for favorable outcome, whereas abnormal EEG background has relatively low PPV (<70) for predicting poor outcome in hypothermia treated neonates (Csek} o et al., 2013, Thoresen et al., 2010).Furthermore, multimodal combination measures of EEG and MRI (Nevalainen et al., 2017, Swarte et al., 2012), or even somatosensory evoked potentials (Nevalainen et al., 2017, Nevalainen et al., 2020, Swarte et al., 2012) may further improve the outcome pre-dictions.The prior studies have shown also that severity of EEG background correlates to worse MRI findings (Massaro et al., 2012, Nash et al., 2011); here we show further that the EEG background measured with BSN relates to MRI categories during early hours after birth, while the EEG background in all infants tends to normalize towards the end of the five day observation period.This highlights the clinical gain from the very early (a)EEG assessments using quantitative measures, such as BSN.Notably, all prior  studies are based on discrete EEG and time categories, which introduces a significant potential source of noise from the arbitrary (time) and ambiguous (EEG severity) category boundaries.Our present work, together with another recent study (Kota et al., 2024), shows that outcome predictions can be substantially improved with the continuous measurement of BSN.Moreover, the results show that the optimal BSN levels for outcome predictions change substantially over time, highlighting the inherent limitations of visual EEG analysis, and opening a novel window to an individual level dynamic and quantitative analysis of cerebral recovery.
Our work has some potential limitations: First, the work presents an exploratory study about clinical outcome prediction by scanning through the whole BSN scale (0-100) over long observation times.Such results cannot directly claim any optimal values for prospective clinical trials; however, the present results can be used to define parameters (time and BSN levels) for the future validation studies with different datasets.Second, this work does not assess sensitivity to possible technical variations in the EEG data that might results from using different electrode types, EEG devices or institutional protocols.Our unpublished experience shows, however, that BSN is remarkably robust to such technical factors.Yet, further work is needed to assess the impact of electrode placement, drug treatments, or other care procedure.Our choice of using AUC as one prediction measure was reasoned by clinical traditions, however the results may be biased by the typical imbalance of subject numbers between outcome categories and over time.This concern was partly alleviated by using alternative metrics and by presenting analyses of PPV/NPV throughout the BSN levels and monitoring periods.Finally, this work is limited to a dataset from only one center, hence critical next steps are validations of the BSN using larger multicenter datasets with more diverse patient material and variable recording systems and environments.

Role of the funding source
The funder of the study had no role in study design, data collection, data analysis, data interpretation, writing of the report, or the decision to submit for publication.

Data sharing
Original patient data is not published within this article, but classifier outputs (BSN time courses) and their associated clinical information can be made available by reasonable request from the principal investigators (SV or PN).We provide the BSN solution available via our cloud service (https://babacloud[dot]fi/).The cloud interface needs credentials that are available at request from the corresponding authors.The system will not store the EEG files, and the user is encouraged to use only pseudonymized files for maximal data protection.

Fig. 1 .
Fig. 1.The overview of the study design.Arrows indicate the directions of data flow between study components.BSN: Brain State of the Newborn.PPV: Positive predictive value.NPV: Negative predictive value.ROC: Receiver operating characteristic curve.MRI: Magnetic resonance imaging.

Fig. 2 .
Fig. 2. Re-validation of BSN by comparison to visual (a)EEG interpretation.(A) Distributions of the hourly average BSN in each visually determined (a)EEG background category.The colors in A are used for illustration purposes only and do not hold any clinical significance; they are employed to enhance the representation of the BSN range, where 0 is depicted by red and 100 is represented by light blue.(B) Comparison of BSN values at the key milestones of (a)EEG recovery from the initially inactive (a)EEG to burst-suppression (yellow dots), to continuous activity (green dots), and to the emergence of SWC (blue dots).In this figure, there are two dots for each subject in the cohort for each milestone, provided by the two human experts.BSN distributions are compared using the Kruskal Wallis test and post hoc pairwise comparison using a Wilcoxon test for each pair, with a Bonferroni correction for multiple hypothesis testing.Asterisks indicate adjusted p-values (p < 0Á05/3 for *, p < 0Á01/3 for **, and p < 0Á001/3 for ***).BSN: Brain State of the Newborn.SWC: Sleep-wake cycling.

Fig. 3 .
Fig. 3. Comparison of BSN trends to clinical (A, B) and MRI (C) outcomes.(A) Individual BSN time courses in infants with a favorable outcome (upper panel) or with other outcomes (lower panel, colored for the outcome groups).Each dot in the lines depicts an hourly median BSN.Periods of rejected segments by the automated artifact and seizure detectors are shown in gray.The horizontal dotted lines depict two fixed BSN thresholds (red = 35 and blue = 80) chosen from the ROC curves (asterisks) at 12 hours of age (see Fig. 4, and Fig. 5).In the Supplemental File page 6, similar BSN time courses are presented, distinguishing those who received hypothermia from those who did not.(B) BSN distributions at designated times (±1 hour) in each outcome group.(C) BSN distributions at designated times (±1 hour) in infant groups colored for the neonatal MRI findings.Note the consistent group differences during the first two days, which disappear towards the end of the five days.The dataset is sparser at the end of the five days due to clinically indicated discontinuation of (a)EEG monitoring in individual infants.The comparisons shown in this graph are computed using two-tailed t-test and Asterisks indicate adjusted p-values for significant group differences (gradient colors depict the respective groups) (*: p < 0Á05, **: p < 0Á01, ***: p < 0Á001).See TableS2and TableS3for group comparisons using Generalized Estimating Equations (GEE).BSN: Brain State of the Newborn.SWC: Sleep-wake cycling.MRI: Magnetic resonance imaging.

Fig. 4 .
Fig. 4. Evolution of the BSN-based outcome prediction during the first two days of life.(A) Prediction of favorable outcome at different BSN levels, considering BSN over the given threshold as true positive.(B) Prediction of the poorest outcomes (death or CP with epilepsy) at different BSN levels, considering BSN under the given threshold as true positive.In all PPV and NPV plots, the x-axis depicts hours after birth, the y-axis depicts the BSN threshold, while the gradient background color represents the PPV/NPV value for the given prediction.White represents 0, transitioning through a gradient from navy to black, which represents 1.The purple lines indicate change, or evolution, in the optimal BSN thresholds over time in each plot.The horizontal red and blue lines indicate the fixed BNS thresholds defined from the ROC curves and used in the PPV/NPV plots in Fig. 5.These two fixed values are selected as the optimal operating points at 12 hours in the ROC curves, illustrated as a purple star for each outcome prediction scenario.The ROC curves and their AUC estimates are shown for fixed time points between 6 and 48 hours of postnatal age.These metrics were computed utilizing the binary one-vsrest scheme.Please note that the ROC plots only show values 0-0Á5 on the x-axis and 0Á5-1 on the y-axis in order to better visualize the differences between time epochs.BSN: Brain State of the Newborn.PPV: Positive predictive value.NPV: Negative predictive value.ROC: Receiver operating characteristic curve.CP: Cerebral Palsy.

Fig. 5 .
Fig. 5. Clinical outcome prediction using two fixed BNS thresholds.The plots show temporal change in PPV and NPV during the first two days of life for (A) favorable, (B) poor (death or CP with epilepsy), and (C) death outcome.The fixed BSN thresholds, BSN = 85 (blue) and BSN = 35 (red) were chosen from the ROC curves showing peak AUC at 12 hours (Fig. 4).BSN: Brain State of the Newborn.PPV: Positive predictive value.NPV: Negative predictive value.

Table 1
Baseline characteristics of the study population.